What does the AI drug discovery startup landscape look like today?

What actually counts as an AI drug discovery startup today?

Today, we would count a company as an AI drug discovery startup when AI is central to finding targets, designing therapeutic molecules or predicting biology, rather than simply another tool used somewhere in R&D.

That definition keeps the landscape useful. Insilico Medicine uses AI across target discovery and small-molecule design. Chai Discovery builds models that generate proteins and antibodies. Xaira Therapeutics is working on biological foundation models and virtual cells. Iambic Therapeutics combines machine learning with medicinal chemistry to design small molecules. Isomorphic Labs is building an AI drug-design engine across small molecules, antibodies, peptides and other modalities.

Recursion and Generate Biomedicines are now public companies, but we still need them in the analysis because they show what happens when an AI-native startup reaches the clinical and public-market stage. Recursion combines automated experiments, large biological datasets and drug development. Generate uses generative models to design therapeutic proteins and already has a program in Phase III.

We would exclude the much larger universe of pharmaceutical companies that happen to use machine learning, CROs adding AI features, clinical-trial software vendors and generic AI tools used by researchers. Otherwise almost every serious drug company eventually ends up inside the category.

The definition of an “AI-discovered drug” also needs care. AI can identify a target, generate the molecule, optimize an existing molecule, predict its structure or help select patients. Those represent very different levels of AI involvement. A drug where AI designed a molecule against a known, validated target gives us evidence about molecular engineering. A drug where AI found a new target and created the molecule gives us a much harder test of whether AI understands disease biology.

Are investors still funding AI drug discovery startups, or only a few winners?

AI drug discovery funding is running very hot, but the money is concentrating around a surprisingly small group of companies.

The Financial Times recently reported that roughly $2.2 billion went into AI drug discovery companies during 2025, up from about $599 million in 2023. Two private rounds announced this year already exceed that entire 2025 figure: Isomorphic Labs raised $2.1 billion and Chai Discovery raised $400 million. Using the FT figure as a rough benchmark, those two financings alone amount to $2.5 billion, about 14% more than the previous year's sector-wide total.

The concentration is extreme. Isomorphic's round was one of the largest private biotech financings ever. Chai's $400 million Series C valued a company founded in 2024 at $3.8 billion. Meanwhile, Dimension closed an $800 million fund focused heavily on computational biology, and Andreessen Horowitz has also put hundreds of millions behind AI-biotech investing.

The other side of the market looks much harsher. BenevolentAI, once one of Europe's flagship AI drug discovery companies, went through repeated restructurings and eventually delisted. Exscientia reached the public market at a multibillion-dollar valuation before merging into Recursion. Generate managed to reopen the IPO route for the category this year, raising $400 million, but that success came with a Phase III program and a much more mature pipeline than most private AI startups can show.

We are past the phase where adding “AI” to a biotech story lifts everyone. Investors are paying extraordinary prices for companies they think could own a foundational model, unique dataset or valuable drug pipeline. The middle of the market is a lot less comfortable.

Company	Recent capital event	What it says about the market
Isomorphic Labs	$2.1B Series B	Frontier AI drug design can attract mega-rounds normally seen only in exceptional biotech companies
Chai Discovery	$400M Series C at $3.8B valuation	A two-year-old molecular-model company can already command a multi-billion-dollar valuation
Generate Biomedicines	$400M IPO	Public investors will fund AI-native biotech when clinical assets are already advanced
Dimension	$800M new fund	Specialist investors are building large pools specifically around computational biology

If you want more recent data on this point, please see our latest AI in drug discovery market report.

Who's actually winning the AI drug discovery race right now?

Right now, Insilico Medicine has the clearest lead in end-to-end clinical validation, while Generate leads generative protein design in the clinic and Isomorphic, Chai and Xaira lead different parts of the frontier-model race.

Insilico's advantage is straightforward: rentosertib has now entered Phase III for idiopathic pulmonary fibrosis. AI contributed to identifying TNIK as the target and designing the molecule, so this is a much stronger test of the AI-discovery thesis than simply optimizing a drug against an already proven mechanism.

