Role of Generative AI in Drug Discovery

The Role of Generative AI in Drug Discovery

The role of generative AI in drug discovery is to propose scientifically relevant candidates for investigation. It extends the range of structures or sequences that researchers can consider, while relying on predictive tools and experiments to evaluate their properties.

Generation and prediction serve different purposes. A generative model may propose a molecular structure. A predictive model may estimate that structure’s activity or solubility. Molecular docking assesses possible binding arrangements within a target structure. None of these outputs, considered alone, demonstrates that a compound will treat a disease.

For small molecules, generative methods can explore alternatives to a known chemical series or suggest structures under specified constraints. For proteins, they can propose backbones or sequences for a desired function. The appropriate representation and evaluation method depend on the research question.

Early work on continuous molecular representations demonstrated how a learned representation could support generation and property-guided exploration. It illustrates a design capability, rather than evidence of therapeutic benefit.

A useful distinction is between a proposed structure, a confirmed hit, a lead and a development candidate. A hit shows activity in a defined experimental system. A lead undergoes further optimisation and characterisation. Candidate nomination is a programme decision based on a broader evidence package; it is not a designation supplied by the model.

Applications Across the Discovery Workflow

Generative AI is most useful within an iterative scientific workflow. Researchers define the objective, generate proposals, assess practical constraints, test selected candidates and use the results to guide further work. The relevant output is a better-supported decision about the next experiment or candidate.

Target selection can draw on biological data, literature and predictive analyses. Generative methods may contribute hypotheses, but identifying a plausible target and validating its disease relevance remain distinct scientific tasks. Similarly, a predicted interaction should not be described as confirmed binding.

Swipe horizontally to view the full table.

This table is a practical synthesis, not a fixed regulatory sequence. Workstreams overlap, and an unexpected finding may require changes to an earlier design assumption. Research on reinforcement learning for molecular design demonstrates how generation can be guided by selected objectives; the objectives themselves still require scientific justification.

Generative Models Used in Drug Discovery

Model families differ in how they represent molecules and generate proposals. Selection should reflect the available data, the output required and the evaluation method. A model’s technical sophistication does not, by itself, establish the quality of the candidates it produces.

• Variational autoencoders: learn a continuous representation from which molecular proposals can be reconstructed or sampled.
• Autoregressive and transformer models: generate sequences progressively. Chemical and protein applications require domain-specific representations and assessment.
• Diffusion models: learn to recover structured outputs from noise and can be adapted to molecular or protein design tasks.
• Generative adversarial networks: train a generator alongside a discriminator that evaluates its outputs against training examples.

Reinforcement learning is often used to steer generation towards specified objectives; it should not be treated as interchangeable with a particular model architecture. Graph representations, sequence representations and three-dimensional structures also describe different aspects of a modelling system.

In RFdiffusion, researchers adapted a structure-prediction network for generative protein design. This is distinct from using a structure predictor only to estimate how an existing sequence may fold.

Published Evidence and Its Interpretation

Evidence for generative AI in discovery ranges from computational benchmarks to laboratory experiments and clinical studies. These evidence levels answer different questions. A successful design benchmark cannot establish patient benefit, and a clinical result for one molecule cannot establish the performance of all generative methods.

Protein design: RFdiffusion

The RFdiffusion study reported generative protein-design tasks and experimental characterisation of selected designs, including protein binders. Its significance lies in connecting computational proposals with measured molecular behaviour. The reported results apply to the designs and conditions evaluated; therapeutic suitability requires further investigation.

Clinical investigation: rentosertib

A randomised phase 2a study published in 2025 evaluated rentosertib, a generative AI-designed small molecule, in 71 adults with idiopathic pulmonary fibrosis over 12 weeks. The primary endpoint concerned treatment-emergent adverse events. Lung-function measurements were among the secondary endpoints.

The authors reported findings that supported further investigation. The study’s size and duration limit conclusions about long-term safety and efficacy. It provides an example of an AI-origin candidate undergoing clinical evaluation, rather than proof that AI removes development risk or establishes regulatory approval.

The study design also matters when interpreting productivity claims. This trial compared treatment groups; it did not compare AI-assisted discovery with a conventional discovery programme. It therefore cannot establish a general improvement in discovery cost, speed or clinical success rate.

Benefits and Practical Limitations

Generative methods can expand design options and help researchers consider several objectives together. Potential value includes exploring alternative structures, identifying property trade-offs and prioritising experiments. Benefits should be assessed against the programme’s actual results and resources, rather than the number of structures generated.

For example, improving predicted target activity may be unhelpful if the same changes reduce solubility or make synthesis impractical. Absorption, distribution, metabolism, excretion and toxicity (ADMET) estimates can support prioritisation, but their reliability depends on the endpoint, training data and candidate chemistry.

Data quality and applicability

Training data may contain inconsistent assay conditions, missing negative results or uneven representation of chemical series. A model can perform well on familiar examples and less reliably on new chemistry. Evaluation should therefore examine relevant independent data and the limits of the intended application.

