drug discovery

Drug discovery is the multidisciplinary scientific process of identifying and validating novel therapeutic compounds that modulate disease-relevant biological targets, spanning target identification, hit discovery, lead optimisation, and preclinical candidate nomination. The modern discipline applies machine learning, deep learning, graph neural networks, and generative modelling to predict molecular properties such as binding affinity, selectivity, and ADMET profiles — dramatically accelerating virtual screening and de novo molecular design. Structural biology tools including protein structure prediction (e.g. AlphaFold) underpin structure-based drug design by supplying accurate 3D target models at proteome scale. The field sits at the intersection of cheminformatics, structural biology, computational chemistry, and clinical informatics, with AI methods increasingly embedded throughout the entire development pipeline from target discovery through biomarker identification and patient stratification.

Semantic Classification

Content

Compositional Relationships (Components)

SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:hasPart ai:TargetIdentification))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:hasPart ai:VirtualScreening))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:hasPart ai:LeadOptimisation))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:hasPart ai:ADMETPrediction))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:hasPart ai:GenerativeMolecularDesign))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:hasPart ai:BiomarkerDiscovery))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:hasPart ai:Retrosynthesis))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:hasPart ai:MolecularDynamicsSimulation))

Dependency Relationships

SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:requires ai:Cheminformatics))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:requires ai:StructuralBiology))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:requires ai:BiomedicalData))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:requires ai:ProteinStructurePrediction))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:requires ai:MultiOmics))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:dependsOn ai:ActiveLearning))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:dependsOn ai:DeepLearning))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:dependsOn ai:MolecularDynamicsSimulation))

Capability Relationships

SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:enables ai:ClinicalTrials))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:enables ai:PersonalisedMedicine))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:enables ai:BiomarkerDiscovery))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:enables ai:PrecisionMedicine))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:enables ai:DrugRepurposing))

Implementation Relationships

SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:uses ai:GraphNeuralNetwork))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:uses ai:GenerativeModel))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:uses ai:Transformer))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:uses ai:ReinforcementLearning))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:uses ai:DiffusionModel))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:uses ai:VariationalAutoencoder))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:uses ai:FederatedLearning))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:uses ai:KnowledgeGraph))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:implements ai:ComputerAidedDrugDesign))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:implements ai:StructureBasedDrugDesign))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:implements ai:LigandBasedDrugDesign))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:implements ai:ADMETPrediction))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:implements ai:Retrosynthesis))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:implements ai:ProteinStructurePrediction))

Reduction Relationships

SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:reducesTo ai:MolecularOptimisation))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:reducesTo ai:VirtualScreening))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:reducesTo ai:TargetIdentification))
SubClassOf(ai:DrugDiscovery
  ObjectSomeValuesFrom(ai:reducesTo ai:LeadOptimisation))

About

Drug discovery is one of the most resource-intensive knowledge-creation endeavours in applied science: the traditional path from target identification to an approved medicine consumes on average ten to fifteen years and USD 2–3 billion per successful drug, with failure rates above 90% driven primarily by late-stage attrition on safety and efficacy grounds. The canonical pipeline proceeds through target identification and validation, assay development and high-throughput screening of synthesised compound libraries, hit-to-lead chemistry to identify tractable starting points, lead optimisation to improve potency, selectivity, and physicochemical properties, ADMET profiling to assess metabolic and safety liabilities, and finally preclinical development in animal models before the Investigational New Drug (IND) filing that initiates human clinical trials across three regulatory phases. At each stage, the cost and timescale of experiments has historically forced sequential, hypothesis-driven exploration of a chemical space estimated at between 10^23 and 10^60 synthesisable drug-like molecules — a space so vast that traditional random or combinatorial library screening can explore only an infinitesimal fraction.

The entry of machine learning and artificial intelligence into this pipeline has fundamentally restructured which stages remain rate-limiting by replacing or augmenting expensive experimental steps with rapid, accurate computational predictions. Deep Learning models trained on large curated datasets such as ChEMBL (containing over 2.4 million bioactive molecules and 19 million activity measurements), BindingDB, PubChem BioAssay, and the Protein Data Bank can score molecular binding affinity orders of magnitude faster than physics-based free-energy perturbation calculations while achieving competitive accuracy for congeneric compound series. Graph Neural Network architectures, which represent molecules as attributed graphs with atoms as typed nodes and bonds as edge features, and propagate information through iterative message-passing layers to produce molecular-level embeddings, have displaced Morgan circular fingerprints and physicochemical descriptors as the standard molecular featurisation for ADMET and activity prediction tasks — consistently outperforming fingerprint-based approaches on MoleculeNet and Therapeutics Data Commons (TDC) benchmark datasets. More transformatively, Generative Model systems shift the paradigm from selecting among molecules in an existing library to creating novel molecules with specified properties, enabling exploration of chemical space that would never be accessible through synthesis and screening of physical compound collections.

Protein Structure Prediction via AlphaFold 3, released by DeepMind in May 2024 and extending the original AlphaFold 2 breakthrough to model protein-ligand, protein-nucleic acid, protein-antibody, and multi-component complexes through a diffusion-based architecture, has effectively unlocked structure-based virtual screening for hundreds of thousands of previously undruggable or structurally uncharacterised targets across the human proteome and the proteomes of major disease pathogens. The AlphaFold Protein Structure Database provides structures for over 200 million proteins, enabling structure-based drug design campaigns to launch without the multi-year crystallography programmes previously required. As of June 2026, the clinical validation of the AI drug discovery paradigm is firmly established: over 173 AI-discovered programmes are in clinical development globally, Insilico Medicine’s rentosertib became the first compound where both target (TNIK kinase, identified computationally as relevant to idiopathic pulmonary fibrosis biology) and drug molecule were fully AI-designed to receive a United States Adopted Name (USAN), with positive Phase IIa results in patients published in Nature Medicine in June 2025 and Phase III planning underway; Recursion Pharmaceuticals, following its USD 688 million acquisition of Exscientia in late 2024, operates one of the largest AI-native pipeline portfolios globally; and the FDA’s January 2025 draft guidance on AI in drug development has established a seven-step credibility assessment framework for AI-generated evidence in regulatory submissions.

The economics of AI-enabled drug discovery are reshaping the pharmaceutical industry’s competitive dynamics. Traditional large-cap pharmaceutical companies are challenged by AI-native startups that can run discovery programmes with dramatically smaller teams and budgets — Insilico Medicine developed its Phase IIa rentosertib candidate with a team smaller than a typical medicinal chemistry department, leveraging AI at every stage from target identification through clinical candidate nomination. This competitive pressure is accelerating industry adoption: major pharmaceutical companies have established large internal AI drug discovery programmes, acquired AI-native companies (e.g. Sanofi’s acquisition of Exscientia collaborations, Recursion’s acquisition of Exscientia itself), and entered large-scale research partnerships with AI companies (e.g. Pfizer’s collaboration with Insilico, AstraZeneca’s partnership with BenevolentAI). The convergence of abundant clinical and biological data (electronic health records, whole-genome sequencing at clinical scale, proteomics from bio-banking), improving generative AI methods, and falling compute costs is expected to further accelerate this structural transformation over the 2026–2030 period.

Components / Architecture

Target Identification and Validation

• Target Identification identifies disease-modifying biological macromolecules — proteins, nucleic acids, protein-protein interaction interfaces — whose modulation produces therapeutic benefit.

• Multi-Omics data integration (genomics, transcriptomics, proteomics, metabolomics, epigenomics) across disease and healthy tissue profiles identifies differentially expressed molecular entities linked to disease phenotypes.

• Knowledge Graph and network medicine approaches represent disease-gene-protein-pathway relationships, enabling link prediction and community detection to prioritise druggable targets.

• Causal Inference via Mendelian randomisation (GWAS genetic data as instrumental variables) distinguishes causal drivers from downstream correlates, reducing non-causal target risk.

• Functional CRISPR screening, proteomics from PTM profiling, and single-cell RNA sequencing validate target expression in disease-relevant cell populations at single-cell resolution.

