AI in Pharma: Who Is the Hero, and Who Is the Gimmick

Abstract

AI drug discovery puts its effort in the wrong place. Almost all of a drug program's money goes to the clinic. Everything before the first human dose, designing the molecule included, is only about nine percent of the cost. Most programs die on the target, not the molecule: the target turns out not to matter in people, and only a trial can show that. So the money flows to whoever pays for the trials. Within a class of drugs, the latecomer often wins. Chinese drugmakers sell their molecules to buyers who can afford Phase 3. And AI drug companies are caught in a bind: without a pipeline they are priced like a tool, and with one they become just another biotech. Recursion, the textbook case among listed companies, has spent the past few years turning from a platform into a biotech. Two earlier booms, genome sequencing in the 1990s and IBM's Watson in the 2010s, ended the same way. The technology stayed in use, and the companies valued on it were sold or shrank.

The computation works, but on the cheapest stretch of the process. The decisions that cost real money (which target to pursue, how to design the trial, when to cut losses) rest on experience that nobody has written down or ever scored, and drugmakers keep both those decisions and the data behind them under tight control. Training models on papers and knowledge bases does not solve this either. Many published results cannot be reproduced, the literature is full of filler and some outright fraud, and the reliable data lives in the wet lab. The benchmarks that model companies use to show off their scientific ability amount to little more than knowledge quizzes and leaderboard grinding. The companies worth backing, then, are the ones that can produce their own gold-standard data cheaply and, in the end, make their own drugs. And to tell a hero from a gimmick, look at only two things: real substance and real money.

The money goes to the clinic, not to discovery

Almost all of a drug program’s money is spent before it earns anything, but once the drug is approved it earns for years. A patent lasts twenty years from the filing date, which usually leaves ten to fifteen years of exclusive sales after approval. Each pill costs next to nothing to make, and insurers and patients rarely turn a drug down because of its price. A program that stops short of approval earns nothing. Anyone selling to drugmakers should start by asking which part of a program’s costs their product touches. Most of the cost is in the clinic, that is, in human trials. Trials come in three phases: Phase 1 checks safety in a small group of people, Phase 2 checks whether the drug works in a larger group, and Phase 3 is the large trial that supports approval.

Figure 1

Phase 3 alone costs more than seven times everything before human testing

Out-of-pocket cost at each phase, US$ millions in 2026 dollars (converted from the report’s 2018 dollars with the US consumer price index), all therapeutic areas.

• The stage where AI works
• In humans and at the FDA

For a program that makes it to the first human dose, everything up to that point costs about US$15.6 million in 2026 dollars (all figures here are in 2026 dollars). That includes target work, screening (sifting through large numbers of compounds for the ones that bind the target), lead optimization (reworking a promising molecule until it is fit to test in people) and the animal studies needed to apply for a trial. Phase 3 alone costs US$118 million, more than seven times as much.

Now imagine a perfect discovery engine. It is free and instant, and for any target it returns the best molecule that exists. Even so, in a program of about US$174 million (the phases in the chart added together), it could save at most US$15.6 million. Some programs do die on chemistry, from a breakdown product in the body that nobody saw coming, say, or a heart side effect that nobody checked. But this engine is already assumed to be perfect at chemistry, and it still cannot touch the thing that sinks most programs: the target turns out not to matter in people. At best, the engine saves about nine percent of the cost and leaves the risk almost where it was.

