journal of drug design and research

Editorial Board

• Special Issues
• Submit Manuscript
• Articles in Press
• Current Issue
• Author Guidelines
• Indexing Completed

QUICK LINKS

• SUBMIT MANUSCRIPT
• RECOMMEND THE JOURNAL
• SUBSCRIBE FOR ALERTS

RELATED JOURNALS

• Advances in Vaccines & Vaccination Research (ISSN: 2688-6545)
• Journal of Pharmacy and Health Science Research
• Journal of Pharmaceutics and Drug Research (ISSN:2640-6152)

Prof. Kwang-Hee Lim, Ph. D. Professor Department of Chemical Engineering Daegu University Korea

Dr. Bontha Venkata Subrahmanya Lokesh, B.Pharm.; M.Pharm.; Ph.D.; Assistant Professor Faculty of Pharmaceutical Sciences UCSI UNIVERSITY (South Campus) Malaysia

Dr. Plato A. Magriotis Associate Professor Laboratory of Pharmaceutical Chemistry, Department of Pharmacy University of Patras Greece

Dr. Rajasekharreddy Pala, PhD Researcher Department of Medicine Chapman University and University of California Irvine USA

Dr. Yasmine Gomaa, Ph.D. Research Scientist & Associate Director Laboratory of Drug Delivery Georgia Institute of Technology USA

Prof. Angus George Dalgleish Foundation Chair Oncology St. George’s University of London UK

Prof. Carmela Saturnino Professor Department of Science University of Basilicata Italy

Prof. Vandamme Thierry Professor Faculty of Pharmacy (PHARM) University of Strasbourg France

Dr. Ramanpreet Walia Associate Professor & Registered Patent Agent Amity Institute of Pharmacy Amity University India

Dr. Maryam Malekigorji Researcher China Medical University CQC UK

Prof. Dhrubo Jyoti Sen Professor Department of Pharmaceutical Chemistry Gujarat Technological University India

Ms. Kavita R. Garg Researcher Advisory Board Member Saliegral Global Pvt Ltd, Singapore

Prof. Sabu Thomas, D.Sc., Ph. D., FRSC Professor Vice Chancellor Mahatma Gandhi University India

Dr. Sagadevan Pattiyappan Assistant Professor Department of Biotechnology KSG College of Arts and Science India

Dr. B. Anupama Associate Professor and HOD Pharmaceutical Chemistry Department K V S R Siddhartha College of Pharmaceutical Sciences India

Dr. Prithviraj Karak Researcher Bankura Christian College Bankura India

Dr. Vivek Kumar Dwivedi (M.Sc. &Ph.D ) Head of R&D (Pharmacology and Toxicology) Panchkula Interstaller Testing & Research Centre Pvt. Ltd. India

Prof. Mohamed Mahrous Amer. Professor Poultry Diseases- Faculty of Veterinary Medicine Cairo University Egypt

Dr. Rahul Shivaji Adnaik Principal Anandi Pharmacy College Kalmbe Tarf Kale India

Dr. Seyed Farshad Heidari Researcher Department of Emergency Medicine Ilam University of Medical Sciences Iran

Prof. Doaa Youssef Mohammed El Sheikh Professor Pediatrics and Pediatric Nephrology Zagazig University Egypt

Dr. Rajesh Singla, M.Sc., Ph.D. Associate Professor & Principal SSD College of Professional Studies Bathinda India

Dr. Mojtaba Mafi M.D. Medical Doctor/ Physician Director Fatty Liver and Related Metabolic Disorders Iran

Dr. Mamata Mishra Senior Research Scientist Jai Research Foundation JRF Global, Vapi, Valsad Gujarat, India

Newsletter signup

• Open Access
• A-Z Journals
• Editor in Chiefs

Journal Groups

• Engineering
• Life Sciences

• All subject areas
• Agricultural and Biological Sciences
• Arts and Humanities
• Biochemistry, Genetics and Molecular Biology
• Business, Management and Accounting
• Chemical Engineering
• Computer Science
• Decision Sciences
• Earth and Planetary Sciences
• Economics, Econometrics and Finance
• Engineering
• Environmental Science
• Health Professions
• Immunology and Microbiology
• Materials Science
• Mathematics
• Multidisciplinary
• Neuroscience
• Pharmacology, Toxicology and Pharmaceutics
• Physics and Astronomy
• Social Sciences
• All subject categories
• Acoustics and Ultrasonics
• Advanced and Specialized Nursing
• Aerospace Engineering
• Agricultural and Biological Sciences (miscellaneous)
• Agronomy and Crop Science
• Algebra and Number Theory
• Analytical Chemistry
• Anesthesiology and Pain Medicine
• Animal Science and Zoology
• Anthropology
• Applied Mathematics
• Applied Microbiology and Biotechnology
• Applied Psychology
• Aquatic Science
• Archeology (arts and humanities)
• Architecture
• Artificial Intelligence
• Arts and Humanities (miscellaneous)
• Assessment and Diagnosis
• Astronomy and Astrophysics
• Atmospheric Science
• Atomic and Molecular Physics, and Optics
• Automotive Engineering
• Behavioral Neuroscience
• Biochemistry
• Biochemistry, Genetics and Molecular Biology (miscellaneous)
• Biochemistry (medical)
• Bioengineering
• Biological Psychiatry
• Biomaterials
• Biomedical Engineering
• Biotechnology
• Building and Construction
• Business and International Management
• Business, Management and Accounting (miscellaneous)
• Cancer Research
• Cardiology and Cardiovascular Medicine
• Care Planning
• Cell Biology
• Cellular and Molecular Neuroscience
• Ceramics and Composites
• Chemical Engineering (miscellaneous)
• Chemical Health and Safety
• Chemistry (miscellaneous)
• Chiropractics
• Civil and Structural Engineering
• Clinical Biochemistry
• Clinical Psychology
• Cognitive Neuroscience
• Colloid and Surface Chemistry
• Communication
• Community and Home Care
• Complementary and Alternative Medicine
• Complementary and Manual Therapy
• Computational Mathematics
• Computational Mechanics
• Computational Theory and Mathematics
• Computer Graphics and Computer-Aided Design
• Computer Networks and Communications
• Computer Science Applications
• Computer Science (miscellaneous)
• Computer Vision and Pattern Recognition
• Computers in Earth Sciences
• Condensed Matter Physics
• Conservation
• Control and Optimization
• Control and Systems Engineering
• Critical Care and Intensive Care Medicine
• Critical Care Nursing
• Cultural Studies
• Decision Sciences (miscellaneous)
• Dental Assisting
• Dental Hygiene
• Dentistry (miscellaneous)
• Dermatology
• Development
• Developmental and Educational Psychology
• Developmental Biology
• Developmental Neuroscience
• Discrete Mathematics and Combinatorics
• Drug Discovery
• Drug Guides
• Earth and Planetary Sciences (miscellaneous)
• Earth-Surface Processes
• Ecological Modeling
• Ecology, Evolution, Behavior and Systematics
• Economic Geology
• Economics and Econometrics
• Economics, Econometrics and Finance (miscellaneous)
• Electrical and Electronic Engineering
• Electrochemistry
• Electronic, Optical and Magnetic Materials
• Emergency Medical Services
• Emergency Medicine
• Emergency Nursing
• Endocrine and Autonomic Systems
• Endocrinology
• Endocrinology, Diabetes and Metabolism
• Energy Engineering and Power Technology
• Energy (miscellaneous)
• Engineering (miscellaneous)
• Environmental Chemistry
• Environmental Engineering
• Environmental Science (miscellaneous)
• Epidemiology
• Experimental and Cognitive Psychology
• Family Practice
• Filtration and Separation
• Fluid Flow and Transfer Processes
• Food Animals
• Food Science
• Fuel Technology
• Fundamentals and Skills
• Gastroenterology
• Gender Studies
• Genetics (clinical)
• Geochemistry and Petrology
• Geography, Planning and Development
• Geometry and Topology
• Geotechnical Engineering and Engineering Geology
• Geriatrics and Gerontology
• Gerontology
• Global and Planetary Change
• Hardware and Architecture
• Health Informatics
• Health Information Management
• Health Policy
• Health Professions (miscellaneous)
• Health (social science)
• Health, Toxicology and Mutagenesis
• History and Philosophy of Science
• Horticulture
• Human Factors and Ergonomics
• Human-Computer Interaction
• Immunology and Allergy
• Immunology and Microbiology (miscellaneous)
• Industrial and Manufacturing Engineering
• Industrial Relations
• Infectious Diseases
• Information Systems
• Information Systems and Management
• Inorganic Chemistry
• Insect Science
• Instrumentation
• Internal Medicine
• Issues, Ethics and Legal Aspects
• Leadership and Management
• Library and Information Sciences
• Life-span and Life-course Studies
• Linguistics and Language
• Literature and Literary Theory
• LPN and LVN
• Management Information Systems
• Management, Monitoring, Policy and Law
• Management of Technology and Innovation
• Management Science and Operations Research
• Materials Chemistry
• Materials Science (miscellaneous)
• Maternity and Midwifery
• Mathematical Physics
• Mathematics (miscellaneous)
• Mechanical Engineering
• Mechanics of Materials
• Media Technology
• Medical and Surgical Nursing
• Medical Assisting and Transcription
• Medical Laboratory Technology
• Medical Terminology
• Medicine (miscellaneous)
• Metals and Alloys
• Microbiology
• Microbiology (medical)
• Modeling and Simulation
• Molecular Biology
• Molecular Medicine
• Nanoscience and Nanotechnology
• Nature and Landscape Conservation
• Neurology (clinical)
• Neuropsychology and Physiological Psychology
• Neuroscience (miscellaneous)
• Nuclear and High Energy Physics
• Nuclear Energy and Engineering
• Numerical Analysis
• Nurse Assisting
• Nursing (miscellaneous)
• Nutrition and Dietetics
• Obstetrics and Gynecology
• Occupational Therapy
• Ocean Engineering
• Oceanography
• Oncology (nursing)
• Ophthalmology
• Oral Surgery
• Organic Chemistry
• Organizational Behavior and Human Resource Management
• Orthodontics
• Orthopedics and Sports Medicine
• Otorhinolaryngology
• Paleontology
• Parasitology
• Pathology and Forensic Medicine
• Pathophysiology
• Pediatrics, Perinatology and Child Health
• Periodontics
• Pharmaceutical Science
• Pharmacology
• Pharmacology (medical)
• Pharmacology (nursing)
• Pharmacology, Toxicology and Pharmaceutics (miscellaneous)
• Physical and Theoretical Chemistry
• Physical Therapy, Sports Therapy and Rehabilitation
• Physics and Astronomy (miscellaneous)
• Physiology (medical)
• Plant Science
• Political Science and International Relations
• Polymers and Plastics
• Process Chemistry and Technology
• Psychiatry and Mental Health
• Psychology (miscellaneous)
• Public Administration
• Public Health, Environmental and Occupational Health
• Pulmonary and Respiratory Medicine
• Radiological and Ultrasound Technology
• Radiology, Nuclear Medicine and Imaging
• Rehabilitation
• Religious Studies
• Renewable Energy, Sustainability and the Environment
• Reproductive Medicine
• Research and Theory
• Respiratory Care
• Review and Exam Preparation
• Reviews and References (medical)
• Rheumatology
• Safety Research
• Safety, Risk, Reliability and Quality
• Sensory Systems
• Signal Processing
• Small Animals
• Social Psychology
• Social Sciences (miscellaneous)
• Social Work
• Sociology and Political Science
• Soil Science
• Space and Planetary Science
• Spectroscopy
• Speech and Hearing
• Sports Science
• Statistical and Nonlinear Physics
• Statistics and Probability
• Statistics, Probability and Uncertainty
• Strategy and Management
• Stratigraphy
• Structural Biology
• Surfaces and Interfaces
• Surfaces, Coatings and Films
• Theoretical Computer Science
• Tourism, Leisure and Hospitality Management
• Transplantation
• Transportation
• Urban Studies
• Veterinary (miscellaneous)
• Visual Arts and Performing Arts
• Waste Management and Disposal
• Water Science and Technology
• All regions / countries
• Asiatic Region
• Eastern Europe
• Latin America
• Middle East
• Northern America
• Pacific Region
• Western Europe
• ARAB COUNTRIES
• IBEROAMERICA
• NORDIC COUNTRIES
• Afghanistan
• Bosnia and Herzegovina
• Brunei Darussalam
• Czech Republic
• Dominican Republic
• Netherlands
• New Caledonia
• New Zealand
• Papua New Guinea
• Philippines
• Puerto Rico
• Russian Federation
• Saudi Arabia
• South Africa
• South Korea
• Switzerland
• Syrian Arab Republic
• Trinidad and Tobago
• United Arab Emirates
• United Kingdom
• United States
• Vatican City State
• Book Series
• Conferences and Proceedings
• Trade Journals

• Citable Docs. (3years)
• Total Cites (3years)

Follow us on @ScimagoJR Scimago Lab , Copyright 2007-2024. Data Source: Scopus®

Cookie settings

Cookie Policy

Legal Notice

Privacy Policy

• +1 408-816-2779
• [email protected]

About the Journal

Journal of Drug Research and Development (ISSN 2470-1009 ) presents up-to-date coverage of advanced drug systems and their innovative service to human health. Furthermore, it brings the latest developments in fast-paced areas such as new drug approvals, covering all aspects of theory, research and application of different disciplines about the effects of drugs on Preclinical and Clinical Pharmacology, drug discovery and design.

The Editorial Committee actively pursues to be a leading source of vital knowledge by promoting excellence in the analysis of the molecules and pathways that fail or are destabilised during a dominant disorder and the interaction of these with chemical compounds and biologics. Moreover, it promotes discoveries that are anticipated to have the first impact on medication based therapeutics. Finally, it pursues the analysis and the opportune adaptation of that science to the needs of the drug discovery process.

Journal Indexing

SCI FORSCHEN JOURNALS

All Sci Forschen Journals are Open Access

New Journals

• Company Information |
• Open Access |
• Peer Review |
• Submit Manuscript |

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

• Advanced Search
• Journal List
• Pharmaceutics

A Guide to In Silico Drug Design

Yiqun chang.

1 Sydney Pharmacy School, Faculty of Medicine and Health, The University of Sydney, Camperdown, NSW 2006, Australia

Bryson A. Hawkins

Jonathan j. du.

2 Department of Biochemistry, Emory University School of Medicine, Atlanta, GA 30322, USA

Paul W. Groundwater

David e. hibbs, associated data.

Not applicable.

The drug discovery process is a rocky path that is full of challenges, with the result that very few candidates progress from hit compound to a commercially available product, often due to factors, such as poor binding affinity, off-target effects, or physicochemical properties, such as solubility or stability. This process is further complicated by high research and development costs and time requirements. It is thus important to optimise every step of the process in order to maximise the chances of success. As a result of the recent advancements in computer power and technology, computer-aided drug design (CADD) has become an integral part of modern drug discovery to guide and accelerate the process. In this review, we present an overview of the important CADD methods and applications, such as in silico structure prediction, refinement, modelling and target validation, that are commonly used in this area.

1. Introduction

New drugs with better efficacy and reduced toxicity are always in high demand, however the process of drug discovery and development is costly and time consuming and presents a number of challenges. The pitfalls of target validation and hit identification aside, a high failure rate is often observed in clinical trials due to poor pharmacokinetics, poor efficacy, and high toxicity [ 1 , 2 ]. A study conducted by Wong et al. that analysed 406,038 trials from January 2000 to October 2015 showed that the probability of success for all drugs (marketed and in development) was only 13.8% [ 3 ]. In 2016, DiMasi and colleagues [ 4 ] estimated a research and development (R&D) cost for a new drug of USD $2.8 billion based upon data for 106 randomly selected new drugs developed by 10 pharmaceutical companies. The average time taken from synthesis to first human testing was estimated to be approximately 2.6 years (31.2 months) and cost approximately USD $430 million, and from the start of a clinical testing to submission with the FDA was 6 to 7 years (80.8 months). In comparison to a study conducted by the same author in 2003, the R&D cost for a new drug had increased drastically by more than two-fold (from USD $1.2 billion) [ 5 ]. A possible reason for the increase in R&D cost is that regulators, such as the FDA have become more risk averse, tightening safety requirements, leading to higher failure rates in trials and increased costs for drug development. It is therefore important to optimise every aspect of the R&D process in order to maximise the chances of success.

The process of drug discovery starts with target identification, followed by target validation, hit discovery, lead optimisation, and preclinical/clinical development. If successful, a drug candidate progresses to the development stage, where it passes through different phases of clinical trials and eventually submission for approval to launch on the market ( Figure 1 ) [ 6 ].

Stages of drug discovery and development.

Briefly, drug targets can be identified using methods, such as data-mining [ 7 ], phenotype screening [ 8 , 9 ], and bioinformatics (e.g., epigenetic, genomic, transcriptomic, and proteomic methods) [ 10 ]. Potential targets must then be validated to determine whether they are rate limiting for the disease’s progression or induction. Establishing a strong link between the target and disease builds up confidence in the scientific hypothesis and thus greater success and efficiency in later stages of the drug discovery process [ 11 , 12 ].

Once the targets are identified and validated, compound screening assays are carried out to discover novel hit compounds (hit-to-lead). There are various strategies that can be used in this screening, involving physical methods, such as mass spectrometry [ 13 ], fragment screening [ 14 , 15 ], nuclear magnetic resonance (NMR) screening [ 16 ], DNA encoded chemical libraries [ 17 ], high throughput screening (HTS) (such as protein or cells) [ 18 ] or in silico methods, such as virtual screening (VS) [ 19 ].

After hit compounds are identified, properties, such as absorption, distribution, metabolism, excretion (ADME), and toxicity should be considered and optimised early in the drug discovery process. Unfavourable pharmacokinetic and toxicity profile of a drug candidate is one of the hurdles that often leads to failure in the clinical trials [ 20 ].

Although physical and computational screening techniques are distinct in nature, they are often integrated in the drug discovery process to complement each other and maximise the potential of the screening results [ 21 ].

Computer-aided drug design (CADD) utilises this information and knowledge to screen for novel drug candidates. With the advancement in technology and computer power in recent years, CADD has proven to be a tool that reduces the time and resources required in the drug discovery pipeline. The aim of this review is to give an overview of the various in silico techniques that are used in the drug discovery process ( Figure 2 ).

Various in silico techniques used in the drug design and discovery process discussed in this review. (Abbreviations: CADD: computer-aided drug design; DFT: density functional theory; MM: molecular mechanical; MM-GBSA: molecular mechanics with generalised Born and surface area; QM: quantum mechanical; QSAR: quantitative structure activity relationship).

2. Structure-Based Drug Design

The functionality of a protein is dependent upon its structure, and structure-based drug design (SBDD) relies on the 3D structural information of the target protein, which can be acquired from experimental methods, such as X-ray crystallography, NMR spectroscopy and cryo-electron microscopy (cryo-EM). The aim of SBDD is to predict the Gibbs free energy of binding (ΔG bind ), the binding affinity of ligands to the binding site, by simulating the interactions between them. Some examples of SBDD include molecular dynamics (MD) simulations [ 22 ], molecular docking [ 23 ], fragment-based docking [ 24 ], and de novo drug design [ 25 ]. Figure 3 describes a general workflow of molecular docking that will be discussed in greater detail.

General workflow of molecular docking. The process begins with the preparation of the protein structure and ligand database separately, followed by molecular docking in which the ligands were ranked based on their binding pose and predicted binding affinity. (Abbreviations: LBDD: Ligand-based drug design; ADME: absorption, distribution, metabolism and excretion; MD: molecular dynamics; MM-GBSA: molecular mechanics with generalised Born and surface area).

2.1. Protein Structure Prediction

The advancements in sequencing technology led to a steep increase in recorded genetic information thus rapidly widening the gap between the amounts of sequence and structural data available. As of May 2022, the UniprotKB/TrEMBL database contained over 231 million sequence entries, yet there are only approximate 193,000 structures recorded in the Protein Data Bank (PDB) [ 26 , 27 ]. To model the structures of those proteins where structural data is not available, homology (comparative) modelling or ab initio methods can be used.

2.1.1. Homology Modelling

Homology modelling involves predicting the structure of a protein by aligning its sequence to a homologous protein that serves as a template for the construction of the model. The process can be broken down into three steps: (1) template identification, (2) sequence-template alignment, and (3) model construction.

