AI Drug Discovery: what is getting real adoption now?

What counts as “real adoption” in AI drug discovery?

Real adoption in AI drug discovery means scientists are repeatedly using AI to decide what to test, what to make and what to advance, with real R&D money riding on those decisions.

That definition is stricter than counting AI partnerships. Pharma companies announce exploratory collaborations all the time, and many disappear quietly. We learn much more when AI becomes part of a workflow that researchers keep using, when a partner moves an AI-generated target or molecule forward, or when the resulting drug reaches patients.

Today, the evidence falls into several distinct layers. Computational prediction is already embedded in drug research. AI-assisted molecule design is getting close to that point. AI-driven target discovery has produced several serious programs but still has a short clinical track record. Autonomous discovery remains early. And the ultimate test—AI consistently producing medicines that survive late clinical development better than conventional drugs—has barely begun.

This distinction explains much of the confusion around the sector. Someone looking at pharma laboratories can reasonably say AI drug discovery is already here. Someone looking only at approved medicines can reasonably wonder where the revolution went.

What we can measure	Where AI drug discovery stands now	What it tells us
AI used in research workflows	Common and expanding	Real workflow adoption
Pharma paying for AI discovery	Widespread	Real commercial demand
AI-enabled drugs entering trials	117 identified in one industry-wide study	Discovery can reach humans
Assets completing Phase I	60 in the same study	A meaningful clinical cohort now exists
Assets completing Phase II	8	Efficacy evidence is still thin
Repeated late-stage superiority	Very little evidence yet	The biggest claim remains open

Where is AI drug discovery actually being used now?

AI drug discovery is getting its clearest adoption today in molecular prediction, virtual screening, lead optimization, protein modeling, ADME prediction and deciding which experiments deserve to happen next.

These tasks fit the technology unusually well. Medicinal chemists can theoretically explore an enormous number of molecules, while laboratory capacity forces them to make only a tiny fraction. A system that ranks those options more accurately can save real experiments without needing to solve an entire disease.

Schrödinger gives us a useful commercial read on this. In its latest reported quarter, annual contract value reached $29.6 million, up 27% year over year, while trailing four-quarter ACV reached $208 million. The company also expects full-year ACV of $218 million to $228 million. That is recurring software usage from organizations paying to run computational discovery, rather than a theoretical market estimate.

Lilly gives us the internal-pharma version. Its TuneLab platform exposes models trained on more than 500,000 preclinical datapoints collected over two decades, including pharmacokinetic and toxicology data. Lilly says the same underlying models are used by its own R&D teams every day.

Those examples point in the same direction: AI is becoming part of the machinery scientists use before they commit laboratory time and money.

Are big pharma companies actually changing how they discover drugs?

Yes. Several major pharma companies are currently building AI directly into their research systems, and the scale of those changes goes well beyond small pilot projects.

Lilly is one of the clearest cases. Its LillyPod supercomputer uses 1,016 NVIDIA Blackwell GPUs. The company has also committed with NVIDIA to invest up to $1 billion over five years in an AI research laboratory where computational models, scientific agents and automated experiments can continuously feed one another.

Roche is moving in a similar direction. Its new NVIDIA-powered AI factory supports Genentech's Lab-in-the-Loop approach, where models generate biological or chemical predictions, experiments test them, and the resulting data flow back into the models.

The interesting part is how much infrastructure surrounds the models. These companies are building computing capacity, automated laboratories, internal datasets, model-development teams and interfaces that ordinary researchers can use. That is a much harder commitment to unwind than signing an AI partnership.

We still need clinical results before saying these systems make pharma fundamentally more productive. Inside the R&D organization, though, the adoption question is getting easier to answer. AI has moved into the operating model at some of the world's largest drug companies.

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

Is AI molecule design becoming normal medicinal chemistry?

AI-assisted molecule design is currently one of the closest parts of AI drug discovery to becoming normal medicinal chemistry.

Drug design is an optimization problem with brutally clear feedback. Researchers may need potency against one target, low activity against another, acceptable solubility, the right half-life, low toxicity and enough exposure in the relevant tissue. Every laboratory result tells the team something about those properties.

