AI Drug Discovery: what is actually working now?
In our AI in drug discovery market deck, you will find everything you need to understand the market
SUMMARY
AI drug discovery is already working today in structure prediction, molecular design, lead optimization and protein engineering, while its ability to produce medicines that succeed more often in patients remains unproven.
The biggest change is that AI-originated programs are no longer stopping at promising molecules. Rentosertib and GB-0895 have reached Phase 3, putting AI-heavy discovery workflows in front of the same clinical tests as conventional drugs.
The strongest evidence currently sits at the molecule level. AI-native companies have repeatedly produced compounds that survive preclinical development and Phase 1, and one analysis found unusually high Phase 1 success rates of roughly 80% to 90%.
That advantage becomes much less obvious in Phase 2. Once trials start testing whether a biological intervention actually changes a human disease, early AI-discovered programs have performed much closer to normal pharmaceutical benchmarks.
Rentosertib is especially important because AI influenced both sides of the project: Insilico used computational biology to prioritize TNIK and generative chemistry to help design the molecule. Its encouraging Phase 2a result was small, but reaching Phase 3 makes the program difficult to dismiss as a demonstration project.
GB-0895 tests something slightly different. TSLP is already a clinically validated asthma target, so Generate:Biomedicines is mainly testing whether generative protein design can produce a competitive antibody with properties such as very infrequent dosing.
AI target discovery is progressing more slowly than molecular design. The new Recursion-Genentech neuroscience program is encouraging because the target survived experimental validation in human neuronal models before Genentech committed further resources.
AlphaFold-class models have already changed structural discovery by giving researchers useful protein and protein-ligand predictions much earlier. Their value is real, although a good structure still leaves most of the difficult pharmacology and disease biology unanswered.
AI also appears capable of compressing early discovery timelines sharply. Rentosertib moved from target discovery to a nominated candidate in about 18 months, while other programs have reported candidate nomination in less than a year; large clinical trials remain far harder to accelerate.
Big pharmaceutical companies are starting to provide another form of validation. Expanded partnerships, milestone payments, exercised options and programs moving into active discovery show that several AI platforms have survived internal scrutiny beyond the original partnership announcement.
The pattern across the field is becoming clearer: AI performs best where experiments produce measurable feedback that models can learn from. Predicting whether a complex human disease will respond to a particular intervention remains the much tougher problem.
The next decisive evidence will come from late-stage clinical outcomes. If programs such as rentosertib and GB-0895 reach approval, AI drug discovery will have crossed from improving parts of pharmaceutical R&D into demonstrating that those improvements can survive the whole development process.
This market map, featured in our AI in drug discovery market deck, highlights top companies and startups in the AI in drug discovery market
Why does AI drug discovery look more serious now?
AI drug discovery looks much more serious today because a few AI-originated or AI-engineered drugs have finally reached the stages where patients, rather than benchmarks, decide whether the technology is useful.
The biggest change is clinical. Insilico Medicine’s rentosertib has entered Phase 3 for idiopathic pulmonary fibrosis after producing an encouraging signal in a randomized Phase 2a trial. Generate:Biomedicines is also running Phase 3 studies of GB-0895, an AI-engineered long-acting antibody for severe asthma, across roughly 1,600 patients.
Behind those two late-stage programs, a larger group has reached earlier human testing. Schrödinger has SGR-1505 and SGR-3515 in Phase 1. Recursion has already completed a Phase 2 trial of REC-994 and is advancing several other clinical programs. Insilico says its platform has nominated more than 30 preclinical candidates since 2021.
There is fresh evidence on target discovery too. Genentech recently advanced the first neuroscience target found through its collaboration with Recursion into a joint small-molecule discovery program. According to Recursion, that decision came after pathway, functional and disease validation in human neuronal models, rather than from a computational ranking alone.
A few years ago, much of the case for AI drug discovery rested on models generating structures, ranking molecules and producing plausible candidates. Today we can ask the harder question: are these systems helping create medicines that survive real pharmaceutical development?
| What AI has shown so far | How convincing is the evidence? |
|---|---|
| Predicting molecular structures | Strong |
| Finding and optimizing drug-like molecules | Strong |
| Reaching human trials quickly | Strong |
| Finding genuinely useful new targets | Promising |
| Improving Phase 2 efficacy success | Still unclear |
| Producing an approved AI-discovered drug | Still unproven |
If you want more recent data on this point, please see our latest AI in drug discovery market report.