Generate is testing something different. Its lead anti-TSLP antibody GB-0895 is in two Phase III asthma studies. Generate's platform designed the protein, but TSLP was already a validated therapeutic target. If the drug succeeds, it would be a major validation of AI-driven protein engineering, although it would tell us less about AI's ability to discover new disease biology.

Iambic sits further back clinically. Its AI-designed HER2 inhibitor IAM1363 remains in Phase I/1b, with an active international study and earlier data showing antitumor activity in heavily pretreated patients. The company says it moved IAM1363 into the clinic in under two years.

The model-first companies are harder to rank because their proof is mostly preclinical. Isomorphic now works with Lilly, Novartis and Johnson & Johnson across multiple drug-design programs. Chai has added Pfizer, Lilly, Novartis and argenx as partners while moving from Chai-1 to antibody-focused Chai-2 and Chai-3 models. Xaira recently released X-Cell, its first virtual-cell model trained on a very large genome-wide perturbation dataset.

Recursion remains unusual because of the scale of the infrastructure underneath its pipeline: automated laboratories, phenotypic experiments, chemistry, clinical development and tens of petabytes of biological data. Its problem is no longer showing that it can build a powerful discovery machine. Investors now need to see enough valuable drugs come out of it.

Company	Where it leads today	Best evidence	What still needs proving
Insilico Medicine	End-to-end AI drug discovery	AI-originated rentosertib in Phase III	Phase III efficacy and eventual approval
Generate Biomedicines	Generative protein therapeutics	GB-0895 in Phase III	Whether designed proteins translate into successful approved medicines
Recursion	Large-scale experimental AI biology	Huge proprietary biological dataset and multiple pharma programs	Consistent clinical output from the platform
Iambic Therapeutics	AI-guided small-molecule design	IAM1363 showing early human antitumor activity	Larger and more mature efficacy data
Isomorphic Labs	Frontier AI drug design	Lilly, Novartis and J&J collaborations	Human clinical validation
Chai Discovery	Generative molecular models	Adoption by several top-20 pharma companies	Whether model performance creates valuable drugs
Xaira Therapeutics	Virtual cells and causal biology	X-Cell and large perturbation datasets	Whether predicted biology improves drug decisions

Has an AI-discovered drug actually reached Phase III now?

Yes. AI drug discovery now has a genuine Phase III test: Insilico Medicine has started a late-stage trial of rentosertib, while Generate is separately running Phase III studies of an AI-designed protein therapeutic.

This is one of the clearest changes in the landscape lately. Insilico initiated its Phase III program after rentosertib produced positive randomized Phase IIa results in idiopathic pulmonary fibrosis. In the 71-patient study published in Nature Medicine, patients receiving the highest dose gained an average 98.4 mL in forced vital capacity over 12 weeks, while the placebo group lost 20.3 mL. The gap was therefore about 119 mL over a relatively short study period.

The result deserves attention because rentosertib combines several layers of AI involvement. Insilico's platform helped identify TNIK as a potential fibrosis target and subsequently generated the molecule that inhibits it. The program survived medicinal chemistry, preclinical testing, Phase I and a randomized Phase II trial before entering Phase III.

Generate's GB-0895 gives us a second late-stage test from another branch of AI drug discovery. Its two SOLAIRIA Phase III trials are evaluating a long-acting anti-TSLP antibody in severe asthma. Earlier clinical work showed a roughly three-month half-life and prolonged suppression of relevant biomarkers, which could eventually allow much less frequent dosing than existing biologics.

The industry's biggest trophy is still unclaimed. No AI-originated medicine has yet received full regulatory approval. Phase III failures are common even after convincing earlier results, and the rentosertib Phase IIa study was small.

Still, the sector has clearly moved beyond the old question of whether AI-generated molecules can reach humans. We now have late-stage trials capable of producing a much harder answer.

Are AI-discovered drugs actually doing better in clinical trials?

AI-discovered drugs currently look unusually good in Phase I and much more ordinary in Phase II. So far, AI appears better at designing molecules than at choosing disease biology.