Firstly, the protein sequence is obtained, either experimentally or from databases, such as the Universal Protein Resource (UniProt) [ 28 ], and this is followed by identifying modelling templates that have high sequence similarity and resolution by performing a BLAST [ 29 ] search against the Protein Data Bank [ 30 ]. PSI-BLAST [ 29 ] uses profile-based methods to identify patterns of residue conservation, which can be more useful and accurate than simply comparing raw sequences, as protein functions are predominately determined by the structural arrangement rather than the amino acid sequence. One of the biggest limitations of homology modelling is that it relies heavily upon the availabilities of suitable templates and accurate sequence alignment. A high sequence identity between the query protein and the template normally gives greater confidence in the homology model. Generally, a minimum of 30% sequence identity is considered to be a threshold for successful homology modelling, as approximately 20% of the residues are expected to be misaligned for sequence identities below 30%, leading to poor homology models. Alignment errors are less frequent when the sequence identity is above 40%, where approximately 90% of the main-chain atoms are likely to be modelled with a root-mean-square deviation (RMSD) of ~1 Å, and the majority of the structural differences occur at loops and in side-chain orientations [ 31 ].

Pairwise alignment methods are used when comparing two sequences and they are generally divided into two categories—global and local alignment ( Figure 4 ). Global alignment aims to align the entire sequences and are most useful when sequences are closely related or of similar lengths. Tools such as EMBOSS Needle [ 32 ] and EMBOSS Stretcher [ 32 ] use the Needleman–Wunsch algorithm [ 33 ] to perform global alignment. In comparison to using a somewhat brute-force approach, the Needleman–Wunsch algorithm uses dynamic programming to find the best alignment by reducing the number of possible alignments that need to be considered and guarantees to find the best alignment. Dynamic programming aims to break a larger problem (the entire sequence) into smaller problems which are then solved optimally. The solutions to these smaller problems are then used to construct an optimal solution to the original problem [ 34 ]. The Needleman–Wunsch algorithm first builds a matrix that is subjected to a gap penalty (negative scores in first row and column), and the matrix is used to assign a score to every possible alignment (usually positive score for match, no score or penalty for mismatch and gaps). Once the cells in the matrix are filled in, traceback starts from the lower right towards the top left of the matrix to find the best alignment with the highest score.

Example of global and local alignment using Needle [ 32 ] and LALIGN [ 32 ]. Global alignment aims to find the best alignment across the two entire length of sequences. Local alignment finds regions of high similarity in parts of the sequences.

Local alignment, on the other hand, aims to identify regions that share high sequence similarity, which is more useful when aligning sequences that are dissimilar or distantly related. EMBOSS water [ 32 ] and LALIGN [ 32 ] are tools that use the Smith–Waterman algorithm [ 35 ] for local alignment. The Smith–Waterman algorithm, such as the Needleman-Wunsch algorithm, uses dynamic programming to perform sequence alignment. However, there is no negative score assigned in this algorithm, and the first row and column are set to 0. Traceback begins with the matrix cell from the highest score and travels up/left until it reaches 0 to produce the highest scoring local alignment.

When searching for templates used for homology modelling, including multiple sequences will improve accuracy of the alignment in regions where there is a low sequence homology, hence multiple sequence alignment (MSA) is essential. The global alignment method for multiple sequences is generally too computationally expensive; modern MSA tools (e.g., ClustalW [ 36 ], T-Coffee [ 37 ] and MUSCLE [ 38 ]) commonly use a progressive alignment approach that combines global and/or local alignment methods, followed by the branching order of a guide tree. This technique aims to achieve a succession of pairwise alignments, first aligning the most similar sequences and then progressing to the next most similar sequence until the entire query set has been incorporated.

For example, MSA was used during the construction of the homology models for Alanine-Serine-Cysteine transporter (SLC1A5) by Garibsingh et al. At the time, there was limited structural information on SLC1A5 due to the lack of an experimentally determined structure of human SCL1 family proteins. Most of the knowledge on the human SLC1 family protein therefore came from the study of prokaryotic homologs, which share low sequence identity. Using the structural information of the recently solved human SLC1A3, Garibsingh et al. carried out a phylogenetic analysis by generating MSA of the human SCL1 family and its prokaryotic homologs using MUSCLE and Promals3D [ 39 ], and built two different conformations of SLC1A5 homology models for the design of SLC1A5 inhibitors [ 40 ].

Once the alignment is complete, the model can be constructed starting with the backbone, then loops and lastly side-chains. The polypeptide backbone of the protein is first created by copying the coordinates of the residues from the template to create the model backbone. Gaps between the alignment of the sequence and the template are then taken care of through insertions and deletions in the alignment. It is important to remodel gaps accurately, as any error introduced here, will be amplified in later stages, thus leading to structural changes that can be critical for protein functionality and protein–protein interactions. Loop modelling, via knowledge-based methods or energy-based methods, can be used to generate predictions of the conformations of the loop. Knowledge-based methods look for experimental data on loops with high sequence similarity to the target from databases, such as PDB, and then insert them into the model. Yang et al. used FREAD [ 41 ] to predict the structure of a missing loop and construct a model of a monoclonal antibody, Se155-4, to study its antibody–antigen interactions with Salmonella Typhimurium O polysaccharide [ 42 ]. On the other hand, energy-based methods predict protein folding using ab initio methods with scoring function optimisation. For example, the Rosetta Next-Generation Kinematic Closure protocol [ 43 ], which employs the ab initio method, was used in loop prediction calculations to construct parts of the leucine-rich repeat kinase 2 (LRRK2) model, as the homology model template had missing loop sections. Mutations in the catalytic domains of LRRK2 are associated with familial and sporadic Parkinson’s disease, yet little is known about its overall structure and the mutations, which alter LRRK2 function and enzymatic activities. Combining homology models with experimental constraints, Guaitoli and co-workers constructed the first structural model of the full length LRRK2 that includes domain engagement and contacts. The model provided insight into the roles that the different domains play in the pathogenesis of Parkinson’s disease and will serve as a basis for future drug design on LRRK2 [ 44 ].

Lastly, side-chains are built onto the backbone model according to the target sequence. Most side-chain types in proteins have a limited number of conformations (rotamers) and programs such as SCWRL [ 45 ] predict these in order to minimise the total potential energy. Upon completion, the model is optimised using molecular mechanics force fields to improve its quality.

A ligand-based approach can be utilised to further optimise homology models with low sequence identity between query sequence and structural template. Moro et al. first presented ligand-based homology modelling, also known as ligand-guided or ligand-supported homology modelling, as a tool to inspect G protein-coupled receptors (GPCRs) structural plasticity [ 46 ]. GPCRs comprise a superfamily of membrane proteins with over 800 members; they play a significant role in cellular signalling in the human body. As such, GPCRs are associated with numerous biological processes, making them important therapeutic targets [ 47 ]. Unfortunately, crystallisation of membrane proteins is known to be challenging, especially in the case of GCPRs, and there were few structural data of GPCRs available until the last decade.

Given that the GPCRs are a diverse family, additional optimisation is required to refine homology models built for those with low sequence identity to the structural template to increase the level of accuracy. In this approach, an initial homology model is first developed using the conventional method. Active ligands are then docked into the binding site for optimisation. The receptor is reorganised and refined based upon the ligand binding in order to better accommodate ligands with higher affinity. Moro et al. first introduced this approach to construct a homology model of the human A 3 receptor based on the structure of bovine rhodopsin in 2006, the only known GPCR structure at the time. A set of structurally related class of pyrazolotriazolopyrimidines with known binding affinities was docked into a conventional rhodopsin-based homology model to induce receptor reorganisation [ 46 ].

The ligand-based homology modelling approach has been used extensively since then in studies of GPCRs, including serotonin receptors [ 48 ], dopamine receptors [ 49 ], cannabinoid receptors [ 50 ], neurokinin-1 receptor [ 51 ], γ-aminobutyric acid (GABA) receptor [ 52 ] and histamine H3 receptors [ 53 ].

2.1.2. Ab Initio Protein Structure Prediction

Historically, the homology modelling approach has been the ‘go-to’ method when it comes to protein structure prediction because it is less computationally expensive and produces more accurate predictions. One of the biggest limitations, however, is that it relies on existing known structures, so that the prediction of more complex targets, such as membrane proteins with little known structural data, is almost impossible. Another solution to this problem is the use of template-free approach, also known as ab initio modelling, free modelling, or de novo modelling [ 54 , 55 ]. As the name implies, this approach predicts a protein structure from amino acid sequences without the use of a template. In addition, the ab initio approach can model protein complexes and provide information on complex formation and protein-protein interaction. This is significant as some proteins exist as oligomers and hence performing docking on monomeric structures may be ineffective [ 56 ]. The principle behind ab initio modelling is based on the thermodynamic hypothesis proposed by Anfinsen, which states that ‘the three-dimensional structure of a native protein in its normal physiological milieu is the one in which the Gibbs free energy of the whole system is lowest; that is that the native conformation is determined by the totality of the inter atomic interactions, and hence by the amino acid sequence, in a given environment [ 57 ].

Ab initio protein structure prediction is traditionally classified into two groups, physics-based and knowledge-based, although recent approaches tend to incorporate both. Purely physics-based methods such as ASTRO-FOLD [ 58 , 59 ] and UNRES [ 60 ] are independent of structural data and the interactions between atoms are modelled based on quantum mechanics. It is believed that all the information about the protein, including the folding process and its 3D structure, can be deduced from the linear amino acid sequence. This approach is often coupled with molecular dynamics refinement which also gives valuable insight into the protein folding process. The Critical Assessment of Methods of Protein Structure (CASP) is a biennial double-blinded structure prediction experiment that assesses the performance of various protein structure prediction methods. ASTRO-FOLD 2.0 successfully predicted a number of good quality structures that are comparable to the best model in CASP9 [ 59 ]. Unfortunately, one of the major drawbacks of pure physics-based approaches is that, due to the enormous amount of conformational space needed to cover, it is often accompanied with high computational cost and time requirement and is only feasible to predict the structures of small proteins.

Bowie and Eisenberg first proposed the idea of assembling short fragments derived from existing structures to form new tertiary structures in 1994 [ 61 ]. The idea behind this process is that the use of low-energy local structures from a fragment library provides confidence in local features as these structures are experimentally validated. Furthermore, significantly reduced computational resources are required as the conformational sampling space is reduced. Rosetta, one of the best-known knowledge-based programs, utilises a library of short fragments that represent a range of local structures by splicing 3D structures of known protein structures. The query sequence is then divided into short ‘sequence window’; the top fragments for each sequence window are identified, on the basis of factors, such as sequence similarity and secondary structure prediction for local backbone structures, and these fragments are assembled to build a pool of structures with favourable local and global interactions (known as decoys) via a Monte Carlo sampling algorithm [ 62 ]. During the assembly process, the representation of the structure is simplified (only includes the backbone atoms and a single centroid side-chain pseudo-atom) in order to sample the conformational space efficiently. It starts off with the protein in a fully extended conformation. A sequence window is selected and one of the top ranked fragments for this window is randomly selected to have its torsion angles replace those of the protein chain. The energy of the conformation is then evaluated by a course-grained energy function and the move accepted or rejected according to the Metropolis criterion. In the Metropolis criterion, a conformation with a lower energy than the previous one is accepted, whereas a conformation with a higher energy (less favourable) is kept based on the acceptance probability [ 63 ]. The whole process repeats until the whole 3D structure is generated. Following this, side-chains are constructed and structures are refined using an all-atom energy function to model the position of every atom in the structure and generate high resolution models [ 64 ]. Other knowledge-based ab initio approaches include I-TASSER [ 65 ] and QUARK [ 66 ].

Another method to improve the accuracy of de novo protein structure prediction is the use of co-evolutionary data for targets with many homologs. The structure of a protein is the key to its biological function, and through the evolutionary process, amino acids in direct physical contact, or in proximity, tend to co-evolve together in order to maintain these interactions and hence preserve the function of the protein. Furthermore, residues that have a high number of evolutionary constraints could indicate important functionalities. Based upon this principle, evolutionary and co-variation data that are obtained from databases such as Pfam [ 67 ] can be harnessed to predict residue contacts and protein folding [ 68 ]. This method works by performing MSA on a large and diverse set of homolog sequences to the query protein, information on amino acids pairs that co-evolve, also known as evolutionary couplings, are then extracted to determine the location of each residues [ 69 ].

The application of neural network-based deep learning approaches to integrate co-evolutionary information has revolutionised the technology used in protein structure prediction and made a huge impact. There are currently a few prediction approaches using deep learning methods to guide protein structure prediction, such as Raptor X [ 70 ], ProQ3D [ 71 ], D-I-TASSER [ 72 ], D-QUARK [ 72 ], and trRosetta [ 73 ]. The impact of using deep learning methods is showcased by AlphaFold, an Artificial Intelligence (AI) system developed by DeepMind and RoseTTAFold [ 74 ], a similar program built using a 3-track neural network from the Baker lab, which has taken the protein modelling community by storm in the two most recent CASPs, CASP13 and CASP14. In CASP13, Alphafold 1 [ 75 ] was placed first in the rankings with an average of Global Distance Test Total Score (GDT_TS) of 70%. The GDT_TS is a metric that corresponds to the accuracy of the backbone of the model, the higher the value, the higher the accuracy [ 76 ]. Subsequently in CASP14, the newer version, Alphafold 2, was placed first again and outperformed all other programs by a huge margin with a median GDT_TS of 92.4 over all categories [ 77 ]. Additionally, the updated version of trRosetta, RoseTTaFold, was ranked second and demonstrated a superior performance than AlphaFold 1 in CASP13, and that all top 10 ranking methods in CASP14 use deep learning-based approaches, signifying the progression in protein prediction accuracy. High accuracy models predicted by AlphaFold 2 are also published in AlphaFold Protein Structure Database ( https://alphafold.ebi.ac.uk/ , accessed on 7 May 2022), providing an extensive structural coverage of known protein sequences [ 78 ].

Knowledge-based methods, such as I-TASSER and QUARK were not tested in CASP14 [ 72 ], however variants of these approaches which integrated deep-learning into protein structure prediction algorithms ranked 8th and 9th, respectively. Physics-based methods, such as UNRES (previously described above), using 3 different approaches (UNRES-template, UNRES-contact and UNRES) achieved GDT_TS scores of 56.37, 39.3 and 29.2, respectively. These results ranked 32nd, 109th and 117th [ 77 ]. The large majority of the top ranking algorithms in CASP14 utilised deep learning approaches, further affirming the utility of deep learning in protein structure prediction approaches [ 72 ].

2.1.3. Protein Model Validation

The accuracy and quality of the predicted structures can be validated and verified using different methods. The stereochemistry of the model can be verified by analysing bond lengths, torsion angles and rotational angles with tools, such as WHATCHECK [ 79 ] and Ramachandran plots [ 80 ]. The Ramachandran plot examines the backbone dihedral angles ϕ and ψ, which represents the rotations made by N—Cα and Cα—C bond in the polypeptide chain, respectively ( Figure 5 ). Torsion angles determine the conformation of each residue and the peptide chain; however, some angle combinations cause close contacts between atoms, leading to steric clashes. The Ramachandran plot determines which torsional angles of the peptide backbone are permitted, and thus assesses the quality of the model. Spatial features, such as 3D conformation and mean force statistical potentials, can be validated using Verify3D [ 81 ], which measures the compatibility of the model to its own amino acid sequence. Each residue in the model is evaluated by its environment, which is defined by the area of the residue that is buried, the fraction of side-chain area that is covered by polar atoms (oxygen and nitrogen) and the local secondary structure. Other structure validation tools include MolProbity [ 82 , 83 ], NQ-Flipper [ 84 ], Iris [ 85 ], SWISS-MODEL [ 86 ] and Coot [ 87 , 88 , 89 ]. In addition to in silico validation, experimental validation of the predicted complexes may also be used to aid selection of a model for future in silico studies. Cross-linking mass spectrometry (XL-MS) provides experimental distance constraints, which can be checked against the predicted models [ 90 ].

( A ) Protein backbone with dihedral angles. ( B ) An example of a Ramachandran plot of crystal structure of human farnesyl pyrophosphate synthase (PDB ID: 4P0V) [ 91 ]. White: disallowed region; yellow: allowed region; red: favourable region.

2.2. Docking-Based Virtual Screening

Docking-based virtual screening aims to discover new drugs by predicting binding modes of both ligand and receptor, studying their interaction patterns, and estimating binding affinity. Some examples of the many docking programs include AutoDock [ 92 ], GOLD [ 93 ], Glide [ 94 , 95 ], SwissDock [ 96 ], DockThor [ 97 ], CB-Dock [ 98 ] and Molecular Operating Environment (MOE) [ 99 ] ( Table 1 ). Due to limitations of X-ray crystallography and NMR spectroscopy, experimentally derived structures often have problems, such as missing hydrogen atoms, incomplete side-chains and loops, ambiguous protonation states and flipped residues. It is therefore essential to prepare the 3D structures accordingly in order to fix these issues before the docking process [ 100 ].

List of common docking programs.

Abbreviations: ASP: Astex Statistical Potential; BRD4: Bromodomain 4; CDK5: Cyclin dependent kinase 5; ChemPLP: Piecewise Linear Potential; HTVS: high throughput virtual screening; MDM2: Mouse double minute 2 homolog; PKB: Protein kinase B; PLK1: Polo-like Kinase 1.

The three main goals of molecular docking are: (1) pose prediction to envisage how a ligand may bind to the receptor, (2) virtual screening to search for novel drug candidates from small molecule libraries and (3) binding affinity prediction using scoring functions to estimate the binding affinity of ligands to the receptor [ 101 ]. Search algorithms and scoring functions are essential components for molecular docking programs.

A good search algorithm should explore all possible binding modes, and this can be a challenging task. The concept of molecular docking originated from the ‘lock and key’ model proposed by Emil Fischer [ 102 ], and early docking programs treated both the protein and ligands as rigid bodies. It was known that protein and ligands are both dynamic entities and that their conformations play an important role in ligand–receptor binding and protein functions, but historically this was too computationally expensive to implement. Modern docking programs treat both protein and ligand with varying degrees of flexibility in order to address this issue.

2.2.1. Binding Site Detection

In docking-based virtual screening, the location of the binding site within the protein must be identified. Most of the protein structures in the PDB are ligand-bound ( holo ) structures, which defines the binding pocket and provides us with its geometries. In cases where only ligand-free ( apo ) structures available, there are traditionally three main types of method to identify potential druggable binding sites. Template-based methods such as firestar [ 103 ], 3DLigandSite [ 104 ] and Libra [ 105 , 106 ] utilise protein sequences to locate residues that are conserved and important for binding. Geometry-based methods, such as CurPocket [ 98 ], Surfnet [ 107 ] and SiteMap [ 108 , 109 ], search for clefts and pockets based on the size and depths of these cavities. Energy-based methods such as FTMap [ 110 ] and Q-SiteFinder [ 111 ] locate sites on the surface of a protein that are energetically favourable for binding. Hybrid methods, such as ConCavity [ 112 ] and MPLs-Pred [ 113 ], as well as machine-learning methods, such as DeepSite [ 114 ], Kalasanty [ 115 ], and DeepCSeqSite [ 116 ] are some of the newer approaches that are under rapid development in recent years.

Beyond locating the orthosteric binding site, these tools are also valuable in identifying potential allosteric binding sites to modulate protein function, hot spots on protein surface to alter protein–protein interactions and also analysing known binding sites to design better molecules that complement the binding pocket. Furthermore, proteins are dynamic systems, and their conformations may change upon ligand binding. Hidden binding pockets, known as cryptic pockets, which are not present in a ligand-free structure, can result from conformational changes upon ligand binding. Detection of cryptic pockets can be a solution to target proteins that were previously considered to be undruggable due to the lack of druggable pockets [ 117 , 118 ].

In addition to the location of the binding site, the evaluation of its potential druggability is equally important. Druggability is the likelihood of being able to modulate a target with a small molecule drug [ 119 ]. It can be evaluated on the basis of target information and association, such as protein sequence similarity or genomic information [ 120 ]. However, this approach only works for well-studied protein families and homologous proteins may not necessarily bind to structurally similar molecules [ 121 ].

Various efforts have been made to evaluate druggability using structure-based approaches. Cheng et al. developed the MAP POD score, one of the first methods to evaluate druggability, using a physics-based method. MAP POD model is a binding free energy model combined with curvature and hydrophobic surface area to estimate the maximal achievable affinity for passively absorbed drugs [ 119 ]. Halgren developed Dscore, which is a weighted sum of size, enclosure and hydrophobicity [ 108 , 109 , 122 ]. Other methods to predict druggability include Drug-like Density (DLID) [ 123 ], DrugPred [ 124 ], DoGSiteScorer [ 125 ], FTMap [ 126 ] and PockDrug [ 127 ].