What does it actually mean for AI drug discovery to “work”?
AI drug discovery is already working if we mean finding and optimizing promising molecules faster; it still has much more to prove if “working” means producing better medicines more often.
A model can predict a protein structure extremely well and still tell us little about whether changing that protein will help someone with Alzheimer’s disease. A generative chemistry system can produce a potent molecule quickly while the biological hypothesis behind the drug turns out to be wrong.
The useful tests are fairly simple. Can AI help teams find targets and development candidates faster? Can those candidates survive preclinical development and Phase 1? And, hardest of all, do they actually help patients and survive Phase 2 and Phase 3 more often?
The evidence is already fairly strong on the first two. The last one remains wide open.
A recent Nature Reviews Drug Discovery review reached much the same conclusion: AI capabilities have improved quickly, while evidence that those capabilities improve clinically relevant decisions remains limited.
As this chart shows, and as featured in our AI in drug discovery market deck, search interest in AI drug discovery has grown rapidly
Has an AI-discovered drug actually worked in patients?
Rentosertib currently gives us the best clinical case that AI can help originate a genuinely new drug and get it far enough to show an effect in patients.
Insilico Medicine used its biology platform to prioritize TNIK as a fibrosis target, then used its generative chemistry system to help design the compound originally known as INS018_055. The project moved from target discovery to a nominated preclinical candidate in roughly 18 months and reached human testing in less than 30 months.
In the randomized Phase 2a study published in Nature Medicine, 71 people with idiopathic pulmonary fibrosis received placebo or one of three doses of rentosertib for 12 weeks. The study mainly tested safety, so the efficacy analysis was small and exploratory.
Still, the 60 mg once-daily group showed an average 98.4 mL increase in forced vital capacity after 12 weeks. That caught attention because IPF progressively damages lung function, and existing treatments mainly aim to slow that decline.
The study was small. Each treatment group contained only around 17 or 18 patients, 12 weeks is short for IPF, and testing several doses increases the chance that one group looks unusually good.
Yet Insilico has now started a Phase 3 trial of rentosertib, which moves the program well beyond a computational success story.
Rentosertib also shows why it is difficult to say exactly how much of a drug was “designed by AI.” Insilico’s models helped identify TNIK and explore chemical space, while medicinal chemists, biologists and pharmacologists kept testing and refining the results.
A recent Nature Reviews Drug Discovery critique makes the same attribution problem explicit. Drug projects contain many decisions, from target choice and assay design to medicinal chemistry, toxicology and clinical trial design. Success at the end cannot be cleanly assigned to one computational method.
The useful proof point is simpler: an AI-heavy workflow helped produce a new target-and-molecule combination that has now reached Phase 3. That already counts for a lot, even if AI did not somehow invent the drug by itself.
Is AI actually better at designing drug molecules?
AI-assisted molecular design is currently the most convincing part of AI drug discovery.
The early clinical numbers are unusually good.
A Drug Discovery Today analysis of clinical programs from AI-native biotechnology companies found an 80% to 90% Phase 1 success rate for AI-discovered molecules. Phase 1 success across the pharmaceutical industry has historically been much lower, generally somewhere around 40% to 65% depending on the dataset and definition used.
The AI sample was still small, and the comparison needs care. AI companies may choose unusually favorable compounds for clinical development, and different companies use “AI-discovered” in different ways.
Even with those caveats, the result fits what the technology is best at doing.
Phase 1 exposes many molecule-level problems: poor pharmacokinetics, unacceptable toxicity, weak exposure or difficulty reaching an effective dose. Those are exactly the areas where large chemical datasets, physics-based calculations and machine-learning models can help teams reject weak candidates earlier.
Schrödinger gives us a separate real-world example. Its drug discovery system mixes machine learning with detailed molecular simulation. SGR-1505, its MALT1 inhibitor, is currently in Phase 1 and has shown preliminary activity in several B-cell cancers alongside a tolerable safety profile. SGR-3515, a Wee1/Myt1 inhibitor, is also in Phase 1. Schrödinger reported a 65% disease-control rate among evaluable patients treated at doses of at least 100 mg in its early clinical data.
