The history of drug discovery is littered with confident target calls that failed in the clinic, and artificial intelligence (AI) target identification has not eliminated that risk, only changed how targets get nominated. Machine learning drug targets now emerge from multi-omics mining, knowledge graphs, and genetic-support scoring long before a program reaches the bench. This guide covers how those methods work and where biological validation still has the final say.
Key takeaways
- Drug mechanisms with human genetic support are roughly 2.6 times more likely to reach approval than those without it, a core rationale for genomics-driven target identification.
- A machine learning model trained on curated genome-wide association study loci prioritizes genes that are more than eight times as likely to be known approved drug targets as chance would predict.
- Knowledge graphs that integrate genetic, pathway, and literature evidence have produced experimentally validated biological hypotheses in peer-reviewed pipelines.
- A drug candidate against an AI-nominated target reached a randomized Phase 2a trial in 2025, reporting a measurable clinical benefit for idiopathic pulmonary fibrosis (IPF).
- Independent benchmarking has found that some sophisticated gene-prioritization algorithms perform only modestly better than simple heuristics at predicting which genes make good drug targets, a genuinely contested point in the field.
What makes a good drug target
A good drug target is not simply a gene associated with disease; it has to be druggable, safe to modulate, and differentiated enough to justify years of development investment. Consensus frameworks from major pharmaceutical companies have formalized this as a multi-dimensional assessment rather than a single yes-or-no genetic signal.
The stakes behind that assessment are high enough to shape an entire company's pipeline. A target call that turns out to be wrong does not fail cheaply: it typically fails only after years of medicinal chemistry, toxicology, and clinical investment have already been spent chasing it, which is precisely why so much AI investment in drug discovery has concentrated on getting this one decision right earlier and more cheaply.
One widely cited analysis of AstraZeneca's pipeline attrition proposed five dimensions for target assessment: the right target, right tissue, right safety, right patient, and right commercial potential. That framework remains a standard reference for why so many biologically plausible targets still fail to become approved drugs.
Genetic evidence has emerged as one of the strongest predictors of eventual success. Drug mechanisms with human genetic support are roughly 2.6 times more likely to reach approval than mechanisms without it, a documented link between a gene variant and disease risk that makes targets nominated this way substantially more likely to survive drug development than those built purely on cell-based or animal-model biology.
Genetic support is necessary but not sufficient on its own, however. A target can carry strong genetic evidence and still fail for reasons the genetics alone cannot predict, including toxicity from hitting the target in tissues where it also performs an essential function, which is exactly why the five-dimension framework above treats genetic evidence as one input among several rather than a standalone approval signal.
How AI mines multi-omics data for targets
Multi-omics mining combines genomic, transcriptomic, and proteomic data layers to flag genes that look like plausible targets across multiple independent lines of evidence, rather than relying on any single dataset. Large population studies have made this approach dramatically more powerful over the past decade simply by expanding the amount of linked genetic and molecular data available to search.
A more detailed treatment of how these data layers combine statistically, and where the approach still runs into trouble, appears in a companion piece on multi-omics integration for drug target discovery. The short version relevant here is that no single omics layer is authoritative on its own, and AI-driven mining tools earn their keep by flagging where several layers agree rather than by trusting any one signal in isolation.
Large-scale population exome sequencing efforts illustrate the scale involved, generating millions of exonic variants that can be cross-referenced against disease phenotypes to flag candidate genes worth pursuing. Tissue-specific gene expression atlases add a second layer, helping researchers confirm that a candidate target is actually expressed in the tissue where a disease originates.
Public target-prioritization platforms have made this kind of multi-omics evidence usable outside of dedicated bioinformatics teams. Open-access resources now integrate genetics, functional genomics, and literature evidence into a single searchable interface, letting a bench scientist check the multi-omics case for a target candidate in minutes rather than weeks.