• AI patient stratification identifies genetically or molecularly defined responder subpopulations, de-risking clinical translation of targets relevant to biologically heterogeneous diseases.

• Human genetic evidence prioritisation: targets supported by protective loss-of-function human variants, disease-associated pQTLs (protein quantitative trait loci), or rare variant associations have substantially higher Phase II and III success rates.

• Network proximity score in the human protein-protein interactome quantifies topological closeness between a candidate target’s first-degree protein neighbours and disease-associated gene modules, providing a genome-wide druggability and disease-relevance metric.

• Deep learning models trained on whole-exome sequencing data from biobank cohorts (UK Biobank: 500,000 participants; AstraZeneca’s 10x Genomics dataset) identify novel target-disease associations by detecting rare variant burden signatures missed by GWAS in common variants.

• Druggability prediction models assess whether a protein has structural features compatible with small molecule binding (pocket volume, depth, hydrophobicity) or biologic targeting (surface accessibility, known epitopes), helping prioritise targets for small molecule versus biologic drug discovery programmes.

  Virtual Screening and Molecular Docking

• Virtual Screening computationally filters large compound libraries to identify candidates likely to bind a biological target, replacing or preceding physical High-Throughput Screening.

• Structure-based virtual screening docks ligands into binding pockets modelled by Protein Structure Prediction; classical docking (AutoDock Vina, Glide) computes physics-inspired binding scores at moderate throughput.

• Deep Learning docking surrogates (DiffDock, Uni-Mol, EquiBind) predict binding poses at millisecond timescales per molecule, enabling screening of ultra-large virtual libraries (Enamine REAL: 10^10–10^12 compounds) in days on GPU clusters.

• Graph Neural Network models trained on protein-ligand complex datasets predict binding affinity from molecular graph features with accuracy competitive with physics-based methods for congeneric series.

• Ligand-based virtual screening trains Machine Learning models from known active/inactive compound profiles, applicable when high-quality structural information is unavailable.

• Active Learning guides iterative screening campaigns by selecting the most informative compounds to test experimentally, maximising information gain per assay dollar spent.

• Consensus scoring: combining predictions from multiple docking programs and ML models improves selectivity for true actives over false positives in virtual screening campaigns.

• Hot-spot analysis and water-displacing fragment screening: AI methods identify energetically favourable binding hotspots on target surfaces and predict the desolvation energetics of displacing bound water molecules in binding pockets.

• Pharmacophore-guided virtual screening filters compound libraries for 3D spatial arrangement of hydrogen bond donors/acceptors, hydrophobic centres, and charged groups matching target binding requirements before computationally expensive docking.

• Scaffold filtering and PAINS (Pan Assay Interference Compounds) removal: AI-assisted filtering of virtual hits removes promiscuous false-positive compounds containing reactive functional groups or aggregation-prone structural features that produce apparent activity in HTS assays.

  Generative Molecular Design

• Variational Autoencoder (VAE) models learn continuous latent molecular representations, enabling gradient-based property optimisation by moving through latent space and decoding back to molecular structures.

• Normalising flow models provide exact likelihood estimates with invertible transformations, enabling precise conditional sampling from target property distributions.

• Transformer language models trained on SMILES and SELFIES (ChemBERTa pre-trained on 77M SMILES, MolGPT, Mol-Instructions) generate syntactically valid, chemically meaningful molecular strings.

• 3D Diffusion Model approaches (DiffSBDD, TargetDiff, MODA) generate molecular conformations conditioned on binding pocket geometry, bypassing SMILES linearisation and enabling pocket-native design.

• Reinforcement Learning with multi-objective reward functions (potency, selectivity, solubility, SA score, metabolic stability, hERG toxicity, BBB permeability) guides de novo molecule optimisation via Pareto-front exploration.

• Multi-objective Bayesian optimisation provides superior sample efficiency over RL when individual property evaluations are expensive or slow.

• Fragment-based generative design: generative models assemble molecules from pre-validated fragment building blocks with known synthetic accessibility, combining the diversity of de novo generation with the practical tractability of fragment-based approaches.

• Multi-target drug design: generative models conditioned on multiple protein structures simultaneously produce molecules predicted to bind two or more disease-relevant targets, enabling polypharmacology by design for oncology and CNS indications.

• Property-guided scaffold hopping: Transformer models conditioned on molecular activity and property fingerprints generate structurally diverse compounds with matched activity profiles, enabling intellectual property differentiation of existing pharmacophores.

• Insilico Medicine’s PandaOmics and Chemistry42 workflow for rentosertib: PandaOmics identified TNIK kinase as the target from multi-omics analysis; Chemistry42 generated candidate structures; experimental validation confirmed in vitro activity; the full design cycle completed in 18 months versus a typical 4–5 year timeline.

  ADMET Prediction and Safety Profiling

• ADMET prediction applies Machine Learning models trained on curated experimental datasets to predict pharmacokinetic and safety-relevant properties from molecular structure alone, enabling triage before synthesis.

• Key endpoints predicted: Caco-2 permeability, P-gp efflux, CYP450 inhibition (1A2/2C9/2D6/3A4), hERG cardiac inhibition, microsomal half-life, plasma protein binding, CNS penetration (LogBB), Ames mutagenicity.

• Graph Neural Network and Transformer models achieve competitive performance on 22 TDC ADMET benchmarks with standardised dataset splits and evaluation metrics.

• Multi-task learning across correlated ADMET endpoints shares molecular representations, improving predictions for data-sparse endpoints by leveraging transfer from data-rich related tasks.

• Active Learning frameworks iteratively select the most informative compounds for wet-lab ADMET assay, building richer training datasets through directed exploration and compounding model improvement.

• Integration of ADMET predictors into generative model reward functions filters computationally proposed molecules for acceptable pharmacokinetic profiles before committing to synthesis.

• PK/PD modelling: quantitative pharmacokinetics-pharmacodynamics models linking predicted ADMET properties (absorption, oral bioavailability, CYP clearance) to predicted plasma concentration-time curves, enabling computational dose-regime simulation before any in vivo experiment.

• Adverse drug event prediction from electronic health records: retrospective ML analysis of EHR datasets links drug-drug interactions and patient genetic or phenotypic covariates to adverse events, guiding prospective safety risk stratification in clinical development.

• Organ-on-chip and 3D organoid integration: AI analysis of complex biological response data from microphysiological systems provides more translatable ADMET data than 2D cell assays, with ML models trained on organoid responses replacing higher-cost animal studies.

• Regulatory-grade ADMET models: industry consortia (Predictive Safety Testing Consortium, DILI-sim initiative) developing AI models for drug-induced liver injury (DILI), cardiotoxicity, and genotoxicity that meet FDA and EMA reproducibility and documentation standards for regulatory submissions.

  Retrosynthesis and Synthetic Accessibility Assessment

• Retrosynthesis planning tools (AiZynthFinder, ASKCOS, Chemist-AI, IBM RXN for Chemistry) predict multi-step synthetic routes from commercially available starting materials using reaction templates and Deep Learning on reaction SMILES.

• Synthetic accessibility scoring functions (Ertl SA score, SCScore, RetroScore) provide rapid, differentiable estimates of synthetic complexity incorporated directly into generative model reward functions.

• Biasing generative models toward synthetically accessible chemical space prevents wasted synthesis effort on molecules that are computationally attractive but experimentally intractable.

• Coley et al. (Science 2019) demonstrated closed-loop AI retrosynthesis-directed robotic flow synthesis, a proof-of-concept for self-driving pharmaceutical laboratories now being industrialised by specialist CROs.

• Reaction prediction models trained on USPTO and Reaxys reaction databases complement retrosynthesis planning with forward synthesis feasibility estimation.

  Physics-Based Simulation and AI Force Fields

• Molecular Dynamics Simulation and enhanced sampling (metadynamics, replica exchange, adaptive sampling) provide physically rigorous quantification of protein-ligand binding free energies and conformational flexibility.

• Free energy perturbation (FEP) and thermodynamic integration with GPU-accelerated classical MD (GROMACS, AMBER, OpenMM, Schrodinger FEP+) achieve near-experimental accuracy binding affinity rankings for lead series.