DoGSiteScorer is a webserver that supports the prediction of potential pockets, characterisation and the druggability estimation. The algorithm first maps a rectangular grid onto the protein; grid points are labelled as either free or occupied depending on whether they lie within the vdW radius of any protein atom. Free grid points are merged to form pockets and subpockets, and neighbouring subpockets are then merged to form pockets. A 3D Difference of Gaussian (DoG) filter is then applied to identify pockets that are favourable to accommodate a ligand. These pockets are characterised global and local descriptors, such as pocket volume, surface, depth, ellipsoidal shape, types of amino acids, presence of metal ions, lipophilic surface, overall hydrophobicity ratio, distances between functional group atoms and many more [ 125 , 128 ].

To predict druggability, a machine learning technique (support vector machine model) trained on a set of known druggable proteins is used to identify druggable pockets based on a subset of these descriptors and to provide a druggability score between 0 to 1, where the higher the score the more druggable is the pocket. A SimpleScore, a linear regression based on size, enclosure and hydrophobicity, is also available to predict druggability [ 129 ].

Michel and co-workers used DoGSite, along with FTMap, CryptoSite, as well as SiteMap to predict ligand binding pockets and evaluate druggability of the nucleoside diphosphates attached to sequence-x (NUDIX) hydrolase protein family. Using a dual druggability assessment approach, the authors identified several proteins that are druggable out of the 22 that were studied. This in silico data was also found to correlate well with experimental results [ 130 ].

Sitemap locates binding sites by placing ‘site points’ around the protein and each site point is analysed for the proximity to the protein surface and solvent exposure. Site points that fulfil the criteria and are within a given distance of each other are combined into subsites, then subsites that have a relatively small gap between them in a solvent-exposed region are merged to form sites. Distance-field and van der Waals (vdW) grids are then generated to characterise the binding site into three basic regions: hydrophobic, hydrophilic (further separates into H-bond donor, acceptor, and metal-binding region) and neither. Sitemap also evaluates the potential binding sites and computes various properties such as size of the site measured by number of site points, exposure to solvent, degree of enclosure by protein, contact of site points with the protein, hydrophobic and hydrophilic character of the site, and the degree to which a ligand can donate hydrogen bonds. These properties contribute to the calculation of the SiteScore (to distinguish drug-binding and non-drug binding sites) and Dscore (druggability score), which helps to recognise druggable binding sites for virtual screening [ 108 , 109 ].

The transient receptor potential vanilloid 4 (TRPV4) is a widely expressed non-selective cation channel involved in various pathological conditions. Despite the availability of several TRPV4 inhibitors, the binding pocket of TRPV4 and the mechanism of action was not well understood. Doñate-Macian and coworkers used Sitemap to search and assess the binding pocket for one of the known TRPV inhibitors HC067047 based on the crystal structure of Xenopus TRPV4 ( Figure 6 ). This group also further characterised the binding pocket and inhibitor–protein binding interactions with the aid of molecular docking, molecular dynamics and mutagenesis studies. The information was then employed to run a structure-based virtual screening to discover novel TRPV4 inhibitors [ 131 ].

Binding site of TRPV4 detected using Sitemap by Doñate-Macian et al. [ 131 ]. Yellow: hydrophobic region; blue: H-bond donor region; red: H-bond acceptor region; white sphere: site point.

2.2.2. Ligand Flexibility

Ligand structures for virtual screening can be obtained from small molecule databases, which are free (e.g., ZINC [ 132 ], DrugBank [ 133 ] and Pubchem [ 134 ]) or commercial (e.g., Maybridge, ChemBridge and Enamine). Conformational sampling of ligands can be performed in several ways. Systematic search generates all possible ligand conformations by exploring all degrees of freedom of the ligand [ 135 ]. Carrying out a systematic search using a brute-force approach (exhaustive search) can easily overwhelm the computing power, especially for molecules with many rotatable bonds and therefore rule-based methods have been the more favoured approaches in recent years. Rule-based methods, such as the incremental construction algorithm (also known as anchor and grow method), generate conformations based on known structural preferences of compounds by limiting the conformational space that is being explored. Usually, a knowledge base of allowed torsion angles and ring conformations (e.g., data from the PDB), and possibly a library of 3D fragment conformations, is used to guide the sampling [ 136 , 137 ]. These break the molecule into fragments that are docked into different regions of the receptor. The fragments are then reassembled together to construct a molecule in a low energy conformation.

Conformer generator OMEGA [ 138 ] employs a prebuilt library of fragments as well as a knowledge base of torsion angles to generate a large set of conformations, which are sampled by geometric and energy criteria to eliminate conformers with internal clashes. Likewise, ConfGen [ 139 ] divides ligands into a core region and peripheral rotamer groups. The core conformation is first generated using a template library, followed by the calculation of the potential energy of rotatable bonds with the torsional term of the OPLS force field, and lastly positioning peripheral groups in their lowest energy forms. To eliminate undesirable conformations or to limit the number of conformations, filtering approaches are applied. Conformations that are too similar are removed based on an energy filter, RMSD, and dihedral angles involving polar hydrogen atoms. Compact conformers are also removed by an empirically derived heuristic scoring method [ 94 , 139 ].

On the other hand, a stochastic search randomly changes the degrees of freedom of the ligand at each step and the change is either accepted or rejected according to a probabilistic criterion such as the Metropolis criterion [ 140 ]. Sampling of conformational space can be performed using different techniques in a stochastic search, including Monte Carlo (MC) sampling [ 62 ], distance geometry sampling [ 141 ] and genetic algorithm-based sampling [ 142 , 143 ]. Balloon [ 142 ], a free conformer generator, uses distance geometry to generate an initial conformer for a ligand, followed by a multi-objective genetic algorithm approach to modify torsion angles around rotatable bonds, stereochemistry of double bonds, chiral centres, and ring conformations. Some other tools that were developed for ligand preparation include Prepflow [ 144 ], VSPrep [ 145 ], Gypsum-DL [ 146 ], Frog2 [ 147 ] and UNICON [ 148 ].

2.2.3. Protein Flexibility

Protein flexibility is essential for their biological function and subtle changes, such as side-chain rearrangements, can alter the size and shape of the binding site and thus bias docking results [ 149 ]. Methods to handle protein flexibility can be divided into four groups: soft docking [ 150 , 151 ], side-chain flexibility [ 152 ], molecular relaxation [ 153 ], and protein ensemble docking [ 154 , 155 ]. Soft docking allows small degrees of overlap between the protein and the ligand by softening the interatomic vdW interactions in docking calculations [ 151 ]. These are the simplest methods and are computationally efficient, but they can only account for small changes. Side-chain flexibility allows the sampling of side-chain conformations by varying their essential torsional degrees of freedom, while the protein backbones are kept fixed [ 156 ]. The molecular relaxation method involves both protein backbone flexibility and side-chain conformational changes; it first uses rigid-protein docking to place the ligand into the binding site then relaxes the protein backbone and the nearby side-chain atoms, usually employing methods, such as MC and MD [ 157 , 158 , 159 ]. Protein ensemble docking methods dock the ligand on a set of rigid protein structures, with different conformations which represent a flexible receptor. The docking results for each conformation are then re-analysed [ 160 ].

Most contemporary docking approaches treat proteins with partial or complete flexibility. For instance, Schrödinger offers a range of docking methodologies with different treatment of protein flexibility. Glide [ 94 , 95 ], with standard precision (SP) and extra precision (XP) is a docking strategy, which allows conformational flexibility for the ligands but treats the receptor as a rigid entity. It softens the active site via vdW scaling (soft docking) with the option of rotamer configuration sampling. Meanwhile, a superior method, Induced Fit Docking, uses Glide for docking to account for ligand flexibility, and Prime [ 161 , 162 ] for side-chain optimisation to account for receptor flexibility [ 163 ]. The ligand is docked into the receptor using Glide with vdW scaling and flexible side-chains are temporarily mutated to alanine to reduce steric clashes and the blocking of the binding site. Once the docking poses are generated, the mutated residues are restored to their original residues and Prime (a program for protein structure predictions) [ 161 , 162 ] is used to predict and reorient the side-chains with each ligand pose. The ligand–receptor complex is then minimised to afford a low-energy protein conformation, which is used for ligand resampling with Glide.

Water molecules have a crucial role in biological systems and interactions, such as stabilising protein–ligand complex, biomolecular recognition and participating in H-bond networks. Water molecules can participate in ligand–protein interactions by acting as bridging waters, and their displacement from the binding site upon ligand binding can also contribute to binding affinity, playing a significant role in the thermodynamics of protein-ligand binding [ 164 ]. The retention or removal of water molecules during virtual screening can have a direct impact on the size, shape and chemical properties of the binding site, which can influence binding geometries and affinity calculations.

Due to the ability of a water molecule to act as both an H-bond donor and acceptor, as well as its highly mobile nature, predicting the location and contribution of water molecules in protein–ligand binding is a challenging task. Crystal structures or cryo-EM structures of proteins can sometimes capture the placement of water molecules in the protein matrix, but the information is not always accurate due to the low resolution of the structural data, and the sample preparation conditions do not reflect the biological environment [ 165 , 166 , 167 , 168 ].

Many approaches were developed to simulate and predict the behaviour of water molecules. Implicit models, also known as continuum models, treat water molecules as a uniform and continuous medium. The free energy of solvation is traditionally estimated based on three parameters, the free energy required to form the solute cavity, vdW interactions and electrostatic interactions between solute and solvent. This method is less computationally demanding but neglects details at the solute–solvent interface [ 167 , 168 ]. Explicit models are computationally more expensive, but the molecular details of each water molecule are considered. Water molecules are normally described using a three-, four-, or five-point model.

In protein–ligand docking, water can be treated explicitly or in an approach involving a combination of implicit and explicit (hybrid), and they can be separated into four categories: (1) Empirical and knowledge-based methods (e.g., Consolv [ 169 ] and WaterScore [ 170 ]), (2) statistical and molecular mechanics methods (e.g., GRID [ 171 , 172 ], 3D-RISM [ 173 , 174 ], SZMAP [ 175 ]), (3) MD simulation methods (e.g., WaterMap [ 176 ], GIST [ 177 ], SPAM [ 116 ]) and, lastly, (4) Monte Carlo simulation methods (e.g., JAWS [ 178 ]).

2.2.4. Scoring Functions

After searching for all possible binding modes, a scoring function is used to evaluate the quality of the docking poses. Scoring functions determine the binding mode and estimate binding affinity, which assists in identifying and ranking potential drug candidates. There are three main categories of scoring functions: force field-based, empirical-based, and knowledge-based methods.

Force field-based scoring functions generally use standard force field parameters taken from force fields, such as AMBER [ 179 ], which consider both the intramolecular energy of the ligand and the intermolecular energy of the protein–ligand complex [ 180 ]. The Δ G estimated using this scoring function is the sum of these energies, which is generally composed of vdW and electrostatic energy terms. An example of program that uses this method is DOCK, which utilises the following equation: [ 181 , 182 ]

where r i j is the distance between protein atom i and ligand atom j , A i j and B i j are vdW components (repulsive and attractive vdW), q i and q j are atomic charges and ε ( r i j ) is the distance-dependent dielectric constant.

Empirical-based functions estimate binding affinity based upon a set of weighted energy terms that are described in the following equation:

The energy terms ( Δ G i ) represents energy terms such as vdW energy, electrostatic energy, hydrogen (H) bond interactions, desolvation, entropy, hydrophobicity, etc., whereas the weighting factors ( W i ) are determined via regression analysis by fitting the binding affinity data of a training set of protein–ligand complex with known 3D structures [ 94 ]. The first empirical scoring function (SCORE) was developed by Böhm in 1994 [ 183 ] based upon a dataset of 45 protein–ligand complexes, and the scoring function considers four energy terms: hydrogen bonds, ionic interactions, the lipophilic protein–ligand contact surface and the number of rotatable bonds in the ligand. Over time, the empirical scoring function has evolved by expanding the data set and considering more energy terms. For example, ChemScore, developed by Eldridge et al. [ 184 ], also considers metal atoms contribution and Glide XP score includes terms to account for desolvation effects [ 94 ].

In knowledge-based functions, structural information is extracted from experimentally determined structures of protein–ligand complexes from databases, such as the PDB [ 30 ] and Cambridge Structural Database (CSD) [ 185 , 186 ]. Boltzmann law is employed to transform the protein–ligand atom pair preferences into distance-dependent pairwise potentials, and the favourability of the binding modes of atom pairs is related to the frequency observed in known protein–ligand structures [ 187 , 188 ]. The potentials are calculated using the following equation:

where w(r) is the pairwise potential between protein and ligand, K B is the Boltzmann constant, T is the absolute temperature of the system, ρ ( r ) is the number density of the protein–ligand atom pair at distance r , and ρ ∗ ( r ) is the pair density in a reference state where the interatomic interactions are zero.

3. Ligand-Based Drug Design

When there is limited structural knowledge on the target protein, biological and chemical information is drawn from known active ligands to identify key features that are responsible for biological activity and this information can be used for ligand-based drug design (LBDD). Common LBDD methods include similarity searches, scaffold hopping, quantitative structure–activity relationship (QSAR) and pharmacophore models. Although CADD approaches are generally classified as structure-based and ligand-based approaches, it should be noted that virtual screening strategies often integrate and combine the two to improve the success rate in hit identification [ 215 ].

3.1. Similarity Search

The underlying hypothesis of molecular similarity is that molecules with similar molecular structures have similar physical properties and biological activities. Two key components in similarity analysis are structural representations and quantitative measurements of similarity between the two structural representations.

Different molecular fingerprints can be used to represent the chemical properties of a molecule, and similarity measurements can rely on the use of 1D, 2D and 3D descriptors. This involves dividing the molecule into a sequence of bits; so, the common bits between molecules can be compared to assess similarity. Some common molecular fingerprints include structural keys, topological fingerprints, circular fingerprints and pharmacophore fingerprints [ 216 ]. Structural key fingerprints, such as the MACCS fingerprint [ 217 ] and TGD fingerprint [ 218 ], search for the presence of structures/features of the molecules based on a pre-defined list of structural keys. This method is most useful when the molecules contain a lot of structural keys. Topological fingerprints (e.g., Daylight fingerprint) [ 219 ] analyse the fragments of the molecule following a connectivity path (usually linear) up to a certain number. The algorithm generates a pattern for each atom in the molecule, then a pattern for each atom and its nearest neighbours and connecting bonds, followed by a pattern that represents each group of atoms and bonds connected by paths up to two bonds long, and the process continues with longer bond paths. Circular fingerprints, such as Molprint2D [ 220 ] and extended-connectivity fingerprints (ECFP) [ 221 ], look at the environment of each atom in the molecules up to a certain radius. Every heavy atom of a molecule is sequentially used as a starting point and is assigned an atom type. This is followed by the assignment of atom types to neighbouring atoms of the central heavy atoms (first layer). This process is repeated with each distance/layer from the central heavy atom and the number of atoms with each given atom type are recorded to calculate descriptor values [ 222 ]. In addition to the common molecular fingerprints mentioned that are mostly used to describe synthetic compounds, the Natural Compound Molecular Fingerprint (NC-MFP) was developed by Seo et al. to better represent natural products [ 223 ].

There are different metrics that can be used to assess and quantify the similarity between two molecules (A and B). Most metrics have the range from 0 (completely dissimilar) to 1 (identical). Some of the common metrics are listed below:

Tanimoto coefficient (range: 0–1): [ 224 ]

Dice index (range: 0–1): [ 225 ]

Cosine coefficient (range: 0–1): [ 226 ]

Euclidean distance (range: 0–1): [ 226 ]

where a is the number of bits present in molecule A, b is the number of bits present in molecule B and c is the number of bits present in both molecule A and B. S denotes similarities and D denotes distances where S = 1 1 + D . The cut-off values for the similarity metrics depend on both the fingerprints and metrics used and hence cannot be compared directly. For example, WebCSD, the online portal to CSD, offers both the Tanimoto coefficient and the Dice index for similarity search and the default cut-off values were set as 0.7 and 0.975, respectively [ 227 ].

Wang and co-workers used a combination of docking-based and 2D similarity search techniques to identify novel CDK8 inhibitors [ 228 ]. A small molecule library was first subjected to molecular docking against multiple crystal structures of CDK8 to account for the protein conformation change. Of the 50 candidates selected from the docking study, 7 showed more than 30% inhibition against CDK8 based on in vitro binding competition assay. Similarity search using Discovery Studio [ 229 ] was performed on W-18 and W-37, two of the most potent candidates, to find similar structures with high CDK8 inhibitory effects. Using the Tanimoto coefficient to calculate the similarities of molecules based on the ECFP_6 fingerprints, WS-2 which shares 0.28 and 0.32 similarity with W-18 and W-37, respectively, was identified and it is significantly more potent than both of the parent molecules ( Figure 7 ).

Discovery of the CDK8 inhibitor WS-2 from W-18 and W-37 using similarity search [ 228 ].

3.2. Quantitative Structure-Activity Relationship (QSAR)

A QSAR model is a computational or mathematical model that derives correlation between the calculated molecular properties of a group of compounds and their experimentally determined activity. QSAR methodology was first proposed by Hansch and Fujita in 1964 who published a method for the correlation of biological activity and chemical structure [ 230 ], and QSAR methodology has evolved a lot since then. 1D- and 2D-QSAR models are classified as ‘classical’ QSAR methodologies, where 1D-QSAR correlates biological activity with molecular properties, such as pKa and logP [ 231 ], and 2D-QSAR correlates biological activity with the structure of the ligands on a 2D basis and considers descriptors, such as topological and constitutional descriptors [ 230 , 232 ]. Topological descriptors are based on the connectivity of atoms in the molecule, including molecular size, shape, branching, heteroatoms and multiple bonds but with no information on the 3D spatial arrangement of the atoms [ 181 ]. Constitutional descriptors simply describe the molecular composition of a molecule, such as molecular weight, number of atoms and bonds, types of atoms, and ring counts.

3D-QSAR takes into account the 3D spatial representation of molecules, such as different conformations and stereo-isomerisation. Two of the most popular 3D-QSAR methodologies are the Comparative Molecular Field Analysis (CoMFA) proposed by Cramer et al. [ 233 ] and the Comparative Molecular Similarity Indices Analysis (CoMSIA) proposed by Klebe et al., a modified version of CoMFA [ 234 ]. The primary goal of 3D-QSAR is to establish a relationship between biological activity and spatial properties of the ligands, therefore data quality and structural diversity are particularly important to construct a good quality 3D-QSAR model. 3D-QSAR is often used for lead optimisation and biological activity prediction for novel compounds as it can quantitatively correlate modifications in 3D chemical structures and the respective changes in biological effects.

For example, the 3D-QSAR method was applied in the structure–activity relationship (SAR) analysis of maslinic acid analogues and the identification of its anti-cancer target. Maslinic acid analogues are known to be anti-cancer compounds but there was no structural information about its molecular target. A common pharmacophore model on five analogues was first constructed, then field points-based descriptors were used to build a 3D-QSAR model after aligning 74 analogues to the pharmacophore model. A field point-based similarity search on maslinic acid was performed on the ZINC database, followed by screening through the 3D-QSAR model for bioactivity prediction and SAR field point’s compliance. Additional filters (Lipinski’s rule of five, absorption, distribution, metabolism, and excretion (ADME) and synthetic accessibility) were also applied and eventually 39 compounds were listed. The compounds were docked against a series of potential cellular targets of maslinic aid analogues (predicted by STITCH) [ 235 ] and identified NR3C1 as a major anti-cancer target of maslinic acid analogue as well as compound P-902 as a potential lead compound [ 236 ].

3.3. Pharmacophores

It is widely believed that Paul Ehrlich came up with the concept of pharmacophore: a molecular framework that carries ( phoros ) the essential features responsible for a drug’s ( pharmacon ) biological activity in the early 1900s [ 237 , 238 ]. However, some consider the concept of modern pharmacophore was in fact proposed by Schueler in 1960 [ 239 ], which was then extended by Beckett and co-workers who introduced the first pharmacophore model with identified distance ranges in 1963 [ 240 ] and Kier who proposed the first computed pharmacophore model in 1967 [ 241 ]. Nowadays, pharmacophore models are extractions of electronic and steric features from ligands in a 3D spatial arrangement that is relevant for interactions to the target protein and the relative biological responses. The features are purely abstract concepts and do not represent chemical functional groups or a typical structural skeleton [ 242 ]. The six classical pharmacophore features classified are H-bond donors, H-bond acceptors, negative ionic, positive ionic, hydrophobic regions, and aromatic regions ( Figure 8 ). On top of that, less common features can also better characterise the chemical functionalities, such various metal binding locations are supported by LigandScout [ 243 , 244 , 245 , 246 ]. Constraints and restrictions can also be applied by introducing excluded volumes to the model to prevent ligands from occupying certain spaces (ligand-inaccessible) [ 247 ]. Pharmacophores can be divided into two sub-categories: ligand-based and structure-based pharmacophores.

