How is AI used in drug discovery?

AI is used across the entire drug discovery pipeline rather than in any single place. In target identification, machine learning integrates multi-omics and network data to propose and prioritise disease targets. In molecular design, generative models and protein structure prediction create and optimise candidate compounds. In ADMET and safety, models predict absorption, metabolism, toxicity, and drug interactions to triage candidates. In clinical translation, AI supports patient stratification and trial design. The maturity varies sharply by stage: AI is strongest at generating hypotheses and predicting molecular properties, and least reliable at predicting whether a molecule will ultimately succeed in patients.

What are the limitations of AI in drug discovery?

The central limitation is the translational gap: AI has improved discovery speed and, in early trials, drug safety, but has not yet demonstrably improved clinical efficacy, and no novel AI-discovered drug has reached full approval to date. AI target identification generates strong hypotheses but cannot replace experimental validation, and many clinical failures stem from targets that turn out not to drive human disease. ADMET models are excellent for triage but imperfect as final answers. Predictions are only as good as their training data, so poor or unrepresentative data produces unreliable results. The pool of late-stage AI-discovered drugs is still small, so firm conclusions cannot yet be drawn.

How does machine learning improve target identification?

Machine learning improves target identification by integrating the vast, heterogeneous biological data relevant to disease, genomics, transcriptomics, proteomics, and clinical records, at a scale no individual could manage manually. Multi-omics models surface molecular signatures and candidate targets that single-data-type analysis would miss. AI-powered knowledge graphs link disparate datasets to evaluate large sets of candidate genes rapidly, compressing work that once took weeks into days. Phenotypic approaches can identify active compounds and work backward toward mechanism. The key boundary is that AI accelerates hypothesis generation, but confirming that a target genuinely drives disease still requires experimental validation, which remains the most decisive and expensive step.

What is generative AI in drug design?

Generative AI in drug design refers to models that create novel candidate molecules with desired properties, rather than screening existing libraries. Given a target and a set of property goals, these models propose new chemical structures, and newer three-dimensional approaches design molecules directly into a target's binding pocket rather than proposing them and checking fit afterward. Combined with accurate protein structure prediction, generative design has made structure-based approaches possible for targets whose structures were previously unavailable, and has compressed early medicinal chemistry timelines substantially. The critical checks are whether the proposed molecules are actually synthesisable and developable, which is where medicinal chemistry judgement remains essential.

How does FDA view AI-generated drug candidates?

The FDA's posture is best described as cautious openness. In 2025 it issued draft guidance on using AI to support regulatory decision-making for drugs and biologics, signalling that AI-derived evidence is welcome in submissions while setting expectations for how it must be justified. The central themes are model credibility and validation on a risk-based, context-specific basis; the quality and provenance of training data; transparency and explainability of AI outputs; and attention to bias and generalisability across populations. The direction is enabling rather than obstructive, but the burden of demonstrating that a model is fit for its specific purpose rests with the sponsor. EMA and ICH frameworks are developing along similar lines.

Articles

AI in drug discovery: From target identification to clinical translation

AI is now embedded across the drug discovery pipeline. This guide maps the full arc, where the technology delivers, where it still falls short, and what the science actually shows.

Written byTrevor J Henderson

| 8 min read

Drug discovery scientist reviewing an AI-generated molecular structure and target pathway on screen, illustrating AI across the drug discovery pipeline

AI now touches every stage of the drug discovery pipeline, from target identification to clinical translation. The science is real, and so are the limits

Flow (2026)

Listen with Speechify

0:00

8:00

Highlights in this article

Key takeaways
What AI actually does in drug discovery
Target ID and validation with AI
AI-driven compound design and optimization
ADMET prediction: Where ML has changed the game
The translational gap: AI's persistent challenge
Regulatory considerations for AI-generated drug candidates
What this means for drug discovery teams

AI in drug discovery has moved from a speculative tool to a working part of the pipeline at nearly every major pharmaceutical company. Machine learning now contributes to how targets are identified, how molecules are designed, how their properties are predicted, and how clinical trials are planned, and by early 2024 each of the top 20 pharmaceutical companies had announced activity in the space. But the honest picture is more nuanced than either the enthusiasts or the sceptics suggest: AI has measurably accelerated parts of discovery and improved some early success rates, while the hardest problem, reliably translating a preclinical prediction into a drug that works in patients, remains largely unsolved.