What the chart leaves out

1. Most of the money goes to programs that fail. The report behind the chart puts the average out-of-pocket cost (the cash actually spent) at about US$229 million per drug in development. Add the programs that failed and the cost of capital over ten years of waiting, and the figure becomes about US$1.16 billion. Most programs fail along the way, and the approved drugs have to earn that money back for them.
2. Most approved drugs never become blockbusters. A blockbuster is a drug that sells more than US$1 billion a year, and drugs like Keytruda, which sell more than US$10 billion a year, are rarer still. Of the 487 drugs the FDA approved between 2014 and 2023, Evaluate forecasts that 193 (about 39%) will top US$1 billion in annual sales at their peak, and only 17 will top US$10 billion.
3. A cheaper front end does help, but only the side that pays for the clinic gets paid. The front end is the discovery stage before trials. At the same success rate, more candidate molecules mean more approvals, so a cheaper front end is worth something. But failure is concentrated in the clinic. Every extra candidate means another round of clinical spending, and the payoff goes to whoever puts up that money. A discovery engine that doubles the number of candidates earns its maker nothing if nobody pays for the trials.
4. These absolute figures run low, but the ratio holds. The figures in this model are lower than the large-pharma numbers usually quoted. DiMasi and colleagues estimate Phase 3 out-of-pocket costs at about US$364 million per drug in development in 2026 dollars (US$255 million in the 2013 dollars they used). Deloitte’s latest count, on 2025 data, puts the average cost of taking a drug from discovery to launch at the 20 biggest drugmakers at US$2.67 billion, with the cost of failures spread across the winners. The estimates disagree on the levels, but in every one of them Phase 3 is the most expensive phase, and by a wide margin.
5. A small share of the cost does not mean a small share of the value. An input that accounts for just nine percent of a program’s spending can still capture far more than nine percent of the revenue, as long as it is scarce enough. Software businesses do this every day. Discovery in drug development cannot, and one reason is that there are too few buyers. Two Phase 3 programs cost roughly US$240 million to more than US$700 million in 2026 dollars. Apart from twenty or so big drugmakers, only about a hundred biotechs can pay for two at once on their own: 72 listed biotechs in the US and Europe have more than US$500 million in annual sales, and a number of clinical-stage companies hold a lot of cash. Meanwhile, nearly 7,000 companies develop drugs. Every one of these buyers has its own chemistry team, and none of them is short of molecules.

The second drug often wins

A patent is a trade: the company gets a monopoly, and the price is making the molecule public. The company then runs its trials, and the trials tell everyone watching whether the target holds up. So within a class of drugs (drugs that act on the same target count as one class), the first company pays for the discovery, carries the target risk and ends up handing the answer to its competitors. The second company can start from three things that already exist: a validated target (trials have shown it really matters), a published molecular structure and a disclosed clinical endpoint (the measure the trial used to judge whether the drug works). From there it builds a better drug.

In every row, the drug that entered the class later either became the bigger product or beat the first drug in a head-to-head trial (a trial comparing the two drugs directly), and each time the first company had paid for the discovery.

Read with the cost figures above, this should worry anyone selling discovery. If the second drug often wins, then finding a new target is just a cost, ideally one paid by someone else’s shareholders. Big drugmakers make their money on portfolio decisions (choosing which programs get funded, such as which validated target to chase) and on running trials well, not on discovery.

Survivorship bias in this table

1. The winners in this table are the industry’s biggest products. Atorvastatin was the fifth statin (a class of cholesterol-lowering drugs), and its own company’s executives nearly killed it because they thought the market was already taken. Adalimumab was the third drug against TNF (an inflammation signal), and it went on to become the best-selling drug in history. Zanubrutinib was the third BTK inhibitor (a drug that blocks the activity of the BTK enzyme) and beat the first one in a head-to-head trial. Zasocitinib, which comes up again later, was filed for approval four years after the first TYK2 inhibitor was approved, and still beat it head to head.
2. The molecule decided the winner, but the company that ran the trials got paid. Within a validated class, the molecule is the main difference, which is why chemistry is not going away. But in every row the winning molecule came from a chemistry team the winner already had, and the profit went to the company that won the trial.
3. This table only includes classes where a second drug came along. Most targets never get a second drug: the first one fails, and nobody touches the target again. And nobody makes a separate table of the classes where a fast follower had nothing to follow. The rule is narrower than it looks. Latecomers do better only when the target holds up, and someone else has usually spent ten years and a billion dollars finding that out.
4. Following is a strategy for big companies. It only works if you can afford the trials that following requires, and then have the factories and the sales force to win the market. In obesity drugs, semaglutide from Novo Nordisk (NVO) reached the market first, but it spent more than 1,000 days on the FDA’s drug shortage list, which blunted its head start. Tirzepatide, from Lilly (LLY), spent about 300 fewer days on the list. Since then Lilly has steadily taken share from Novo Nordisk, and supply may be one reason.