Knowledge graphs and network medicine
Network medicine treats disease not as a single broken gene but as a disruption to a connected module within the broader biological network, a framing that has reshaped how researchers think about disease mechanism. Knowledge graphs operationalize that idea by connecting genes, proteins, pathways, and published literature into a single queryable structure that can surface non-obvious target hypotheses.
A recent peer-reviewed pipeline demonstrated this directly: an experimentally validated knowledge-graph approach generated and then tested biological hypotheses in drug discovery, closing the loop between computational hypothesis generation and wet-lab confirmation rather than stopping at the prediction stage. That validation step is what distinguishes a rigorous knowledge-graph pipeline from one that simply produces plausible-sounding but unconfirmed leads.
One of the best-known real-world examples of network-based drug repurposing involved researchers using a knowledge graph to flag an existing anti-inflammatory compound as a candidate for a respiratory disease based on its known effect on a cellular entry pathway, a hypothesis that was later followed up in clinical use. Examples like this illustrate the technology's real strength: finding connections across a knowledge base too large for any single researcher to hold in their head at once.
That strength is also its main limitation. A knowledge graph can only surface connections that already exist somewhere in its underlying data and literature, which means it is very good at repurposing well-characterized compounds against newly understood mechanisms and considerably less good at nominating a genuinely novel biological pathway that no prior publication has described. Discovery teams get the most value from knowledge graphs when they treat a surfaced connection as a lead worth investigating, not as evidence the connection is real.
AI-driven target prioritization
Once multiple candidate targets clear an initial multi-omics screen, prioritization models rank them by how likely each one is to succeed in development. These models typically combine genetic association strength, expression specificity, and structural tractability into a single score that a discovery team can use to allocate limited experimental resources.
This is the step where AI-driven target identification most directly changes day-to-day resource allocation, since a discovery budget can only fund a fraction of the candidates a multi-omics screen or knowledge graph might surface. Getting the ranking right, or at least well-calibrated enough to avoid wasting a validation slot on a weak candidate, is arguably where these methods earn their keep more than at the initial discovery stage itself.
One widely used method trains a machine learning model on a curated set of gold-standard genetic loci to prioritize the most likely causal gene at each disease-associated locus. That locus-to-gene model prioritizes genes that are more than eight times as likely to be known drug targets as would be expected by chance, a concrete and reproducible accuracy signal for genetics-driven prioritization.
That said, prioritization accuracy varies considerably by disease area and by which genetic evidence type is used, and no single headline accuracy figure generalizes across every prioritization tool on the market. Some independent benchmarking analyses have found that certain sophisticated scoring methods outperform simple nearest-gene heuristics only modestly, a genuinely contested question that discovery teams should weigh before treating any prioritization score as definitive.
That contested finding is a useful check on how much weight any single prioritization score should carry. A model that edges out a simple heuristic by a small margin on average still leaves plenty of room for individual targets to be mis-ranked, which is exactly why the evidence-type table below is meant to be read as inputs to a judgment call rather than a formula that outputs a final answer on its own.
Prioritization models rarely rely on a single evidence type, and the table below summarizes how the main categories complement one another.
Evidence type | What it captures | Typical role in prioritization |
Genetic association | Whether a gene variant is statistically linked to disease risk in a real population | Primary signal for causal-gene scoring models |
Expression specificity | Whether a gene is active in the tissue where disease originates | Filters out genetically linked genes that are unlikely to be relevant in the affected tissue |
Structural tractability | Whether a protein has a binding site suitable for a small molecule or biologic | Determines whether a prioritized gene is realistically druggable |
Literature and pathway context | Whether existing biology supports a plausible disease mechanism | Adds qualitative confidence alongside quantitative scores |
Validation: Where biology still leads
No amount of computational prioritization removes the need for functional confirmation in a real biological system. CRISPR-based genetic screens, including knockout, interference, and activation variants, remain the standard method for testing whether disrupting a candidate target actually changes a disease-relevant cellular phenotype, and single-cell versions of these screens now let researchers test many hypotheses in parallel. That ordering matters: the screen's job is to confirm mechanism after a target has already been computationally nominated, testing a specific, falsifiable hypothesis about a single candidate rather than searching broadly the way a multi-omics or knowledge-graph platform does earlier in the pipeline.