This guide maps that full arc for working scientists, from early target biology through to the clinic. It complements DDN's ongoing coverage of how AI is transforming drug discovery, and stays deliberately science-first: focused on the biology and chemistry rather than the software, and on what the evidence supports rather than what the press releases claim.


Key takeaways AI now spans the full discovery pipeline: target identification, molecular design, ADMET and safety prediction, and clinical translation. It is a set of tools applied across stages, not a single technology. The clearest wins are in speed and early-stage success. AI-discovered candidates have compressed preclinical discovery timelines dramatically and, by some analyses, improved Phase 1 pass rates, though these figures come with important caveats. The translational gap remains the central challenge. Improved Phase 1 safety has not yet translated into demonstrably more effective drugs, and no novel AI-discovered drug has reached full approval to date. Generative models, protein structure prediction, and multi-omics target discovery are the areas of most active scientific progress, each with real capability and real limits. Regulators including the FDA are actively engaging with AI in drug development, signalling openness alongside expectations for validation, transparency, and data quality.

Key takeaways

AI now spans the full discovery pipeline: target identification, molecular design, ADMET and safety prediction, and clinical translation. It is a set of tools applied across stages, not a single technology.
The clearest wins are in speed and early-stage success. AI-discovered candidates have compressed preclinical discovery timelines dramatically and, by some analyses, improved Phase 1 pass rates, though these figures come with important caveats.
The translational gap remains the central challenge. Improved Phase 1 safety has not yet translated into demonstrably more effective drugs, and no novel AI-discovered drug has reached full approval to date.
Generative models, protein structure prediction, and multi-omics target discovery are the areas of most active scientific progress, each with real capability and real limits.
Regulators including the FDA are actively engaging with AI in drug development, signalling openness alongside expectations for validation, transparency, and data quality.

What AI actually does in drug discovery

AI in drug discovery is not one capability but a family of methods applied at different points in a long pipeline, each addressing a distinct scientific problem. Understanding what AI does is easier when organised by the stage of discovery it acts on, because the maturity, the evidence, and the limits differ sharply from one stage to the next.

Discovery Stage	What AI Contributes	Maturity and Evidence
Target identification	Mines multi-omics and network data to propose and prioritise disease targets	Advancing; strong at hypothesis generation, still needs biological validation
Molecular design	Generates and optimises candidate structures, predicts protein structure	Rapidly maturing; generative and structure models in routine use
ADMET and safety	Predicts absorption, metabolism, toxicity, and interactions in silico	Established; useful for triage, not a full replacement for assays
Clinical translation	Supports patient stratification, trial design, and outcome prediction	Emerging; promising but the translational gap persists
Regulatory	Informs submissions; subject to evolving FDA and EMA guidance	Early; frameworks actively developing

The practical implication is that a blanket claim about whether AI works in drug discovery is meaningless. It works well for some tasks, generating and prioritising hypotheses, predicting structures and properties, and much less reliably for others, particularly predicting whether a molecule will succeed clinically. The rest of this guide walks through the pipeline stage by stage.

A blanket claim about whether AI works in drug discovery is meaningless. It works well for generating hypotheses and predicting structures, and far less reliably for predicting whether a molecule will succeed in patients.

Target ID and validation with AI

Target identification, deciding which biological molecule a drug should act on, is where discovery begins, and it is one of the areas where AI has made the most credible early contribution. The appeal is straightforward: the biological data relevant to target discovery, genomics, transcriptomics, proteomics, and clinical records, is vast, heterogeneous, and impossible for any individual to integrate manually, which is exactly the kind of problem machine learning is suited to.