AI can narrow the next round of molecules.

Recursion's REC-7735 shows how aggressive that narrowing can become. The company says it went from a novel hit to a development candidate in ten months after making 242 compounds. The resulting PI3Kα H1047R inhibitor was designed for more than 100-fold selectivity over wild-type PI3Kα. Regulators have cleared the program to begin a Phase I/II study.

Recursion has also said across its broader small-molecule work that its platform can advance candidates while synthesizing roughly 90% fewer compounds than an industry-average benchmark it uses internally. That company-reported comparison will need outside validation, but it shows what AI-native chemistry teams are actually trying to optimize: the number of physical experiments needed to reach a viable molecule.

Insilico Medicine provides another useful case. During development of rentosertib, the company says it went from therapeutic hypothesis to preclinical candidate in around 18 months while synthesizing 78 molecules.

We should avoid turning 78 versus 242 molecules into a league table because the programs started from different targets and different chemistry. The broader point is simpler. AI-heavy teams are increasingly trying to get more information out of every molecule they physically make.

Can AI actually discover new drug targets?

AI can already generate drug targets that pharma companies are willing to pursue, although we still have limited proof that AI finds better targets than conventional biology.

A recent Recursion milestone makes this much harder to dismiss as theory. Genentech has now exercised its first option on a neuroscience target found through the companies' collaboration and moved that target into a joint early-discovery program.

The path behind that decision is unusually interesting. Recursion built whole-genome CRISPR maps using disease-relevant neural cells, AI identified candidate relationships, and the companies then put those candidates through pathway, functional and disease validation. Genentech eventually selected one previously unexplored neuroscience target for a small-molecule program.

That is a better test than asking whether an algorithm can output a list of genes. A sophisticated pharmaceutical partner looked at the resulting biology, ran validation work and chose to spend additional resources on one of the targets.

Roche and Genentech have so far accepted six Recursion Phenomaps and initiated at least one small-molecule program from the collaboration. Recursion has received $213 million in upfront and milestone payments from that partnership. Sanofi has separately advanced multiple programs through milestone stages with Recursion.

Target discovery is getting operational adoption. The uncertainty starts one step later. Human biology has a long history of producing beautifully validated targets that fail once a drug reaches patients, and AI has not accumulated enough clinical outcomes to show that it escapes that problem.

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

Do any AI-discovered drugs actually work in humans yet?

Yes, a few AI-linked drugs have now produced real human efficacy data, although the clinical sample is still small enough that one or two failures could materially change the picture.

Rentosertib is the cleanest example because AI was involved both in finding the TNIK target and in designing the molecule.

Its randomized Phase IIa trial enrolled 71 people with idiopathic pulmonary fibrosis. After 12 weeks, patients receiving placebo had an average 20.3 mL decline in forced vital capacity, while patients on the highest rentosertib dose showed an average 98.4 mL improvement. That creates an observed separation of about 119 mL.

The trial was small. Only 18 patients started in the highest-dose group, 16 of the 71 participants discontinued before the end of treatment, and the study was designed primarily to assess safety. The efficacy result therefore needs a larger and longer trial.

Something interesting has happened since those original results. A new Nature Biotechnology analysis of serum samples from the trial found that six separate proteomic aging clocks all moved toward lower predicted biological age in the rentosertib-treated groups. That secondary analysis cannot prove an anti-aging effect, but it adds another biological observation to a program that had already shown changes in fibrosis-related proteins.

Recursion's REC-4881 adds a different kind of human evidence. In familial adenomatous polyposis, 75% of 12 evaluable patients had lower total polyp burden after 12 weeks, with a median decline of 43%. Twelve weeks after treatment stopped, nine of 11 evaluable patients still had lower polyp burden, and the median reduction reached 53%.

The REC-4881 dataset is even smaller and lacks a randomized placebo group, so confidence should stay lower there. Still, we have moved beyond the stage where every AI-discovery program stops at preclinical charts.