The most widely cited quantitative analysis was published in Drug Discovery Today by researchers affiliated with BCG. Looking at the first cohort of AI-discovered molecules to reach clinical development, they estimated an 80–90% Phase I success rate. Historical industry success rates are usually much lower.

That fits what AI should already be good at. Machine-learning systems can help medicinal chemists optimize potency, selectivity, pharmacokinetics, solubility and other properties before a molecule enters humans. Better candidates should fail less often because of basic drug-like properties.

The Phase II result is much less spectacular. The same analysis found a success rate of roughly 40%, based on a smaller sample, around the historical range for conventional drug development. Phase II is where researchers begin asking whether changing the selected biological mechanism actually improves the disease.

A newly published Nature Reviews Drug Discovery assessment makes the same problem hard to ignore. After reviewing progress across the field, its authors concluded that clinically relevant evidence remains limited despite years of increasingly impressive AI benchmarks. They argue that AI evaluation needs to focus much more on whether models improve real drug-development decisions.

For us, this is probably the most important dataset in the sector right now. AI seems capable of sending cleaner, more drug-like molecules into clinical development. We still do not have convincing evidence that it consistently picks better therapeutic hypotheses.

If you want more recent data on this point, please see our latest AI in drug discovery market report.

Is AI really making drug discovery much faster now?

AI is already making the discovery part of drug development much faster, with some programs moving from target or design work to a clinical candidate in roughly one to two years.

Insilico has reported candidate-generation timelines of roughly 12–18 months for several AI-driven programs. Iambic says IAM1363 reached the clinic in under two years. Protein-design models can now propose large numbers of new binders or antibody structures before researchers synthesize a small subset for laboratory testing.

The biggest time saving comes from reducing physical iteration. Traditional medicinal chemistry can involve repeated cycles of design, synthesis and testing. A computational system can rank far larger chemical spaces before synthesis, learn from experimental results and narrow the next round.

That changes early discovery considerably. A team that once spent years finding and optimizing a lead can sometimes reach a development candidate in a fraction of that time.

Clinical development moves on a different clock. Toxicology studies still have to be run. Patients still have to be recruited. Chronic diseases require enough treatment time to measure durable effects. Phase III programs may need hundreds or thousands of participants. Regulators still expect real experimental and human evidence.

The more useful speed metric is starting to change. Generating a molecule six months earlier is useful. Killing a bad drug program two years earlier could be worth far more.

Is AI actually making drug discovery cheaper?

AI is lowering the cost of searching for and optimizing drug candidates, but AI drug discovery companies are still expensive businesses because experiments and clinical trials quickly become the bigger bill.

Recursion shows the scale clearly. The company has built automated laboratories, massive biological datasets, its own computational infrastructure and a clinical pipeline. That gives its models unusual amounts of proprietary data, while also creating a cost base that looks much closer to biotech than software.

Generate illustrates the same issue from the protein side. After its IPO, the company reported more than $500 million in cash and marketable securities, yet its runway guidance still stretched only a few years because Phase III trials and multiple clinical programs consume capital quickly.

The $2.1 billion financing raised by Isomorphic also tells us something about the economics. Frontier models may make individual design cycles dramatically more efficient, but building an integrated AI drug company requires elite researchers, large-scale compute, experimental laboratories and eventually human trials.

The economics become especially unforgiving later in development. If AI saves tens of millions during discovery and then a Phase III drug fails after hundreds of millions have been spent, most of the theoretical efficiency disappears.

The larger opportunity is to spend less money on candidates that were going to fail anyway. Better early decisions can reduce the number of expensive failures entering clinical development. We have good evidence that AI speeds candidate creation; evidence that it substantially lowers the cost per approved drug remains thin.

What are pharma companies really paying AI drug discovery startups for?

Pharma companies are currently paying AI drug discovery startups for access to better drug candidates and discovery engines, with most of the headline money released only when actual medicines progress.