Scoring functions and experimental relevance

A generator can improve its numerical score without improving the property that matters experimentally. Published work on the limits of molecular generation with reinforcement learning illustrates why optimisation objectives and evaluation design require scrutiny.

GuacaMol benchmarks provide structured ways to assess molecular-generation performance. Such benchmarks support comparison, but do not replace testing of synthesis feasibility, biological activity or candidate performance in the intended setting.

Synthesis, novelty and reproducibility

A proposed structure must be practically accessible before its properties can be tested. Synthetic-accessibility estimates do not establish a reproducible process, acceptable impurity profile or reliable supply. Structural novelty within a dataset also does not establish patentability or freedom to operate; those require separate assessment.

Traceability should connect the model version, input data, candidate identity, selection rationale and experimental results. Without that record, teams may struggle to reproduce a result or understand why a design decision changed.

These limitations explain why the role of generative AI in drug discovery should be assessed through experimentally supported progress. Faster computation is useful when it improves the quality or efficiency of the scientific decisions that follow.

Development Requirements After Candidate Selection

An AI-origin molecule must undergo evaluation of exposure, quality, stability and reproducible supply, as required for other development candidates. Formulation, analytical and process evidence must develop alongside the programme’s safety and clinical evidence.

Preformulation and formulation development

For a small-molecule candidate, early assessment may examine solubility, solid-state properties, degradation pathways and excipient compatibility. These findings inform formulation development and process choices. The appropriate work depends on the molecule, intended route, dose and stage of development.

The International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use (ICH) develops harmonised technical guidelines. Its Q8(R2) pharmaceutical-development guideline provides principles for understanding product and process performance. Application remains subject to regional implementation and product context.

Analytical methods and stability

Analytical development establishes how relevant quality attributes will be measured. Depending on the product, this may include identity, assay, related substances and dissolution. Methods and their performance evidence should be suitable for the decision they support.

ICH Q14 addresses science- and risk-based analytical procedure development. Stability studies separately examine how quality changes over time under relevant conditions. The ICH Q1A(R2) guideline describes the role of these data in supporting storage conditions and shelf-life or retest-period decisions.

Integrated CMC, clinical and manufacturing planning

Chemistry, manufacturing and controls (CMC) work connects material attributes, formulation, process, analytical methods and specifications. Nonclinical and clinical planning must account for the material being evaluated. A change in formulation or process may require additional assessment of its effect on quality or performance.

These activities may progress in parallel. Early analytical findings can change formulation choices; stability results can alter packaging requirements; scale-up can reveal process sensitivities. The broader pharmaceutical drug development process depends on maintaining continuity across these decisions.

Responsibilities should be explicit when work involves external partners. Technology transfer needs documented process and method knowledge. Manufacturing or clinical activities performed by qualified partners should be distinguished from work undertaken directly by the development organisation.

AI Evidence and Regulatory Expectations

Regulatory relevance depends on how an AI output is used. Exploring candidate ideas internally is different from relying on an AI-derived result to support a regulatory decision about safety, effectiveness or quality. The intended use determines the evidence and oversight that need consideration.

The United States Food and Drug Administration (FDA) has published draft guidance on AI supporting drug and biological-product regulatory decisions. It proposes assessing model credibility within a defined context of use through a risk-based framework. The cited document is draft guidance, not a final universal requirement for exploratory discovery models.

The joint FDA–European Medicines Agency (EMA) principles for good AI practice address human oversight, data governance, context of use, performance assessment and lifecycle management. These principles support responsible practice; they do not create a single international approval route.

For a specific programme, teams should document which outputs influence decisions, what limitations apply and who reviews them. Current authority expectations should be checked for the product, development stage and market. The role of regulatory affairs in drug development includes connecting those expectations with the planned evidence package.

Considerations for Programmes in India

For India-based pharmaceutical teams, AI-assisted discovery should connect to an evidence and development plan appropriate to the intended markets. Model origin does not establish that the candidate, its manufacturing process or its clinical programme meets applicable requirements.

The Central Drugs Standard Control Organization (CDSCO) and the applicable New Drugs and Clinical Trials Rules, 2019, and amendments are relevant to new-drug and clinical-trial planning in India. Required permissions and documentation depend on the product and proposed activity. Teams should verify the current pathway rather than infer an exemption or accelerated route from the use of AI.

For programmes involving multiple organisations, project agreements and quality arrangements should define data access, confidentiality, deliverables and review responsibilities. Candidate records should remain connected to source data and experimental reports as work moves into formulation, analytical and regulatory development.

Where other markets are intended, the India plan should be coordinated with the relevant authorities’ requirements. Scientific principles may be shared across jurisdictions, while procedures and submission expectations differ.

Assessment Priorities for Development Teams

A development assessment should separate what is proposed, what has been measured and what remains uncertain. The following priorities help organise that review without assuming that every programme requires the same tests or acceptance criteria.

Frequently Asked Questions