Those drugs are nowhere near proven medicines yet. They do add to a pattern we now see across multiple companies: computationally intensive discovery platforms can repeatedly produce molecules with the basic qualities needed to survive the jump from laboratory testing into humans.
| Clinical evidence | What we have seen |
|---|---|
| AI-discovered Phase 1 success in the Drug Discovery Today analysis | 80%–90% |
| Historical Phase 1 success | Roughly 40%–65% |
| AI-discovered Phase 2 success in the same analysis | Roughly 40% |
| Historical Phase 2 performance | Roughly similar |
If you want more recent data on this point, please see our latest AI in drug discovery market report.
This chart, featured in our AI in drug discovery market deck, shows annual venture capital investment in AI drug discovery startups
Why does AI's advantage seem to shrink in Phase 2?
AI drug discovery currently looks much less exceptional once trials start asking whether a drug actually changes a human disease.
The same Drug Discovery Today analysis that found an 80% to 90% Phase 1 success rate estimated roughly 40% success in Phase 2. The Phase 2 sample was even smaller, but the result sits much closer to normal pharmaceutical performance.
That gap points to the real bottleneck. A medicinal chemistry system can improve solubility, potency, selectivity and other measurable properties. Phase 2 asks whether researchers chose the right biological intervention in the first place.
Human diseases make that much harder. Patients with the same diagnosis can have different biology, animal models can mislead, and biomarkers can divide patients badly.
Recursion’s REC-994 trial in cerebral cavernous malformation shows how ambiguous this stage can become. The randomized Phase 2 study met its primary safety endpoint. At the higher 400 mg dose, 50% of patients showed a reduction in total lesion volume compared with 28% on placebo, and several MRI and functional measures moved in an encouraging direction.
The trial was exploratory and was not designed to provide definitive statistical proof of efficacy.
So far, AI looks considerably better at building credible molecules than at proving that the biological idea behind those molecules will help patients.
Can AI really find new drug targets?
AI target discovery is starting to produce serious drug programs, although we currently have far less clinical proof here than we do for molecular design.
TNIK remains the most advanced example. Insilico’s system prioritized the target from disease and aging biology, researchers then validated it experimentally, and rentosertib has now reached Phase 3.
A fresh example from Recursion and Genentech gives us a useful second case.
Their collaboration built large disease-specific biological maps from perturbation experiments in human cells. Recursion says more than one trillion induced-pluripotent-stem-cell-derived neuronal cells have been produced for the partnership since 2022.
Genentech has now advanced the collaboration’s first previously unexplored neuroscience target into a joint small-molecule discovery program.
According to Recursion, the target had to survive three stages of experimental testing: pathway validation, functional testing in human neurons and disease-phenotype validation. Only then did Genentech exercise its option and move the target into drug discovery.
That is much stronger evidence than a model simply ranking a target highly. A major pharmaceutical company committed resources after substantial experimental validation.
The ceiling is still clear. A recent Nature Reviews Drug Discovery review on AI target identification points out that a target is only fully validated once targeting it produces a successful medicine.
AI can already generate biological hypotheses worth serious investment. Whether those hypotheses succeed more often than human-generated ones is still unknown.
If you want more recent data on this point, please see our latest AI in drug discovery market report.
This chart, featured in our AI in drug discovery market deck, shows how Shrödinger is positioned in AI drug discovery
Is AlphaFold actually useful for discovering drugs?
AlphaFold-class models are already useful in real drug discovery because they can give researchers structural information that used to be slow, expensive or sometimes impossible to obtain.
AlphaFold 3 pushed that usefulness beyond isolated protein structures. The model can predict complexes involving proteins, DNA, RNA, small molecules, ions and modified residues.
In the Nature paper introducing AlphaFold 3, the system substantially outperformed traditional docking tools on the PoseBusters protein-ligand benchmark. It could also predict many ligand binding poses without receiving an experimentally measured binding pocket beforehand.
For a medicinal chemistry team, that can change how quickly a project starts. Researchers who lack a crystal structure may still get a plausible picture of where a molecule binds and how the surrounding protein is arranged. That gives them a better starting point for deciding what to make next.