Insilico's recent Lilly agreement is a useful example. The deal is worth up to roughly $2.75 billion, but the upfront payment is $115 million. About 96% of the advertised value therefore depends on future development, regulatory and commercial milestones.

That structure shows how sophisticated pharma buyers view the technology. Lilly is comfortable putting more than $100 million on the table today, which is meaningful validation. It still wants most of the economics tied to programs that survive real drug development.

Recursion provides another useful comparison because some of its headline collaboration value has already converted into cash. In its latest annual filing, Recursion said the Sanofi collaboration had generated $134 million in upfront and progress-based payments after five milestones. That is much more informative than a theoretical multi-billion-dollar maximum because programs had to keep moving for the payments to occur.

Isomorphic's early Lilly and Novartis agreements followed a similar pattern, with tens of millions paid upfront and much larger potential milestone pools. Chai's growing list of pharma customers suggests a more platform-oriented form of demand: drugmakers want direct access to molecular models that their internal teams can use on difficult targets.

Across these agreements, pharma is behaving like a buyer of scientific productivity. Software access has value, but the biggest checks still follow compounds, targets and programs that move closer to medicines.

Which AI drug discovery business model is actually working?

The strongest AI drug discovery business model today is a hybrid: keep ownership of some drugs, partner others with pharma, and use the platform to create more programs than the company could finance alone.

Insilico is probably the clearest example. The company develops proprietary drugs internally while signing discovery and licensing agreements with companies such as Lilly, SK Biopharmaceuticals and other pharmaceutical groups. That gives Insilico several ways to make money: upfront payments, research funding, milestones, royalties and the eventual value of drugs it retains.

Recursion follows a comparable logic through large partnerships with Sanofi and Roche/Genentech alongside its own pipeline. Generate owns clinical assets while also working with pharmaceutical partners. These businesses accept the cost of drug development because successful medicines capture vastly more value than a software subscription.

Chai is currently leaning much more toward the model/platform side. Pharma partners use its frontier molecular-design systems to create antibodies and other therapeutic candidates. Latent Labs has also positioned itself primarily around giving partners access to generative protein-design capabilities.

Isomorphic sits somewhere between the two. Its pharmaceutical collaborations validate and fund the technology, while the company is simultaneously advancing an internal pipeline. With enough capital, that architecture gives it a chance to collect both platform economics and drug economics.

Pure software looks harder to defend as strong biological models spread. Owning drug rights, proprietary data or a unique experimental loop gives the company something valuable even when the underlying model architecture improves elsewhere.

Model	Examples	Main advantage	Main weakness
Asset-led hybrid	Insilico, Recursion, Generate	Keeps drug upside while using partnerships to share risk	Expensive clinical development
Frontier model + internal pipeline	Isomorphic	Can monetize partnerships and build proprietary drugs	Requires huge amounts of capital
Partner-first platform	Chai, Latent Labs	Less clinical-development burden	More exposed if molecular models become easier to reproduce

If you want more recent data on this point, please see our latest AI in drug discovery market report.

Why are AI protein-design startups getting so much attention now?

AI protein design is currently one of the hottest parts of drug discovery because models are moving from predicting protein structure to creating new therapeutic proteins from scratch.

Generate gives this trend clinical credibility. GB-0895 is already in Phase III, while the company is also moving other computationally designed programs into human studies. The platform works through a design-build-test-learn loop where models propose proteins, experiments measure what actually happened and those results feed future design.

Chai represents the newer model-first wave. After releasing Chai-2 and then Chai-3, the company signed collaborations with Lilly, Pfizer and Novartis. More recently, argenx also agreed to use Chai's platform for de novo antibody discovery. Several of those deals arrived within months of each other, suggesting that large drugmakers are actively testing generative protein design rather than treating it as a distant research project.

Latent Labs is another company worth watching. It was founded by Simon Kohl, who previously co-led work on AlphaFold2, and raised $50 million to build generative models for protein design. Its strategy includes experimental validation, which is essential because a protein that looks plausible computationally can still fail when researchers try to express, manufacture or use it.