• AI force fields (ANI, MACE, NequIP) achieve DFT-quality potential energy surfaces at molecular mechanics cost, enabling quantum-mechanical-quality simulations on microsecond timescales previously inaccessible ab initio.

• The convergence of AlphaFold structure prediction, AI force fields, and enhanced sampling enables fully computational lead optimisation cycles before any synthesis, radically compressing design-make-test timelines.

• Alchemical free energy calculations rank closely related compound series with sufficient accuracy to guide lead optimisation chemistry without requiring individual experimental affinity measurements for every compound.

  Biomarker Discovery and Patient Stratification

• Biomarker Discovery identifies molecular signatures predicting drug response, disease progression, or safety risks, enabling efficient Clinical Trials through better patient selection and adaptive design.

• Multi-Omics integration across patient cohorts (genomics, transcriptomics, proteomics, metabolomics, epigenomics) provides richer multimodal views of disease biology than any single data type.

• Federated Learning across hospital datasets enables predictive biomarker model training without centralised data sharing, preserving patient privacy and institutional data governance compliance.

• The MELLODDY consortium demonstrated federated multi-task learning across ten major pharmaceutical companies’ proprietary bioactivity datasets, producing models superior to any single-company data.

• Digital Twin patient models — parameterised from multi-omics, imaging, and electronic health records — enable in silico simulation of individual drug responses for Personalised Medicine in oncology and rare metabolic disease.

• AI imaging biomarker discovery using Deep Learning on histopathology and radiology images identifies response-predictive features invisible to manual pathological assessment.

  Formal Pipeline and Computational Workflow

  The computational drug discovery pipeline can be formalised as a sequential multi-stage decision process with feedback loops between stages:

  Stage 1: Target Identification and Validation

• Input: disease omics datasets, genetic association data (GWAS), literature knowledge graphs

• AI methods: multi-omics integration, Knowledge Graph embedding, Causal Inference (Mendelian randomisation)

• Output: ranked list of candidate biological targets with confidence scores and human genetic validation evidence

• Validation: CRISPR functional screen, proteomics, target-disease association literature support

  Stage 2: Hit Discovery via Virtual Screening

• Input: 3D target structure (experimental or Protein Structure Prediction), compound library (ChEMBL, ZINC20, Enamine REAL)

• AI methods: Graph Neural Network binding affinity predictors, Deep Learning docking surrogates (DiffDock, Uni-Mol), Active Learning compound selection

• Output: ranked hit list with predicted binding poses and affinity scores, typically filtered to 0.1–1% of input library

• Validation: experimental binding assay (SPR, ITC, fluorescence polarisation), primary activity assay (IC50)

  Stage 3: Lead Generation via Generative Chemistry

• Input: validated hit compounds, target structural information, ADMET constraints

• AI methods: Variational Autoencoder molecular optimisation, 3D Diffusion Model generation (DiffSBDD), Transformer SMILES generation, Reinforcement Learning multi-objective optimisation

• Output: ranked list of novel molecular structures with predicted activity, selectivity, and ADMET profiles

• Validation: synthesis, binding assay, selectivity panel, in vitro ADMET assay

  Stage 4: Lead Optimisation

• Input: lead compound series with experimental activity and ADMET data

• AI methods: QSAR models, Graph Neural Network multi-task property predictors, Molecular Dynamics Simulation and FEP for binding affinity ranking, Retrosynthesis for synthetic route planning

• Output: optimised lead series with improved potency, selectivity, and ADMET profiles

• Validation: in vitro multi-parameter profiling, in vivo PK studies, safety pharmacology

  Stage 5: Preclinical Candidate Nomination

• Input: optimised lead candidates with full in vitro and in vivo data packages

• AI methods: predictive toxicology models, Digital Twin PK/PD modelling, formulation optimisation

• Output: single preclinical candidate (PC) with IND-enabling data package

• Validation: GLP toxicology studies, IND application preparation

  Feedback Loops and Adaptive Design

• Experimental data from each stage retrains and improves predictive models for the next design cycle

• Active Learning continuously identifies the most informative experiments to run at each stage

• Causal Inference on observational clinical data updates target hypotheses when unexpected signals emerge

• Self-driving laboratory integration closes the loop between computational design and experimental validation at Stage 2–4

  Key Paradigm Families

  Structure-Based Drug Design (SBDD)

• Exploits experimentally determined or computationally predicted 3D protein structures to guide molecular design through shape, hydrogen bonding, hydrophobic, and electrostatic complementarity with the binding pocket.

• AlphaFold 2/3 has made SBDD applicable at proteome scale by providing high-accuracy predicted structures for targets that resisted crystallisation or had no prior structural data.

• Fragment-based drug discovery (FBDD) screens small chemical fragments (MW < 300 Da) binding weakly but efficiently, then grows or merges fragments guided by structural information; AI fragment growing models (DeLinker, FREED) automate guided expansion.

• Molecular docking (AutoDock Vina, Glide, DiffDock) and deep learning scoring functions predict binding poses and affinities for candidate ligands in predicted or experimental pocket structures.

• Hot-spot analysis and allosteric site identification complement orthosteric pocket targeting for difficult or flat binding sites.

  Ligand-Based Drug Design (LBDD)

• Builds predictive activity models from activities of known active compounds when structural information is unavailable or insufficient quality for SBDD.

• QSAR modelling uses Machine Learning regression to predict activity from molecular descriptors; pharmacophore modelling encodes 3D spatial arrangements of chemical features essential for target binding.

• Graph Neural Network and Transformer molecular property predictors are the modern LBDD implementation, trained on large activity datasets rather than small hand-curated pharmacophore models.

• Similarity searching (Tanimoto coefficient on Morgan fingerprints) and scaffold enumeration are classical LBDD tools still used for rapid library filtering before ML scoring.

• Matched molecular pair (MMP) analysis identifies consistent SAR patterns across large activity datasets, guiding medicinal chemistry decisions without explicit model training.

  De Novo Molecular Generation

• Creates novel molecules without starting from existing scaffolds, enabling exploration of chemical space beyond synthesised compound libraries.

• Dominant paradigm as of 2025–2026 as generative model capabilities have matured and compute costs fallen sufficiently for practical pharmaceutical deployment.

• Conditional generation guided by target protein embedding, binding pocket geometry, or desired property constraints allows chemists to specify the design problem and receive ranked candidate molecules.

• Convergence of 3D generative models with AlphaFold pocket structure prediction enables end-to-end pipelines from disease target sequence to candidate drug structures with no prior known binders required.

• De novo generation output must be filtered by synthetic accessibility predictors and retrosynthesis planning before synthesis, preventing waste on chemically attractive but practically intractable proposals.

  Drug Repurposing and Network Medicine

• Identifies new clinical applications for existing approved drugs, reducing time-to-proof-of-concept from years to months by leveraging existing safety pharmacology data packages.

• Knowledge Graph embedding methods (TransE, RotatE, ComplEx) trained on drug-target, gene-disease, protein-protein, and pathway networks enable link prediction for new drug-indication pairs.

• Network proximity analysis quantifies topological distance between a drug’s target neighbourhood and a disease module within the human interactome, providing a quantitative repurposing plausibility metric.

• BenevolentAI’s knowledge graph identified Baricitinib for COVID-19 (January 2020), validated in the WHO SOLIDARITY trial and approved under Emergency Use Authorisation — the landmark validation of AI-driven repurposing.

• SARS-CoV-2 drug repurposing screens using drug-target interaction databases and AlphaFold structures identified multiple candidate antivirals within weeks of structural data availability.

  Multi-Target Drug Design

• Addresses diseases driven by multiple redundant pathways (cancer, neurodegeneration, metabolic syndrome) where single-target inhibition produces partial or transient benefit with resistance development.

• Engineers single molecules with polypharmacological profiles combining activity against multiple disease-relevant targets simultaneously, avoiding pharmacokinetic complexity of drug combinations.