The main AI approaches to finding and confirming targets:

Multi-omics integration. Machine learning models integrate genomic, transcriptomic, and proteomic data to identify molecular signatures and candidate targets associated with disease, surfacing connections that single-data-type analysis would miss.
Knowledge graphs and network biology. AI-powered knowledge graphs link disparate biological and clinical data, enabling researchers to evaluate large sets of candidate genes rapidly; in one reported case, evaluating dozens of immune-related genes as potential Alzheimer's targets was compressed from weeks to days.
Phenotypic approaches. Rather than starting from a known target, AI can analyse phenotypic screening data to identify compounds with desired biological effects and then work backward toward mechanism, complementing the dominant target-first paradigm.
Target validation. Machine learning can strengthen the evidence that a target is genuinely disease-relevant, though validation, confirming a target actually drives disease, remains the step where computational prediction most requires experimental confirmation.

The honest boundary here is between hypothesis generation, where AI genuinely accelerates the work, and validation, where the biology still has to be proven experimentally. A target proposed by a model is a well-informed hypothesis, not a confirmed target, and the distinction matters because a flawed target is the single most expensive failure mode in drug discovery. These approaches are examined in depth in the coverage of AI-powered target identification.

AI-driven compound design and optimization

Once a target is chosen, the problem becomes finding and refining a molecule that acts on it, and this is where AI has advanced most visibly in recent years. Generative models and protein structure prediction have changed what is practical in early medicinal chemistry, compressing timelines that were previously measured in years.

Where AI is changing molecular design:

Protein structure prediction. Accurate prediction of protein three-dimensional structure has made structure-based design possible for targets whose structures were previously unavailable, expanding the range of proteins that can be approached rationally.
Generative molecular design. Generative models propose novel candidate molecules with desired properties, and newer three-dimensional approaches design molecules directly into a target's binding pocket rather than proposing them and checking fit afterward.
Lead optimisation. Machine learning accelerates the iterative refinement of a promising molecule, predicting how structural changes will affect potency, selectivity, and properties, and narrowing the synthesise-and-test cycle that dominates medicinal chemistry.
Chemical space exploration. AI can search regions of chemical space far larger than medicinal chemistry intuition alone would reach, though whether the molecules it proposes are synthesizable and developable remains a critical check.

The clearest illustration of what this enables is a case like ISM001-055, a generative-AI-derived inhibitor for idiopathic pulmonary fibrosis, where the preclinical discovery phase from target identification to lead optimisation took roughly 18 months against a typical four to five years. That kind of timeline compression is real and repeatable. What it does not by itself guarantee is clinical success, which is a separate question addressed below. The design methods are covered in detail in the coverage of generative AI in molecular design.

ADMET prediction: Where ML has changed the game

Predicting how a molecule will behave in the body, its absorption, distribution, metabolism, excretion, and toxicity, is where machine learning has become genuinely embedded in routine practice, because the task fits ML well and the payoff is high. Filtering out compounds destined to fail on metabolism or toxicity before they consume synthesis and assay resources is one of the most practical returns AI offers in discovery.

The ADMET and safety properties AI models predict:

Metabolism. Models predict how metabolic enzymes will act on a molecule and what metabolites will form, informing design changes before a compound is made.
Toxicity. Machine learning flags structural features associated with toxicity risk, providing early warning, though the models are better at flagging known liabilities than at predicting genuinely novel toxicities.
Drug-drug interactions. AI predicts the likelihood of clinically relevant interactions, which matters especially for drugs likely to be co-administered with others.
Permeability and distribution. Models predict properties such as blood-brain barrier permeability, which is decisive for central nervous system drug design and hard to optimise by intuition alone.

The mature, honest position on ADMET prediction is that it is excellent for triage and prioritisation and imperfect as a final answer. A model can reliably tell you which compounds are most likely to have a problem, focusing experimental effort where it matters, but a favourable prediction does not remove the need for confirmatory assays. The specific methods and their reliability are examined in the coverage of AI-powered ADMET prediction.

