Drug discovery: virtual screening & molecule generation

Finding a drug candidate means searching a chemical space of far more than 10^60 plausible small molecules. AI narrows it: structure-based and ligand-based models score billions of virtual compounds, generative models propose new ones tuned to potency, selectivity and ADMET, and active-learning loops choose what the lab makes next. The benefit is mainly fewer synthesis-and-test cycles and broader exploration in early discovery, not a guaranteed shortcut through clinical trials.

Problem & value

• Today: high-throughput screening (HTS) tests hundreds of thousands to millions of physical compounds at high cost, with low hit rates. Hit-to-lead and lead optimisation then run through many design-make-test-analyse (DMTA) cycles, each taking weeks.
• AI value: ultra-large virtual screening (docking plus ML rescoring) enriches hits before any wet-lab work; generative design explores novel scaffolds and multi-parameter optimisation; ADMET models flag liabilities early; active learning reduces the number of compounds made.
• Evidence: AI-derived molecules have entered clinical trials (e.g. Insilico Medicine's TNIK inhibitor rentosertib in phase 2a, published in Nature Medicine 2025) [4], but no broad proof yet that overall approval rates or timelines have changed; treat vendor claims on speed as case-specific [5].

Variants

• Structure-based virtual screening: docking and co-folding of ligands to a known or predicted target structure (see SCI-001).
• Ligand-based screening and QSAR: similarity, pharmacophore and graph neural network models when no structure exists.
• De novo / generative design: diffusion, autoregressive and reinforcement-learning generators, plus fragment-based and scaffold-hopping modes.
• ADMET and property prediction: solubility, permeability, hERG, metabolic stability; links to SCI-015 for safety assessment.
• Academic / small biotech: open tools plus cloud GPUs; big pharma: federated or proprietary platforms. No consumer variant.

How it works

target definition (SCI-017) + structure (SCI-001)
  -> library preparation (enumerated, make-on-demand, in-house; filters for PAINS/drug-likeness)
  -> virtual screening (docking -> ML rescoring/co-folding -> free-energy refinement of top hits)
  -> generative design (conditioned on pocket or on actives; multi-objective scoring)
  -> property/ADMET prediction + synthesisability (SCI-004 retrosynthesis)
  -> medicinal chemist review (human-in-the-loop selection)
  -> synthesis + assays (SCI-007 automation) -> data back to models (active learning)

• Scoring: classical docking (fast, imprecise) is combined with learned scoring functions and, for the final few hundred, physics-based free-energy perturbation (FEP).
• Co-folding models (AlphaFold 3-class, Boltz-2) predict ligand-bound structures and, in Boltz-2's case, an affinity estimate approaching FEP accuracy at far lower cost on benchmark sets [2].
• Generative models propose molecules as SMILES strings, graphs or 3D point clouds; a scoring oracle (docking, ML, assay-trained models) steers them.
• Human-in-the-loop: chemists veto unstable, toxic, unsynthesisable or IP-conflicting structures; wet-lab data is the ground truth.

Data preparation

Inputs are compound libraries (SMILES/SDF), historical assay and ADMET results, target structures and literature. They typically sit in ELN/LIMS systems, chemistry databases, Excel files from CROs and PDF reports.

Typical gaps: inconsistent structure representations (salts, tautomers, stereochemistry), duplicate compounds under different IDs, assay values with mixed units and thresholds ("<10 uM"), activity cliffs caused by differing assay protocols, and CRO data delivered only as PDF. Models trained on such data inherit the noise, so data quality limits accuracy more than the model choice.

Preparation steps: 1. Inventory sources and owners (PREP-001, PREP-004, PREP-005). 2. Standardise structures with a fixed RDKit pipeline and deduplicate. 3. Normalise units, censored values and assay metadata into tidy tables (PREP-011, PREP-012). 4. Extract tables from CRO reports and papers (PREP-006, DOC-014). 5. Define a scaffold-split test set with the chemists (PREP-018). 6. Classify compound data as confidential and restrict access (PREP-024).

Effort share: high - typically half or more of the project. See GUIDE-003 for a readiness check.