( A ) Chemical structure of nitrofurantoin. ( B ) Nitrofurantoin superimposed with pharmacophore features. Light red sphere: H-bond acceptor; light blue sphere: H-bond donor; red sphere: negative ionic; blue sphere: positive ionic; orange torus: aromatic ring.

Ligand-based pharmacophores are based on the chemical structures of ligands when there is little structural information about the target protein is available. To construct a ligand-based pharmacophore, the conformational space of flexible active molecules (through conformational sampling) is covered because the molecules should be in their bioactive conformations. The molecules are then aligned, and common features are extracted to generate a pharmacophore model. Alignment techniques are divided into point-based and property-based approaches [ 248 ]. In point-based approach, atoms, fragments, or chemical feature point distances are minimised, and pairs of points are superimposed by minimising distances. Some examples of programs that use point-based alignment include HipHop [ 249 ], Phase [ 250 , 251 ] and Galahad [ 252 ]. In contrast, property-based approaches (e.g., MOE [ 99 ]) generate alignments based on molecular field descriptors, such as electron density, electrostatic potential, molecular shape and volume, etc. [ 248 , 253 ].

A study conducted by Rampogu et al. developed a ligand-based pharmacophore model for the screening of natural compounds against HER2 kinase domain. A total of 82 compounds with various levels of activity were chosen from the literature where 32 of them were used to construct a pharmacophore model using Discovery Studio [ 229 ]. The rest of the compounds, together with decoy set, were employed to validate the model. A total of 197 201 compounds from the Universal Natural Products Database were first filtered for ADME and Lipinski’s Rule of Five to identify compounds with drug-like properties, followed by screening against the pharmacophore model. The resulting compounds were subjected to molecular docking and MD simulations and eventually identified two potential leads against HER2 breast cancers [ 254 ].

One of the biggest drawbacks in ligand-based pharmacophore modelling is the selection of training set ligands. Searching for active ligands to form a training set from the literature could be a difficult task as biological assays were conducted under different experimental conditions. Performing biological assays, such as enzyme kinetic assays, under consistent conditions can be useful to investigate relative biological activities of the ligands, and hence the direct ligand–protein interactions. Although highly different pharmacophore models can increase diversity and cover wider chemical space, training set ligands with larger structural differences might require other experimental validation (e.g., X-ray crystal structure) to confirm they share the same binding site. Nevertheless, the real problem of ligand-based pharmacophore modelling lies in defining if a ligand is active or inactive, particularly in the case of defining qualitative pharmacophores. The diversity of the dataset could hugely affect the pharmacophore model generated, including feature types, locations and excluded volumes [ 255 ].

Compared to ligand-based pharmacophores, structure-based approaches are less likely to be biased by the chemical structures of existing active compounds and thus yield more diverse molecules. Structure-based pharmacophore models are constructed from either a protein–ligand complex or from the 3D structure of the receptor alone (receptor-based). The protein–ligand complex approach evaluates the key interactions between the ligand and the binding site and then transforms this information into a pharmacophore model [ 256 ]. For cases where the structural information of the ligand is lacking, the receptor-based approach can be applied. Pharmacophore hypotheses can be derived from protein structures using two methods: geometric constraints [ 257 ] and binding site analysis using virtual probe atoms [ 258 ].

MurG is one of the enzymes involved in the biosynthesis of the peptidoglycan layer in Mycobacterium tuberculosis and inhibition of MurG could be useful for the treatment of tuberculosis. Saxena et al. built a pharmacophore model based on the protein–ligand interactions of the homology model of Mycobacterium tuberculosis MurG due to the lack of available crystal structures. The pharmacophore model, along with molecular docking and MD simulations, was used and identified three lead compounds that were potential Mycobacterium tuberculosis MurG inhibitors [ 259 ].

3.3.1. Pharmacophore Validation

Before employing the pharmacophore model for virtual screening, it is essential to validate the model to evaluate the predictivity of the model. Decoys databases such as DUD-E [ 260 ], MUV [ 261 ] and DEKOIS [ 262 ] are often used to test the model’s ability to differentiate active and inactive compounds. Multiple refinements are frequently performed to result a better model after testing with different metrics. Examples of some of the commonly used metrics are listed below:

Yield of actives ( Y a ) shows the retrieved true positive compounds ( H a ) in relation to the number of hits retrieved ( H t ) [ 263 ].

Sensitivity ( S e ) is the ratio Ha to all the actives compounds ( A ) in the database. The closer the number is to 1, the higher number of active compounds returned from the search. It gives an insight into the ability of the model to select truly active compounds [ 264 ].

Specificity ( S p ) is the ratio of rejected true negatives ( T N ) to all the–inactive compounds ( D − A ), where D is the number of entries in the database). When S p = 1, all the inactive compounds have been correctly rejected. Specificity tells us the ability of the model to discard inactive compounds [ 264 ].

Enrichment factor ( E F ) measures Y a proportional to the ratio of A in the whole database [ 264 ].

The Goodness of Hit list ( G H   s c o r e ) is a combination of sensitivity, specificity, and yield of actives of different weightings. It considers both true actives ratio and true inactives ratio, which makes it a very powerful tool [ 265 ]. The GH score ranges from 0 (null model) to 1 (ideal model), a model with a GH score > 0.6 is generally expected to be reliable [ 266 ].

The Receiver operative characteristic (ROC) curve displays the increase of false positives that results with increased true positives. On the Y -axis the true-positive rate ( S e ) is represented, and on the X -axis the false-positive rate (1 −   S p ) is represented. The area under the curve (AUC) is normally used to measure the performance of the model. The greater the AUC (ideal value is 1), the better is the model. An AUC of 0.5 indicates a random database search and thus a poor model ( Figure 9 ) [ 267 ].

An example of Receiver operative characteristic (ROC) curve. Black: random classifier; orange: ideal curve; red: ROC curve.

3.3.2. Pharmacophore Screening

Once the model is validated, databases of small molecules are screened against the pharmacophore model (query) and molecules that match with the features in the model will be extracted and identified as hit compounds. Other than ligand alignment, conformational flexibility is the other big challenge that is encountered in pharmacophore-based virtual screening. In general, conformations of ligands can either be pre-enumerated before the screening process, or conformation search is performed on-the-fly in the pharmacophore fitting process [ 268 ]. Pre-enumeration is less computationally expensive but requires bigger storage space, whereas conformation search on-the-fly is time consuming and requires more computer power. Some common pharmacophore screening programs include Catalyst [ 249 ], Phase [ 250 , 251 ], LigandScout [ 269 ], PharmID [ 270 ].

Recently, Dong and co-workers discovered an anti-fungal inhibitor that can inhibit both squalene cyclooxygenase and CYP51 using pharmacophore modelling. First, a ligand-based common feature pharmacophore model was generated for squalene cyclooxygenase based on seven known inhibitors with diverse scaffolds. Next, a structure-based pharmacophore model for CYP51 was generated from the crystal structure of CYP51 and its interaction with the co-crystalised ligand itraconazole (PBD ID: 5V5Z). Fragments were selected and superimposed onto the pharmacophore features of each of the model and one was constructed by linking different fragments from each of the two models generated, and it was found to inhibit both enzymes simultaneously ( Figure 10 ) [ 271 ].

Chemical structure of 1 . Naphthyl, phenyl and imidazole fragments that match with pharmacophore features of the squalene cyclooxygenase model (blue) and CYP51 model (red), respectively, were connected to give 1 with dual-target inhibition [ 271 ].

Unlike molecular docking, which has well developed scoring functions to estimate binding affinity, pharmacophore screening only assesses how well the ligand matches with the pharmacophore model and that is commonly done by calculating RMSD. Some programs also implement penalties and weightings based on different features, such as the fitness score from phase is based on RMSD, vector terms and volume terms [ 250 , 251 ]. Nevertheless, visual inspection and other criteria such as ADME [ 272 ] and pan-assay interference compounds (PAINS) [ 273 ] are often required to filter inappropriate hits.

3.4. Scaffold Hopping

Scaffold hopping (lead hopping) is a technique that identifies iso-functional molecular structures with significantly different molecular backbones [ 274 ]. The process normally starts with a known active compound, and by replacing with different “cores” (scaffold hopping), a structurally novel compound with similar biological activity is created. The search for alternative cores can be carried out using other LBDD methodologies, such as pharmacophore searching, shape screening and similarity searching using 2D or 3D fingerprints.

In scaffold hopping, the degree of change of the new molecule compared to the original parent molecule ranges from minor changes, such as heterocycle replacement to extensive modifications like topology-based hopping which creates molecules with significantly different scaffold. Sun et al. classified scaffold hopping into four categories based on the degree of modification [ 275 ]. Heterocycle replacement is defined as 1° hopping. Even though there are limited changes in properties of the molecule, it often accompanies a high success rate and an increase in binding affinity to the target protein. 2° hopping involves ring opening and closure which could be useful for adjusting molecular flexibility. 3° hopping are a substitution of pseudopeptides or peptidomimetics that replaces the peptide backbone of the parent molecule with nonpeptic moiety. 4° hopping is topology-based and produces molecules with new chemical backbones to the parent drug, which could present novel properties.

Scaffold hopping is particularly useful in optimising known ligands to improve their efficacy and ADME profile [ 276 ]. Blaquiere et al. discovered novel NF-κB inducing kinase (NIK) inhibitors with improved selectivity and pharmacokinetic properties using the scaffold hopping method. By replacing the oxepin ring in their previously discovered benzoxepine class NIK inhibitors with different cores, novel molecules with reduced nonoxidative metabolism (glutathione conjugation and amide hydrolysis) and thus reduced in vitro clearance were identified [ 277 ]. Scaffold hopping is also an effective strategy to optimise natural products with insufficient levels of activity and high structural complexity to increase their potency and synthetic accessibility. By changing the connectivity of the piperidine ring of natural product evodiamine, Wang and co-workers identified a novel indolopyrazinoquinazolinone scaffold 2 with anti-tumour properties, bringing the IC 50 value from over 200 µM to 47.5 µM when tested against HCT116 cells. Further structural optimisation resulted in a molecule 3 with an IC 50 value of 2 nM ( Figure 11 ) [ 278 ].

Design and optimisation of indolopyrazinoquinazolinone derivatives from evodiamine using scaffold hopping [ 278 ].

4. De Novo and Fragment-Based Drug Design

De novo drug design allows the generation of novel molecules with new scaffolds, especially when majority of small molecule libraries have been exhausted for virtual screening. Before performing de novo drug design, the primary target constraints must be determined first. In SBDD where the structure of the receptor is known, molecular shapes, sub-molecular physical and chemical properties that are important for binding to the active site are extracted to derive shape constraints and interaction sites (normally divided into H-bonds, electrostatic and hydrophobic interactions). In LBDD, pharmacophore features can be used directly in a similarity design method or treated as interaction sites and generate a pseudo-receptor model [ 25 ].

Building blocks used for the generation of molecules can either be atoms or organic fragments. Early programs mainly used atom-based approach that is more likely to encounter issues with synthetic accessibility, but the molecules generated would be more diverse as all chemical space can be sampled. Newer programs use a fragment-based approach that is generally more synthetically feasible but the resulting molecules are relatively less diverse [ 279 ]. Furthermore, using fragments obtained by cleaving drug molecules had shown to generate ligands that are more likely to have drug-like properties [ 280 ]. Some examples of de novo /fragment-based drug design programs include LUDI [ 257 ], LigBuilder [ 281 ], ACFIS [ 282 ] and SEED [ 283 ].

Structure sampling can be carried out in various methods: linking, growing and lattice-based sampling. The linking approach links the building blocks that are positioned at the interaction sites with linker to form a complete molecule [ 257 , 284 ]. The growing approach starts off with one building block that is positioned at one of the interaction sites (starting point), then the structure grows from the starting point, trying to fit suitable interactions for the interaction sites as well as the regions of the receptor between interaction sites [ 285 , 286 ]. The lattice strategy places the binding pocket with lattice points and the ligands are formed from the lattice points that lie along the shortest path that connects the interaction points [ 287 ]. Once the molecules are generated, they can either be assessed with structure-based methods to predict the binding affinity, or with ligand-based methods where the molecules are compared to known active compounds. The ligands are then optimised until a promising drug candidate is produced.

Ni and co-workers discovered a new class of Cyclophilin A (CypA) inhibitors using de novo drug design approach with LigBuilder 2.0. Analysis of existing CypA inhibitors shows that potent inhibitors contain an amide fragment as a linker that forms H-bond interactions with residues between the two sub-binding pockets. Using an acylurea linker as the starting point, new molecules are generated by growing structures from both ends of the structures to occupy the two sub-binding pockets of CypA. Out of the top 98 molecules that were generated, a common scaffold 4 was identified. Compound 4 was found to be potent and was further optimised based on SAR information to give 5 that was 20 times more potent ( Figure 12 ) [ 288 ].

Discovery process of 4 and 5 .

5. Hierarchical Virtual Screening (HLVS)

Both structure-based and ligand-based virtual screening methods have their own strengths and weaknesses. Structure-based methods are dependent on the availability of protein structures and it could be computationally demanding and time consuming for methods, such as MD and flexible molecular docking. Docking-based methods also demonstrate varying performance depending on the nature of the target binding sites [ 289 ]. Ligand-based screening methods on the other hand, rely heavily on the knowledge and information of active ligands and as a result are more biased towards the chemical scaffolds of available active compounds and generate less diverse results. Unlike docking studies where there are well established scoring functions used to approximate binding affinity and to rank molecules, pharmacophore methods lack a reliable and general scoring system. There can also be a lot of variations in models generated in ligand-based approaches, for example a slight difference in ligand selection in the training set could generate a very different QSAR model.

There are clear benefits to combine and integrate different approaches in CADD and the most common way is to use ligand- and structure-based methods in a sequential order, commonly known as hierarchical virtual screening (HLVS) ( Figure 13 ). Generally, ligand-based filters are first applied because they are fast and less computationally expensive. Once the number of candidates is reduced, structure-based methods are applied to further filter inappropriate drug candidates before taking them for biological testing [ 290 ]. The hierarchical combination of pharmacophore modelling and molecular docking are the two most extensively employed methods in HLVS and there are numerous successful examples using this approach, such as the identification of matrix metalloproteinase 2 (MMP2) inhibitors by Di Pizio et al. [ 291 ] as well as the discovery of novel PKR-like endoplasmic reticulum kinase (PERK) inhibitors by Wang et al. [ 292 ].

HLVS approach. A series of filters are sequentially applied to a database of small molecules to reduce the number of molecules to be taken to biological testing and extract lead compounds for further investigation and optimisation.

6. Molecular Mechanical/Generalised Born Surface Area (MM-GBSA)

A more robust way to estimate the binding free energy of ligands to protein is to use the combined Molecular Mechanical/Generalised Born Surface Area (MM-GBSA) approach [ 293 ]. MM-GBSA is a force-field based method that computes the free energy of binding from the difference between the free energies of the protein, ligand, and the complex in solution. The free energy is calculated by using a combination of gas-phase molecular mechanics (MM) energy, electrostatic solvation energy (GB) and non-electrostatic contribution to solvation energy (SA). It provides a more accurate prediction because it can treat both the ligand and protein as flexible, allowing structural rearrangements required for the induced-fit pose. For the same reason, MM-GBSA is more computationally expensive compared to conventional docking studies, therefore they are generally implemented after a completed docking study to re-score selected ligands.

The total binding free energy ΔG bind can be calculated using the following equation: [ 293 ]

where E complex is the energy of the optimised complex, E ligand and E receptor   are the energy of the optimised free ligand and receptor, respectively. This equation can be further broken down into different components of the contributing energies:

where Δ E M M is the changes in the gas-phase molecular mechanics (MM) energy, including changes in the internal energy ( Δ E int ), electrostatic energy ( Δ E ele ) and vdW energy ( Δ E vdW ). Δ G s o l is the sum of electrostatic solvation energy with polar contribution ( Δ G P B / G B ) and non-polar contribution ( Δ G S A ) between the solute and continuum solvent. Δ G P B / G B is calculated using either the Poisson-Boltzmann (PB) or generalised Born (GB) model. Δ G S A is estimated using the solvent-accessible surface area ( SASA ), where γ is the surface tension constant and b is a correction constant. The change in conformational entropy ( − T Δ S ) is calculated by normal-mode analysis [ 294 , 295 , 296 ].

7. Molecular Dynamics

Molecular dynamics (MD) is an in silico simulation method based on molecular mechanics (MM), to study the individual particle motions of model systems over time [ 297 ]. MD can provide insights into biomolecular processes, such as protein folding, conformational changes, ligand binding and disassociation by simulating the interactions between atoms and molecules at an atomic level [ 298 , 299 , 300 , 301 ]. In the context of drug design, simulating the responses of proteins to various perturbations, including mutation [ 302 ], phosphorylation [ 303 ], protonation [ 304 ] and ligand binding [ 305 ], can be observed in well-established models, making MD a powerful tool in understanding the mechanisms for pathogenic or therapeutical processes. Since initially being applied in macromolecules [ 306 ], the application of MD simulation has been extensively developed, both in algorithms and force field parameters. A variety of MD software packages are available such as Gromacs [ 307 ], AMBER [ 308 ], Lammps [ 309 ], NAMD [ 310 ], CHARMM [ 311 ] and Desmond [ 312 ]. These mainstream programs for MD simulation share similar functionalities and have achieved high performance by utilising the compute power and speed of graphics processing units (GPUs). MD has been gradually accepted and is now widely used in pharmaceutical science especially with recent breakthroughs in both structural biology techniques (leading to a larger number of experimentally obtained protein structures) and computational hardware. Currently, MD simulation is integrated into the crucial pilot stages of drug discovery [ 313 ]. Two major usages of MD in recent novel drug design are: (1) to provide dynamic structural insights of biomolecules and (2) to provide precise energetic information of receptor–ligand complexes, key information in lead identification and lead optimisation.

In this context, MD provides valuable time-dependent information on drug targets and their ligands [ 314 ]. MD simulations calculate the position and motion of each atom at each timestep. With accurately controlled simulation conditions, MD can capture the binding processes in action, which are difficult to observe experimentally, provide the details, such as the path in which ligand slides into the binding pocket [ 315 ], how the protein–ligand intermediate state forms and evolves [ 316 ], giving explanation of the binding mechanism at atomic resolution.

For the ligand binding processes, MD always works hand-in-hand with molecular docking [ 317 ]. As previously mentioned, the flexibility of a protein structure is a fundamental factor in both protein biological function and the shape of a binding pocket. However, the initial protein structure used in SBDD are usually the state of the protein acquired from experimental methods, such as X-ray crystal diffraction or cryo-EM [ 313 ]. In reality, different states of the protein exist and the protein dynamics profoundly affect the binding process. Docking into a single static structure would likely retrieve only one subset of promising ligands.

There are two main hypotheses of the ligand recognition: conformational selection and induced fit mechanism, which may coexist in most cases [ 318 ]. MD combined with ensemble docking is one solution to address receptor flexibility by conducting simulations to explore the conformational space and select representative conformations as a receptor ensemble into following dockings. This method is usually integrated in virtual screening workflows to enrich the structural diversity of lead candidates and possible rational binding poses [ 319 ]. Many successful practices of MD-based ensemble docking have been published. Li et al. conducted unrestrained MD simulations on estrogen-related receptor α (ERRα) to obtain structural ensembles for a virtual screening scheme which combines similarity search and ensemble docking. Seven novel scaffolds different from known agonists with remarkable activity were identified [ 320 ]. Recently, machine learning (ML) methodologies were also introduced to boost ensemble docking both on ensemble optimisation [ 321 ] and ligand score aggregation [ 322 ]. On the other hand, methods based on the induced fit mechanism, are also powered by MD simulation. Induced fit docking methods that aim to address the flexibility issue in ligand binding have been successfully utilised in many drug discovery projects [ 323 , 324 ]. However, the poses sampling step of classic IFD still leaves worries on robustness and accuracy. Thus, MD is introduced into the upgraded methodology called IFD-MD to overcome these challenges [ 325 ]. Compared to the traditional IFD protocol, in IFD-MD, short MD simulations are first applied in the rescoring procedure to equilibrate the trial binding models, then metadynamic simulations [ 326 , 327 ] are conducted to assess the local stability. This new method showed promising outcome both in efficiency and accuracy. Zhang et al. discovered dual agonist with nanomolar affinity on both orexin-1 and orexin-2 receptors and performed comprehensive computational modelling studies, including IFD-MD and conventional MD to explore the binding interactions [ 328 ].