• Reinforcement Learning and multi-objective Pareto optimisation handle the complex multi-dimensional objective landscape of simultaneous activity, selectivity, and ADMET constraints across multiple targets.

• Dual-acting inhibitors (e.g. EGFR/HER2 dual inhibitors, GSK-3β/HDAC dual inhibitors) exemplify validated multi-target drug design achieving clinical benefit beyond single-agent profiles.

• Computational fragment merging and scaffold hybridisation guided by structural biology of multiple target binding pockets enable systematic multi-target design.

  Use Cases / Major Families

  Oncology

• Most active AI drug discovery application area, driven by cancer genomics consortia (TCGA, ICGC) data abundance and clinical urgency for improved cancer therapeutics.

• Graph Neural Network models predict kinase inhibitor binding affinity against oncology-relevant kinase panels from molecular graph representations, enabling rapid selectivity profiling before synthesis.

• PROTAC (Proteolysis-Targeting Chimera) degrader design: AI optimises ternary complex geometry between target protein, E3 ubiquitin ligase, and bifunctional degrader molecule — a three-body problem beyond classical docking.

• Recursion REC-4881 (AI-discovered PIK3CB inhibitor): 75% polyp burden reduction, 82% durable response at week 25, in familial adenomatous polyposis patients (2025 clinical data).

• Immuno-oncology biomarker discovery: Deep Learning models identify tumour microenvironment features from histopathology images predicting checkpoint immunotherapy response.

• AI-designed degraders, covalent inhibitors, and stapled peptides target previously “undruggable” transcription factors and protein-protein interactions in oncology.

  Infectious Disease

• Stokes et al. (Cell 2020): Deep Learning trained on 2,335 molecules with measured antibacterial activity screened 107 million drug-like compounds in silico, identifying halicin (activity against pan-drug-resistant Acinetobacter baumannii).

• AlphaFold-enabled structure-based virtual screening against SARS-CoV-2 main protease and other viral targets within weeks of pathogen emergence during COVID-19 pandemic.

• Fragment-growing generative models explore antiviral chemical space for HIV, influenza, hepatitis C, and emerging respiratory virus targets leveraging conserved active site geometry.

• Drug repurposing via Knowledge Graph link prediction identifies approved antivirals with predicted activity against novel viral protein targets faster than de novo discovery timelines.

• AI antimicrobial discovery addresses the antimicrobial resistance (AMR) crisis by exploring novel antibiotic chemical space against priority WHO pathogens intractable to conventional approaches.

  Rare and Neglected Diseases

• Rare diseases (<200,000 US patients or <5/10,000 EU) and neglected tropical diseases have enormous unmet need but insufficient market size to support traditional HTS programmes.

• Generative molecular design enables economic virtual library exploration because screening cost is primarily computational rather than synthesis and assay-based.

• Federated Learning across rare disease registries and academic centres aggregates statistical strength across individually too-small datasets without centralised data sharing.

• Orphan drug designation and priority review vouchers incentivise AI-first drug discovery companies to explore rare disease targets with compressed development timelines.

• Patient-derived induced pluripotent stem cell (iPSC) disease models combined with AI phenotypic scoring enable unbiased compound screening without prior target knowledge.

  CNS Drug Discovery

• Blood-brain barrier penetration, P-gp efflux avoidance, CYP450 metabolism, hERG inhibition, and CNS receptor off-target binding must all be addressed simultaneously in CNS lead optimisation.

• Deep Learning models jointly predict multi-parameter CNS ADMET profiles (BBB penetration, microsomal half-life, P-gp ratio, hERG inhibition, CNS receptor selectivity) from molecular structure.

• Multi-task Graph Neural Network architectures share representations across correlated CNS ADMET endpoints, improving data-sparse endpoint predictions through cross-task transfer.

• Blood-brain barrier-specific generative constraints incorporated into molecular design reward functions prevent synthesis of otherwise potent molecules with unacceptable CNS liabilities.

• Target identification in neurodegeneration (Alzheimer’s, Parkinson’s) integrates single-cell transcriptomics, proteomics from patient brain tissue, and GWAS genetic data to identify disease-relevant targets expressed in specific neuronal subtypes.

  Antibody and Biologic Design

• Sequence-to-structure protein language models (ESMFold, RoseTTAFold, AlphaFold 3) and inverse folding models (ProteinMPNN) enable antibody and biologic design from sequence space without full experimental library screening.

• RFdiffusion (Baker Lab, 2023) demonstrated de novo design of functional protein binders, enzyme scaffolds, and symmetric protein assemblies validated experimentally.

• Antibody CDR engineering via Deep Learning optimises affinity maturation, stability, and reduced immunogenicity without labour-intensive phage display campaigns.

• Bispecific antibody and antibody-drug conjugate (ADC) linker design: AI models optimise geometrical and physicochemical linker properties for stable conjugation and conditional payload release.

• Nanobody (single-domain antibody) design against GPCRs and other difficult targets benefits from AI sequence design guided by structural predictions of target epitopes.

  Drug Repurposing

• BenevolentAI knowledge graph identified Baricitinib (JAK1/JAK2 inhibitor) for COVID-19 in January 2020, based on predicted AAK1 and BIKE kinase inhibition regulating coronavirus endocytosis.

• Baricitinib validated in WHO SOLIDARITY trial, received FDA Emergency Use Authorisation, and entered standard COVID-19 treatment protocols — the landmark clinical AI repurposing validation.

• Knowledge Graph link prediction (DrugBank, STRING, DisGeNet, OpenTargets) identifies drug-indication pairs for systematic clinical validation in basket or umbrella trials.

• AI-identified repurposing hypotheses have shorter safety evaluation timelines due to existing ADMET and human safety data, enabling faster proof-of-concept trials than de novo programmes.

• Network proximity scores in the human protein interactome guide rational combination therapy design by identifying drug pairs with complementary target neighbourhoods in disease modules.

  Academic Context

• 1964: Hansch and Fujita publish QSAR (J. Am. Chem. Soc.), establishing physicochemical molecular properties correlate quantitatively with biological activity — the foundation of computational drug design.

• 1988–1995: HIV protease crystal structures enable structure-based design of the first generation of HIV protease inhibitors; Saquinavir receives FDA approval (1995) as the first structure-based drug.

• 2007: Irwin and Shoichet publish ZINC database paper; free online library of purchasable compounds enables large-scale computational virtual screening.

• 2012: Merck molecular activity prediction challenge on Kaggle won by Hinton’s Toronto group using deep neural networks, demonstrating representation learning outperforms hand-crafted fingerprint descriptors — the launch of the deep learning era in molecular ML.

• 2014: DeepTox challenge validates deep neural networks for toxicity prediction across multiple endpoints simultaneously, establishing multi-task learning as superior to single-task QSAR for data-sparse endpoints.

• 2015: Duvenaud et al. (NeurIPS 2015) introduce convolutional networks on molecular graphs for learning molecular fingerprints — the first molecular graph neural network applied to property prediction.

• 2017: Gilmer et al. (ICML 2017) formalise the neural message passing framework for molecular property prediction, establishing the Graph Neural Network paradigm now dominant in molecular ML.

• 2018: Gómez-Bombarelli et al. (ACS Cent. Sci.) introduce the first VAE-based continuous molecular representation space enabling gradient-based property optimisation — the foundational generative chemistry paper.

• 2018: Sanchez-Lengeling and Aspuru-Guzik (Science) review inverse molecular design, articulating the generative chemistry paradigm for the broader scientific community.

• 2019: Zhavoronkov et al. (Nature Biotechnology) demonstrate AI-designed DDR1 kinase inhibitors synthesised and biologically validated in 46 days — the first published AI-to-experiment cycle.

• 2020: Jumper et al. (AlphaFold 2, Nature 2021 / CASP14 2020) — characterised by CASP organisers as “a solution to protein folding”; atomic-level accuracy from sequence alone transforms structural biology and SBDD.

• 2020: Stokes et al. (Cell) identify halicin against drug-resistant Acinetobacter baumannii using deep learning screening — landmark antibiotic discovery paper demonstrating clinical relevance of AI drug discovery.