Drug	Where AI entered the process	Human evidence so far	How much weight we give it
Rentosertib	Target discovery and molecule design	Randomized Phase IIa signal in IPF	Important, still needs larger confirmation
REC-4881	AI-driven phenotypic discovery	Phase I/II activity in FAP	Promising, very small cohort
REC-7735	AI-native molecular optimization	Cleared for Phase I/II	Good discovery proof, no patient efficacy yet
Zasocitinib	Heavy computational design	Successful pivotal Phase III psoriasis studies	Strong late-stage validation of computational drug design

Are AI-discovered drugs doing better in clinical trials?

AI-enabled drugs appear unusually strong in Phase I so far, while Phase II results are much less impressive.

This is one of the most important findings in the whole field because AI may already be improving one part of drug discovery without solving the next one.

An earlier analysis in Drug Discovery Today estimated Phase I success for AI-discovered molecules at around 80% to 90%, comfortably above commonly cited historical industry rates. Phase II performance in the same analysis was around 40%, much closer to conventional drug-development experience.

We now have a larger denominator. An industry-wide analysis presented at the 2026 ASCO meeting identified 117 AI-enabled therapeutic assets from 63 companies that had entered interventional clinical trials. Sixty had completed Phase I by the end of 2025. Only eight had completed Phase II.

The ASCO dataset also tells us something about selection. Oncology represented 69 of the 117 assets, or 59%. Small molecules represented 96, or 82%. Only about 36% were aimed at novel biological targets.

So the attractive Phase I numbers may reflect several effects at once. AI could genuinely be producing cleaner molecules. Many programs are also operating in areas with established targets, rich datasets and well-understood chemistry.

For now, the Phase I advantage looks credible enough to take seriously. The claim that AI improves clinical efficacy across diseases has far weaker support.

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

Why can AI design a good molecule and still pick a bad drug?

AI is currently much better at optimizing chemistry than predicting whether changing a biological target will actually help a patient.

The difference sounds obvious once we separate the two problems.

A molecule gives researchers rapid feedback. They can measure binding affinity, metabolic stability, solubility, permeability, selectivity, exposure and dozens of other properties. Those experiments create relatively structured datasets, which are exactly what machine-learning systems like.

Disease biology gives far messier feedback. A target can look compelling in cells, animals, genetics and patient datasets and still fail in Phase II. Tumors evolve. Biological pathways compensate. Different patients with the same diagnosis can have different disease drivers.

The current clinical numbers fit that distinction surprisingly well. Strong early safety and pharmacokinetic performance is exactly where improved molecular design should show up first. Phase II forces the program to answer the harder biological question.

That is also why we should be careful when people talk about an "AI drug" as though AI did one homogeneous thing. AI may have designed an excellent molecule against a target humans selected years earlier. Another program may use AI to discover the target itself. The second program is asking much more of the technology.

Has AlphaFold-style protein prediction turned into real drug discovery?

Protein-structure AI has clearly crossed into real drug-discovery work, and companies are now building entire discovery businesses around that capability.

The change is easy to underestimate because AlphaFold itself became famous as a scientific breakthrough. Its practical value is more mundane and potentially more important: researchers can start many projects with structural information that previously required expensive or difficult experiments to obtain.

Predicted structures can help identify binding pockets, compare mutations, prioritize targets and guide molecular design. Experimental structures and physical simulations remain crucial because proteins move, binding pockets change shape and a static prediction cannot capture every drug interaction.

Isomorphic Labs shows where this technology is heading commercially. Its Novartis collaboration originally covered three difficult small-molecule targets and was later expanded by up to three additional programs. The company has also built programs with Lilly and, more recently, Johnson & Johnson across multiple targets and modalities including small molecules, antibodies, peptides and molecular glues.

Isomorphic then raised $2.1 billion in its latest financing round to expand its drug-design engine and advance its own pipeline.

Funding tells us what investors believe. The repeated pharma programs tell us more about adoption. Novartis had experience working with the system and chose to expand the number of programs. Johnson & Johnson subsequently added another major pharmaceutical user.

Protein prediction has become part of drug design rather than remaining a scientific side tool. We are still waiting to see whether Isomorphic's own molecules can carry that advantage through the clinic.