Another important objective of MD is capturing conformational changes, particularly those related to important functional processes. As these biomolecular processes usually take place on a larger timescale than conventional MD can sample (within reasonable time and computational cost), several sophisticated MD schemes such as steered MD (sMD) [ 329 ], accelerated MD (aMD) [ 330 ], replica-exchange MD (REMD) [ 331 ] and coarse-grained MD [ 332 ] were developed to overcome the barrier [ 333 ]. In drug discovery, MD is widely used to explain biomolecular mechanisms, such as drug resistance caused by mutations [ 302 , 334 , 335 , 336 ]. Compared to time-consuming experimental method, which only gives static structural information, MD can rapidly provide detailed explanations of the interactions between the ligand and the receptor, including drug–protein or protein–protein interactions and not only structural and dynamical information but also energetic insights. Many studies verified the feasibility of using MD simulations in studying virus resistance mechanisms especially on recent COVID-19 topics. Liu et al. performed an all-atom MD simulation and free energy calculation to explain the resistance mechanisms of SARS-COV-2 variants Delta and Lambda to bamlanivimab [ 337 ].

The design of drugs targeting allosteric sites is another application of MD [ 305 , 338 ]. Allosteric binding sites are usually not as obvious as orthosteric sites from experimentally obtained structures, often due to their reliance on ligand binding and the induced conformational changes [ 333 ]. The formation of cryptic pockets is also considered being adjusted by both mechanisms [ 339 ]: conformational selection based on the flexibility of the cryptic pockets first and then stabilised by ligand as induced fit [ 340 ]. MD simulations have been shown to be of great use in identifying cryptic binding pockets and distinguishing allosteric and orthosteric sites [ 341 , 342 ]. Mixed solvents MD simulation, which uses small molecules/fragments with water as probe, have been successfully applied to detect and characterise allosteric sites [ 343 , 344 ]. Zuzic et al. used molecular dynamics simulations with benzene probes to detect the cryptic pockets in the SARS-CoV-2 spike glycoprotein and successfully identified a potentially druggable cryptic [ 345 ].

Protein misfolding is also an important topic that MD method is deeply involved. Unlike regular protein folding processes which have plenty of well-established solutions including homology modelling [ 58 , 59 , 60 ] and ab initio modelling [ 70 , 71 , 72 , 73 ], the high-resolution dynamic misfolding procedures of intrinsically disordered proteins (IDPs) are extremely difficult to be investigated in experiments for their heterogeneity [ 346 , 347 ]. Among all the cases, the pathological misfolding and aggregation of Alzheimer’s disease (AD) related amyloid-β (Aβ) peptide and tau protein are the most pressing areas for novel therapeutic agent development. Man et al. evaluated the effects of MM force fields on amyloid peptide assembly based on the experimental observation [ 348 , 349 ]. Liu et al. constructed the Markov state model based on the microsecond time scale MD simulation to explore the mechanism of VQIVYK (PHF6) peptide for tau protein aggregation [ 350 ].

MD simulations can also contribute to lead candidates’ optimisation after the initial identification effort. Even though the structural information obtained by molecular docking provides insights into understanding the receptor-ligand interaction, the scoring functions suffered from their approximation in descripting desolvation, entropic penalties and conformational strains [ 351 ], leading to inaccurate energetic results in affinity prediction. The accurate evaluation of receptor–ligand interactions, along with the refinement of the binding complex structures, are needed, and they are becoming a standard protocol at the post-docking stage [ 352 , 353 ]. The purpose of MD optimisation is to fix clashes and stabilise and correct the binding complex, as well as to provide substantially accurate value of binding affinities by MD-based free energy calculations. Regular methods in this field consist of the alchemical approaches, such as thermodynamic integration (TI), free energy perturbation (FEP) [ 354 , 355 , 356 , 357 ] and endpoint approximation methods, such as molecular mechanics Poisson–Boltzmann (generalised Born) surface area (MM/PB(GB)SA) and linear interaction energy (LIE) ( Figure 14 ) [ 295 , 358 , 359 , 360 ].

The distribution of the regular free energy calculation methods in accuracy/efficiency scale and their applications in drug discovery.

TI and FEP are theoretically rigorous methods with highly precise result. However, the sampling of the calculation requires a large amount of computing resources, and the system of computing is also limited because of the complicated simulations setup [ 361 ]. Usually, these methods are used to compare the free energy difference between two given systems with minor modifications, specifically, the lead optimisation process in drug design [ 362 ]. TI and FEP calculate the free energy difference between systems of similar chemical constitutions where the experimental data is not available. To accomplish the calculation, a thermodynamic cycle is introduced to connect the results from a series of TI calculations to experimental observables ( Figure 15 ).

Thermodynamic cycle for relative binding free energy calculation. ΔG 1 and ΔG 2 are the binding energy of reference ligand L 1 and modified ligand L 2 , respectively, ΔG 3 is the free energy difference of two ligands in solution, ΔG 4 is the free energy difference of two ligand–receptor complex in solution.

In this closed thermodynamic cycle, the free energy difference between two ligand binding ΔΔG can be calculated precisely as it is identical to ΔG 4 –ΔG 3 .

For two systems in TI approach, A and B, with potential energies U A and U B , λ is introduced as the coupling parameter with value between 1 and 0, the new potential energy function is defined as

In canonical ensemble, the partition function of the system is

where K B is the Boltzmann constant, N V T means constant number ( N ), volume ( V ), and temperature ( T ).

The free energy of the system is defined as

The free energy between A and B is calculated as

For FEP method, a series of small perturbations, in our cases minor chemical structural modifications, are conducted to link the starting and ending state. The derivation is similar to TI

where q → is variable for coordinates and momentum.

Sophisticated solutions to drug discovery problems are provided by the application of TI and FEP. Nowadays most of the simulation packages support relative free energy simulations, including FEP plus within the Schrödinger suite [ 363 ], AMBER TI [ 364 , 365 ], CHARMM [ 366 ], Gromacs [ 367 ], Q open-source [ 368 ] MD package and so forth. Tang et al. utilised FEP to guide the discovery of novel D-amino acids oxidase inhibitors, with good consistency shown between bioassay results and the energy calculations [ 369 ]. Zou et al. developed a method for scaffold hopping transformations via alchemical free energy calculations, which broaden the usage of such approaches in lead modification and optimisation [ 370 ].

One issue that limits the utility of TI/FEP is the complicated set up for the system. Although the results of TI and FEP calculation are exact in theory, the accuracy is dependent on sampling/studying sufficient intermediate states that provide enough overlaps in each λ window. Different choices of λ are implemented, such as fixed value, slow growth and dynamic modified growth [ 371 ], improving the accuracy of the result, while also significantly increasing the computational cost. Even though GPU accelerated techniques have been applied in TI/FEP calculations [ 372 ], they are still not economically accessible for large datasets. Though MD software offered limited support to setup TI/FEP systems, much effort has been made to assist the preparation for TI/FEP calculations. Automated workflow tools, such as FESetup [ 373 ] for AMBER and Gromacs, PyAutoFEP [ 367 ] for Gromacs, FEPrepare [ 374 ] for NAMD, QligFEP [ 375 ] and QresFEP [ 376 ] for Q and many others [ 377 , 378 , 379 ] provide convenience for researchers to conduct alchemical free energy simulations in drug design.

MM/PB(GB)SA, as an endpoint method which depend on the sampling of the final states of the system, is a good trade-off for computational cost and accuracy in calculating the binding free energy [ 380 ]. The balanced performance makes it popular in broader utilities than the alchemical free energy methods [ 380 , 381 , 382 ].

As the system is always in solvent, Δ G 0 b i n d , s o l v is almost impossible to calculate directly in explicit model because the majority of energetic contribution is made by the solvents instead of the complex, and the fluctuation of total energy is far beyond the binding energy [ 380 ]. Thus, the MM/PB(GB)SA also calculate through the thermodynamic cycle to avoid the problem ( Figure 16 ).

Thermodynamic cycle of binding free energy calculations for protein–ligand complex. ΔG 0 bind,solv is the free energy of interest, solvation energy and binding energy in vacuum are directly calculated terms.

In which the total binding energy can be divided as solvation energy and gas phase MM energy ( Δ G 0 b i n d , v a c u u m ).

For the MM energy

where Δ H 0 is the enthalpy changes in the gas-phase molecular mechanics (MM) energy which is calculated statistically based on the trajectories produced by MD, Δ E 0 i n t stands for the internal energy including the bond, angle and dihedral, Δ E 0 v d W for vdW energy and Δ E 0 e l e for electrostatic energy.   T Δ S 0 is the contribution of entropy, which can be obtained by normal mode analysis, quasi-harmonic analysis or quasi-Gaussian approach.

As for the solvation energy, which consists of the electrostatic, vdW and cavity effects, can be represented as nonpolar and polar terms which is calculated in a different manner.

In which the nonpolar energy Δ G 0 s o l v , n o n p o l a r is easy to estimate, the value is linearly proportional to the solvent-accessible surface area (SASA) because it is basically determined by the interaction with the first layer of solvents. The equation is as follows

where γ (0.00542 kcal/mol Å) and b (0.92 kcal/mol) are constants fitted to experimental data [ 383 ].

The polar solvation energy in implicit solvent Δ G 0 s o l v , p o l a r is estimated by Poisson–Boltzmann (PB) model or Generalised Born (GB) model. In the PB model, a solute is represented by an atomic-detail model as in a MM force field, while the solvent molecules and any dissolved electrolyte are treated as a structure-less continuum [ 308 ]. The continuum treatment represents the solute as a dielectric body whose shape is defined by atomic coordinates and atomic cavity radii [ 384 ]. The electrostatic field can be computed by solving the PB equation: [ 385 ]

where ε ( r ) is the dielectric constant, φ(r) is the electrostatic potential, ρ ( r ) is the solute charge, λ ( r ) is the Stern layer masking function, z i is the charge of ion type i , c i is the bulk number density of ion type i far from the solute, K B is the Boltzmann constant, and T is the temperature; the summation is over all different ion types. The salt term in the PB equation can be linearised when the Boltzmann factor is close to zero but in highly charged systems the PB equation cannot accurately describe the ionic interactions and correlation enhancement. In such systems, full nonlinear PB equation solvers are more appropriated [ 308 ]. The solvation free energy in PB model is represented as [ 386 ]

PB is an approach with standard numerical solution, obtaining results of better accuracy. However, the Poisson–Boltzmann equation needs to be solved every time the conformation changes, and hence the computational costs are relatively high in MD application [ 386 ].

The GB model is an alternative approach with reasonable approximates and good efficiency. Analytic generalised Born method is used to obtain the estimate of the electrostatic energy of solvation, each atom in molecule is represented as a sphere of radius R i with a charge q i in the centre, dielectric constant ε for solute and solvent are 1 and 80, respectively [ 294 , 387 ]. The equation is as below [ 381 , 388 ]

where r i j is the distance between atoms i and j , the R i are the effective Born radii, and f G B () is a certain smooth function of its arguments. The electrostatic screening effects of (monovalent) salt are incorporated via the Debye–Huckel screening parameter κ [ 308 ].

The common representation of f G B ( r i j , R i , R j ) is [ 389 ]

The advantages of MM/PB(GB)SA, including desired balance of accuracy/efficiency and the capability of computing absolute binding energy, give this method a much wider use in drug design. Current computational resources allow for the MM/PB(GB)SA to be implemented into the virtual screening workflow as a re-scoring tool to improve the hit rate [ 390 ]. MM/PB(GB)SA can also help to investigate the binding free energy of many two-component systems such as protein–ligand [ 391 ], protein–protein [ 350 ], protein–DNA systems [ 392 ] and many more. Moreover, binding free energy decomposition and the contribution of each residue, can be estimated in such a method, which gives key residue-specific information of the binding process [ 393 , 394 ].

8. QM/MM and DFT Approaches

Quantum mechanical (QM) and molecular mechanical (MM) calculations can be employed during the drug design process to explore the interaction between ligands and proteins and also how it is processed within the body (ADME). These calculations use a molecular descriptor approach allowing for prediction of ADME properties and modulation in the design process [ 395 ]. All of these factors are a consequence of the electronic interactions within a system. The use of molecular mechanics (MM) approaches has been discussed above. Quantum mechanical (QM) approaches provide more realistic results, often in agreement with experimental data, however at significantly greater computational cost when compared to MM [ 396 ]. QM approaches can be used to not only study the binding poses but also explore the energy landscape of natural processes or drug-receptor processes [ 397 , 398 ]. QM and MM approaches are used in two main ways to study ligand–protein interactions the first of which only utilises QM to analyse a small region of interest such as the binding site, while the second method also uses QM to analyse the region of interest while using the less computationally expensive MM approach to model the remainder of the system [ 397 , 398 ]. The application of pure Density Functional Theory (DFT) or ab initio work is limited due to the expensive computational cost, and as such it is limited to small systems, or for exploring derivable properties [ 399 , 400 ]. However, the application of the hybrid approach allows larger systems to be partitioned with the area of interest (i.e., the active site) being analysed with QM [ 401 ]. DFT is a well-established technique, and the experimental design needs to be in line with size, and property being explored for the system. DFT is computationally more efficient and accurate relative to QM ( ab initio ) methods. QM attempts to solve the Schrodinger equation to model the behaviour of the system and this is a non-trivial task for systems where N > 1 (where N is the number of electrons in the system). This can be highly accurate depending on the method employed (i.e., Moller–Plesset vs. Hartree Fock(HF)), however the equation cannot be fully solved as electron correlation effects ( E X C ) are unaccounted for. In contrast, DFT explores electronic behaviour of a molecule or system as function of the electronic density, with the energy being directly relatable [ 402 ]. This approach allows for much faster generation of a wavefunction to review and the accuracy is dependent on the functional applied. Recently, Bursch et al. provided a thorough review of the functionals and basis set selection for DFT application [ 403 ]. Commonly, for DFT in drug design setting, a hybrid functional (Equation (38)) is used

In Equation (38) the α coefficient determines the amount of the exact exchange ( E X H F ) derived from first principles that is mixed with the semi-local exchange ( E X G G A ) . This combined approach was proposed by Becke in 1993, with the first approach being a 50/50 mix for HF and semi-local E X energies [ 404 ]. Since, the hybrid functional have grown significantly with an area of HF percentage amounts, commonly, it is between 20–30% [ 403 ]. The most commonly applied hybrid is the B3LYP functional containing a scalable 20% HF component [ 404 , 405 , 406 ]. Alongside a functional is a basis set, which provides numerical functionals for the molecular orbital shape and the occupational. Most commonly a split basis sets such as the 6-311 family is employed [ 407 , 408 , 409 , 410 ]. Albeit currently, functionals and basis sets of much higher complexity are being benchmarked and tested (i.e., Coupled-Cluster (CCSD(T), aug-cc-PVDZ, respectively) [ 411 ].

The hybrid approaches are more computationally efficient with the trade-off of reduced accuracy in regions away from the active site. However, these methods allow for entire system to be reviewed. QM/MM considers the whole system as conceptually two parts. The active/model region, which uses QM (DFT, commonly). The remaining region is studied using MM (force field approaches), with the boundary between both sites being the QM/MM interactions. The resultant energy of the system takes the form,

Here, the system energy ( E s y s ) is the summative total of the QM, MM and interface region, respectively. It is obvious that the QM region is more computationally demanding, whilst the peripheries are much more efficient. This approach has been present since the 70s and its impact resulted in a Nobel prize being awarded to Karplus, Levitt and Warshel. [ 412 , 413 ] Since the two-part method, the QM/MM scheme has developed further to the currently more applied, which is our own n-layered Integrated molecular Orbital and Molecular mechanics (ONIOM) and comparative approaches. The ONIOM approach splits a system up to N-layers, with the inner layers closer to the active-site having an electronic density/energy calculated at a higher level of computational theory [ 401 , 414 ]. The MM analyses can be further enhanced by the addition of polarisation terms, solvating the system and even searching for excited states. A thorough review of the ONIOM is provided by Chung et al. and its vast application can be read there [ 401 ]. ONIOM has been applied in many computational packages, such as Gaussian and ORCA [ 415 , 416 ]. The boundary selection can be cumbersome, and considering the residue type and possible interactions it produces to influence the level of theory applied [ 401 ]. For drug design, ONIOM can be applied to provide energetic information in both structure- and ligand-based approaches. For structure-based approaches, the application of QM or QM/MM can be used to study enzymatic processes, when considering it as an outcome of energy [ 417 , 418 ]. At the core, the understanding of the Michaelis-Menten mechanistic scheme can be used to find rate constants between states [ 417 ]. In QM setting this is found by understanding the change in potential energy surface (PES) between states. This approach extends on how ligands interact with targets to understand how the PES is modified or overcome by generation or outcompeting of bonds [ 400 , 418 , 419 , 420 , 421 ]. Extending from PES alone, application analytical tools related to the properties of the wavefunction can be used to describe, modify and improve ligands. The use of frontier molecular orbital (FMO)s can be used to explore the electron donation ability of the ligand by analysis of the HOMO (highest occupied molecular orbital) and LUMO (lowest unoccupied molecular orbital), this can explore priori and postori energy of a ligand on binding [ 399 ]. Separately, the interactions present between ligand and receptor can be separated into energetic types to understand how bonding variations occur via R-group selection [ 399 ]. Other than ligand interactions and structure-based phenomena, DFT can be utilised to make predictions on binding affinities, pKa, IC 50 , DFT-assisted QSAR, drug-interactions, delivery enhancement and ADME properties [ 401 , 414 , 422 , 423 , 424 , 425 , 426 , 427 , 428 , 429 ]. An example of ADME using DFT is the prediction of pKa [ 430 ]. DFT was used on the SAMPL6 bind test based on DFT alone and the error ranges were quite large (2–4 pKa units) [ 430 ]. Although, when using conceptual DFT (combining molecular descriptors with the DFT results) [ 431 ] in a machine learning model, predictions were improved and allowed for extension of the technique to be used for the prediction of non-acidic compounds as well. This approach overall lowered the errors to ~1.85 pKa units [ 432 ]. ADME predictions can also be made by utilising global reactivity descriptors, such as the Fukui Functions. This approach allows for the electron density of the molecule to be broken into neutral, positive or negative, which correlates to compounds that can cause electrophilic attack processes to aid in understanding toxicity [ 432 , 433 ]. Although less common in its use, QM/MM approaches have been pivotal in understanding many health burdens, such as, bacteria resistance, and HIV virus proteases process as two examples [ 422 , 434 , 435 , 436 , 437 , 438 , 439 , 440 , 441 ]. Noting its importance and success in many aspects, DFT or QM/MM approaches although currently under used, are growing in application due to improvement in computation resources. The application of QM or QM/MM can have ample benefit in drug design exploring how and why a process mechanistically occurs.

9. Conclusions

Recent advances in computational software and hardware have revolutionised the use of in silico methods in drug design, with access to high-performance computers allowing for more complex calculations and larger data sets to be feasibly processed. In this review, we have highlighted a range of in silico methods that are commonly used in the hit identification and lead optimisation stages of the drug design process, yet computational methods are also applied in other areas in the pipeline. Some examples include drug repurposing [ 442 , 443 ], protein–protein docking, de novo protein design, inverse docking [ 444 ], adverse events prediction, physiologically-based pharmacokinetic modelling, and guiding chemical synthesis [ 442 , 443 ].

In addition to classical CADD strategies, such as molecular docking and pharmacophore screening, more accurate and computationally expensive methods, such as MD, DFT, and MM/PB(GB)SA, are now routinely used to further analyse short-listed compounds to better predict binding interactions and docking energies, highlighting compounds which guide us into selecting and optimising the lead compound with the highest success rate.

With the rapid development in artificial intelligence, deep learning-based approaches in drug design have become a trending topic, and various of these strategies were developed for molecular docking [ 445 , 446 ], property prediction [ 447 , 448 ], compound retrosynthesis [ 449 , 450 , 451 ], de novo drug design [ 452 , 453 ] and many more. Although the benefits of incorporating machine learning elements have been highlighted in recent years, there are still certain limitations in these approaches. The training of an algorithm relies heavily on a large amount of data, and therefore the availability of a comprehensive and high-quality dataset directly impacts the performance of the algorithm. Many of the more complex and recent models which utilise machine learning capabilities lack transparency due to their “black box” nature, and the results are not always able to be rationally interpreted and applied, thus limiting the scope of their potential applications in rational drug discovery and design [ 454 ]. Nevertheless, the development machine learning based CADD methodologies will be one of the major focuses in the future to continue improving current strategies and to overcome existing challenging barriers in the drug discovery process.