The limits are concrete too.
A plausible binding pose does not tell us whether a molecule will dissolve properly, survive metabolism, reach the right tissue, avoid related proteins or improve a disease. Flexible proteins can adopt several biologically relevant conformations, and high-confidence structural predictions can still be wrong.
AlphaFold has already changed the structural side of discovery. Extending that success all the way to clinical outcomes would be a much bigger claim.
For drug developers, the practical win is straightforward: better structural information arrives earlier, which can make the next experiment much smarter.
Are AI-designed antibodies working too?
AI-designed biologics now provide one of the strongest tests of generative drug design, with Generate:Biomedicines already taking GB-0895 into Phase 3.
GB-0895 targets TSLP, a cytokine involved in severe asthma. The biology itself is already validated: AstraZeneca and Amgen’s tezepelumab blocks the same target and is approved.
Generate is testing a different question. Can its computational protein-design platform build an antibody with unusually good properties against a target that already works?
GB-0895 was engineered for very strong target binding and a long half-life. Generate is now studying twice-yearly dosing, which would be much less frequent than many current injectable biologic treatments.
The company has moved the molecule into two global Phase 3 studies, SOLAIRIA-1 and SOLAIRIA-2, covering roughly 1,600 adults and adolescents with severe asthma. Both trials are designed around clinically significant asthma exacerbations over 52 weeks.
That design makes GB-0895 a cleaner test of AI molecular engineering than rentosertib. With rentosertib, AI influenced both the biological target and the molecule, so failure could come from either side. With GB-0895, we already know TSLP is druggable and clinically useful. The main question is whether Generate created a competitive therapeutic molecule around that biology.
Generate has already shown that its platform can create functioning human antibodies in other programs. Its earlier GB-0669 COVID antibody reached a first-in-human randomized study and behaved pharmacologically as intended before the company deprioritized further development.
Together, those programs make protein engineering one of the more interesting areas to watch. Proteins offer enormous design spaces, their properties can be tested experimentally at scale, and each new experiment gives the model more information for the next round.
This chart, featured in our AI in drug discovery market deck, shows annual funding in AI drug discovery startups
Is AI really making drug discovery faster and cheaper?
AI can already cut years and a lot of experimental work from early drug discovery, while the expensive clinical part still moves much closer to its old speed.
Rentosertib gives us the cleanest published example. The project went from target identification to preclinical candidate nomination in roughly 18 months and reached human testing in less than 30 months.
Conventional small-molecule discovery commonly takes several years before a development candidate reaches the clinic.
Insilico has also repeated short discovery cycles across other projects. The company says it has nominated more than 30 preclinical candidates since 2021. In a recent partnership with Hisun Pharmaceutical, the companies reported nominating a preclinical candidate only eight months after starting the program.
The cost numbers are harder to compare cleanly, but they point in the same direction. Insilico has reported spending around $2.6 million on the discovery work that produced the original rentosertib candidate.
AI can reduce that early bill by proposing compounds, predicting their properties and helping teams decide which few molecules deserve synthesis.
The limit appears later. Patients still have to be recruited, toxicology still has to be assessed, and large Phase 3 studies still take years.
Today we have good evidence that AI can make candidate discovery faster and cheaper. We still do not know whether it lowers the cost of producing an approved medicine.
| Part of development | What AI can change today |
|---|---|
| Target and molecule search | Potentially much faster |
| Lead optimization | Fewer synthesis-and-test cycles |
| Candidate nomination | In some cases, cut to well under two years |
| Phase 1–3 trial duration | Much harder to compress |
| Cost per approved drug | Still unknown |
If you want more recent data on this point, please see our latest AI in drug discovery market report.
Are big pharmaceutical companies actually trusting AI drug discovery?
Big Pharma is now giving AI discovery companies repeat business, milestone payments and additional programs, which tells us more than the original wave of experimental partnerships.
Novartis gives us one of the clearest examples. Its original Isomorphic Labs collaboration covered three difficult small-molecule targets. After roughly a year of working together and seeing results privately, Novartis expanded the partnership by up to three additional programs.
Isomorphic has since added a multi-target, cross-modality collaboration with Johnson & Johnson while continuing its work with Eli Lilly.