China sells the molecules, but the buyer still controls the path to approval

If a drug is only worth something once it gets through the clinic, then what is scarce is the ability to get a molecule past the FDA: enough cash to fund Phase 3, a track record of working with the regulator, and data the FDA is willing to accept. Companies that have all three keep most of the profit. Companies that have none of them can only be asset suppliers (sellers of drug candidates, which the industry calls assets): they make molecules cheaply and sell them to companies that can carry them to approval.

Chinese drugmakers now do this at scale. In 2025 they signed 157 out-licensing deals (out-licensing means selling the rights to a drug candidate) with a headline value of US$135.7 billion, up from 94 deals worth US$51.9 billion in 2024. The average cash paid at signing has more than tripled since 2022. Two things have not changed, though. Most of the headline value is milestone payments (money paid only if later trials succeed or sales targets are hit), and most of that money will never be paid. And it is still the buyer who decides whether the molecule ever gets a chance to reach patients.

Behind the deal numbers

1. Prices are up, but the deals are still built the old way. By 8 October 2026, the year’s cumulative headline value of out-licensing deals had reached US$139.4 billion, more than in all of 2025. Upfront payments (the cash paid at signing) came to US$8 billion in the first three quarters, compared with US$7.2 billion for all of 2025. According to Evaluate, the average upfront payment in licensing deals for Chinese assets rose from US$52 million in 2022 to US$172 million in early 2026. The business works like this: a company sells a preclinical molecule (one not yet tested in people) for fifty million dollars, about the average upfront in 2022, and after that, each time the molecule changes hands, the value it gains goes to the next buyer. On 30 January 2026, AstraZeneca (AZN) agreed to pay CSPC (1093.HK) US$1.2 billion in upfront cash for the rights outside China to eight once-monthly injectable obesity programs, none of which had finished Phase 1. Eight molecules at that price would have fetched around four hundred million dollars, a third of what CSPC got. Pfizer (PFE) agreed to pay US$1.25 billion upfront for a single bispecific antibody (an antibody drug that binds two targets at once) from 3SBio (1530.HK), which was already in several clinical trials in China.
2. Analysts disagree on how much of the discount remains. By the count of Clarivate (CLVT), Chinese companies accounted for 32% of global out-licensing deals in the first half of 2025, up from 5% in 2020. Yet upfront payments for Chinese assets were still 60 to 70% lower than for comparable Western assets, and total deal values 40 to 50% smaller. In early 2026, though, an Evaluate analyst said upfronts for Chinese companies were catching up with deals elsewhere. Either way, the seller’s money comes mostly at signing, and what the drug earns once it reaches the market goes to the buyer.
3. Data from a single country is the weak point. Sintilimab, a PD-1 inhibitor (a cancer immunotherapy) from Innovent (1801.HK), was licensed by Lilly and had been tested only in China. In February 2022, the FDA’s Oncologic Drugs Advisory Committee voted 14 to 1 that the FDA should require additional trials showing that the results apply to American patients before making a decision. The committee only advises, and the vote was about whether to require more trials, not whether to approve the drug. But the FDA’s own briefing document listed three conditions under which foreign trial data alone can support an approval, and found that the trial did not meet them. The FDA then rejected the application. Surufatinib was rejected on similar grounds, even though it came with a US bridging study (a small extra trial run to show that results carry over to American patients). In May 2026, the acting head of the FDA office that reviews cell and gene therapies said that running the first human trials outside the United States might pay off in the short term, but could backfire in the long run.
4. The rule is not absolute. Toripalimab was approved on 27 October 2023 for nasopharyngeal carcinoma on the strength of a Phase 3 trial in a single region and a Phase 2 trial in a single country, after being turned down earlier over manufacturing problems. But multiregional trials are now the default expectation, and their cost either falls on the seller or shows up as a discount in the price.
5. The current test is ivonescimab. It is a bispecific antibody from Akeso (9926.HK) that targets both PD-1 and VEGF (a growth factor that tumors use to grow new blood vessels), and Akeso has licensed the rights for the United States, Europe and several other markets to Summit Therapeutics (SMMT). Akeso has reported positive results from three Phase 3 lung cancer trials run only in China, and one of them, HARMONi-A, showed a statistically significant survival benefit (one unlikely to be due to chance). But Summit’s application to the FDA rests on HARMONi, a Phase 3 trial run across several regions, and the FDA’s decision is due on 14 November 2026. The FDA had explicitly asked for a statistically significant benefit in overall survival (how long patients live), and HARMONi did not deliver one. A hazard ratio compares the risk of death between the two groups of patients: 1.0 means no difference, and the lower the number, the better the new drug. Among HARMONi’s 165 patients in North America and Europe, the figure was 0.98 at the main analysis in April 2025, against 0.79 for the trial as a whole. These patients enrolled more slowly than expected, so at that point their median follow-up was only 9.2 months. In an update released in July 2026, their hazard ratio was 0.76, close to the 0.78 and 0.79 reported for the whole trial in earlier analyses. The FDA now has to decide whether a survival result that falls short of statistical significance is good enough, and whether the results apply to American patients.