Funding Statement

This research was supported by the Faculty of Medicine and Health, University of Sydney.

Author Contributions

F.L., Y.C., B.A.H. and J.J.D. wrote the manuscript under the supervision of P.W.G. and D.E.H. All authors have read and agreed to the published version of the manuscript.

Institutional Review Board Statement

Informed consent statement, data availability statement, conflicts of interest.

The authors declare no conflict of interest.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

• - Google Chrome

Intended for healthcare professionals

• Access provided by Google Indexer
• My email alerts
• BMA member login
• Username * Password * Forgot your log in details? Need to activate BMA Member Log In Log in via OpenAthens Log in via your institution

Search form

• Advanced search
• Search responses
• Search blogs
• Multiple adverse...

Multiple adverse outcomes associated with antipsychotic use in people with dementia: population based matched cohort study

Linked editorial.

Use of antipsychotics in adults with dementia

• Related content
• Peer review
• Pearl L H Mok , research fellow 1 2 ,
• Matthew J Carr , research fellow 1 2 3 ,
• Bruce Guthrie , professor 4 ,
• Daniel R Morales , Wellcome Trust clinical research fellow 5 ,
• Aziz Sheikh , professor 6 7 ,
• Rachel A Elliott , professor 3 8 ,
• Elizabeth M Camacho , senior research fellow 8 ,
• Tjeerd van Staa , professor 9 ,
• Anthony J Avery , professor 3 10 ,
• Darren M Ashcroft , professor 1 2 3
• 1 Centre for Pharmacoepidemiology and Drug Safety, Division of Pharmacy and Optometry, University of Manchester, Manchester, M13 9PT, UK
• 2 Manchester Academic Health Science Centre, Manchester, UK
• 3 NIHR Greater Manchester Patient Safety Research Collaboration, University of Manchester, Manchester, UK
• 4 Advanced Care Research Centre, Usher Institute, College of Medicine and Veterinary Medicine, University of Edinburgh, Edinburgh, UK
• 5 Population Health and Genomics, University of Dundee, Dundee, UK
• 6 Usher Institute, College of Medicine and Veterinary Medicine, University of Edinburgh, Edinburgh, UK
• 7 Nuffield Department of Primary Care Health Sciences, University of Oxford, Oxford, UK
• 8 Manchester Centre for Health Economics, Division of Population Health, Manchester, UK
• 9 Division of Informatics, Imaging and Data Sciences, University of Manchester, Manchester, UK
• 10 Centre for Primary Care, School of Medicine, University of Nottingham, Nottingham, UK
• Correspondence to: P L H Mok pearl.mok{at}manchester.ac.uk
• Accepted 29 February 2024

Objective To investigate risks of multiple adverse outcomes associated with use of antipsychotics in people with dementia.

Design Population based matched cohort study.

Setting Linked primary care, hospital and mortality data from Clinical Practice Research Datalink (CPRD), England.

Population Adults (≥50 years) with a diagnosis of dementia between 1 January 1998 and 31 May 2018 (n=173 910, 63.0% women). Each new antipsychotic user (n=35 339, 62.5% women) was matched with up to 15 non-users using incidence density sampling.

Main outcome measures The main outcomes were stroke, venous thromboembolism, myocardial infarction, heart failure, ventricular arrhythmia, fracture, pneumonia, and acute kidney injury, stratified by periods of antipsychotic use, with absolute risks calculated using cumulative incidence in antipsychotic users versus matched comparators. An unrelated (negative control) outcome of appendicitis and cholecystitis combined was also investigated to detect potential unmeasured confounding.

Results Compared with non-use, any antipsychotic use was associated with increased risks of all outcomes, except ventricular arrhythmia. Current use (90 days after a prescription) was associated with elevated risks of pneumonia (hazard ratio 2.19, 95% confidence interval (CI) 2.10 to 2.28), acute kidney injury (1.72, 1.61 to 1.84), venous thromboembolism (1.62, 1.46 to 1.80), stroke (1.61, 1.52 to 1.71), fracture (1.43, 1.35 to 1.52), myocardial infarction (1.28, 1.15 to 1.42), and heart failure (1.27, 1.18 to 1.37). No increased risks were observed for the negative control outcome (appendicitis and cholecystitis). In the 90 days after drug initiation, the cumulative incidence of pneumonia among antipsychotic users was 4.48% (4.26% to 4.71%) versus 1.49% (1.45% to 1.53%) in the matched cohort of non-users (difference 2.99%, 95% CI 2.77% to 3.22%).

Conclusions Antipsychotic use compared with non-use in adults with dementia was associated with increased risks of stroke, venous thromboembolism, myocardial infarction, heart failure, fracture, pneumonia, and acute kidney injury, but not ventricular arrhythmia. The range of adverse outcomes was wider than previously highlighted in regulatory alerts, with the highest risks soon after initiation of treatment.

Introduction

Dementia is a clinical syndrome characterised by progressive cognitive decline and functional disability, with estimates suggesting that by 2050 around 152.8 million people globally will be affected. 1 Behavioural and psychological symptoms of dementia are common aspects of the disease and include features such as apathy, depression, aggression, anxiety, irritability, delirium, and psychosis. Such symptoms can negatively impact the quality of life of patients and their carers and are associated with early admission to care. 2 3 Antipsychotics are commonly prescribed for the management of behavioural and psychological symptoms of dementia, despite longstanding concerns about their safety. 4 5 6 During the covid-19 pandemic, the proportion of people with dementia prescribed antipsychotics increased, possibly owing to worsened behavioural and psychological symptoms of dementia linked to lockdown measures or reduced availability of non-pharmaceutical treatment options. 7 According to guidelines from the UK’s National Institute for Health and Care Excellence, antipsychotics should only be prescribed for the treatment of behavioural and psychological symptoms of dementia if non-drug interventions have been ineffective, if patients are at risk of harming themselves or others or are experiencing agitation, hallucinations, or delusions causing them severe distress. 8 Antipsychotics should at most be prescribed at the lowest effective dose and for the shortest possible time. Only two antipsychotics, risperidone (an atypical, or second generation, antipsychotic) and haloperidol (a typical, or first generation, antipsychotic), are licensed in the UK for the treatment of behavioural and psychological symptoms of dementia, 9 although others have been commonly prescribed off-label. 5 10

Based on evidence from clinical trials of risperidone, the US Food and Drug Administration (FDA) first issued a warning in 2003 about the increased risks of cerebrovascular adverse events (eg, stroke, transient ischaemic attack) associated with use of atypical antipsychotics in older adults with dementia. 11 A meta-analysis of 17 trials among such patients subsequently found a 1.6-1.7-fold increased risk of mortality with atypical antipsychotics compared with placebo, which led the FDA to issue a “black box” warning in 2005 for all atypical antipsychotics. 11 This warning was extended to typical antipsychotics in 2008, after two observational studies reported that the risk of death associated with their use among older people might be even greater than for atypical antipsychotics. 12 13 14 The increased risks for stroke and mortality have been consistently reported by many observational studies and meta-analyses since, 11 15 16 17 18 19 20 21 and they have led to regulatory safety warnings and national interventions in the UK, US, and Europe, aiming to reduce inappropriate prescribing of these drugs for the treatment of behavioural and psychological symptoms of dementia. 8 11 22 23 24 25 26 Other adverse outcomes have also been investigated in observational studies, 27 28 29 although, with the exception of pneumonia, 14 30 31 32 the evidence is less conclusive or is more limited among people with dementia. For example, inconsistent or limited evidence has been found for risks of myocardial infarction, 33 34 ventricular arrhythmia, 35 36 venous thromboembolism, 37 38 39 40 fracture, 41 42 43 and acute kidney injury. 44 45 46 Most studies also reported only one outcome or type of outcomes. Examining multiple adverse events in a single cohort is needed to give a more comprehensive estimate of the total potential harm associated with use of antipsychotics in people with dementia.

Using linked primary and secondary care data in England, we investigated the risks of a range of adverse outcomes potentially associated with antipsychotic use in a large cohort of adults with dementia—namely, stroke, venous thromboembolism, myocardial infarction, heart failure, ventricular arrhythmia, fracture, pneumonia, and acute kidney injury. We report both relative and absolute risks.

Data sources

The study used anonymised electronic health records from Clinical Practice Research Datalink (CPRD). In the UK, residents are required to be registered with a primary care general practice to receive care from the NHS. The NHS is a publicly funded healthcare service, free at the point of use. More than 98% of the UK population are registered with a general practice, and their electronic health records are transferred when they change practice. 47 48 Community prescribing is most often done by the general practitioner, including antipsychotic treatment recommended by specialists. CPRD data are sourced from more than 2000 general practices covering around 20% of the UK population, and include information on diagnoses, primary healthcare contacts, prescribed drugs, laboratory test results, and referrals to secondary healthcare services. 47 48 CPRD contains two databases: Aurum and GOLD. CPRD Aurum includes data from contributing general practices in England that use the EMIS Web patient management software, and CPRD GOLD consists of patient data from practices across all four UK nations that use the Vision system. Both datasets are broadly representative of the UK population. 47 48 49 Primary care data from general practices in England can be linked to other datasets, including hospital admissions in Hospital Episode Statistics, and mortality and index of multiple deprivation data from the Office for National Statistics (ONS). Individual patients can opt-out of sharing their records with CPRD, and individual patient consent was not required as all data were deidentified.

Study population

We delineated two cohorts, one each from Aurum and GOLD. For the latter, we included patients from English practices only because linkage to hospital admission and mortality data were required in our analyses. To ensure that the study dataset would not contain any duplicate patient records, we used the bridging file provided by CPRD to identify English practices that have migrated from the GOLD to the Aurum dataset, and removed such practices from the GOLD dataset. For both cohorts, we included patients who had a first dementia diagnosis code between 1 January 1998 and 31 May 2018. Dementia was identified from Read, SNOMED, or EMIS codes used in the databases (see supplementary appendix). We defined the date of first dementia diagnosis as the date of first dementia code. Patients needed to be aged 50 years or over at the time of dementia diagnosis, have been registered with the CPRD practice for at least a year, not be prescribed an antipsychotic in the 365 days before their first dementia code, and have records that were eligible for linkage to Hospital Episodes Statistics, mortality, and index of multiple deprivation data. In addition, because anticholinesterases (such as donepezil, rivastigmine, and galantamine) may sometimes be prescribed to patients showing signs of dementia before their first dementia code, we excluded patients with an anticholinesterase prescription before their first dementia code. Supplementary figures S1 and S2 show how the two cohorts for Aurum and GOLD, respectively, were delineated.

Study design

Matched cohort design —We implemented a matched cohort design. Supplementary figure S3 shows the study design graphically. 50 For the Aurum and GOLD cohorts separately, patients who used antipsychotics were defined as patients in each cohort issued with an antipsychotic prescription after (or on the same day as) the date of their first dementia diagnosis, with the date of first antipsychotic prescription being the index date after which outcomes were measured. For each outcome, follow-up began from the date of the first antipsychotic prescription (the index date) and ended on the earliest of date of first diagnosis of outcome (ie, the earliest recording of the outcome whether it was from the patient’s primary or secondary care or mortality records), death, transfer out of the general practice, last data collection date of the general practice, two years from the date of antipsychotics initiation, or 31 May 2018. Because patients who have experienced an outcome were potentially at higher risk of subsequently experiencing the same event, which could confound any risks associated with antipsychotic use, we excluded those with a history of the specific outcome under investigation before the index date from the analysis of that outcome. For example, we excluded patients with a record of stroke before the index date from the analysis of stroke, but they would still be eligible for the study of other outcomes. For the analysis of acute kidney injury, patients with a diagnosis of end stage kidney disease before the index date were also excluded, and a diagnosis of end stage kidney disease after the index date was an additional condition for end of follow-up. 44

Matched comparators —Each patient who used antipsychotics on or after the date of their first dementia diagnosis was matched using incidence density sampling with up to 15 randomly selected patients who had the same date of first dementia diagnosis (or up to 56 days after) and who had not been prescribed an antipsychotic before diagnosis. Incidence density sampling involves matching on sampling time, with each antipsychotic user in our study being matched to one or more comparators who were eligible for an antipsychotic but had not become a user at the time of matching. 51 The selection of comparators was done with replacement—that is, an individual could be used as a comparator in multiple matched sets. In our study, this meant that patients were eligible to be a non-user matched comparator up to the date of their first antipsychotic prescription. We excluded matched comparators with a history of the specific outcome under investigation before the index date from the analysis of that event. For each outcome, follow-up of matched comparators began on the same day as the patient to whom they were matched (the index date) and ended on the earliest of date of their first antipsychotic prescription (if any), or date of one of the end of follow-up events described earlier for the antipsychotic users.

Use of antipsychotics

We included both typical and atypical antipsychotics, identified by product codes in Aurum and GOLD (see supplementary appendix for list of drugs included). Senior author DMA (pharmacist) reviewed the code lists. As previous studies have shown a temporal association between antipsychotic use and development of adverse outcomes, 30 31 52 we treated use of antipsychotics as a time varying variable, classified as current, recent, and past use. Current use was defined as the first 90 days from the date of an antipsychotic prescription, recent use as up to 180 days after current use ended, and past use as the time after the recent use period had ended. If a patient was issued another prescription during the 90 days after their last prescription, their current use period would be extended by 90 days from the date of their latest prescription. For example, if a patient had two prescriptions and the second was issued 60 days after the first, their current use period would be a total of 150 days: 60 days after the first prescription plus 90 days after the second. At the end of the 150 days current use period, the next 180 days would be the recent use period, and the time after this recent use period would be past use. As patients could have multiple prescriptions over time, they could move between the three antipsychotic use categories during follow-up, and they could therefore be defined as current, recent, or past users more than once. See the supplementary appendix for further information on how this definition is applied.

In post hoc analyses, we also investigated typical versus atypical antipsychotics, and specific drug substances: haloperidol, risperidone, quetiapine, and other antipsychotics (as a combined category).

Outcomes were stroke, venous thromboembolism (including deep vein thrombosis and pulmonary embolism), myocardial infarction, heart failure, ventricular arrhythmia, fracture, pneumonia, and acute kidney injury. With the exceptions of pneumonia and acute kidney injury, outcomes were identified by Read, SNOMED, or EMIS codes in the primary care records, and by ICD-10 (international classification of diseases, 10th revision) codes from linked secondary care data from Hospital Episodes Statistics, and cause of death data from the ONS mortality records. For pneumonia and acute kidney injury, we only included those that were diagnosed in hospitals or as a cause of death, ascertained from Hospital Episodes Statistics and ONS data.

We also investigated appendicitis and cholecystitis combined as an unrelated (negative control) outcome to detect potential unmeasured confounding. 53 These outcomes were chosen because evidence of an association with antipsychotic use is lacking from the literature. We identified appendicitis and cholecystitis from Read, SNOMED, EMIS, and ICD-10 codes. Clinicians (BG, AJA, DRM) checked all code lists (see supplementary appendix).

We used propensity score methods to control for imbalances in measurable patient characteristics between antipsychotic users and their matched non-users, with personal characteristics, lifestyle, comorbidities, and prescribed drugs included in the propensity score models. A counterfactual framework for causal inference was applied to estimate the average treatment effect adjusting for inverse probability of treatment weights generated from the propensity score models. 54 55 Selection of covariates was informed by the literature, based on their potential associations with antipsychotic initiation and study outcomes. 31 34 44 56 57 All variables were assessed before the index date (see supplementary figure S3). Variables for personal characteristics included sex, age at dementia diagnosis, age at start of follow-up, ethnicity, and index of multiple deprivation fifths based on the location of the general practice. Comorbidities were derived as dichotomous variables and included a history of hypertension, types 1 and 2 diabetes mellitus, chronic obstructive pulmonary disease, rheumatoid arthritis, moderate or severe renal disease, moderate or severe liver disease, atrial fibrillation, cancer, and serious mental illness (bipolar disorders, schizophrenia, schizoaffective disorders, and other psychotic disorders). Lifestyle factors included smoking status and alcohol use. Medication covariates were represented as dichotomous indicators, defined by at least two prescriptions for each of the following drugs in the 12 months before the index date: antiplatelets, oral anticoagulants, angiotensin converting enzyme inhibitors or angiotensin II receptor blockers, alpha blockers, beta blockers, calcium channel blockers, diuretics, lipid lowering drugs, insulin and antidiabetic drugs, non-steroidal anti-inflammatory drugs, antidepressants, benzodiazepines, and lithium. We also included the following potential confounders for the investigations of venous thromboembolism and fracture: prescriptions for hormone replacement therapy and selective oestrogen receptor modulators (for venous thromboembolism), 58 59 a history of inflammatory bowel disease (for pneumonia and fracture), 60 61 and prescriptions for immunosuppressants, oral corticosteroids, and inhaled corticosteroids (for pneumonia). 62 63

Statistical analysis

For each patient included in the study, we derived a propensity score representing the patient’s probability of receiving antipsychotic treatment. Propensity scores were estimated using multivariable logistic regression, with antipsychotic use as the dependent variable. Predictors included personal characteristics, lifestyle, comorbidities, and prescribed drugs. Patients with missing information on ethnicity, index of multiple deprivation, smoking, or alcohol use were grouped into an unknown category for each of these variables and included in the propensity score models. We used the Hosmer-Lemeshow test and likelihood ratio test to test the fit of the models, and interaction terms were included to improve the model fit. 64 The derived scores were used as inverse probability of treatment weights to reweigh the data, balancing the distribution of baseline covariates between antipsychotic users and non-users (matched comparators)—that is, standardised differences <0.1 after weighting. 65 Propensity score models were run for each outcome, and for the Aurum and GOLD cohorts separately. For further information, see the supplementary appendix section on propensity score methods to control for potential confounding.

Analyses for estimating harms were then conducted after combining (appending) the Aurum and GOLD datasets. We used Cox regression survival analyses to estimate the risks of each outcome associated with antipsychotic use relative to the comparator cohort, and we report the results as hazard ratios. Use of an antipsychotic was treated as a time varying variable. To account for the matched design, we fitted stratified models according to the matched sets and used robust variance estimation. In all models, we also included a covariate indicating whether the patient was from the Aurum or GOLD cohort and calculated hazard ratios with adjustments for inverse probability of treatment weights. Cox regression assumes proportional hazards—that is, the relative hazard of the outcome remains constant during the follow-up period. 66 We assessed this assumption using the Grambsch-Therneau test based on the Schoenfeld residuals. 67 Because this assumption did not hold for all outcomes examined, in addition to reporting the hazard ratios pertaining to the whole follow-up period, we estimated hazard ratios separately for the several time windows: the first seven days, 8-30 days, 31-180 days, 181-365 days, and 366 days to two years (see supplementary appendix for an illustration of stratification of follow-up time). For each outcome, we calculated the incidence rate and the number needed to harm (NNH) over the first 180 days as well as two years after start of follow-up. The NNH represents the number of patients needed to be treated with an antipsychotic for one additional patient to experience the outcome compared with no treatment. We also calculated cumulative incidence percentages (absolute risks) for each outcome accounting for competing mortality risks based on previous recommendations. 68 These were calculated at 90 days, 180 days, 365 days, and two years after start of follow-up for antipsychotic users and their matched comparators separately. We also reported the difference in cumulative incidence between antipsychotic users and their matched comparators at these time points. Analyses were conducted using Stata/MP v16.1.

Sensitivity analyses

We investigated two other definitions of antipsychotic use as sensitivity analyses: the first 60 days as current use followed by 120 days of recent use, and a current use period of 30 days followed by a recent use period of 60 days. We also conducted the following post hoc sensitivity analyses. Firstly, as levomepromazine is often prescribed in palliative care to treat distressing symptoms in the last days of life, 69 we censored individuals at the time of their first levomepromazine prescription. Secondly, we used Fine-Gray subdistribution hazard regression models to estimate the hazard of each adverse outcome, accounting for the competing risks of death. 70 These results were reported as subhazard ratios. Thirdly, we compared the incidence rates and hazards of adverse outcomes for male versus female individuals. For these sex specific analyses, we modified the existing matched cohort by excluding non-user comparators who were of a different sex from the antipsychotic user to whom they were matched. We then derived a new propensity score for each individual by excluding sex as a covariate in the propensity score models. Incidence rate ratios and corresponding 95% confidence intervals (CIs) for male versus female individuals were calculated using the ‘iri’ command in Stata. To investigate whether hazards of each adverse outcome associated with antipsychotic use differed by sex, we fitted Cox regression models with sex, antipsychotic use, and their interaction as covariates. Sex specific hazard ratios and ratios of male to female hazard ratios were reported.