• 2021: Huang et al. launch Therapeutics Data Commons (TDC), standardising benchmarking across 22 ADMET endpoints and 66 ML tasks enabling rigorous cross-method comparison.

• 2022: MELLODDY consortium demonstrates federated multi-task learning across ten pharmaceutical companies’ proprietary data, producing superior models while preserving data confidentiality — proof-of-concept for privacy-preserving pre-competitive collaboration.

• 2023: Corso et al. (DiffDock, ICLR 2023) demonstrate diffusion-based blind molecular docking predicting binding poses for novel targets without prior pocket specification, achieving accuracy competitive with physics-based methods for unseen target-ligand pairs.

• 2023: Hayes et al. (ESM3, Meta) demonstrate protein language models as a unified representation for sequence, structure, and function prediction, enabling structure-conditioned de novo protein generation.

• 2024: Abramson et al. (AlphaFold 3, Nature) extend framework to model protein-ligand, protein-nucleic acid, and multi-component complexes using diffusion architecture — enabling SBDD for previously inaccessible complex targets.

• 2024: Recursion Pharmaceuticals acquires Exscientia for USD 688 million, creating the largest AI-native pharmaceutical company and signalling consolidation of the AI drug discovery sector.

• 2025: Insilico Medicine publishes positive Phase IIa rentosertib results in Nature Medicine — first USAN-named compound where both target and drug were AI-designed.

• 2026: UK OpenBind initiative announced (June 2026): government-backed platform for open AI drug discovery infrastructure targeting £100 billion cost reduction through national compute and pre-competitive biomedical data sharing.

• 2026: First AI-designed drug approval anticipated by analysts; if achieved, this milestone will catalyse a fundamental restructuring of pharmaceutical R&D investment toward AI-first pipelines.

  Current Landscape (2026)

• 173+ AI-discovered drug programmes in clinical development globally as of June 2026; 15–20 expected to enter pivotal Phase III trials in 2026.

• Insilico Medicine’s rentosertib (ISM001-055): AI-designed TNIK inhibitor for idiopathic pulmonary fibrosis — positive Phase IIa results published in Nature Medicine (June 2025); first USAN-named compound where both target and molecule were fully AI-designed; Phase III planning underway.

• Insilico raised USD 293 million Hong Kong IPO and USD 100 million Eli Lilly partnership for pipeline expansion.

• Recursion Pharmaceuticals completed USD 688 million acquisition of Exscientia (late 2024), creating the largest AI-native pharmaceutical company by pipeline breadth.

• Recursion’s REC-4881 (PIK3CB inhibitor for familial adenomatous polyposis): 75% polyp burden reduction, 82% durable response at week 25 (2025 Phase II data).

• FDA January 2025 draft guidance on AI in drug development: seven-step credibility assessment framework for AI-generated evidence in regulatory submissions — the most significant regulatory signal for the field.

• First AI-designed drug approval projected 2026–2027 with ~60% probability by industry analysts.

• AstraZeneca, GSK, Pfizer, Roche, and Merck KGaA have all established substantial internal AI drug discovery programmes, signalling the technology is no longer experimental in the pharmaceutical mainstream.

• AlphaFold 3 public release (2024): triggered proliferation of SBDD campaigns against previously intractable targets; hundreds of research groups running AlphaFold-enabled SBDD pipelines.

• UK OpenBind initiative (June 2026): government commitment to making UK a global AI drug discovery leader, targeting £100 billion cost reduction through open AI platforms, national compute, and pre-competitive data sharing.

• BenevolentAI’s baricitinib COVID-19 repurposing (2020) validation in randomised controlled trial (ACTT-2, NEJM 2021) provided the first rigorous clinical proof-of-concept for AI-driven drug repurposing at the level of a randomised controlled trial publication.

• ATOM (Accelerating Therapeutics for Opportunities in Medicine) consortium: GlaxoSmithKline, UCSF, and national laboratories collaborating on open ML models for ADMET and activity prediction; demonstrated pre-competitive data sharing at industrial scale outside the EU.

• AMR (Antimicrobial Resistance) drug discovery: AI methods gaining traction for identifying novel antibiotic scaffolds against WHO priority pathogens (Acinetobacter baumannii, Klebsiella pneumoniae, MRSA) where conventional discovery economics are insufficient; Stokes et al. halicin demonstration catalysed field.

• Rare disease drug discovery: AI-enabled patient stratification and biomarker identification reduces the required clinical trial size for rare diseases (Orphan Drug Act designation), improving commercial viability for conditions with small patient populations.

• Major AI drug discovery platform companies operational at scale (2026): Insilico Medicine, Recursion (post-Exscientia), Schrödinger, Relay Therapeutics, Nurix Therapeutics, EQT-backed Elix, and Valo Health; collectively operating 400+ AI-active pipeline assets.

• CRO (Contract Research Organisation) adoption: WuXi AppTec, Evotec, Covance, and other major CROs have embedded AI drug discovery capabilities into client-facing services, democratising AI-augmented drug discovery to smaller biotechs without internal AI infrastructure.

• Foundation model transfer to drug discovery: pharmaceutical companies fine-tuning general-purpose LLMs (GPT-4o, Claude 3.5 Sonnet) on proprietary bioassay data and scientific literature for interactive compound suggestion, SAR interpretation, and scientific hypothesis generation.

• Computational efficiency: GPU clusters dedicated to drug discovery (Insilico’s internal cluster, Recursion’s Phoenix supercomputer) now rank among the highest-utilised scientific computing infrastructure, reflecting the compute intensity of deep generative model training and inference at pharma scale.

  UK Context

• UK accounts for ~10% of global branded pharmaceutical R&D expenditure; globally prominent in AI drug discovery combining world-leading structural biology research with a concentrated biopharma ecosystem.

• EMBL-EBI (Hinxton, Cambridge): maintains ChEMBL, UniProt, Protein Data Bank in Europe (PDBe), and European Phenome-Genotype Archive — globally unique concentration of open biomedical training data under UK institutional stewardship.

• Wellcome Sanger Institute (Hinxton): large-scale genomics, epigenomics, and functional genomics datasets underpinning AI target identification and biomarker discovery.

• Francis Crick Institute (London): structural biology, CRISPR functional genomics, and immunology research supporting AI drug target validation.

• MRC Laboratory of Molecular Biology (Cambridge): foundational protein structure and biology research; multiple Nobel laureate-associated groups.

• AstraZeneca (Cambridge): Centre for Genomics Research and AI Medicines Accelerator applying Machine Learning throughout discovery pipelines; active in multi-omics target identification and generative chemistry.

• GSK (London/Stevenage): AI-Medicines Accelerator and collaboration with Exscientia before its acquisition; AI applied to target identification, ADMET prediction, and clinical trial design.

• BenevolentAI (London, founded 2013): pioneered knowledge-graph-based drug repurposing; provided landmark COVID-19 baricitinib repurposing validation (January 2020).

• Exscientia (Oxford, founded 2012): demonstrated AI-designed molecules in Phase I clinical trials; first GMP-manufactured AI-designed drug candidate; acquired by Recursion for USD 688 million (2024).

• Imperial College London: Institute for Digital Molecular Design and Fabrication and UKRI Centre in AI for Healthcare; partnering with World Economic Forum in OpenBind initiative (June 2026).

• Francis Crick / Medicines Discovery Catapult partnership: KQ Labs national accelerator for up to 10 data-driven drug discovery start-ups annually.

• Medicines Discovery Catapult (MDC) (Alderley Park, Cheshire — Manchester pharmaceutical corridor): integrated computational and wet-laboratory infrastructure for SME drug discovery; translational bridge between AI models and experimental validation.

• University of Edinburgh: Usher Institute and Wellcome Centre for Cell Biology contribute structural biology expertise; National Phenome Centre collaboration supports metabolomics ADMET data.

• University of Manchester (Henry Royce Institute): AI methods applied to materials-drug delivery interfaces; ELIXIR UK node coordination.