The technical opportunity is unusually large. Protein sequence space is far too vast to search physically, while modern models have learned useful relationships between sequence, structure and function. Even modest improvements in experimental hit rates can remove a huge amount of lab work.

We should still separate protein engineering from disease discovery. Designing a better antibody against a proven target is commercially valuable, but the harder biological question remains untouched. The strongest companies will eventually need to show that their models can create proteins with properties that conventional engineering rarely finds.

Can virtual-cell models fix AI drug discovery's biggest problem?

Virtual-cell models are currently the industry's most ambitious attempt to predict disease biology before testing drugs in patients, and they could attack the Phase II failure problem much more directly than molecule generation does.

Xaira's X-Cell shows where the field is going. The company trained the model using billions of genomic data points from a large perturbation dataset. The objective is to predict how cells respond when genes or biological pathways are changed across different contexts.

That moves AI closer to the question drug developers actually struggle with: if we interfere with this target, what happens to the biological system?

The scale of current efforts is already substantial. Arc Institute's virtual-cell work has attracted thousands of researchers through its challenges, while recent models have been trained on datasets containing information from tens or hundreds of millions of cells. Researchers are now testing whether these systems can predict unseen perturbations and, crucially, transfer what they learned into cellular contexts that were absent from training.

Generalization is where the technology becomes difficult. A model can perform well when the training data contains similar cell types, genes and experimental conditions. Drug discovery often asks for the opposite: predict what happens in a disease state, patient population or intervention we have never observed before.

If virtual cells eventually make those predictions reliable, the economics could be huge. A company could reject weak targets before chemistry, animal studies and clinical trials begin. For now, though, virtual cells remain an important research frontier rather than a proven source of successful medicines.

Is proprietary biology data becoming the real moat in AI drug discovery?

Proprietary biological data is becoming one of the strongest moats in AI drug discovery because model architectures can spread quickly while unique experimental results remain much harder to copy.

Recursion built its company around this idea before foundation models became fashionable. Following the Exscientia combination, it has described access to more than 60 petabytes of proprietary and licensed biological and chemical data. Automated laboratories generate more data, models analyze it, and the next experiments can be chosen from what the system learned.

Lilly's TuneLab gives us an unusually revealing benchmark from inside Big Pharma. The platform offers biotech companies models trained on more than 500,000 preclinical data points collected across more than 20 years, including pharmacokinetic and toxicology experiments. Lilly already uses those models internally.

Those numbers explain why simply having a good neural network will become less impressive. Lilly can train against decades of private experiments. Recursion can generate new phenotypic measurements at industrial scale. Xaira is deliberately creating perturbation data for its biological models. Protein-design companies increasingly maintain wet labs so every failed or successful design improves the next generation of models.

The quality of that data matters more than raw volume. Drug discovery needs information tied to causality, experimental conditions and biological context. A carefully controlled perturbation dataset can be more useful than a much larger collection of loosely connected observations.

Companies building a loop between prediction and proprietary experiments should get stronger as models improve. Companies depending mainly on an algorithm that competitors can eventually reproduce face a much less comfortable future.

If you want more recent data on this point, please see our latest AI in drug discovery market report.

Can Big Pharma and Big Tech squeeze AI drug discovery startups?

Big Pharma and Big Tech are moving deeper into AI drug discovery right now, which will squeeze generic AI platforms but should increase the value of startups that own unusual data, models or drug assets.

Lilly has become the clearest example of the build-and-partner approach. It already operates TuneLab, has built major internal AI infrastructure and has now joined NVIDIA in a co-innovation laboratory where the two companies plan to invest up to $1 billion over five years in compute, talent and research. The lab is designed around continuous learning between computational systems and automated wet-lab experiments.

At the same time, Lilly keeps signing external AI partnerships. That combination looks a lot like the rest of biotech. Large pharmaceutical companies maintain thousands of internal scientists and still license drugs from startups because no research organization discovers everything itself.

Big Tech is pushing from the other direction. Alphabet already owns a strategic position through Isomorphic Labs. NVIDIA is becoming part of the infrastructure underneath computational biology. OpenAI was an early investor in Chai and has worked with Lilly. Anthropic has gone further lately by launching Claude Science and discussing plans to develop drugs itself, including work around neglected diseases.