Patient and public involvement

This study is part of a National Institute of Health and Care Research funded programme (RP-PG-1214-20012): Avoiding patient harm through the application of prescribing safety indicators in English general practices (PRoTeCT). Two patient and public involvement members in the project team contributed to the study design and protocol of this study. Our study was not, however, coproduced with people with dementia or their carers.

Characteristics of study population

A total of 173 910 adults (63.0% women) with dementia were eligible for inclusion in the study: 139 772 (62.9% women) in the Aurum dataset and 34 138 (63.4% women) in GOLD. The mean age at dementia diagnosis for individuals in both cohorts was 82.1 years (standard deviation (SD) 7.9 years), and the median age was 83 years (interquartile range (IQR) 78-88 years in Aurum and 78-87 years in GOLD). A total of 35 339 individuals (62.5% women; 28 187 in Aurum, 62.6% women; 7152 in GOLD, 62.5% women) were prescribed an antipsychotic during the study period, and a matched set was generated for each of these individuals. The mean number of days between first dementia diagnosis and date of a first antipsychotic prescription was 693.8 ((SD 771.1), median 443 days) in Aurum and 576.6 ((SD 670.0), median 342 days) in GOLD. A total of 544 203 antipsychotic prescriptions (433 694 in Aurum, 110 509 in GOLD) were issued, of which 25.3% were for a typical antipsychotic and 74.7% for an atypical antipsychotic. The most prescribed antipsychotics were risperidone (29.8% of all prescriptions), quetiapine (28.7%), haloperidol (10.5%), and olanzapine (8.8%), which together accounted for almost 80% of all prescriptions (see supplementary table S1).

Since we excluded people with a history of the event before the start of follow-up, the number of individuals and matched sets included in analysis varies by outcome. Table 1 shows the baseline characteristics of patients for the analysis of stroke, before and after inverse probability of treatment weighting. Antipsychotic users were more likely than their matched comparators to have a history of serious mental illness and to be prescribed antidepressants or benzodiazepines in the 12 months before start of follow-up. After inverse probability of treatment weighting, standardised differences were <0.1 for all covariates. Baseline characteristics of individuals included in the analyses of other outcomes were similar to those reported for stroke (see supplementary tables S2-S9).

Baseline characteristics of antipsychotic users and matched comparators included in the analysis of stroke (CPRD Aurum and GOLD combined data). Values are number (percentage) unless stated otherwise

• View inline

Incidence rates and relative hazards of adverse outcomes

All antipsychotics.

In the two years after initiation of antipsychotics, the highest incidence rates of adverse outcomes were for pneumonia, fracture, and stroke, and ventricular arrhythmias were rare ( table 2 ). Figure 1 shows the hazard ratios of adverse outcomes associated with current, recent, past, and any use of antipsychotics versus non-use (ie, matched comparators). Except for ventricular arrhythmia, any use of antipsychotics was associated with increased risks for all adverse outcomes, ranging from a hazard ratio of 2.03 (95% CI 1.96 to 2.10) for pneumonia to 1.16 (1.09 to 1.24) for heart failure. Current use (ie, prescribed in the previous 90 days) was associated with high risks for pneumonia (2.19, 2.10 to 2.28), acute kidney injury (1.72, 1.61 to 1.84), venous thromboembolism (1.62, 1.46 to 1.80), and stroke (1.61, 1.52 to 1.71). Recent antipsychotic use (ie, in the 180 days after current use ended) was also associated with increased risk for these outcomes, as well as for fracture, but past use of antipsychotics (ie, after recent use ended) was not associated with increased risks of the adverse outcomes examined, except for pneumonia. For the negative control outcome (appendicitis and cholecystitis), no significant associations were found with current, recent, or any antipsychotic use, but a statistically significant association was observed with past use (1.90, 1.01 to 3.56).

Incidence rate (per 10 000 person years) and number needed to harm of adverse outcomes associated with antipsychotic use during the first 180 days and two years of follow-up period

Hazard ratios (adjusted for inverse probability of treatment weights) of adverse outcomes associated with current, recent, and past antipsychotic use; with current use being defined as the first 90 days from the date of an antipsychotic prescription, recent use as up to 180 days after current use ended, and past use as after recent use. CI=confidence interval

• Download figure
• Open in new tab
• Download powerpoint

Table 2 shows that the NNH ranged from 9 (95% CI 9 to 10) for pneumonia to 167 (116 to 301) for myocardial infarction during the first 180 days after initiation of antipsychotics, and from 15 (14 to 16) for pneumonia to 254 (183 to 413) for myocardial infarction after two years. These figures suggest that over the 180 days after drug initiation, use of antipsychotics might be associated with one additional case of pneumonia for every nine patients treated, and one additional case of myocardial infarction for every 167 patients treated. At two years, there might be one additional case of pneumonia for every 15 patients treated, and one additional case of myocardial infarction for every 254 patients treated.

Table 3 shows hazard ratios stratified by follow-up time (except for ventricular arrhythmia and the negative control where the number of patients was very low). For almost all outcomes, relative hazards were highest in the first seven days after initiation of antipsychotic treatment. Risks for pneumonia were particularly increased in the first seven days (9.99, 8.78 to 11.40) and remained substantial afterwards (3.39, 3.04 to 3.77, 8-30 days). No increased risks for heart failure were found for current users after 180 days from start of treatment, nor for myocardial infarction one year after drug initiation. However, risks for stroke, venous thromboembolism, fracture, pneumonia, and acute kidney injury remained increased among continuous antipsychotic users up to two years after initiation of treatment.

Hazard ratios (adjusted for IPT weights) of adverse outcomes associated with current, recent, and past antipsychotic use stratified by follow-up period

Types of antipsychotics

During the current use period of 90 days after a prescription, both typical and atypical antipsychotics were associated with increased risks of all adverse outcomes compared with non-use, except for ventricular arrhythmia and the negative control (see supplementary table S10). Hazards were higher when current use of typical antipsychotics was directly compared with atypical antipsychotics for stroke (1.23, 1.09 to 1.40), heart failure (1.18, 1.01 to 1.39), fracture (1.22, 1.08 to 1.38), pneumonia (1.92, 1.77 to 2.08), and acute kidney injury (1.22, 1.05 to 1.42), but no significant differences between the two types of drug were found for the risks of venous thromboembolism or myocardial infarction.

Supplementary table S11 shows the risks of adverse outcomes associated with haloperidol (the most prescribed typical antipsychotic) and with risperidone and quetiapine (the two most prescribed atypical antipsychotics). Current use of risperidone and haloperidol compared with non-use was associated with increased risks of all adverse outcomes except for ventricular arrhythmia and the negative control. Current use of quetiapine compared with non-use was associated with increased risks for fracture, pneumonia, and acute kidney injury. Among current users of haloperidol or risperidone, risks for fracture, pneumonia, and acute kidney injury were higher for haloperidol versus risperidone, but risks for stroke, venous thromboembolism, myocardial infarction, and heart failure were similar for both drugs. With the exceptions of myocardial infarction, ventricular arrhythmia, and the negative control, risks of all adverse outcomes were higher for haloperidol than for quetiapine, especially for pneumonia (2.53, 2.21 to 2.89) and venous thromboembolism (1.99, 1.33 to 2.97). Among current users of quetiapine compared with risperidone, there were no significant differences in risks for myocardial infarction, heart failure, or fracture. However, risks for stroke (0.64, 0.53 to 0.78), venous thromboembolism (0.49, 0.36 to 0.68), pneumonia (0.72, 0.63 to 0.81), and acute kidney injury (0.81, 0.67 to 0.96) were lower for quetiapine than for risperidone.

Absolute risks of adverse outcomes

Cumulative incidence for all outcomes examined was higher for antipsychotic users versus matched comparators, except for ventricular arrhythmia and the negative control ( table 4 ). The absolute risk, as well as risk difference, was particularly large for pneumonia. In the 90 days after initiation of an antipsychotic, the cumulative incidence of pneumonia among antipsychotic users was 4.48% (95% CI 4.26% to 4.71%) v 1.49% (1.45% to 1.53%) in the matched cohort of non-users (difference 2.99%, 95% CI 2.77% to 3.22%). At one year, this increased to 10.41% (10.05% to 10.78%) for antipsychotic users compared with 5.63% (5.55% to 5.70%) for non-users (difference 4.78%, 4.41% to 5.16%).

Cumulative incidence of adverse outcomes associated with antipsychotic use at 90, 180, and 365 days and at two years after start of follow-up

Similar results were found in sensitivity analysis using two other definitions of antipsychotic use (see supplementary figures S4 and S5). Of the 544 203 antipsychotic prescriptions issued, 1.3% were for levomepromazine (see supplementary table S1). Results remained similar when patients were censored at the time of their first levomepromazine prescription (see supplementary figure S6). Results of the Fine-Gray models accounting for the competing risks of death also showed broadly similar patterns of hazards to those from the Cox models (see supplementary table S12 and figure S7). Sex specific analyses showed that male patients had higher incidence rates of all adverse outcomes than female patients, except for fracture and venous thromboembolism where incidence was higher for female patients than for male patients (see supplementary table S13). Compared with female antipsychotic users, male users had increased hazards for pneumonia and acute kidney injury (male to female hazard ratio 1.16, 95% CI 1.08 to 1.25 for pneumonia and 1.22, 1.08 to 1.37 for acute kidney injury), but lower hazards for stroke (0.81, 0.73 to 0.91). No significant differences were found by sex in the hazards for venous thromboembolism, myocardial infarction, heart failure, ventricular arrhythmia, or fracture (see supplementary table S14).

In this population based cohort study of adults (≥50 years) with dementia, use of antipsychotics compared with non-use was associated with increased risks for stroke, venous thromboembolism, myocardial infarction, heart failure, fracture, pneumonia, and acute kidney injury. Increased risks were observed among current and recent users and were highest in the first week after initiation of treatment. In the 90 days after a prescription, relative hazards were highest for pneumonia, acute kidney injury, stroke, and venous thromboembolism, with increased risks ranging from 1.5-fold (for venous thromboembolism) to twofold (for pneumonia) compared with non-use. No increased risk was found for ventricular arrhythmia or the negative control outcome (appendicitis and cholecystitis). Absolute risk differences between antipsychotic users and their matched comparators were substantial for most adverse events, and largest for pneumonia. In the 90 days after a prescription, risks of stroke, heart failure, fracture, pneumonia, and acute kidney injury were higher for typical antipsychotics versus atypical antipsychotics, whereas no significant differences between these two drug classes were found for risks of venous thromboembolism or myocardial infarction. Haloperidol was associated with higher risks for fracture, pneumonia, and acute kidney injury than risperidone, but no significant differences between the two drugs were found for the other outcomes. Risks of almost all adverse outcomes were higher for haloperidol than for quetiapine. No significant differences were found between risperidone and quetiapine for risks of myocardial infarction, heart failure, or fracture, but risks for stroke, venous thromboembolism, pneumonia, and acute kidney injury were lower for quetiapine versus risperidone.

Comparison with other studies

A population based study in Wales reported no increased risks for non-fatal acute cardiac events associated with antipsychotic use in patients with all cause dementia, although those with Alzheimer’s disease showed increased risks. 37 Systematic reviews and meta-analyses of studies not limited to patients with dementia have also reported inconsistent evidence for myocardial infarction, or lack of robustness of these data. 33 34 71 Our findings for myocardial infarction were similar to those in a study that first documented a modest and time limited increase in risk of this outcome associated with antipsychotic use among patients with dementia. 56 In a study of nursing home residents in the US, users of typical, but not atypical, antipsychotics were more likely than non-users to be admitted to hospital for ventricular arrhythmia or cardiac arrest, 35 and a study not limited to older people reported increased risks for ventricular arrhythmia or sudden cardiac death associated with both typical and atypical antipsychotics. 36 We did not find any association with ventricular arrhythmia, but the number of events was low and we did not examine cardiac arrest or sudden death.

Increased risks of venous thromboembolism associated with antipsychotic use have been reported in the general population, 38 but meta-analyses found increased risks of venous thromboembolism only among younger users. 39 40 Our findings are consistent with those of the Welsh study, which reported increased risks of venous thromboembolism in the 12 months after drug initiation (prior event rate ratio 1.95, 95% CI 1.83 to 2.0). 37 In absolute terms, however, these risks were relatively low compared with other outcomes examined in this study.

We found that both the relative and the absolute risks for pneumonia were highest among all outcomes examined. Current users of antipsychotics had a twofold increased risk compared with non-users ( fig 1 ), and although this magnitude of increased risk was comparable to previous reports, 14 31 32 we additionally observed that risks were greater in the first week after drug initiation. One study also reported a particularly high risk for patients with hospital diagnosed pneumonia in the first week, but the magnitude of increase (odds ratio 4.5, 95% CI 2.8 to 7.3) was much lower than our observation. 30 The mechanisms linking antipsychotic use and development of pneumonia is not well understood, and substantial heterogeneity exists among the drug substances, but antipsychotic induced extrapyramidal symptoms, sedation, xerostomia (dry mouth), and dyskinesia or impaired swallowing are commonly considered as potential risk factors. 72 In addition, because elderly people with pneumonia may be less likely than younger patients to present with respiratory symptoms but more likely to show signs of delirium, 73 it is possible that reverse causality might have contributed to the high risks observed in the early days after drug initiation, as delirium from the onset of pneumonia might have been treated with antipsychotics before pneumonia was diagnosed. 30 However, although causality cannot be inferred, the particularly high increased risks observed for a range of outcomes and not only for pneumonia in the early days after drug initiation are consistent with other studies. 28 This could be partly explained by further prescriptions being given only to patients who tolerated the first days of drug use.

The use of atypical antipsychotics in older adults (≥65 years) has been shown to be associated with increased risk of acute kidney injury. 44 45 46 Two studies reported significantly increased risks in users compared with non-users in the 90 days after initiation of atypical antipsychotics. 44 45 In contrast, another study observed no increased risks from use of the broad category of atypical antipsychotics, although a significantly increased risk was found with olanzapine. 46 In our study, we found increased risks of acute kidney injury with both typical and atypical antipsychotics, with risks being higher for haloperidol than for risperidone and quetiapine.

In a meta-analysis of observational studies, antipsychotic use was associated with increased risks of hip fracture among people with dementia. 41 A self-controlled case series study of older adult patients (≥65 years) also reported increased risks of falls and fracture after initiation of antipsychotics, but incidence was found to be even higher in the 14 days before treatment started. 43 Similar findings were also reported in another study, suggesting that the risks observed during the treatment periods might not be attributable to the antipsychotics alone. 42 Although we cannot eliminate confounding in our study, we minimised this risk by adjusting for a large number of both clinical and non-clinical characteristics that might have influenced treatment assignment. We also found no increased risks associated with current or recent antipsychotic use for the negative control outcome (appendicitis and cholecystitis).

Our study found that the risks of stroke and heart failure were higher for typical antipsychotics than for atypical antipsychotics, but risks of venous thromboembolism and myocardial infarction were similar between the two drug classes. We also found no significant differences between haloperidol and risperidone in risks of these four outcomes, but significantly increased risks for stroke, venous thromboembolism, and heart failure for haloperidol versus quetiapine. Previous studies of elderly patients have reported similar risks for cardiovascular or cerebrovascular events associated with use of typical and atypical antipsychotics, 17 74 75 76 but risks of these outcomes and of all cause mortality were increased with haloperidol versus risperidone. 21 76 For fracture and pneumonia, we found that risks were higher in association with typical antipsychotics than atypical antipsychotics and for haloperidol versus risperidone or quetiapine. The findings from previous studies comparing these risks by antipsychotic types have been inconsistent. 30 31 32 74 75

Strengths and limitations of this study

A key strength of this study was the investigation of a wide range of adverse events in a large population based cohort, and the reporting of both relative and absolute risk differences over multiple periods. Previous studies commonly focused on a single outcome or type of outcome, such as cerebrovascular events, and on the reporting of relative risks. By examining the same cohort at risk, we were able to directly compare the hazards of multiple outcomes without differential biases between the cohorts. In addition, we only included patients with a clinician recorded diagnosis of dementia, and we adjusted for many variables that might have influenced the probability of antipsychotic initiation, seeking to minimise confounding by indication. CPRD is one of the largest primary care databases in the world, and it is broadly representative of the UK population. 47 48 49 The database includes all prescriptions issued in participating primary care practices in the UK, and it is recognised as a high quality resource to support international pharmacovigilance. 77 The longitudinal nature of CPRD, with linked data from secondary care and mortality records, enabled us to capture the study outcomes from multiple sources, as well as information on prescribing and comorbidities. 78 79 Our findings were also robust to different classifications of usage periods and we found no associations between current and recent antipsychotic use with the development of the negative control outcome (appendicitis and cholecystitis). However, a significant association with past use was observed that we are unable to explain.

As with all observational studies, residual confounding cannot be excluded. For example, polypharmacy is common among elderly people, which could lead to drug-drug interactions and potentially confound our findings. 80 81 We also did not have information on indications for antipsychotics treatment. We minimised the risk of confounding using propensity score methods to control for imbalances in measurable patient characteristics between antipsychotic users and their matched comparators. However, unlike randomised control trials, which, if properly conducted, could account for both observed and unobserved differences between treated and untreated groups, the propensity score method can only adjust for the observed differences between two groups. Additionally, our choice of covariates was based on the literature and discussions with clinical experts and was not formally structured using, for example, a directed acyclic graph. Although the strong associations with pneumonia in the first seven days of antipsychotic initiation may partially be attributed to reverse causality, however, it is less likely to explain associations over longer periods. We also found no increased risk for appendicitis and cholecystitis during current and recent use—our negative control outcome that was included to detect potential unmeasured confounding. 53 Another limitation of our study is that although prescriptions issued in primary care are reliable in CPRD, information on dosage is not well recorded and information on drug adherence or prescriptions issued while patients are in hospital is not available. 48 Misclassification of drug use is therefore a potential problem. As with other electronic health data that are routinely collected for administrative rather than research purposes, potential issues exist with coding errors, missing or incomplete information, and variations in data quality between practices and healthcare settings. Although the data undergo quality checks before being released and our use of the linked data would have helped to deal with such problems, we were restricted to data coded in patients’ electronic health records. In addition, despite the representativeness of the CPRD data, care should be taken in making inferences beyond the population studied. Our sex specific investigations were also conducted as post hoc analyses. By using existing matched sets but restricting the comparators to those of the same sex as the antipsychotic user to whom they were matched, the number of comparators was greatly reduced. Although we found some evidence of differences in hazards for stroke, pneumonia, and acute kidney injury between male and female antipsychotic users, further research is needed to validate these findings.

Policy implications

The mechanisms underlying the links between antipsychotics and the outcomes in our study are not fully understood. In the UK, US, and Europe, current regulatory warnings for using antipsychotics to treat behavioural and psychological symptoms of dementia were mostly based on evidence of increased risks for stroke and mortality. 8 11 22 23 24 25 26 We found a considerably wider range of harms associated with antipsychotic use in people with dementia, and the risks of harm were highest soon after initiation. Our findings must be seen in the context of trial evidence of at best modest benefit on behavioural and psychological symptoms of dementia. The efficacy of antipsychotics in the management of behavioural and psychological symptoms of dementia remains inconclusive. 82 83 84 85 Atypical antipsychotics, including risperidone, which is one of two antipsychotics licensed in the UK for the treatment of behavioural and psychological symptoms of dementia, have the strongest evidence base, but the benefits are only modest. 82 85

Any potential benefits of antipsychotic treatment therefore need to be weighed against the risk of serious harm across multiple outcomes. Although there may be times when an antipsychotic prescription is the least bad option, clinicians should actively consider the risks, considering patients’ pre-existing comorbidities and living support. The NNH reported in this study can help to inform clinical judgements on the appropriateness of treatments, taking account of the modest potential benefits reported in clinical trials. When prescriptions of such drugs are needed, treatment plans should be reviewed regularly with patients and their carers to reassess the need for continuing treatment. 9 In addition, given the higher risks of adverse events in the early days after drug initiation, clinical examinations should be taken before, and clinical reviews conducted shortly after, the start of treatment. Our study reaffirms that these drugs should only be prescribed for the shortest period possible. 9 Although regulators have made efforts to limit the use of these drugs to people with the most severe behavioural and psychological symptoms of dementia, 8 82 86 antipsychotic prescribing in dementia remains common and has even increased in recent years. 4 5 87 88 If such trends continue, further communication on the associated risks could be considered by guideline developers or regulators after a review of the totality of evidence. Greater accountability and monitoring in the use of these drugs may be called for, and additional legal reforms may be required to regulate adherence. 89 In recent years, other psychotropic drugs such as antidepressants, benzodiazepines, mood stabilisers, and anticonvulsants have been prescribed instead of antipsychotics for the treatment of behavioural and psychological symptoms of dementia. 28 90 91 These drugs, however, also pose their own risks. Further research is needed into safer drug treatment of behavioural and psychological symptoms of dementia and more efficacious, easy to deliver, initial non-drug treatments.