• University of Cambridge: MRC Cancer Unit, Milner Therapeutics Institute, and Cavendish AI/ML groups active in ML for target identification, clinical data integration, and molecular design.

  Future Directions (2026-2030)

• Self-driving laboratories: closed-loop AI-directed robotic synthesis and assay platforms iteratively explore and optimise chemical space with minimal human intervention in design-make-test-analyse cycles; proof-of-concept platforms being industrialised at pharmaceutical CROs.

• First fully AI-designed and AI-trialled clinical approval anticipated 2026–2028 — a milestone expected to catalyse a wave of pharmaceutical industry investment and restructuring.

• Multimodal foundation models jointly reasoning over molecular sequence, 3D structure, biological assay data, patient omics profiles, and clinical outcome data in a unified representational framework.

• Protein language models (ESM3, ProGen, ProteinGPT) expected to achieve sufficient functional accuracy for full de novo biologic and enzyme design without experimental library screening.

• AI-driven agentic laboratory orchestration: AI agents managing multi-step experimental workflows including compound synthesis, purification, biological assay, and data analysis as integrated automated pipelines.

• Regulatory convergence: EU AI Act, UK MHRA adaptive pathway, and FDA’s seven-step credibility framework expected to converge on AI-specific evidence standards for clinical submissions by 2027–2028.

• Federated Learning across pharma company and hospital networks under UK Data Protection and Digital Information Act and EU AI Act governance frameworks will enable larger, more diverse training datasets without data centralisation.

• Mechanistic AI models integrating kinetic models, signalling pathway simulators, and pharmacodynamics models with data-driven ML for physiologically grounded drug response prediction.

• Patient-specific digital twin models for individual drug response personalisation in oncology, rare metabolic diseases, and autoimmune conditions reaching clinical deployment.

• AI-designed prodrugs, nanomedicines, and novel drug delivery systems extending AI design from active molecule to complete therapeutic product.

• Bispecific antibody and ADC (Antibody-Drug Conjugate) design: AI methods for multi-domain biologic optimisation combining antigen-binding arm design, linker chemistry, and payload selection for next-generation oncology therapeutics.

• Covalent and targeted covalent inhibitor design: AI methods for identifying electrophilic warhead-compatible binding pockets and designing selective covalent probes against previously undruggable targets.

• RNA-targeted drug discovery: Deep Learning models for predicting small molecule binding to RNA secondary structures (riboswitches, splice sites, viral RNA) opening up the unexplored RNA druggable space estimated at 80% of the transcriptome.

  Standards, Regulation, and Data Infrastructure

  Regulatory Frameworks for AI-Driven Drug Development

• FDA Draft Guidance on AI/ML (January 2025): seven-step credibility assessment framework establishing how AI-generated evidence must be documented, validated, and presented in IND and NDA regulatory submissions.

• EMA Reflection Paper on AI in Drug Development (2023): articulates principles for AI/ML model documentation, data quality, algorithm transparency, and lifecycle management in European marketing authorisation applications.

• MHRA (UK) Adaptive Regulatory Pathway: UK Medicines and Healthcare products Regulatory Agency has established adaptive pathways for AI-enabled drug candidates, coordinating with the FDA and EMA through the International Council for Harmonisation (ICH).

• ICH S7A/S7B guidelines: govern safety pharmacology data requirements for cardiovascular, CNS, and respiratory systems still required empirically even when AI prediction informs lead selection.

• FDA Complex Innovative Trial Designs (CID) programme: facilitates adaptive Clinical Trials designs informed by Biomarker Discovery and computational disease models, reducing sample size requirements through better patient stratification.

• EU AI Act (2024): classifies AI systems used for drug development as high-risk, requiring registration, conformity assessment, and ongoing monitoring post-deployment.

  Molecular Representation Standards

• SMILES (Simplified Molecular Input Line Entry System): line notation representing molecular structure as a string; de facto standard input for Transformer language models and cheminformatics tools.

• InChI (International Chemical Identifier): canonical IUPAC molecular identifier enabling unambiguous cross-database molecule identification; standard for data integration across ChEMBL, PubChem, ZINC.

• SELFIES (Self-Referencing Embedded Strings): alternative molecular string representation where every string is a valid molecule, eliminating syntactic validity as a constraint for generative models.

• Molecular graph with RDKit featurisation: standard input format for Graph Neural Network models, encoding atom and bond features as node and edge attributes.

• SDF/MOL2 format: 3D molecular structure file formats for docking and molecular dynamics simulation inputs.

• mmCIF/PDB format: protein structure file formats for structural biology applications; output format of AlphaFold and experimental crystallography.

  FAIR Data Principles and Open Science Infrastructure

• FAIR (Findable, Accessible, Interoperable, Reusable) data principles increasingly mandated for AI training data in regulated pharmaceutical contexts by FDA and EMA guidance.

• ELIXIR European infrastructure provides FAIR biomedical data platform including ChEMBL, UniProt, PDBe, and cross-linking services enabling reproducible molecular AI research at European scale.

• Open Targets (EMBL-EBI, Sanger, Biogen, Bristol Myers Squibb) integrates genetic, drug, and disease data for open target identification and validation across human diseases.

• Pistoia Alliance and Structural Genomics Consortium promote open-source software and pre-competitive data sharing to accelerate the field beyond proprietary data silo constraints.

• MELLODDY consortium demonstrated that federated pre-competitive data sharing across pharma can be technically implemented and produces models superior to any single organisation’s proprietary data alone.

  Limitations and Open Challenges

• Distribution shift: models trained on synthesisable, drug-like chemical space may extrapolate poorly when exploring structurally novel regions with properties outside the training distribution; confidence calibration methods and uncertainty quantification are active research areas.

• Data scarcity and imbalance: ADMET endpoint datasets are small (hundreds to low thousands of molecules), imbalanced (rare adverse events), and noisy (assay variability across laboratories); addressed partially by Active Learning, transfer learning, and data augmentation.

• Explainability and interpretability: black-box deep learning models are difficult to interpret mechanistically; regulatory bodies increasingly expect chemically meaningful rationale supporting AI predictions, driving development of attention-based attribution, GNNExplainer, and concept-based explanations.

• Synthesis accessibility gap: generative models propose molecules that are chemically attractive in silico but synthetically intractable or prohibitively expensive; addressed by Retrosynthesis planning integration and synthetic accessibility scoring in reward functions.

• Clinical translation gap: even optimal preclinical AI predictions cannot substitute for empirical human safety and efficacy data required by regulators; the majority of Phase II and Phase III failures reflect biology and patient heterogeneity rather than prediction model quality.

• Reproducibility: heterogeneous experimental conditions, assay variability, and data provenance issues in public datasets complicate fair benchmarking; TDC and MoleculeNet provide standardised splits but cannot eliminate all experimental irreproducibility.

• Multi-target activity prediction: most models predict single-endpoint activity; predicting the full activity profile of a molecule across the proteome (selectivity, off-target binding, metabolic enzymes) simultaneously remains an unsolved challenge.

• Protein flexibility and induced fit: AlphaFold 3 and most structure prediction tools provide single or small ensembles of conformations; accounting for binding-induced protein flexibility (induced fit) and allosteric effects in docking and generative design remains an open problem.

• Long-range clinical prediction: connecting molecular properties to clinical efficacy in complex, heterogeneous patient populations requires models spanning molecular pharmacology, PK/PD, disease biology, and clinical response — a multi-scale integration challenge beyond current single-domain models.

• AI-designed molecule IP: patentability of AI-generated molecular structures and the inventorship question (can an AI be an inventor?) remain unresolved in UK, EU, and US patent law, creating uncertainty for AI drug discovery companies.

  Benchmark Datasets and Evaluation Standards

  Molecular Property Prediction Benchmarks

• MoleculeNet (Wu et al., 2018): Standardised collection of 17 molecular property prediction datasets spanning quantum chemistry (QM7, QM9), physical chemistry (ESOL, FreeSolv, Lipophilicity), biophysics (BACE, PCBA), and physiology (HIV, ClinTox, Tox21, ToxCast) with predefined splits and evaluation metrics. The standard reference benchmark for evaluating Graph Neural Network and Transformer molecular property models.