The pipeline question has no good answer

Sooner or later, every AI drug discovery company, and more and more AI-for-science companies, has to answer one question: should it build its own pipeline (a set of drug candidates the company takes into trials itself)? Both obvious answers are bad. Answer no, and the company is a software business selling a model, which the industry calls a platform. It can publish results on benchmarks (standard tests for comparing models) and even top them, and that is enough to carry it to around the second funding round. Then investors ask where the revenue will come from, and the answer is a handful of research contracts. They had valued it as a software business.

Answer yes, and the company becomes a biotech. A biotech is valued on its pipeline, a pipeline needs clinical money, and so each extra dollar spent on the platform is a dollar taken from the drug candidates that support the valuation. The asset wins this argument every time, because a drug candidate has a clear date (the day the trial reads out) and a platform does not. There is a third answer: sell molecules early to fund the platform. That keeps the company alive, but only as an asset supplier. Schrödinger (SDGR), Recursion (RXRX) and Isomorphic Labs have all answered this question in public. Recursion’s story is the most complete, and the next section tells it in full.

Figure 2

Cash at signing: a platform on its own, and a platform with assets attached

Guaranteed cash at signing or closing, US$ millions. Hover or tap a bar for the headline value and what was bought.

• Platform alone, no assets
• Assets attached, with a clinical path

Buyers paid only US$82.5 million in cash to partner with the drug discovery company owned by Alphabet (GOOGL), but US$4 billion in cash for a single molecule that had finished Phase 2b (the larger, later stage of Phase 2).