The translational gap: AI's persistent challenge

The hardest and most important question in AI drug discovery is whether it produces better drugs, not just faster candidates, and here the evidence is genuinely mixed and worth stating carefully. On one hand, an analysis reported in Nature Biotechnology found AI-discovered drugs advancing through Phase 1 trials at an 80 to 90 percent success rate against an industry average nearer 40 to 65 percent, a striking early signal. On the other hand, that advantage has not yet carried through to later trials.

A sober counterpoint in npj Drug Discovery notes that while AI-derived molecules trend toward better Phase 1 safety and tolerability, their Phase 2 proof-of-concept success rates so far look comparable to historical norms of around 40 percent, meaning that AI's measurable gains in speed and safety have not yet produced demonstrably more effective drugs. Reinforcing the caution, a 2025 analysis in Clinical Pharmacology and Therapeutics observes that no novel AI-discovered drug has yet reached full clinical approval, and that the number of molecules reaching later-stage trials remains small.

Why the translational gap persists:

Better Phase 1 does not mean better efficacy. Improved safety and tolerability, likely reflecting better molecular properties, is a real gain, but Phase 1 primarily tests safety. Whether a drug works is decided in Phase 2 and beyond, where AI's advantage has not yet been demonstrated.
Target choice still dominates outcomes. Many clinical failures stem from choosing a target that turns out not to drive disease in humans, and AI target identification does not yet solve this, because the underlying human biology is often simply not yet understood.
Preclinical models remain imperfect. AI predictions are trained and tested against preclinical data, and when those models and assays do not represent human disease well, a confident prediction can still fail to translate.
Small numbers, early days. The pool of AI-discovered drugs in late-stage trials is still small, so firm conclusions about clinical success rates cannot yet be drawn in either direction.

The balanced reading is that AI has demonstrably improved the speed and, in early trials, the safety profile of drug candidates, while the question of whether it improves the ultimate clinical success rate remains open and genuinely unresolved. This is the central scientific challenge of the field, and it is examined further in the coverage of AI in clinical trials and translation.

Regulatory considerations for AI-generated drug candidates

As AI moves deeper into the discovery pipeline, regulators have begun to engage with it directly, and the posture is best described as cautious openness. In 2025, the FDA issued draft guidance on the use of AI to support regulatory decision-making for drugs and biologics, signalling that AI-derived evidence is welcome in submissions while setting expectations for how it should be justified. Regulators including the FDA have signalled openness to AI, modelling, and simulation as tools to modernise how therapies are designed and evaluated.

The regulatory themes that matter for AI-generated candidates:

Model credibility and validation. Regulators expect a risk-based demonstration that a model is fit for the specific context in which its output is used, rather than general assurances about the technology.
Data quality and provenance. Because AI outputs depend entirely on training data, the quality, relevance, and provenance of that data is a central regulatory concern, echoing the field's own garbage-in, garbage-out caution.
Transparency and explainability. Evidence that can be examined and explained is easier to bring into a regulatory submission than an opaque black-box output, which is why explainability is an active area of both research and regulatory attention.
Bias and generalisability. Whether a model trained on one population or data source generalises fairly to others has both scientific and equity dimensions that regulators are increasingly attentive to.

The regulatory direction is enabling rather than obstructive, but it places the burden of demonstrating credibility on the sponsor. These considerations, and the evolving FDA, EMA, and ICH frameworks, are examined in the coverage of regulating AI in drug discovery.


What this means for drug discovery teams The productive way to think about AI in drug discovery is stage by stage, matching expectations to the maturity of the evidence. Lean on AI where it is strongest and best validated: integrating multi-omics data for target hypotheses, generating and optimising molecules, and triaging candidates on ADMET and safety before committing experimental resources. Treat its outputs as well-informed hypotheses that still require biological and clinical confirmation, particularly at the two points where failure is most expensive, target validation and clinical efficacy. Keep the science in the lead: the teams getting the most from AI are those pairing it with strong biology and rigorous experimental validation, not those treating a prediction as an answer. For DDN's ongoing reporting on the field, how AI is transforming drug discovery tracks where the technology is heading.