Conclusions

Antipsychotic use is associated with a wide range of serious adverse outcomes in people with dementia, with relatively large absolute risks of harm for some outcomes. These risks should be considered in future regulatory decisions, alongside cerebrovascular events and mortality. Any potential benefits of antipsychotic treatment need to be weighed against risk of serious harm, and treatment plans should be reviewed regularly. The effect of antipsychotics on behavioural and psychological symptoms of dementia is modest at best, but the proportion of people with dementia prescribed antipsychotics has increased in recent years. Our finding that antipsychotics are associated with a wider range of risks than previously known is therefore of direct relevance to guideline developers, regulators, and clinicians considering the appropriateness of antipsychotic prescribing for behavioural and psychological symptoms of dementia.

What is already known on this topic

Despite safety concerns, antipsychotics continue to be frequently prescribed for the management of behavioural and psychological symptoms of dementia

Current regulatory warnings for the treatment of behavioural and psychological symptoms of dementia using antipsychotics are based on evidence of increased risks of stroke and death

Evidence for other adverse outcomes is less conclusive or is more limited among people with dementia, and comparisons of risks for multiple adverse events are also difficult owing to different study designs and populations

What this study adds

Antipsychotic use in people with dementia was associated with increased risks of stroke, venous thromboembolism, myocardial infarction, heart failure, fracture, pneumonia, and acute kidney injury, compared with non-use, but not ventricular arrhythmia

Relative hazards were highest for pneumonia, acute kidney injury, stroke, and venous thromboembolism, and absolute risk and risk difference between antipsychotic users and their matched comparators was largest for pneumonia

Risks of these wide ranging adverse outcomes need to be considered before prescribing antipsychotic drugs to people with dementia

Ethics statements

Ethical approval.

This study was approved by the Clinical Practice Research Datalink’s (CPRD) independent scientific advisory committee (protocol 18_168). CPRD also has ethical approval from the Health Research Authority to support research using anonymised patient data (research ethics committee reference 21/EM/0265). 92 Individual patient consent was not required as all data were deidentified.

Data availability statement

Electronic health records are, by definition, considered sensitive data in the UK by the Data Protection Act and cannot be shared via public deposition because of information governance restriction in place to protect patient confidentiality. Access to Clinical Practice Research Datalink (CPRD) data is subject to protocol approval via CPRD’s research data governance process. For more information see https://cprd.com/data-access . Linked secondary care data from Hospital Episodes Statistics, mortality data from the Office for National Statistics, and index of multiple deprivation data can also be requested from CPRD.

Acknowledgments

We thank Hayley Gorton and Thomas Allen for their contribution to the protocol development, Evan Kontopantelis for his statistical advice, and members of our patient and public involvement team, Antony Chuter and Jillian Beggs, for their contributions to this project. This study is based on data from the Clinical Practice Research Datalink obtained under licence from the UK Medicines and Healthcare Products Regulatory Agency (MHRA). The data are provided by patients and collected by the NHS as part of its care and support. Hospital Episode Statistics and Office for National Statistics mortality data are subject to Crown copyright (2022) protection, reused with the permission of The Health and Social Care Information Centre, all rights reserved. The interpretation and conclusions contained in this study are those of the authors alone, and not necessarily those of the MHRA, National Institute of Health and Care Research, NHS, or Department of Health and Social Care. The study protocol was approved by Clinical Practice Research Datalink’s independent scientific advisory committee (reference: 18_168). We would like to acknowledge all the data providers and general practices who make anonymised data available for research.

Contributors: All authors conceived and designed the study and acquired, analysed, or interpreted the data. BG, DRM, TvS, AJA, and DMA reviewed the clinical codes. PLHM conducted the statistical analyses and wrote the first draft of the manuscript. All authors critically revised the manuscript for important intellectual content and approved the final version. AJA, DMA, RAE, BG, DRM, AS, and TvS obtained the funding. PLHM is the guarantor. The corresponding author (PLHM) attests that all listed authors meet authorship criteria and that no others meeting the criteria have been omitted.

Funding: This study was funded by the National Institute for Health and Care Research (NIHR, RP-PG-1214-20012). MJC, AJA, and DMA were supported by the NIHR Greater Manchester Patient Safety Translational Research Centre (PSTRC-2016-003) at the time of this study and are now supported by the NIHR Greater Manchester Patient Safety Research Collaboration (NIHR204295). The funders had no role in the study design; in the collection, analysis, or interpretation of data; in the writing of the report; or in the decision to submit the article for publication. PLHM has full access to all data and all authors have full access to the statistical reports and tables in the study. PLHM takes responsibility for the integrity of the data and the accuracy of the data analysis.

Competing interests: All authors have completed the ICMJE uniform disclosure form at www.icmje.org/coi_disclosure.pdf and declare: BG reports research grants from the National Institute for Health and Care Research (NIHR). DRM was awarded a Wellcome Trust Clinical Research Development Fellowship (214588/Z/18/Z). AS reports a research grant from the NIHR. RAE reports research grants from the NIHR and NHS England, and travel costs to attend a roundtable dinner discussion on medication errors, House of Commons, Westminster, on 29 March 2022. TvS reports research grants from the NIHR. AJA is national clinical director for prescribing for NHS England and reports research grants from the NIHR. DMA reports research grants from the NIHR, AbbVie, Almirall, Celgene, Eli Lilly, Janssen, Novartis, UCB, and the Leo Foundation. All other authors declare no support from any organisation for the submitted work (except those listed in the funding section); no financial relationships with any organisations that might have an interest in the submitted work in the previous three years; no other relationships or activities that could appear to have influenced the submitted work.

Transparency: The lead author (PLHM: the manuscript’s guarantor) affirms that the manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained.

Dissemination to participants and related patient and public communities: This study used anonymised electronic health records from the CPRD and it is therefore not possible to disseminate the findings directly to individuals whose data we used. This study is part of a National Institute for Health and Care Research (NIHR) funded programme (RP-PG-1214-20012): Avoiding patient harm through the application of prescribing safety indicators in English general practices (PRoTeCT). We have experienced patient and public involvement members aligned to the programme who we will consult in the results dissemination. In addition, senior author DMA is director of NIHR Greater Manchester Patient Safety Research Collaboration (GMPSRC), and co-authors MJC and AJA are affiliated with it. The Patient Safety Research Collaboration has a community of public contributors including patients, carers, and people accessing health and social care services. The authors will work with this network to disseminate findings.

Provenance and peer review: Not commissioned; externally peer reviewed.

This is an Open Access article distributed in accordance with the terms of the Creative Commons Attribution (CC BY 4.0) license, which permits others to distribute, remix, adapt and build upon this work, for commercial use, provided the original work is properly cited. See: http://creativecommons.org/licenses/by/4.0/ .

• GBD 2019 Dementia Forecasting Collaborators
• McKeith I ,
• Stocks SJ ,
• Kontopantelis E ,
• Ashcroft DM
• Szczepura A ,
• Collings SL ,
• ↵ National Institute for Health and Care Excellence. Dementia: assessment, management and support for people living with dementia and their carers. NICE guideline NG97, 20 June 2018. www.nice.org.uk/guidance/ng97
• ↵ National Institute for Health and Care Excellence. Dementia: antipsychotics. Clinical knowledge summaries. December 2022. https://cks.nice.org.uk/topics/dementia/prescribing-information/antipsychotics/
• Marston L ,
• Nazareth I ,
• Petersen I ,
• Walters K ,
• Bronskill SE ,
• Normand SL ,
• Schneeweiss S ,
• Setoguchi S ,
• Brookhart A ,
• Dormuth C ,
• Dzahini O ,
• Schneider LS ,
• Dagerman KS ,
• Douglas IJ ,
• Sacchetti E ,
• Turrina C ,
• Valsecchi P
• Seyfried LS ,
• Kheirbek RE ,
• Little JT ,
• ↵ Medicines Healthcare Products Regulatory Agency. New advice issued on risperidone and olanzapine 2004. https://webarchive.nationalarchives.gov.uk/20141205150130/http://mhra.gov.uk/printpreview/defaultsp/con002047
• ↵ Medicines Healthcare Products Regulatory Agency. Antipsychotics: use in elderly patients with dementia. Drug safety update March 2009;2:8. https://webarchive.nationalarchives.gov.uk/20091114175357/http://mhra.gov.uk/Publications/Safetyguidance/DrugSafetyUpdate/CON041211
• ↵ Banerjee S. The use of antipsychotic medication for people with dementia: time for action. An independent report commissioned and funded by the Department of Health 2009.
• ↵ European Medicines Agency. Priorities for Drug Safety Research: Safety aspects of antipsychotics in demented patients. 2010. https://www.ema.europa.eu/en/documents/other/emea-2010-priorities-drug-safety-research-safety-aspects-antipsychotics-demented-patients_en.pdf
• Hessol NA ,
• Lexchin J ,
• Harrington C
• Lapeyre-Mestre M
• Rogowska M ,
• Thornton M ,
• van Marum RJ ,
• Jansen PA ,
• Souverein PC ,
• Schobben AF ,
• Trifirò G ,
• Gambassi G ,
• Tolppanen AM ,
• Koponen M ,
• Tanskanen A ,
• Zivkovic S ,
• Liperoti R ,
• Lapane KL ,
• Jönsson AK ,
• Spigset O ,
• Ostuzzi G ,
• Thabane L ,
• Brännström J ,
• Lövheim H ,
• Gustafson Y ,
• Nordström P
• Schuemie MJ ,
• Ramcharran D ,
• Brøgger Kristiansen S ,
• Baltzer Houlind M ,
• Herrett E ,
• Gallagher AM ,
• Bhaskaran K ,
• Campbell J ,
• Mahadevan P ,
• Fordyce S ,
• Rassen JA ,
• Rothman KJ ,
• Greenland S ,
• Kleijer BC ,
• Egberts AC ,
• Heerdink ER
• Lipsitch M ,
• Tchetgen Tchetgen E ,
• Austin PC ,
• Pariente A ,
• Fourrier-Réglat A ,
• Ducruet T ,
• Rigler SK ,
• Shireman TI ,
• Cook-Wiens GJ ,
• Balkwill A ,
• Million Women Study Collaborators
• Vinogradova Y ,
• Coupland C ,
• Hippisley-Cox J
• Sandler RS ,
• Kappelman MD
• Spoorenberg SMC ,
• Snijders D ,
• Ovidius Study Group ,
• Capisce Study Group ,
• STEP Study Group
• Koivula I ,
• ↵ Lunt M. Propensity analysis in Stata, revision: 1.1. 2014. University of Manchester. http://personalpages.manchester.ac.uk/staff/mark.lunt/propensity_guide.pdf
• Grambsch PM ,
• Therneau TM
• Gooley TA ,
• Leisenring W ,
• Crowley J ,
• Schmitz A ,
• Gastaldon C ,
• Sultana J ,
• Calabró M ,
• Garcia-Serna R ,
• Huybrechts KF ,
• Silliman RA ,
• Brookhart MA ,
• Schneeweiss S
• Gerhard T ,
• Sahlberg M ,
• Gislason GH ,
• Torp-Pedersen C ,
• Andersson C
• Crellin E ,
• Donegan K ,
• Padmanabhan S ,
• Cameron E ,
• Williams R ,
• Strongman H
• Nørgaard A ,
• Jensen-Dahm C ,
• Wimberley T ,
• Hansen ES ,
• Gitlin LN ,
• Lyketsos CG
• Goodarzi Z ,
• Veroniki AA ,
• Frederiksen KS ,
• Frisoni GB ,
• El Helou ML
• McDermid J ,
• Ballard C ,
• Schnier C ,
• McCarthy A ,
• Morales DR ,
• CVD-COVID-UK/COVID-IMPACT Consortium
• ↵ CPRD - Safeguarding patient data https://cprd.com/safeguarding-patient-data

Journal of Drug Design and Research

2022 Volume 9 - Issue 1

Recent Articles

Journals Section

Join our weekly newslettter for the latest Articles of Journal.

Subscribe to Newsletters

And stay informed about our news and events

• STM Journals
• Special Issues
• Conferences
• Editorial Board Members
• Reviewers Board Members
• Advisory Panel
• Indexing Bodies
• For Authors
• For Reviewers
• For Editors
• For Advisory Board
• Special Issue Guidelines
• Peer-Review Policy
• Manuscript Submission Guidelines
• Publication Ethics and Virtue
• Article Processing Charge
• Editorial Policy
• Advertising Policy
• STM Website and Link Policy
• Distribution and dessemination of Research
• Informed consent Policy

"Connect with colleagues and showcase your academic achievements."

"Unleashing the potential of your words"

 "Explore a vast collection of books and broaden your horizons."

 "Empower yourself with the knowledge and skills needed to succeed."

"Collaborate with like-minded professionals and share your knowledge."

"Learn from experts and engage with a community of learners."

• ICDR Group of Companies

• Training Programs

Research & Reviews: A Journal of Drug Design & Discovery

ISSN: 2349-9036

Journal Menu

Editors overview.

Prof. (DR.) Angshu Banerjee

Institutional Profile Link : https://ww. . . View Full Profile

STM Journals, An imprint of Consortium e-Learning Network Pvt. Ltd. A-118, 1st Floor, Sector-63, Noida, U.P. India, Pin – 201301 E-mail: [email protected] (Tel) (+91) 0120- 4781 200 (Mob) (+91) 9810078958, +919667725932

A Comparative Study of Atibalāmūla Kvātha andBhūmyāmalakī Cūrṇa in the Management ofDiabetic Sensory Polyneuropathy

Distinctive Approaches for Improving the Aqueous Solubility of BCS Class II Drugs

A Review on Amino Acids

About the Journal

Research & Reviews: A Journal of Drug Design & Discovery (rrjoddd) : 2349-9036(e) is a peer-reviewed hybrid open access journal launched in 2014 covers the latest research on how drugs work, View Full Focus and Scope…

Journal Particulars

2020: 6.502 2021: 6.199 2022: 6.053

[email protected]

+91 120 478 1209

Vol-11 Issue-01 2024

Fast Dissolving Oral Thin Films: A Review Chenna M Shalini, Asireddy Sathvika Reddy, Vallarapu Nanda Krishna Veni, Madhira Jayaprakash, Rama Rao T Keywords: Fast dissolving films, Polymers, Solvent casting method, Fast Dissolving Tablets, Orally disintegrating films.

Navigating the Horizon: Current Trends in Elevating Quality Assurance in Clinical Research Patil Divyashree Kantilal, Patil Bhagyshri Sandip, Patil Bhushan Ravindra, Patil Divyani Rajendra Keywords: Data integrity, Inspection, Quality systems, Risk-based monitoring, Subject protection.

Protective Effects Of Curcumin Against Osteoporosis And Its Molecular Mechanism Chhatrapalsinh R. Parmar, Pratiksha N. Patil, Riddhi S.Bhavsar, Rohitkumar D.Shah, Riya A.Rajput Keywords: Osteoporosis (OP), Osteoblast, Osteoclast, Molecular Mechanism, Curcumin (CUR).

Simultaneous Determination And Stability Indicating Method For Telmisartan, Metoprolol Succinate And Cilnidipine In Tablet Dosage Form. Dr. Bhaveshkumar Patel Keywords: Telmisartan, Metoprolol, Cilnidipine, HPLC, Validation, Degradation.

For Subscriber Access

Special Issue

WEBSITE DISCLAIMER

Last updated: 2022-06-15

The information provided by STM Journals (“Company”, “we”, “our”, “us”) on https://journals.stmjournals.com / (the “Site”) is for general informational purposes only. All information on the Site is provided in good faith, however, we make no representation or warranty of any kind, express or implied, regarding the accuracy, adequacy, validity, reliability, availability, or completeness of any information on the Site.

UNDER NO CIRCUMSTANCE SHALL WE HAVE ANY LIABILITY TO YOU FOR ANY LOSS OR DAMAGE OF ANY KIND INCURRED AS A RESULT OF THE USE OF THE SITE OR RELIANCE ON ANY INFORMATION PROVIDED ON THE SITE. YOUR USE OF THE SITE AND YOUR RELIANCE ON ANY INFORMATION ON THE SITE IS SOLELY AT YOUR OWN RISK.

EXTERNAL LINKS DISCLAIMER

The Site may contain (or you may be sent through the Site) links to other websites or content belonging to or originating from third parties or links to websites and features. Such external links are not investigated, monitored, or checked for accuracy, adequacy, validity, reliability, availability, or completeness by us.

WE DO NOT WARRANT, ENDORSE, GUARANTEE, OR ASSUME RESPONSIBILITY FOR THE ACCURACY OR RELIABILITY OF ANY INFORMATION OFFERED BY THIRD-PARTY WEBSITES LINKED THROUGH THE SITE OR ANY WEBSITE OR FEATURE LINKED IN ANY BANNER OR OTHER ADVERTISING. WE WILL NOT BE A PARTY TO OR IN ANY WAY BE RESPONSIBLE FOR MONITORING ANY TRANSACTION BETWEEN YOU AND THIRD-PARTY PROVIDERS OF PRODUCTS OR SERVICES.

PROFESSIONAL DISCLAIMER

The Site can not and does not contain medical advice. The information is provided for general informational and educational purposes only and is not a substitute for professional medical advice. Accordingly, before taking any actions based on such information, we encourage you to consult with the appropriate professionals. We do not provide any kind of medical advice.

Content published on https://journals.stmjournals.com / is intended to be used and must be used for informational purposes only. It is very important to do your analysis before making any decision based on your circumstances. You should take independent medical advice from a professional or independently research and verify any information that you find on our Website and wish to rely upon.

THE USE OR RELIANCE OF ANY INFORMATION CONTAINED ON THIS SITE IS SOLELY AT YOUR OWN RISK.

AFFILIATES DISCLAIMER

The Site may contain links to affiliate websites, and we may receive an affiliate commission for any purchases or actions made by you on the affiliate websites using such links.

TESTIMONIALS DISCLAIMER

The Site may contain testimonials by users of our products and/or services. These testimonials reflect the real-life experiences and opinions of such users. However, the experiences are personal to those particular users, and may not necessarily be representative of all users of our products and/or services. We do not claim, and you should not assume that all users will have the same experiences.

YOUR RESULTS MAY VARY.

The testimonials on the Site are submitted in various forms such as text, audio, and/or video, and are reviewed by us before being posted. They appear on the Site verbatim as given by the users, except for the correction of grammar or typing errors. Some testimonials may have been shortened for the sake of brevity, where the full testimonial contained extraneous information not relevant to the general public.

The views and opinions contained in the testimonials belong solely to the individual user and do not reflect our views and opinions.

ERRORS AND OMISSIONS DISCLAIMER

While we have made every attempt to ensure that the information contained in this site has been obtained from reliable sources, STM Journals is not responsible for any errors or omissions or the results obtained from the use of this information. All information on this site is provided “as is”, with no guarantee of completeness, accuracy, timeliness, or of the results obtained from the use of this information, and without warranty of any kind, express or implied, including, but not limited to warranties of performance, merchantability, and fitness for a particular purpose.

In no event will STM Journals, its related partnerships or corporations, or the partners, agents, or employees thereof be liable to you or anyone else for any decision made or action taken in reliance on the information in this Site or for any consequential, special or similar damages, even if advised of the possibility of such damages.

GUEST CONTRIBUTORS DISCLAIMER

This Site may include content from guest contributors and any views or opinions expressed in such posts are personal and do not represent those of STM Journals or any of its staff or affiliates unless explicitly stated.

LOGOS AND TRADEMARKS DISCLAIMER

All logos and trademarks of third parties referenced on https://journals.stmjournals.com / are the trademarks and logos of their respective owners. Any inclusion of such trademarks or logos does not imply or constitute any approval, endorsement, or sponsorship of STM Journals by such owners.

Should you have any feedback, comments, requests for technical support, or other inquiries, please contact us by email: [email protected] .