Implementation paths

Which path for whom

• Enterprise: Path B - elastic cloud GPU capacity in an EU region and managed docking/ADMET services fit screening campaigns that are bursty; move to C when the volume or IP sensitivity of the compound library justifies an own cluster.
• Developer: Path C - open tools (RDKit, docking engines, open generative models) give full control and no per-call costs; cloud APIs only for peaks.

Path A - Buy (platform / CRO)

• Software platforms: Schrodinger (physics + ML, FEP+), Cadence OpenEye/Orion, Chemical Computing Group MOE, BIOVIA (Dassault, EU), Optibrium StarDrop (UK), Cresset (UK), deepmirror (UK), Iktos (France, EU, generative design plus robotics), Exscientia heritage now inside Recursion, BenevolentAI-type knowledge platforms (UK).
• AI-first discovery partners: Isomorphic Labs (partnerships with large pharma), Insilico Medicine, Recursion, XtalPi; EU services include Evotec (Germany) and Nuvisan.
• Choose A when: you are a biotech without a computational team, or want a collaborative programme. Check IP ownership of generated molecules, data-use rights and exit terms.

Path B - Build on cloud APIs (recommended for enterprises)

• Managed model services: NVIDIA BioNeMo (NIM microservices for DiffDock, MolMIM, GenMol, Boltz-2), AWS/Azure/Google Cloud HPC and GPU instances in EU regions. AlphaFold 3 weights are restricted to academic use [1].
• Orchestration: Nextflow/Snakemake or Airflow pipelines, ELN/LIMS integration (Benchling, Dotmatics, Genedata), chemistry cheminformatics via RDKit.
• LLM role: literature mining, patent landscape, assay-report extraction and agentic tool orchestration (see SCI-011); not the core molecular model.

Path C - Self-host (open source)

• Docking / screening: AutoDock Vina, GNINA (CNN rescoring), DiffDock, Uni-Dock (GPU), DOCK, OpenMM for MD, OpenFE for open free-energy calculations.
• Co-folding: Boltz-1/Boltz-2 (MIT licence), Chai-1, Protenix, OpenFold3.
• Generative: REINVENT 4 (AstraZeneca, open source), DiffSBDD, Pocket2Mol-class models, TargetDiff; property models: Chemprop, ADMET-AI, MolFormer-type foundation models.
• Chemistry core: RDKit, Datamol, DeepChem, Open Babel.
• Hardware: HW2 handles a single-pocket project; HW3/HW4 for ultra-large library docking and Boltz-2 at scale. Operations effort is substantial; needs computational chemists.

Path D - Custom training

Worth it when you hold proprietary assay data (typically tens of thousands of measured compounds for a target family or ADMET endpoint). Fine-tune property models or generator oracles on in-house data; federated approaches such as the EU/IMI MELLODDY project showed multi-pharma learning without sharing structures [6]. Poor data quality and assay noise, not model size, usually limit gains.

State of the art (2026)

• Co-folding plus affinity heads (Boltz-2, Chai/AlphaFold 3 descendants) have replaced pure docking as the first-pass structure predictor, but generalisation to novel pockets and chemotypes is weaker than headline benchmarks suggest; training-set leakage of similar complexes is a known critique.
• Ultra-large screening of billions of make-on-demand molecules (Enamine REAL and similar) is routine; ML-guided docking cuts compute drastically.
• Generative design is judged by prospective wet-lab hit rates, not benchmark scores; many published models propose unsynthesisable or unstable compounds without synthesis-aware constraints.
• Limits: activity cliffs, scoring-function errors, toxicity and in-vivo PK remain poorly predicted; the main bottleneck is clinical efficacy in humans, which in-silico chemistry does not fix.

Shared resources & combinations

• Shared: scientific-fm and GPU infrastructure with SCI-001 (structure prediction) and SCI-003 (materials property models); RDKit/graph-ML stack with SCI-003, SCI-004, SCI-015.
• Upstream: SCI-017 target and biomarker discovery, SCI-009 genomics.
• Downstream/parallel: SCI-004 retrosynthesis, SCI-007 self-driving labs for DMTA, SCI-015 in-silico toxicology, HLTH-009 trial design.