• Therapeutics Data Commons (TDC) (Huang et al., 2021): Comprehensive benchmark covering 22 ADMET endpoints, 66 machine learning tasks across drug discovery, development, and manufacturing, with standardised dataset splits and a public leaderboard. Became the primary reference benchmark for the AI drug discovery community from 2021 onward.

• AqSolDB: Aqueous solubility benchmark containing ~10,000 compounds with experimental solubility measurements; used to evaluate solubility prediction models critical for lead optimisation.

• BindingDB: Binding affinity database of ~2.9 million measured binding affinities for ~1.2 million drug-like molecules against 9,000 protein targets; used as primary training and evaluation data for affinity prediction models.

  Virtual Screening and Docking Benchmarks

• DUD-E (Directory of Useful Decoys — Enhanced): 22,886 active compounds and property-matched decoys against 102 protein targets; standard benchmark for virtual screening method evaluation.

• DEKOIS 2.0: Challenging benchmark including 81 target datasets designed to reduce analogue bias in virtual screening evaluation.

• LIT-PCBA: Activity-balanced virtual screening benchmark based on PubChem BioAssay data, designed to address the artifactual enrichment problems in earlier benchmarks.

  Protein Structure and Molecular Generation Benchmarks

• CASP (Critical Assessment of Protein Structure Prediction): Biennial blind prediction competition evaluating computational methods for protein structure prediction from amino acid sequence; CASP14 (2020) and CASP15 (2022) validated AlphaFold 2’s breakthrough accuracy.

• CASF (Comparative Assessment of Scoring Functions): Benchmark for evaluating protein-ligand binding affinity prediction and docking scoring function performance against the PDBbind refined set.

• GuacaMol: Generative chemistry benchmark measuring distribution learning (ability to reproduce property distribution of known drugs), goal-directed generation (ability to optimise molecular properties), and de novo molecular design.

• MOSES (Molecular Sets): Generative chemistry benchmark for evaluating the quality of generated molecular distributions using metrics including validity, uniqueness, novelty, FCD (Fréchet ChemNet Distance), and scaffold diversity.

  Data Repositories

• ChEMBL (EMBL-EBI): Over 2.4 million bioactive molecules with measured biological activity data; the primary open training resource for molecular ADMET and activity models.

• Protein Data Bank (PDB): Over 230,000 experimentally determined protein and nucleic acid structures; training data for protein structure prediction and virtual docking.

• AlphaFold Protein Structure Database (DeepMind/EMBL-EBI): Over 200 million predicted protein structures for virtually all known proteomes; enables SBDD at proteome scale.

• ZINC20: 1.4 billion purchasable compounds; used for ultra-large virtual library screening.

  Key Terminology

• ADMET: Absorption, Distribution, Metabolism, Excretion, Toxicity — the pharmacokinetic and safety properties that determine whether a compound is likely to become a successful drug; predicted computationally from molecular structure to triage candidates before synthesis.

• Target Identification: The process of identifying a disease-relevant biological macromolecule (protein, RNA, protein-protein interaction) whose modulation by a drug would produce a therapeutic benefit in patients.

• Lead Optimisation: The iterative chemistry phase in which a hit compound’s potency, selectivity, and ADMET properties are improved through systematic structural modifications guided by structure-activity relationship (SAR) data.

• Virtual Screening: Computational filtering of large compound libraries by predicted binding affinity, ADMET properties, or other criteria to select candidates for experimental synthesis and assay.

• SMILES (Simplified Molecular Input Line Entry System): A line notation representing molecular structure as a string of characters; widely used as input representation for Transformer language models trained on molecular data.

• Molecular Graph: Graph representation of a molecule with atoms as nodes and bonds as edges, labelled with atom type, formal charge, hybridisation, and bond order features; the canonical input for Graph Neural Network molecular models.

• Binding Affinity: The strength of the interaction between a drug molecule and its target protein, typically measured as IC50, Ki, or Kd; the primary optimisation target in lead optimisation.

• Structure-Based Drug Design (SBDD): Drug design approach that exploits the 3D structure of the biological target (experimentally or computationally determined) to guide molecular design.

• Free Energy Perturbation (FEP): Rigorous physics-based alchemical simulation method for calculating relative binding free energies between closely related compounds with near-experimental accuracy.

• Generative Chemistry: AI paradigm in which machine learning models generate novel molecular structures with desired properties, rather than merely scoring existing compounds.

• Hit-to-Lead (H2L): The drug discovery phase where initial active compounds (hits) from screening are optimised to identify higher-quality lead compounds with improved potency and drug-likeness.

• PROTAC (Proteolysis-Targeting Chimera): Bifunctional molecule that recruits E3 ubiquitin ligase to a target protein to induce its degradation; an emerging drug modality enabled by AI-assisted linker design.

• Scaffold Hopping: Replacement of the core chemical scaffold of a lead compound while retaining biological activity, used to improve pharmacokinetic properties, intellectual property differentiation, or synthetic accessibility.

• Drug Repurposing: Identifying new therapeutic applications for existing approved drugs, leveraging existing safety data to accelerate proof-of-concept; computationally driven by Knowledge Graph link prediction.

• QSAR (Quantitative Structure-Activity Relationship): Statistical or machine learning models relating molecular structural features to biological activity measurements; the foundation of ligand-based drug design.

• HTS (High-Throughput Screening): Automated experimental technique screening hundreds of thousands to millions of compounds against biological targets; computationally complemented and increasingly preceded by virtual screening.

• Pharmacophore: The spatial arrangement of molecular features (hydrogen bond donors/acceptors, hydrophobic centres, aromatic rings, charged groups) necessary for a molecule to exhibit a particular biological activity.

• Blood-Brain Barrier (BBB): Selective permeability barrier of the central nervous system vasculature; molecules must cross it to reach CNS targets, making BBB penetration a critical ADMET constraint for CNS drugs.

• Selectivity: The degree to which a drug acts on its intended target relative to other proteins; predicted by AI models using kinase activity profiles, proteome-wide off-target activity datasets.

• Synthetic Accessibility (SA): Measure of how easily a molecule can be synthesised from commercially available starting materials; incorporated into generative model reward functions to prevent synthesis-intractable proposals.

• Diffusion Model: Generative model class that learns to reverse a gradual noising process; applied to 3D molecular generation (DiffSBDD, TargetDiff) to produce protein-pocket-conditioned molecular structures.

• Multi-Omics: Integration of multiple types of genome-scale biological data (genomics, transcriptomics, proteomics, metabolomics, epigenomics) to provide richer views of disease biology for target identification and stratification.

• Federated Learning: Machine learning approach where models are trained across distributed data sources (e.g. multiple hospital or company datasets) without centralising sensitive data; applied in MELLODDY consortium for cross-pharma bioactivity model training.

• Active Learning: Iterative machine learning methodology where the model actively queries for labels on the most informative unlabelled examples; applied to ADMET and activity modelling to maximise information gain per assay dollar spent.

• Molecular Dynamics (MD) Simulation: Physics-based simulation of atomic motion governed by Newton’s equations of motion and a potential energy function (force field); used to sample protein-ligand binding conformations and compute binding free energies.

• Free Energy Perturbation (FEP): Alchemical simulation method for computing relative binding free energies between chemically related compounds by gradually transforming one molecule into another within an MD simulation; near-experimental accuracy for lead optimisation ranking.

• Protein Language Model: Neural network trained on evolutionary sequence data (UniRef, BFD) treating amino acid sequences analogously to natural language tokens; learns evolutionary conservation patterns that encode structural and functional constraints; examples: ESM-2, ProtTrans, ProGen.

• Chemical Space: The abstract N-dimensional space of all possible chemical compounds, estimated to contain 10^23–10^60 drug-like molecules; AI generative models navigate this space during de novo molecular design.

• Allosteric Site: Binding site on a protein distinct from the orthosteric (active) site; ligands binding allosteric sites modulate protein function indirectly; increasingly targeted by AI-driven structure-based design for previously intractable proteins.