Recursion’s relationship with Roche and Genentech has also moved well beyond an announcement. The original deal came with $150 million upfront and allows Roche and Genentech to initiate as many as 40 programs. Recursion says each successful program could trigger more than $300 million in development, commercial and sales milestones.
Several concrete payments have already occurred. Roche and Genentech accepted Recursion’s first large neuronal map in 2024 and a microglia map in 2025, each triggering a $30 million milestone. Genentech has now selected the first neuroscience target from the collaboration for small-molecule discovery.
Sanofi provides another useful example. Its inherited collaboration with Recursion covers as many as 15 AI-designed small-molecule programs. Recursion reported receiving $134 million in upfront and progress payments from that relationship by early 2026.
The huge headline values still need perspective because most milestone money only arrives if drugs keep progressing.
The more useful evidence is repeat behavior. When a pharmaceutical company expands a collaboration, pays additional milestones or moves an AI-derived target into active drug discovery, the platform has passed internal tests that outsiders usually cannot see.
This chart, featured in our AI in drug discovery market deck, compares the main business model options for AI drug discovery biotech companies
Are automated labs more important than the AI models themselves?
The strongest AI drug discovery platforms increasingly look like closed experimental systems where models decide what to test, laboratories generate new data, and the next model decision gets better.
Recursion shows the scale this can reach. Its neuroscience collaboration with Roche and Genentech has produced more than one trillion human neuronal cells, according to the company, allowing whole-genome perturbation experiments that would be impossible to interpret manually.
Generate uses a similar feedback loop for proteins. Computational models create designs, automated experimental systems make and test them, and those measurements feed into later rounds of protein engineering.
Schrödinger attacks the same problem from another angle by combining machine learning with physics-based simulation, giving its models information about molecular behavior rather than relying entirely on patterns learned from existing datasets.
The reason these loops are valuable is simple: many important drug-discovery answers do not exist until someone runs the experiment.
A model may have read every published paper on a target and still have no way to know what a completely new molecule will do in a particular human cell. The experiment creates new information.
That makes proprietary experimental data unusually valuable. Algorithms spread quickly, model architectures get copied and computing gets cheaper. A company that continually generates biological data competitors do not have can build a much harder advantage.
So what part of AI drug discovery is actually working now?
AI drug discovery is clearly working today in structure prediction, molecular search, lead optimization and increasingly protein engineering; the evidence becomes much weaker when AI has to predict complex human biology.
There is now enough real-world evidence to draw that line quite firmly.
Structure models such as AlphaFold can give researchers useful starting points for targets that previously lacked good structural information. Generative and predictive chemistry systems can search huge molecular spaces and help teams balance potency, selectivity, solubility, metabolism and other properties before synthesizing compounds. Several AI-native pipelines have repeatedly produced molecules capable of entering human trials.
Protein engineering is moving rapidly in the same direction. GB-0895 reaching Phase 3 is currently the clearest example.
Target discovery sits one level lower in confidence. Rentosertib provides a serious clinical case, and the new Recursion-Genentech neuroscience target gives us another experimentally validated example, but the number of mature programs remains small.
Clinical efficacy is where the evidence thins out fast.
As seen above, the early analysis of AI-discovered drugs found unusually high Phase 1 success while Phase 2 performance landed much closer to the industry norm.
A very recent Nature Reviews Drug Discovery assessment reached a similarly skeptical conclusion. Its authors argued that AI’s technical capabilities have advanced much faster than the evidence that AI improves actual clinical decision-making.
That is basically where the field stands now: AI is already strong at problems with measurable feedback and much less proven once the problem becomes predicting what a complex disease will do in a human being.
If you want more recent data on this point, please see our latest AI in drug discovery market report.
This chart, featured in our AI in drug discovery market deck, shows revenue breakdown by customer segment in the AI in drug discovery market
Is AI drug discovery actually working now?
Yes. AI drug discovery is already working in several important parts of the process, although claims that the technology has solved drug discovery still run far ahead of the evidence.
AI-heavy platforms repeatedly generate viable drug candidates. Some projects have moved from target or program start to clinical testing in roughly two to three years. Early AI-discovered molecules have performed unusually well in Phase 1. AlphaFold-class models are useful enough to change real structural work. An AI-engineered antibody is now being tested in two Phase 3 asthma trials. Rentosertib has carried an AI-identified target and AI-assisted molecule all the way into Phase 3 after an encouraging randomized Phase 2a result.