The evidence, company by company

1. Schrödinger had every advantage and still ended up building a pipeline. Founded in 1990, it is the best-known computational chemistry software company, and its software was used to design zasocitinib, discussed later. For decades its answer was no. Then it switched to yes and took three wholly owned cancer programs into the clinic, one of which it dropped in August 2025 after two patient deaths. By mid-2025 it was exploring strategic options for the lead program, which in practice means looking for someone else to pay for the trials. For 2025 it reported revenue of US$256 million and an operating loss of US$167 million. Its net loss narrowed to US$103 million only because of US$65 million of other income, about three-quarters of it a paper gain on equity investments.
2. Without a pipeline, an AI company is a high-tech CRO. A CRO is a contract research organization, the industry’s outsourcing firm. Many AI companies do much the same work: they use algorithms to sift through a large number of options for a drugmaker and tell it which few look most promising, which saves the client time. The business is paid by the project, one job at a time, and cannot be built once and sold again and again like software. By the time a drug reaches the clinic, the factory and the market, the buyer holds all the cards. It has the money, the plants and the sales channels, and it can pay once to buy the patent outright or take a license, after which how much the drug earns has nothing to do with the seller. A seller that wants royalties or milestone payments needs a reason for the buyer to share, and if the buyer can build its own AI team or switch suppliers at any time, there is none.
3. Within a few budget cycles, the pipeline eats the platform’s entire budget. Taking one program from a candidate molecule to a filing for human trials (toxicology studies, process development and the first batches of drug) costs around forty million renminbi, about US$6 million. It is like a chip tape-out (sending a finished chip design to the factory for a trial run): you commit the money, you wait, and only at the end do you learn the result. A company cannot pay for that and a research platform out of the same budget. Two or three budget cycles later, the platform has become an in-house service group, and the company has become a mid-sized biotech with six programs and a few months of cash. Plenty of AI drug companies look like this today, and a better model would not have changed that.
4. The third answer turns the AI company into an asset supplier, like the Chinese drugmakers. It still makes molecules, but sells them early and uses the proceeds to fund the platform. Insilico (3696.HK) discovered the cancer molecule ISM3091 with its own generative AI platform and licensed it to Exelixis (EXEL) in 2023 for US$80 million upfront, a month after the molecule entered Phase 1. Insilico’s chief executive said one reason was an overcrowded pipeline: 31 programs aimed at 29 targets. Chinese drugmakers use a variant of this called a NewCo, in which investors set up a new company outside China that takes the rights to the asset and pays for the trials. The originating company and the investors share ownership, and the originating company still collects an upfront payment and royalties (a percentage of future sales). Hengrui (600276.SH, 1276.HK) used this route to license its obesity drug HRS9531 (in the same class as semaglutide and tirzepatide) to a NewCo called Kailera Therapeutics. Selling early keeps the platform going. The catch is that the AI company sells at early-stage prices and gives up whatever the drug earns later.
5. Recursion and Exscientia merged, then cut programs and staff. They were the two most prominent platform companies in the West, and when they merged in 2024 they promised about ten clinical readouts (the moment a trial’s results come out) within eighteen months. Layoffs began three months later. By the ninth month, two mid-stage programs had been stopped and a third paused. By the tenth, a fifth of the staff had left and research had shrunk to cancer and rare diseases.
6. Two platform deals together brought in one-fiftieth of what one molecule did. Isomorphic Labs is Alphabet’s drug discovery company. It is built on AlphaFold (a model that predicts protein structures) and led by a Nobel laureate. In January 2024 it signed its first two partnerships with drugmakers, with a headline value of about US$3 billion, of which US$82.5 million was cash. In December 2022, Takeda (TAK) bought Nimbus’s TYK2 subsidiary for US$4 billion in guaranteed cash plus up to US$2 billion in sales milestones. What it bought was a single molecule, which had already completed Phase 2b in psoriasis. Nimbus, too, is a computational drug design company. The cash is about fifty times apart, and the plainest difference is that Nimbus had a molecule to sell. Clinical data probably does not explain most of the gap. In the AstraZeneca and CSPC bar, CSPC’s lead program had not been tested in people either, yet it still brought in US$1.2 billion in cash, because eight specific molecules came with it.
7. Isomorphic has not escaped the pipeline dilemma either. A platform lives on investors’ money, because cash from buyers cannot carry it. In May 2026, Isomorphic raised US$2.1 billion in a round led by Thrive Capital, bringing its total to about US$2.7 billion, without having disclosed a single pipeline molecule or put anything into the clinic. The company says it will spend the money developing the platform and using it to fill the pipeline and move programs into the clinic faster. The best-funded discovery platform raised US$2.1 billion to turn itself into a biotech.
8. Most startups cannot name their molecule, and the listed ones are traditional drug companies underneath. Ask an AI drug discovery startup to name one molecule that its own platform produced and that a competent human team could not have. Most cannot. The business plan has a pipeline slide, and underneath the pipeline is conventional chemistry. The listed ones run as two outfits under one name: an AI front that raises the money, and a traditional drug R&D group in the back that does the work. When money gets tight, the AI group tends to go first, though the Recursion timeline in the next section cannot show which group was cut first. It only shows that under pressure, the merged company cut programs and a fifth of its people, with layoffs on both sides of the merger.
9. Two well-known AI drug companies listed in Hong Kong gave opposite answers to the pipeline question. Call them Company A and Company B. Company A essentially said no. It sells drug discovery services, lab robots and what it calls AI-for-science solutions, to customers in new materials and consumer health as well as pharma. It raised more than US$750 million in 2025 by placing new shares, and that year it turned its first full-year profit, RMB 134.6 million on revenue of RMB 802.6 million. But by its own announcement, its two core businesses, robotics and drug discovery, were still losing money, just less than the year before. The profit came mainly from two places. First, financial assets it holds rose in value by at least RMB 500 million. Second, in 2024 it had booked a large paper loss on preferred shares issued before its listing. Those shares converted to ordinary shares at the listing, so the loss did not recur. Set those two aside and Company A is a services company with a large investment portfolio, earning its revenue from clients. Company B said yes. It listed at the end of 2025, in what it called Hong Kong’s largest biotech fundraising of the year. In 2025 it had revenue of US$56.2 million, an adjusted loss of US$43.8 million, and R&D spending of about 145% of revenue. By its own account, its newest batch of eight heart and metabolic disease molecules goes after established, high-confidence targets such as GLP-1R (the target of semaglutide) and Lp(a) (a blood particle linked to heart disease), plus a few newer ones, and aims to stand apart on chemistry. That is the follower strategy, and a sensible one. Its lead program is the exception: the target is new, and its own platform found it. It expected to turn a profit in the first half of 2026, mainly from out-licensing and collaboration income, which is the third answer described above.
10. Why the AI label helps, and why so few companies will pay for the research. The AI label helps because it lowers the cost of raising money: once one drug succeeds, investors treat it as evidence that the platform behind it can produce more. Few will pay for the research itself, because anyone can free-ride. Whatever the first company learns about making the models work, everyone else can learn for free just by looking at its results, so the rational move is to sit on the cash and let someone else go first, just as with a second drug.

Recursion is the textbook case

Recursion shows most clearly how listed AI drug companies have fared. Founded in Salt Lake City in 2013, it looks for drugs by taking microscope pictures of huge numbers of cells with automated equipment, then using machine learning to compare how drugs and genetic changes alter what the cells look like (a method called phenotypic screening). In 2021 it went public at US$18 a share, calling itself “a biotechnology company scaling more like a technology company”. Since then it has followed the path laid out in the last section: it built its own pipeline, merged with Exscientia, then cut programs and staff. In 2026 it also got a new chief executive and a new chair.

Its best human data so far comes from REC-4881, a molecule that originally belonged to Takeda. The platform’s contribution was to pick a disease for it: familial adenomatous polyposis (FAP). Picking a disease for an existing molecule is closer to the expensive decisions described earlier than to designing molecules. The result is positive: after 12 weeks, the total amount of polyps in the gut (the polyp burden) fell by a median of nearly half. But the trial has only 15 patients and no control group, and a registrational trial (the pivotal trial meant to support approval) will not start before the second half of 2027. The molecules the company designed with AI are all still in early trials.

In just over five years, Recursion went from a company promising to scale like a tech business to a clinical-stage company betting on five programs of its own. In November 2026 its shares traded around US$3.80, against US$18 at the IPO.

Money is still short, too. Roche (ROG.SW) and Sanofi (SNY) have paid about US$360 million over the years, roughly what Recursion spends in one year. In the third quarter of 2026 it lost US$119 million, and in October it sold another US$200 million of shares.