Why do pharma companies keep paying outside AI drug-discovery firms?

Pharma companies keep using external AI firms because proprietary biological data and specialist computational expertise are valuable in different places, and combining them is often faster than rebuilding everything internally.

The clearest evidence is repeat behavior.

Novartis expanded its Isomorphic Labs collaboration. Bristol Myers Squibb already used Schrödinger's computational platform and has now expanded that relationship by deploying Bunsen inside its research organization. Genentech moved from generating Recursion biological maps to optioning an actual neuroscience target. Sanofi has repeatedly advanced Recursion programs through discovery milestones.

Those observations are stronger than the multibillion-dollar maximum values frequently attached to AI partnerships. Most headline milestone values will never be paid unless programs travel a very long way.

Lilly has taken another route with TuneLab. Participating biotechs can use selected models trained on Lilly's proprietary drug-development data while contributing additional training information through federated learning. The underlying private datasets stay inside each participant's environment.

Demand arrived quickly. TuneLab's team initially targeted five partners before launch, then received more than 600 inbound inquiries and raised its internal ambition to 50 partners by year-end. The platform is also being integrated into established research environments including Schrödinger LiveDesign, Revvity Signals and CDD Vault.

That integration layer is a big part of the appeal. AI tools become much easier to adopt when researchers can access them inside software they already use.

Is AI actually making drug discovery much faster?

AI is already cutting months and sometimes years from specific discovery steps, but clinical development still moves on a much slower biological clock.

The early-discovery examples are real. Insilico reports around 18 months from hypothesis to preclinical candidate for rentosertib. Recursion reached a development candidate for REC-7735 in ten months from its first novel hit. Computational teams can screen or score enormous chemical libraries before synthesizing anything.

Those gains can add up. A company that avoids weak compounds earlier saves synthesis, assays, animal work and scientist time. A toxicity problem found before candidate selection is dramatically cheaper than the same problem found after a Phase I trial starts.

Clinical development changes the equation. Human trials still require patients to be recruited, treated and followed. Chronic diseases need enough time for outcomes to appear. Regulators need safety data. Large Phase III trials often need hundreds or thousands of participants.

So the credible productivity case today is built from many smaller compressions across discovery.

That can still be worth billions. A pharmaceutical company does not need a ten-year development program to suddenly take two years. Consistently removing failed experiments and shortening candidate selection would already change the economics of R&D.

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

Are AI scientists and autonomous drug-discovery labs actually useful now?

AI agents are beginning to handle real drug-discovery workflows, while the most convincing use today is scientific orchestration under human control.

Schrödinger's Bunsen is the freshest commercial example. Bristol Myers Squibb recently agreed to deploy the AI co-scientist across its research organization after years of using Schrödinger's computational tools.

Bunsen can plan and execute multi-step molecular workflows, run calculations in parallel, interpret results and connect with systems such as Schrödinger's RetroSynth synthesis-planning software. During its latest quarter, Schrödinger also launched Bunsen in early access and said its own therapeutic teams had already been using the system internally.

Lilly is building the same basic idea from another direction. Its collaboration with NVIDIA connects AI agents to automated wet laboratories so computational suggestions can generate experiments continuously. Genentech's Lab-in-the-Loop follows a similar cycle.

This is probably where autonomous systems become useful first. Drug researchers spend substantial time moving between databases, modeling programs, chemical-design tools and experimental results. An agent can automate many of those steps while scientists remain responsible for the scientific judgment.

We have very little evidence that an autonomous system can independently decide which disease hypothesis deserves hundreds of millions of dollars of clinical development. Today's adoption is happening lower down the stack, where agents can execute validated scientific workflows faster and at much higher volume.

Is AI drug discovery making money yet?

AI drug discovery has clearly become a business people will pay for, although we cannot yet show that AI produces a superior return on total pharmaceutical R&D spending.

Schrödinger currently gives us one of the cleanest pieces of evidence because its computational platform has recurring customers. Trailing four-quarter ACV reached $208 million in its latest reported quarter, and quarterly ACV grew 27% year over year.