Effort, cost & KPIs

• Path A: weeks to contract; platform licences or milestone-based collaborations, costs vary widely.
• Path B: 3-6 months to a working pipeline; cost drivers are GPU hours, ELN integration and expert staff.
• Path C/D: 6-12+ months, a dedicated computational chemistry and ML team; cost dominated by people and data generation.
• KPIs: prospective hit rate vs. HTS baseline, enrichment factor, compounds synthesised per lead series, DMTA cycle time, potency/ADMET gains per cycle, synthesisability rate, time from target to candidate nomination.

People & change

Involve early: management and the R&D department lead set the target profile and decision gates. Data owners (medicinal chemistry, assay groups, CRO managers) agree on data release. IT provides compute and integrates ELN/LIMS; IT security and legal/compliance cover IP, CRO contracts and export rules. Medicinal chemists as end users review every shortlist.

Works council: depends. The use case does not monitor staff, but shared platforms with user-level logs can fall under §87(1) no. 6 BetrVG, which the Federal Labour Court applies when the system is merely suited to monitor behaviour or performance. Inform the works council in good time (§90(1) no. 3 BetrVG); public bodies involve the Personalrat. See GUIDE-004.

Data protection: no personal data in the compound data; check only accounts and logs with the DPO. Clinical or patient-derived data would change this and require a DPIA (GUIDE-005).

AI literacy: support chemists with training on applicability domains, uncertainty and the difference between a predicted and a measured value (GUIDE-006).

Communication and pilot: run a retrospective pilot on a finished project: would the model have found the known hits? Share results openly with the project team (GUIDE-009, GUIDE-008).

Typical concerns: loss of chemical intuition and trust in black-box scores. Address them by keeping chemists as decision makers and tracking prediction versus assay outcomes.

Risks, regulation & compliance

• EU AI Act: minimal risk for early-stage research; no Annex III use. AI literacy (Art. 4) applies. Not a medical device itself; downstream clinical AI is covered elsewhere. Research exemption (Art. 2(6)) applies to scientific R&D systems.
• Medicines regulation: EMA's reflection paper on AI across the medicinal product lifecycle expects documented, risk-based use; AI that only supports discovery is generally low regulatory impact, but data used in submissions must be traceable and GxP-compliant [3].
• GDPR: mostly not applicable to chemical data; applies if models use patient-derived data (omics, clinical) or researcher data; use pseudonymisation and DPAs.
• IP and licensing: inventorship for AI-generated molecules, model licences (AlphaFold 3 weights non-commercial), and training on third-party assay or patent data.
• Biosecurity / dual use: generative toxicity-optimisation models can in principle design harmful agents; restrict access, log use and screen outputs against controlled-agent lists (CWC, Australia Group) at policy level.
• Technical risks: benchmark overfitting, hallucinated binding poses, assay noise; mitigate with prospective validation, orthogonal assays and chemist review.

Getting started

1. Pick one target with a structure and a few known actives; define success as prospective hit rate vs. a retrospective HTS baseline.
2. Reproduce a baseline with open tools (RDKit + GNINA/Vina + Boltz-2) on a benchmark such as DUD-E/LIT-PCBA, noting leakage caveats.
3. Add ADMET filters and a synthesisability check; run a small prospective batch (tens of compounds) with a CRO.
4. Set up data capture (ELN/LIMS) so every assay result feeds active learning.
5. Decide buy vs. build after the pilot, and review IP and model licence terms before scaling.

Sources

1. Abramson et al., Accurate structure prediction of biomolecular interactions with AlphaFold 3, Nature 2024 - https://www.nature.com/articles/s41586-024-07487-w
2. Boltz-2 (MIT / Recursion), binding affinity co-folding model - https://github.com/jwohlwend/boltz
3. EMA - Reflection paper on the use of AI in the medicinal product lifecycle - https://www.ema.europa.eu/en/news/use-artificial-intelligence-ai-medicinal-product-lifecycle-scientific-guideline
4. Xu et al., Rentosertib phase 2a trial, Nature Medicine 2025 - https://www.nature.com/articles/s41591-025-03743-2
5. Jayatunga et al., How successful are AI-discovered drugs in clinical trials?, Drug Discovery Today 2024 - https://www.sciencedirect.com/science/article/pii/S135964462400134X
6. MELLODDY federated drug discovery (IMI) - https://www.melloddy.eu/
7. REINVENT 4, generative molecular design - https://github.com/MolecularAI/REINVENT4
8. Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj