Roughly 173 drug programs that trace their origin to AI-driven discovery platforms are currently sitting in clinical trials. Not one of them has crossed the finish line to full FDA approval. That single fact sits awkwardly next to years of press releases promising that AI would compress a decade of drug discovery into months. It's worth asking what's actually happening in the space between those two numbers — because the answer says more about how drug development works than it does about whether the AI is any good.
This isn't a story about AI failing at drug discovery. It's a story about what "AI discovered this drug" actually means, how long clinical trials take regardless of how a molecule was found, and where the real evidence of impact is starting to show up — which is not, yet, in an approval letter.
If you work in pharma or biotech R&D, invest in the sector, or build software for it, the useful question about AI drug discovery clinical trials isn't "does AI work?" but "which stage does it actually change, and what evidence would prove it?" This piece covers what counts as AI-discovered, why speed at one stage doesn't shorten the whole pipeline, what the 173-programs, zero-approvals gap means, practical implications for different teams, and which readouts to watch through 2028.
What "AI-discovered" actually means
The phrase gets used loosely, and that looseness is doing a lot of work in the hype cycle. A drug labeled "AI-discovered" can mean several very different things:
- De novo molecular generation — a generative model proposes novel chemical structures with desired properties, and a candidate emerges that a chemist did not hand-draw first, a technique closely related to how AI protein design generates novel biomolecules from scratch.
- Target identification — AI models mine genomic, proteomic, or literature data to flag a biological target (a protein or pathway) as promising, and a conventional discovery process takes it from there.
- Hit-to-lead optimization — AI narrows a massive virtual chemical library down to a shortlist of candidates for wet-lab testing, accelerating a step that used to take human medicinal chemists months.
- Repurposing and combination prediction — AI models predict that an existing approved drug (or combination of drugs) will work against a new disease, sidestepping de novo chemistry entirely.
Companies like Insilico Medicine, Recursion Pharmaceuticals, Exscientia (now part of Recursion), Isomorphic Labs, and BenevolentAI each anchor their public narrative in different parts of this list. Insilico is probably the closest to the popular image — an AI system (its Chemistry42 and PandaOmics platforms) proposed a novel target and a novel molecule for idiopathic pulmonary fibrosis, and that candidate, rentosertib, reached clinical trials. Recursion leans harder on phenotypic screening — using AI to interpret millions of cellular images and infer which molecules do useful things to diseased cells, without necessarily starting from a known target.
The point is that "AI-discovered drug" is a spectrum, not a category. A drug where AI shaved a few months off target selection is being counted in the same public conversation as a drug whose entire molecular structure was generated by a model. That conflation is part of why the headline numbers — hundreds of AI-associated programs — can coexist with zero full approvals without anyone being obviously wrong. The claims being made are just fuzzier than they sound.
How AI drug discovery clinical trials actually work, and why speed at one stage doesn't mean speed overall
Drug development has four broad phases after a candidate is chosen: preclinical testing (in vitro and animal studies), then Phase 1 (safety, small healthy or patient cohorts), Phase 2 (efficacy signal, larger patient group), and Phase 3 (large-scale confirmatory efficacy, often the longest and most expensive phase). Only after clearing all three clinical phases does a sponsor file for approval, and regulators then take months to over a year to review the submission — a process our broader look at AI in clinical trials covers in more detail beyond just the discovery stage.
Where AI has demonstrably compressed timelines is almost entirely in the earliest, most computational stages:
| Stage | Traditional timeline | AI-assisted timeline (claimed) | What AI is actually doing |
|---|---|---|---|
| Target identification | 1–3 years | Weeks to months | Mining omics/literature data for candidate targets |
| Hit discovery / lead optimization | 2–4 years | Months | Screening virtual libraries, generating novel structures |
| Preclinical candidate nomination | 1–2 years | Similar, sometimes faster | Predicting toxicity/ADME properties in silico |
| Phase 1 (safety) | 1–2 years | Unchanged | Governed by regulatory and biological timelines |
| Phase 2 (efficacy signal) | 2–3 years | Unchanged | Governed by patient recruitment, disease biology |
| Phase 3 (confirmatory) | 2–4 years | Unchanged | Governed by statistical power, endpoints, follow-up |
| Regulatory review | 1–1.5 years | Unchanged | FDA/EMA process, not compressible by discovery method |
The compression is real but structurally limited to roughly the first 3-5 years of what is typically a 10-12 year process. Once a molecule enters human trials, it is subject to the same biology, the same regulatory scrutiny, and the same patient recruitment bottlenecks as any other candidate — AI-originated or not. A model can find a promising molecule in months; it cannot make a Phase 3 cardiovascular outcomes trial finish faster, because that trial's duration is set by how long it takes enough patients to have enough cardiac events to reach statistical significance.
This is the mechanical reason the "zero approvals" number is not itself damning. The earliest AI-native platforms (Exscientia, BenevolentAI, Insilico) only started nominating clinical candidates in volume around 2020-2022. Given a typical clinical-to-approval timeline of 8-12 years even for well-behaved candidates, an approval wave — if it's coming — would be expected to land in the back half of this decade at the earliest, not before it.
Why now: 173 programs, zero approvals, and what that gap means
The ~173 figure matters because it establishes scale: this is no longer a handful of speculative bets from a few well-funded startups. It represents a meaningful slice of the industry's early-stage pipeline. Large pharmaceutical companies have folded AI-discovery partnerships into standard R&D practice rather than treating them as experimental side projects — Isomorphic Labs (spun out of Alphabet's DeepMind) has signed structured discovery deals with major pharma partners, and most top-20 pharma companies now run internal AI discovery units alongside external platform partnerships, part of a broader wave of AI agents entering pharmaceutical operations beyond just discovery.
But the zero-approvals figure is the sobering counterweight, and it deserves to be stated plainly rather than explained away. It means that, as of today, no regulator has yet confirmed through the full clinical gauntlet that an AI-native discovery process reliably produces drugs that are safer or more effective than those found through conventional means. Every claim about AI's superiority in drug discovery is currently a claim about speed and cost at the discovery stage — not about the clinical outcomes of the resulting molecules, because the industry doesn't have enough completed Phase 3 data yet to know.
Some of the earliest and most closely watched candidates are also the best evidence of what a nuanced picture looks like:
- Insilico Medicine's rentosertib (INS018_055) for idiopathic pulmonary fibrosis is widely cited as the field's flagship case — a novel target and novel molecule both nominated by AI, in trials since 2021 and reporting encouraging Phase 2a safety and biomarker data.
- Exscientia/Recursion's oncology candidates have had a mixed record, including at least one high-profile program discontinuation, a useful reminder that AI-originated molecules fail in trials for the same reasons any molecule fails: toxicity, lack of efficacy, or an underlying disease hypothesis that doesn't hold up in humans.
- BenevolentAI's early repurposing work (including its widely publicized baricitinib prediction for COVID-19, later borne out) showed AI's target/indication-matching strength, even though that specific case involved an already-approved drug rather than a novel molecule going through full clinical development.
The honest synthesis: AI is unambiguously changing how candidates get selected and how fast they reach the clinic. It has not yet been shown to change the probability that a candidate survives the clinic — because we don't yet have a large enough completed cohort to measure that probability with any confidence.
Benefits of AI in Drug Discovery
The zero-approvals figure is a reason for caution about clinical claims, not a reason to dismiss the gains AI has already delivered upstream. Those gains are real, and they matter to different parts of the industry in different ways: to R&D teams choosing candidates, to finance teams funding programs, and to patients waiting for treatments in diseases with few options.
Faster movement through early discovery
Target identification and hit-to-lead work have historically taken years. AI-assisted platforms report compressing parts of that work into months by mining omics and literature data for targets and screening virtual libraries instead of physical ones. For R&D teams, that means candidates reach preclinical testing sooner, and weak ideas can be dropped earlier, before they consume lab budget and the attention of medicinal chemists.
Cheaper exploration of chemical space
Testing molecules in the lab is expensive. Narrowing a vast virtual library to a short list before synthesis lets teams explore far more structures for the same budget. Generative models can also propose structures a chemist might not have drawn by hand, which widens the range of options considered at the start of a program.
Earlier warning on toxicity and drug-like properties
Predicting absorption, metabolism, and toxicity liabilities in silico helps teams avoid advancing molecules likely to fail in preclinical safety studies. Each candidate dropped at the computational stage rather than in animal testing saves time and money, even though these predictions still need laboratory confirmation.
New uses for existing drugs
Repurposing models can suggest that an approved drug may work against a new disease, as BenevolentAI's baricitinib prediction did. Because the drug's safety profile is already known, this route can reach patients faster than developing a new molecule, and it makes AI's strength at matching targets and indications directly useful.
Better-informed portfolio decisions
AI platforms give R&D leaders more structured evidence when choosing which targets and candidates to fund. That doesn't guarantee clinical success, but it can make early portfolio decisions more systematic and easier to revisit as new data arrives, and it gives leadership a clearer record of why each program was started.
AI Drug Discovery Use Cases
The definitions above become clearer when you look at how specific companies and programs have applied them. These examples come from the programs discussed in this piece and show the range of what "AI-driven" can mean in practice, from a fully AI-nominated molecule to AI quietly supporting conventional programs.
Novel target and molecule for fibrosis
Insilico Medicine used its platforms to propose both a target and a new molecule for idiopathic pulmonary fibrosis, producing rentosertib. It is the closest real example to the popular image of AI-discovered drugs, and its Phase 2a safety and biomarker data make it the case most observers watch. The outcome that matters, efficacy in larger trials, is still ahead.
Phenotypic screening at scale
Recursion applies AI to millions of cellular images to infer which molecules change diseased cells in useful ways, without necessarily starting from a known target. The approach suits diseases where the biology is poorly mapped, though candidates still face the same clinical attrition as any other molecule, as some discontinued programs have shown.
Repurposing approved drugs
BenevolentAI's prediction that baricitinib could help COVID-19 patients, later supported by trial results, shows AI matching an existing drug to a new indication. This use case skips de novo chemistry, so the main question is whether the predicted effect holds up in patients at the doses already known to be safe.
Partnered structure-based discovery
Isomorphic Labs signs discovery deals with large pharma partners, applying structure prediction and design models to partner-selected targets. For the partner, the value lies in target selection and design quality rather than a de-risked clinical asset, which is why these deals are priced as early-stage bets.
In-house discovery units at large pharma
Most top pharma companies now run internal AI groups that support target selection, virtual screening, and property prediction across existing programs. Here AI is less a headline than a routine part of the discovery toolkit, and its impact shows up as faster, cheaper early-stage work rather than as individual "AI-discovered" drugs. Much of the field's real effect may be hidden in programs nobody labels as AI-native.
Common AI Drug Discovery Mistakes
Organisations adopting or investing in AI discovery tend to make the same errors, most of them rooted in treating early-stage speed as proof of clinical value. Each is understandable given the marketing around the field, and each is avoidable with a few direct questions.
Projecting faster approvals from faster discovery
Shaving years off target and lead selection doesn't shorten Phase 2 or Phase 3, which are governed by biology, recruitment, and statistics. Business plans that assume AI-discovered candidates will reach market much sooner than industry averages are built on a hypothesis the data doesn't yet support, and they set expectations that later look like failure.
Assuming AI candidates will fail less often
There is no published, statistically robust evidence that AI-originated molecules survive the clinic at better rates than conventional ones. Pricing deals or setting portfolio risk as if attrition will be lower exposes investors and R&D budgets to the same roughly 90% failure rate as everyone else, without the contingency to absorb it.
Accepting "AI-designed" without asking which stage
Vendors use the label for anything from a small assist in virtual screening to fully generated molecules. Buyers who don't ask exactly which step the AI touched, and what prospective evidence supports it, can't compare platforms or judge what they are actually paying for.
Trusting retrospective validation alone
Showing that a model rediscovers known drugs or liabilities is necessary but weak evidence. Without prospective results, where predictions are made before outcomes are known, teams can't tell whether a model generalises to genuinely new problems.
Underinvesting in data foundations
Models trained on biased or poorly curated bioactivity and omics data tend to propose more of what is already well studied. Organisations that buy models without investing in data quality get faster versions of familiar ideas rather than new ones.
AI Drug Discovery Best Practices
For organizations deciding how much to bet on AI-native discovery, the current data supports a specific and fairly narrow set of conclusions rather than a blanket endorsement or dismissal.
For pharma and biotech R&D leaders:
- Treat AI discovery platforms as a tool for improving the quality and speed of your candidate funnel, not as a guarantee of downstream clinical success. The evidence supports faster, cheaper hit-to-lead work; it does not yet support lower attrition rates in the clinic.
- Budget for the same clinical trial costs and timelines regardless of discovery method. Any internal projection that assumes AI-discovered candidates will reach approval meaningfully faster than industry-average timelines is not supported by data yet — it's a hypothesis.