The missing proof sits further downstream. We still have no large, mature dataset showing that AI-discovered medicines survive Phase 2 and Phase 3 at much higher rates or reach approval more often than conventionally discovered drugs. The early clinical data warn against assuming that outcome: the large apparent advantage in Phase 1 shrinks once efficacy becomes the main question.
So the answer is fairly clear today. AI is already making drug discovery faster and helping teams build better candidates. Whether those candidates become successful medicines more often is still the open question.
Rentosertib and GB-0895 now carry a lot of that burden. If programs like these begin producing approvals, we can start asking whether AI has changed pharmaceutical R&D all the way through development.
For now, AI drug discovery works. AI drug development still has something much harder to prove.
OUR METHODOLOGY
The question behind this analysis is AI Drug Discovery: what is actually working now? We broke that question into the parts of drug discovery where AI can realistically be tested: structure prediction, target discovery, molecular design and optimization, progression into human trials, clinical efficacy, discovery speed and cost, adoption by established pharmaceutical companies, and the experimental infrastructure behind these platforms.
For each dimension, we looked for recent observable evidence rather than treating every AI milestone as equivalent. A benchmark result, a preclinical candidate, a Phase 1 trial, a randomized Phase 2 result and entry into Phase 3 answer different questions, so we gave more weight as the evidence moved closer to real pharmaceutical validation.
We used conventional drug discovery as a comparison where it helped establish a meaningful baseline, particularly for clinical success rates and discovery timelines. We kept molecular viability separate from clinical efficacy because Phase 1 and Phase 2 test very different parts of the drug-development problem.
We treated “AI-discovered” as a description of a workflow rather than an all-or-nothing label. Modern drug programs combine computational models with medicinal chemistry, biology, pharmacology and clinical development, so the relevant question is whether AI materially improved part of the process and whether the resulting hypothesis or molecule survived the experiments that followed.
For commercial adoption, we prioritized what pharmaceutical companies did after the original partnership announcement. Expansions, exercised options, milestone payments and programs moving into active discovery carry more weight than headline deal values because those decisions happen after partners have seen results that are often unavailable publicly.
We also looked at the experimental systems behind the models. Recursion, Generate:Biomedicines and Schrödinger use different combinations of automated experimentation, proprietary biological data, machine learning and physics-based simulation, so the analysis treats the model-and-experiment loop as part of the discovery platform rather than judging the algorithm in isolation.
Finally, we aggregated the evidence across these dimensions instead of relying on one dramatic proof point. That allows the conclusion to stay strong where the evidence is already strong and more cautious where the field is still waiting for mature clinical outcomes.
Key sources used for this analysis include: Nature Medicine on the randomized Phase 2a trial of rentosertib, Nature Biotechnology on rentosertib’s target discovery, generative chemistry and development timeline, Insilico Medicine on rentosertib entering Phase 3, Drug Discovery Today on Phase 1 and Phase 2 success rates for AI-discovered drugs, Nature Reviews Drug Discovery on the clinically relevant impact of AI drug discovery, Nature Reviews Drug Discovery on AI target identification and validation, Nature on AlphaFold 3 and protein-ligand prediction, ClinicalTrials.gov for the SOLAIRIA-1 Phase 3 study of GB-0895, ClinicalTrials.gov for the SOLAIRIA-2 Phase 3 study, Generate:Biomedicines on GB-0895 and its Phase 3 program, Recursion on the REC-994 Phase 2 study, Recursion on the first neuroscience target advanced with Genentech, Recursion on its Roche, Genentech and Sanofi collaborations, Isomorphic Labs on the expansion of its Novartis collaboration, and Isomorphic Labs on its Johnson & Johnson collaboration.
This chart, featured in our AI in drug discovery market deck, shows how AI drug discovery platform technology has evolved over time
Related blog posts
- AI Drug Discovery: what are the biggest unsolved problems?
- AI Drug Discovery: what is getting real adoption now?
- The latest funding news in AI for drug discovery
- Which AI drug discovery startup is growing the fastest?