The downstream economics have also become harder to ignore. Ajax Therapeutics, co-founded by Schrödinger and built around a molecule designed in collaboration with the company, was acquired by Lilly for up to $2.3 billion. Schrödinger owned about 5.8% of Ajax before the deal.

An earlier Schrödinger-linked program, zasocitinib, gives us a stronger scientific endpoint. The compound came from a computationally intensive collaboration between Nimbus and Schrödinger, was acquired by Takeda, and has now produced positive results across pivotal Phase III psoriasis studies. In a head-to-head Phase III trial, more than 35% of patients taking zasocitinib achieved complete skin clearance at week 16, more than 2.5 times the rate seen with deucravacitinib.

Recursion gives us a different commercial model. Across its collaborations, the company has collected more than $500 million in partner cash inflows, including repeated discovery milestones from Sanofi and Roche/Genentech.

None of these figures gives us the calculation we ultimately want: successful approved medicines per dollar of R&D.

Commercial adoption has still crossed a meaningful threshold. Pharma is paying repeatedly, some AI-linked assets are creating multibillion-dollar transaction value, and computationally designed molecules are reaching late clinical development.

Economic test	Evidence today	Our read
Will pharma pay for AI tools?	$208M trailing ACV at Schrödinger	Clearly yes
Will pharma pay for AI-generated discovery work?	Repeated milestones across several partnerships	Yes
Can AI-linked molecules create major asset value?	Ajax deal up to $2.3B; major prior transactions	Yes
Can computational drugs reach Phase III?	Zasocitinib has done it	Yes
Does AI raise pharma-wide R&D returns?	No mature comparative dataset yet	Still unknown

Where is AI drug discovery still being overhyped?

The biggest overstatement today is the idea that AI has already learned how to reliably invent successful medicines from end to end.

The numbers give us a cleaner view.

The ASCO industry study found 117 AI-enabled clinical assets across 63 companies. Only eight had completed Phase II by the end of 2025. Nearly three-fifths of the pipeline was in oncology, more than four-fifths consisted of small molecules, and roughly two-thirds targeted biology that was already known.

That is a serious clinical pipeline, but it is still a highly selected one.

Autonomous target discovery remains particularly easy to oversell. Finding statistical relationships in biological data can produce thousands of hypotheses. The expensive question is whether manipulating one of those targets changes the course of a human disease. Genentech's recent decision to advance a Recursion-generated neuroscience target is encouraging precisely because so few AI-discovered targets have reached that kind of external validation.

Claims about speed also get stretched. AI can dramatically accelerate virtual screening and molecule optimization. Human trials continue to consume years.

The same goes for clinical success rates. The unusually strong Phase I record deserves attention. Phase II has so far looked much more ordinary.

The useful technology has advanced faster than the story people tell about it.

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

So what is getting real adoption in AI drug discovery now?

AI drug discovery has real adoption today, especially when AI helps researchers design molecules, predict properties, understand protein structures and choose better experiments.

The adoption hierarchy is becoming quite clear.

At the top sits computational chemistry and molecular prediction. The recurring software revenue at Schrödinger, Lilly's daily internal model usage and the integration of TuneLab into common discovery software show that these tools are already becoming part of ordinary research.

AI-assisted molecule design comes next. Programs such as REC-7735 and rentosertib show that teams can move from computational ideas to real clinical candidates with unusually small experimental searches and short discovery timelines.

Target discovery is further behind, although Genentech's decision to advance a Recursion-generated neuroscience target gives the field one of its strongest recent validation points.

Clinical efficacy remains the bottleneck. Rentosertib has produced randomized Phase IIa evidence. REC-4881 has shown encouraging patient activity. Zasocitinib demonstrates how far computational drug design can ultimately travel, with successful pivotal Phase III studies. Yet the broader denominator still contains 117 AI-enabled clinical assets and only eight that had completed Phase II in the latest industry-wide analysis.

So our answer is fairly sharp now: AI has already become a real drug-discovery technology. Its strongest job today is helping scientists make better choices before and during medicinal chemistry.

Whether AI can consistently make the much bigger choice—which biological idea will become a successful medicine—remains the part the industry still has to prove.

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.