- Evaluate AI discovery vendors on the rigor of their validation data (retrospective accuracy on known targets, prospective hit rates) rather than on marketing language about "AI-designed" molecules. Ask specifically which stage of the funnel the AI touched.
For investors and business development teams:
- Attrition math still applies. Industry-wide, roughly 90% of drug candidates that enter Phase 1 fail to reach approval; there is no published, statistically robust evidence yet that AI-originated candidates beat that baseline, though several companies claim internal data suggesting improved hit rates at the preclinical stage.
- Licensing and partnership deals structured around AI platforms (the Isomorphic Labs model, for instance) are effectively bets on target-selection quality, not de-risked clinical assets. Price them accordingly.
- The 2026-2028 window is when the first meaningful wave of AI-native Phase 3 readouts is expected. That is the period to watch for real signal, not press releases about new candidates entering trials.
For builders and technical teams working adjacent to this space, the more immediate opportunity often isn't the generative chemistry itself but the infrastructure around it: cleaning and structuring the multi-modal biological data these models train on, building the MLOps and validation pipelines that make model predictions auditable to a regulator, and integrating AI tooling into the messy reality of existing lab and clinical-data systems. That infrastructure work has clear, near-term ROI regardless of how the novel-molecule story eventually resolves.
Real limitations and open questions
A few limitations are structural, not just a matter of the technology maturing further.
Data quality and bias. Generative and predictive models are trained on existing bioactivity, structural, and omics datasets, which carry the biases of decades of prior drug discovery — overrepresenting certain target classes and chemical scaffolds that were historically easy to study, and underrepresenting biology that's harder to assay. A model trained on this data will tend to propose more of what's already been explored, which cuts against the "AI finds things humans would never find" narrative in some cases.
Validation is retrospective by default. Most published accuracy claims for AI discovery models are backtested against known outcomes — can the model "rediscover" an approved drug or a known toxic liability? That's a necessary but not sufficient test. Prospective validation, where the model's prediction precedes and is confirmed by a real, blinded clinical result, is much rarer and takes years to accumulate.
Biology, not chemistry, is usually the bottleneck. Many Phase 2 and 3 failures happen because the underlying disease hypothesis was wrong — the target didn't matter as much as preclinical models suggested, or human biology diverged from animal models in ways nobody could have modeled. AI is good at chemistry and target-matching within a known hypothesis space. It does not yet solve the harder problem of validating disease biology itself — the same challenge that makes personalized medicine hard to get right even after a drug is approved.
Regulatory frameworks are still catching up. The FDA has issued draft guidance on AI use in drug development, but there's no finalized, AI-specific approval pathway. Every AI-originated candidate currently goes through the identical regulatory process as a conventionally discovered one — which is arguably appropriate (safety and efficacy standards shouldn't change based on how a molecule was found) but means AI gets no regulatory shortcut to offset its earlier-stage speed gains.
Selection effects in reporting. Companies publicize their AI-discovery wins prominently and are far quieter about discontinued programs. The true success rate of AI-originated candidates relative to industry baselines is harder to assess than the marketing suggests, simply because the denominator (quiet failures) is undercounted in public discourse.
What to watch next
The next 24-36 months will produce the first real evidence base, not just pipeline announcements. A few concrete markers to track:
| Signal | Why it matters |
|---|---|
| First Phase 3 readouts from AI-native candidates (2026-2028) | Will show whether AI-originated molecules have comparable or better efficacy/safety rates than industry baselines |
| Any full FDA approval of an AI-discovered novel molecule | The symbolic and practical milestone the field is waiting on |
| Published prospective (not retrospective) validation studies | Needed to assess whether AI predictions hold up before, not after, results are known |
| FDA finalized AI/ML guidance for drug development | Will clarify whether discovery method affects submission requirements |
| Attrition-rate comparisons across large AI-originated cohorts | The real test of whether AI improves candidate quality, not just speed |
| Consolidation or new entrants among AI discovery platforms | Signals whether investor confidence tracks clinical progress or outpaces it |
If the 173 current programs mature at anything close to typical industry attrition rates, expect somewhere in the range of 15-25 to eventually reach approval — a number that would be a genuine milestone for the field, even though it would represent a minority of the total pipeline. That's not a failure of the technology; it's how drug development works for every discovery method that's ever existed. The question worth tracking isn't whether AI produces miracle hit rates — it's whether it produces meaningfully better ones than the conventional process it's being compared against, and whether it does so cheaply enough to matter.