Consensus guidance known as the GOT-IT recommendations lays out the validation criteria a target should meet before it advances, covering disease linkage, safety, differentiation, and assay feasibility. These recommendations exist precisely because computational scores, however sophisticated, cannot substitute for direct evidence that modulating a target changes disease biology in the intended way.
Structural predictions illustrate the same principle from a different angle. Even highly accurate computational structure models still benefit from experimental confirmation before a target is considered fully assessed, particularly for target classes where subtle conformational details determine whether a compound will actually bind as predicted.
The practical rule discovery teams tend to converge on is that computational evidence sets the order in which targets get tested, while wet-lab and clinical evidence decides which ones actually advance. Skipping that ordering, treating a strong prioritization score as if it were equivalent to a confirmed functional result, is the most common way computational target work leads a program astray.
Notable AI-identified targets that have reached the clinic
A handful of AI-nominated targets have now produced real clinical data, giving the field its first concrete signal of how these methods perform once biology meets the clinic. The most complete published example involves a small-molecule inhibitor developed against a target identified through generative AI and machine learning analysis of disease-relevant tissue data.
That candidate, aimed at a kinase called TNIK for IPF, reached a randomized Phase 2a trial in 2025, with the highest dose tested producing a significant improvement in lung function over 12 weeks and a safety profile consistent with earlier-phase data. It stands as one of the clearest published demonstrations that an AI-nominated target and an AI-designed molecule can move together through early clinical testing.
Other AI-nominated targets have reached first-in-human testing more recently, though public, peer-reviewed clinical data for those programs remains limited at this stage. That asymmetry, one well-documented clinical success alongside several earlier-stage programs still awaiting mature trial data, is a fair summary of where AI-driven target identification currently stands in the clinic.
That scarcity of mature examples is itself informative. AI-driven target identification as a widely adopted practice is still young enough that most nominated targets have not yet had time to clear a full clinical development cycle, which means any summary of "how well this works" is necessarily provisional and will keep shifting as more programs mature.
Not every AI-nominated target reaches that bar, and the field's public record includes disappointing results alongside the successes: at least one other AI-nominated target has advanced to early-phase human testing only for its sponsor to later report that efficacy did not support continued development, a reminder that reaching human testing is not the same as confirming clinical benefit, regardless of how the target was originally identified.
Signals that consistently distinguish AI-nominated targets likely to survive validation from those that do not include:
- Independent confirmation from at least two distinct data types, such as genetics plus expression data, rather than a single computational signal.
- A plausible, mechanistically specific hypothesis for how modulating the target changes disease biology, not just a statistical association.
- Confirmatory functional genomics evidence, typically from a CRISPR-based screen, in a disease-relevant cell model.
Why AI target identification still needs biological confirmation
AI target identification has genuinely expanded the search space available to drug discovery teams, surfacing candidates from genomics, knowledge graphs, and prioritization models that traditional literature review alone would likely miss. The technology's real contribution has been breadth and speed at the hypothesis-generation stage, not a replacement for confirmatory biology, and conflating the two is where overconfident claims about AI-driven target discovery tend to originate.
Machine learning drug targets earn their place in a pipeline the same way any other target does: through genetic support, expression evidence, and functional validation that together survive the scrutiny a costly development program demands. The most productive discovery teams are treating AI-generated target lists as a well-informed starting point rather than a finished answer, not as a substitute for the validation work described throughout this guide.
The clinical record so far, one clear success alongside at least one well-documented disappointment, supports that framing better than an unqualified endorsement would. AI target identification has changed how many candidate hypotheses a team can generate and how quickly it can rank them, and that is a genuine, measurable advance; it has not yet changed how many of those hypotheses turn out to be right once biology gets the final vote, and readers evaluating any specific platform's claims should ask which of those two things it is actually improving.
This article was produced under Drug Discovery News' AI Editorial Guidelines.