• USAN (United States Adopted Name): Official non-proprietary name assigned by the USAN Council to a pharmaceutical compound entering clinical development; Insilico Medicine’s rentosertib was the first AI-designed compound to receive a USAN.

• Toxicophore: Structural feature or fragment associated with toxic biological activity; flagging toxicophores in generative model output is a first-line filter preventing synthesis of safety-problematic compounds.

• Matched Molecular Pair (MMP): A pair of compounds differing by a defined structural transformation at a single site; MMP analysis quantifies the average property impact of a structural change, providing interpretable SAR data for lead optimisation.

• Bioisostere: A chemical group that can replace another in a molecule while maintaining similar biological activity; bioisosteric replacement is a key tool in lead optimisation for improving metabolic stability or reducing toxicity while preserving potency.

• Clinical Candidate (CC): A compound selected to enter GLP toxicology studies and IND-enabling studies; achieving CC status represents the culmination of discovery-phase optimisation before the clinical development investment commitment.

• Disease Area Ontology: Structured controlled vocabulary linking diseases, genes, proteins, pathways, and phenotypes; used to anchor Knowledge Graph based target identification and multi-omics pathway analysis in computational drug discovery pipelines.

• Phenotypic Screening: Drug discovery approach testing compounds in disease-relevant cellular or animal models without knowing the molecular target; AI analysis of phenotypic screening images and data enables target deconvolution.

• Target Deconvolution: Process of identifying the molecular target(s) responsible for a phenotypic effect observed in a cell or animal model; AI chemo-proteomic methods (proteome-wide ligand affinity mapping, computational target fishing) accelerate this process.

• Open Targets: European Bioinformatics Institute open platform integrating genetic, drug, literature, and biological pathway data to support systematic target identification; provides scored gene-disease associations with supporting evidence chains.

• Multiparameter Optimisation (MPO): Lead optimisation strategy simultaneously optimising multiple drug-like properties (potency, selectivity, solubility, permeability, metabolic stability) through composite scoring functions or Pareto optimisation.

• Chemical Matter: Collective term for the compound collections (physical libraries, virtual libraries, fragment libraries) screened or designed during hit identification; quality and diversity of chemical matter determines accessible SAR space.

• Natural Product-Inspired Design: Generative model approach drawing inspiration from natural product scaffolds — historically the most prolific source of drug candidates — to design novel synthetic analogues with enhanced pharmacological profiles.

  Research & Literature

  1. Jumper, J. et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596, 583–589. https://doi.org/10.1038/s41586-021-03819-2
  2. Abramson, J. et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630, 493–500. https://doi.org/10.1038/s41586-024-07487-w
  3. Senior, A. W. et al. (2020). Improved protein structure prediction using potentials from deep learning. Nature, 577, 706–710.
  4. Gilmer, J. et al. (2017). Neural message passing for quantum chemistry. Proceedings of ICML 2017, 1263–1272.
  5. Yang, K. et al. (2019). Analyzing learned molecular representations for property prediction. J. Chem. Inf. Model., 59(8), 3370–3388.
  6. Stokes, J. M. et al. (2020). A deep learning approach to antibiotic discovery. Cell, 180(4), 688–702.
  7. Zhavoronkov, A. et al. (2019). Deep learning enables rapid identification of potent DDR1 kinase inhibitors. Nature Biotechnology, 37, 1038–1040.
  8. Gómez-Bombarelli, R. et al. (2018). Automatic chemical design using a data-driven continuous representation of molecules. ACS Central Science, 4(2), 268–276.
  9. Sanchez-Lengeling, B. & Aspuru-Guzik, A. (2018). Inverse molecular design using machine learning. Science, 361(6400), 360–365.
  10. Corso, G. et al. (2022). DiffSBDD: Structure-based drug design with equivariant diffusion models. ICML 2022 Workshop on Structured Prediction.
  11. Davies, M. et al. (2015). ChEMBL web services: streamlining access to drug discovery data and utilities. Nucleic Acids Research, 43(W1), W612–W620.
  12. Huang, K. et al. (2021). Therapeutics Data Commons: machine learning datasets and tasks for drug discovery and development. NeurIPS 2021 Datasets and Benchmarks Track.
  13. Bender, A. & Cortés-Ciriano, I. (2021). Artificial intelligence in drug discovery: what is realistic, what are illusions? Part 1. Drug Discovery Today, 26(2), 511–524.
  14. Schneider, P. et al. (2020). Rethinking drug design in the artificial intelligence era. Nature Reviews Drug Discovery, 19(5), 353–364.
  15. Walters, W. P. & Murcko, M. (2020). Assessing the impact of generative AI on medicinal chemistry. Nature Biotechnology, 38, 143–145.
  16. Méndez-Lucio, O. et al. (2020). De novo generation of hit-like molecules from gene expression signatures using artificial intelligence. Nature Communications, 11, 10.
  17. Insilico Medicine (2023). Identification of preclinical candidate for idiopathic pulmonary fibrosis. Nature Biotechnology, 41, 402–408.
  18. Krishnan, R. et al. (MELLODDY Consortium) (2022). Multi-task learning for drug discovery with private data. Science, 378(6624), 1088.
  19. Gao, W. & Coley, C. W. (2020). The synthesizability of molecules proposed by generative models. J. Chem. Inf. Model., 60(12), 5714–5723.
  20. Coley, C. W. et al. (2019). A robotic platform for flow synthesis of organic compounds informed by AI planning. Science, 365(6453), eaax1566.
  21. Duvenaud, D. et al. (2015). Convolutional networks on graphs for learning molecular fingerprints. NeurIPS 2015.
  22. Lipinski, C. A. et al. (2001). Experimental and computational approaches to estimate solubility and permeability in drug discovery. Advanced Drug Delivery Reviews, 46(1–3), 3–26.
  23. FDA (2025). Artificial Intelligence in Drug Development: Draft Guidance for Industry. U.S. Food and Drug Administration, January 2025. https://www.fda.gov
  24. UK Government / DSIT (2026). UK to become world leader in drug discovery: OpenBind initiative. GOV.UK, June 2026. https://www.gov.uk/government/news/uk-to-become-world-leader-in-drug-discovery
  25. Richardson, P. et al. (2020). Baricitinib as potential treatment for 2019-nCoV acute respiratory disease. The Lancet, 395, e30–e31. [BenevolentAI repurposing validation]
  26. Schneider, G. (2018). Automating drug discovery. Nature Reviews Drug Discovery, 17(2), 97–113.
  27. Noé, F. et al. (2020). Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning. Science, 365(6457), eaaw1147.
  28. Lim, J. et al. (2018). Molecular generative model based on conditional variational autoencoder for de novo molecular design. J. Cheminformatics, 10, 31.
  29. Jensen, J. H. (2019). A graph-based genetic algorithm and generative model/Monte Carlo tree search for the exploration of chemical space. Chemical Science, 10(12), 3567–3572.
  30. Hansch, C. & Fujita, T. (1964). ρ-σ-π Analysis: A Method for the Correlation of Biological Activity and Chemical Structure. J. Am. Chem. Soc., 86(8), 1616–1626. [QSAR foundational paper]
  31. Mayr, A. et al. (2016). DeepTox: Toxicity Prediction Using Deep Learning. Frontiers in Environmental Science, 3, 80. [Multi-task deep learning for toxicity]
  32. Ragoza, M. et al. (2017). Protein-ligand scoring with convolutional neural networks. J. Chem. Inf. Model., 57(4), 942–957. [CNN docking score surrogate]
  33. McNutt, A. T. et al. (2021). GNINA 1.0: molecular docking with deep learning. J. Cheminformatics, 13, 43. [Deep learning docking]
  34. Mao, J. et al. (2025). AI for drug discovery in 2025: landscape and global outlook. ScienceDirect, Drug Discovery Today. https://www.sciencedirect.com/science/article/abs/pii/S0031699725075118
  35. NCBi / PMC (2025). From Lab to Clinic: How AI is reshaping drug discovery timelines and industry outcomes. PMC12298131. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12298131/

Provenance