This raises the bar for startups. Access to powerful language models, protein models or GPU infrastructure will gradually become common. A pitch built around “we use AI to discover drugs faster” will age badly.

A startup becomes much harder to squeeze when it owns something the larger companies cannot quickly recreate: a strong clinical asset, a proprietary experimental dataset, exceptional performance on a difficult biological problem or a lab system that keeps producing new training data.

Which AI drug discovery startups look built to last?

The AI drug discovery companies most likely to last are the ones accumulating several kinds of proof at once: better experiments, pharma money, proprietary data and clinical results.

Clinical evidence deserves the highest weight. Candidate counts will become less impressive as generative models make candidate creation easier. A positive randomized Phase II trial tells us far more than a model generating millions of molecules. A successful Phase III program would move the standard higher again.

Real payments from pharma come next. Huge “up to” deal values can make partnerships look more advanced than they are. Upfront payments and repeated milestones show that another sophisticated organization has reviewed the work and decided to keep paying.

The experimental feedback loop is becoming another dividing line. Companies such as Recursion, Xaira and Generate were designed around repeated interaction between models and laboratories. That creates a chance for the technology to improve from proprietary evidence rather than from the same public datasets available to everyone.

Capital discipline will matter more than it first appears. Faster discovery allows companies to create more programs than they can possibly take through clinical trials. The temptation is to fill a pipeline with dozens of candidates. The better company may be the one that kills most of them early and concentrates money on the few with unusually strong evidence.

We would be much more skeptical today of a startup whose main achievements are benchmark scores, virtual screening numbers or the size of its generated molecular library. The standard is moving toward whether the AI causes better decisions in the real drug-development process.

So what does the AI drug discovery startup landscape look like today?

The AI drug discovery startup landscape today is stronger, richer and scientifically more credible than it was during the first hype cycle, but the central clinical claim remains unproven: AI has yet to show that it can repeatedly produce approved drugs at a meaningfully higher rate than conventional R&D.

Several old uncertainties have already disappeared. AI can design credible small molecules. It can generate therapeutic proteins. It can move some candidates from design to the clinic unusually quickly. Major pharmaceutical companies are willing to pay serious upfront money for access. Investors are again deploying billions of dollars into the strongest companies.

The landscape has also become easier to segment. Insilico and Generate are now testing AI-driven discovery in late-stage clinical development. Recursion represents the industrial biological-data model. Isomorphic and Chai are competing closer to frontier AI model development. Xaira is making an unusually large bet on virtual cells and causal biology. Iambic is showing how AI can be integrated tightly into medicinal chemistry.

Recent developments have raised the scientific bar at exactly the same time as funding has exploded. The newest Nature Reviews Drug Discovery assessment remains skeptical about clinically relevant impact. That skepticism now sits beside genuine progress into Phase III. The models have improved faster than the clinical evidence.

Molecular design already looks permanently changed. Generating and optimizing candidates will become faster, more computational and increasingly integrated with automated experiments. That part of the AI drug discovery thesis no longer looks particularly speculative.

Disease biology is where the real race now sits. The next breakthrough will come from predicting which targets, mechanisms and patients actually matter before companies spend years proving themselves wrong in clinical trials. Virtual cells, perturbation models and proprietary biological datasets are all attempts to solve that harder problem.

For startups, this creates a much tougher landscape than the funding headlines suggest. Generic AI capabilities will become easier to obtain as pharma companies, NVIDIA, OpenAI, Anthropic, Alphabet and open research groups push deeper into biology. Durable value should concentrate around proprietary experiments, strong pharmaceutical relationships and drugs that work in patients.

So our answer today is clear. AI drug discovery has grown into a serious biotechnology industry and has already changed early-stage drug design. The clinical revolution is still being decided. Phase III programs now give us a way to test it with evidence that actually counts.

If you want more recent data on this point, please see our latest AI in drug discovery market report.

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.