Teams evaluating where AI genuinely changes drug discovery outcomes versus where it just changes the marketing copy can get a clearer, hands-on assessment from Woyce Technologies.
FAQ
Has any AI-discovered drug been approved by the FDA?
No AI-originated drug has yet received full FDA approval. Several candidates, including Insilico Medicine's rentosertib, have advanced through Phase 2 trials and are considered the field's furthest-along programs, but full approval — which requires completing Phase 3 and regulatory review — has not yet happened for any of them. That gap is mostly about timing rather than failure: the first AI-native candidates entered trials only around 2020-2021, and clinical development typically takes many years. The first approvals are expected once Phase 3 readouts arrive.
How many AI-discovered drugs are currently in clinical trials?
Estimates put the number at roughly 173 AI-originated drug programs currently in some stage of clinical development, spanning companies like Insilico Medicine, Recursion Pharmaceuticals, Exscientia, and BenevolentAI, among others. Most are in early phases, which is expected given how recently they started. Counts vary between trackers because "AI-originated" has no standard definition: some lists include repurposed drugs or programs where AI assisted only one step, so treat any single number as approximate.
Why hasn't AI sped up the entire drug development timeline?
AI has mainly accelerated the early, computational stages of discovery — target identification and molecule design — which historically take 3-5 years. Clinical trials themselves are governed by biological response times, patient recruitment, and regulatory review, none of which a discovery algorithm can compress. A Phase 3 trial still has to dose enough patients for long enough to show a real effect on outcomes. AI may help indirectly, through better patient matching, site selection, and trial design, but those are separate tools from the discovery models that designed the molecule.
Is AI drug discovery actually better than traditional methods, or just faster?
The current evidence supports faster and cheaper early-stage discovery. Whether AI-originated candidates have better clinical success rates than conventionally discovered ones is not yet established, because too few AI-native candidates have completed the full clinical pipeline to generate statistically meaningful attrition data. Some companies report better preclinical hit rates internally, but those aren't the same as clinical success. The fair summary is that AI is a proven accelerator for early work and an unproven improver of the odds in the clinic.
What's the difference between an AI-assisted drug and an AI-discovered drug?
"AI-assisted" typically means AI sped up one step of a largely conventional process, such as virtual screening or toxicity prediction. "AI-discovered" is often used more loosely, but the strongest form of the claim applies to cases where a generative model proposed a genuinely novel molecular structure and, sometimes, a novel biological target.
Which companies are leading in AI drug discovery?
Insilico Medicine, Recursion Pharmaceuticals, Isomorphic Labs, and BenevolentAI are among the most visible AI-native discovery companies, alongside internal AI units at large pharmaceutical companies that partner with or license technology from these platforms. "Leading" can mean different things: furthest clinical progress, size of pipeline, or quality of partnerships. Insilico's rentosertib is among the furthest-along programs, while companies like Isomorphic Labs are better judged by their partnerships, since their candidates are earlier in development.
When might we see the first approved AI-discovered drug?
Given that the earliest AI-native clinical candidates entered trials around 2020-2021 and typical clinical-to-approval timelines run 8-12 years, the first approvals are plausible in the 2027-2030 window, though this depends entirely on how individual Phase 2 and Phase 3 trials read out. Programs like Insilico's rentosertib, which have cleared Phase 2, are the ones to watch.
Conclusion
The headline numbers, roughly 173 AI-originated programs in clinical trials and no full FDA approvals, describe a pipeline that is young, not one that has failed. AI has clearly changed the front end of drug discovery: target selection, molecule design, and hit-to-lead work are faster and cheaper. Clinical trials, governed by biology, recruitment, and regulatory review, run on the same clock as ever.
The open question is whether AI-originated candidates survive the clinic more often than conventional ones, and there isn't enough completed data to answer it yet. Training data biases, retrospective-only validation, uncertain disease biology, and selective reporting of wins all argue for caution about bold claims in either direction. Typical attrition would still mean a meaningful number of approvals over the coming years, which would be a real milestone.
For R&D leaders and investors, the practical step is to evaluate AI platforms by the specific stage they improve and the prospective evidence behind them, and to budget clinical timelines conventionally. If you're building the data pipelines, validation tooling, or ML infrastructure that pharma R&D depends on, our healthcare AI development team can help.