What this means for drug discovery teams

The productive way to think about AI in drug discovery is stage by stage, matching expectations to the maturity of the evidence. Lean on AI where it is strongest and best validated: integrating multi-omics data for target hypotheses, generating and optimising molecules, and triaging candidates on ADMET and safety before committing experimental resources. Treat its outputs as well-informed hypotheses that still require biological and clinical confirmation, particularly at the two points where failure is most expensive, target validation and clinical efficacy. Keep the science in the lead: the teams getting the most from AI are those pairing it with strong biology and rigorous experimental validation, not those treating a prediction as an answer. For DDN's ongoing reporting on the field, how AI is transforming drug discovery tracks where the technology is heading.

This article was produced in accordance with Drug Discovery News’ Editorial Policies.

Frequently Asked Questions (FAQs)

How is AI used in drug discovery?
AI is used across the entire drug discovery pipeline rather than in any single place. In target identification, machine learning integrates multi-omics and network data to propose and prioritise disease targets. In molecular design, generative models and protein structure prediction create and optimise candidate compounds. In ADMET and safety, models predict absorption, metabolism, toxicity, and drug interactions to triage candidates. In clinical translation, AI supports patient stratification and trial design. The maturity varies sharply by stage: AI is strongest at generating hypotheses and predicting molecular properties, and least reliable at predicting whether a molecule will ultimately succeed in patients.
What are the limitations of AI in drug discovery?
The central limitation is the translational gap: AI has improved discovery speed and, in early trials, drug safety, but has not yet demonstrably improved clinical efficacy, and no novel AI-discovered drug has reached full approval to date. AI target identification generates strong hypotheses but cannot replace experimental validation, and many clinical failures stem from targets that turn out not to drive human disease. ADMET models are excellent for triage but imperfect as final answers. Predictions are only as good as their training data, so poor or unrepresentative data produces unreliable results. The pool of late-stage AI-discovered drugs is still small, so firm conclusions cannot yet be drawn.
How does machine learning improve target identification?
Machine learning improves target identification by integrating the vast, heterogeneous biological data relevant to disease, genomics, transcriptomics, proteomics, and clinical records, at a scale no individual could manage manually. Multi-omics models surface molecular signatures and candidate targets that single-data-type analysis would miss. AI-powered knowledge graphs link disparate datasets to evaluate large sets of candidate genes rapidly, compressing work that once took weeks into days. Phenotypic approaches can identify active compounds and work backward toward mechanism. The key boundary is that AI accelerates hypothesis generation, but confirming that a target genuinely drives disease still requires experimental validation, which remains the most decisive and expensive step.
What is generative AI in drug design?
Generative AI in drug design refers to models that create novel candidate molecules with desired properties, rather than screening existing libraries. Given a target and a set of property goals, these models propose new chemical structures, and newer three-dimensional approaches design molecules directly into a target's binding pocket rather than proposing them and checking fit afterward. Combined with accurate protein structure prediction, generative design has made structure-based approaches possible for targets whose structures were previously unavailable, and has compressed early medicinal chemistry timelines substantially. The critical checks are whether the proposed molecules are actually synthesisable and developable, which is where medicinal chemistry judgement remains essential.
How does FDA view AI-generated drug candidates?
The FDA's posture is best described as cautious openness. In 2025 it issued draft guidance on using AI to support regulatory decision-making for drugs and biologics, signalling that AI-derived evidence is welcome in submissions while setting expectations for how it must be justified. The central themes are model credibility and validation on a risk-based, context-specific basis; the quality and provenance of training data; transparency and explainability of AI outputs; and attention to bias and generalisability across populations. The direction is enabling rather than obstructive, but the burden of demonstrating that a model is fit for its specific purpose rests with the sponsor. EMA and ICH frameworks are developing along similar lines.

Add Drug Discovery News as a preferred Google source to see more of our trusted coverage.

About the Author

Trevor J Henderson
Trevor Henderson is the Creative Services Director for the Laboratory Products Group at LabX Media Group. With over two decades of experience, he specializes in scientific and technical writing, editing, and content creation. His academic background includes training in human biology, physical anthropology, and community health. Since 2013, he has been developing content to engage and inform scientists and laboratorians.
View Full Profile