The programs and the money

1. REC-4881’s molecule came from Takeda, and the platform picked the disease. The molecule was called TAK-733, and Takeda had tested it in solid tumors. The earliest version of Recursion’s platform screened human cells missing the APC gene (a gene that keeps gut cells from overgrowing) and found that blocking MEK (an enzyme that relays growth signals inside cells) made the cells look closer to normal. In 2020 Recursion licensed the molecule from Takeda for US$1.5 million upfront plus up to US$39.5 million in milestone payments. FAP is an inherited disease caused by mutations in APC. Hundreds to thousands of polyps grow in the gut, nearly every untreated patient develops colorectal cancer, and no drug is approved for it. FAP was picked by the earliest version of the platform, so the choice says little about how good today’s system is.
2. The results are encouraging, but early. As of September 2026, 13 of 15 evaluable patients had less polyp burden after 12 weeks, with a median drop of 47%. Twelve weeks after stopping treatment, 11 of 13 were still below where they started. For comparison, in a Dutch registry, 87% of similar patients who took no drug saw their polyp burden rise each year. Of the 26 treated patients, 4 had grade 3 side effects (severe but not life-threatening), and 4 stopped because of adverse events. This is a small open-label trial (doctors and patients know which drug is given) with no control group, so it cannot show much yet.
3. A registrational trial will probably have to measure something else. Another FAP drug is already in Phase 3: eRapa, a formulation of rapamycin from Biodexa (BDRX). Its trial is randomized, double-blind and placebo-controlled, with 100 of a planned 168 patients enrolled by October 2026. According to Biodexa’s chief executive, the FDA and the European Medicines Agency both said polyp burden alone could not serve as the endpoint of a registrational trial. So the eRapa trial uses a composite endpoint and has to wait until enough patients progress (needing major surgery, for example), which could be as early as mid-2028 or as late as the end of 2029. REC-4881’s results so far measure polyp burden. If the FDA takes the same view of it, the registrational trial will have to measure something else and will take longer.
4. REC-1245’s target and molecule both came from the platform. It makes cells break down a protein called RBM39, which helps with RNA splicing (the step that trims freshly copied RNA into its finished template). The company says it took 18 months to go from the biological finding to a drug candidate. In the Phase 1 trial, among patients on 45 mg a day or more, there were three partial responses (tumors shrinking by at least 30%), two confirmed in endometrial cancer and one unconfirmed in liver cancer. Disease was controlled (tumors shrank or stopped growing) in 11 of 12 patients. The first data from an endometrial cancer expansion group are due in the second half of 2027. RBM39 is not a new target, though. A paper published in Science in 2017 found that an old drug, indisulam, works by degrading RBM39, and that it was safe in early trials but fewer than one patient in ten responded. REC-1245 has to show it can do what indisulam could not.
5. The AI-designed molecule Recursion inherited now needs someone else to carry it. REC-617 is a CDK7 inhibitor (CDK7 is an enzyme that helps control cell division and gene transcription). It came from Exscientia, which promoted it as precisely designed with AI. As of September 2025, one of 29 patients with advanced solid tumors had a confirmed partial response, and five more had stable disease. In October 2026 Recursion said it would look for partners to take it and another program, REC-3565, forward. Before that, the three mid-stage programs stopped or paused in May 2025 (see the table above) were all ones Recursion had taken into Phase 2 itself. Counting REC-4881, only one of the four has reached the point of planning a registrational trial.
6. Buyers have paid, but not enough to keep the company going. Roche and its Genentech unit have paid US$216 million to date, and Sanofi US$144 million. In 2026 both picked their first project from the collaborations: Genentech chose a neuroscience target, and Sanofi chose an immunology drug candidate, which Sanofi will pay to develop. That is buyers grading the work with real money. Yet together these payments roughly match one year of Recursion’s cash spending. The company has capped its 2026 cash operating costs at US$375 million, while its third-quarter revenue was only US$18.6 million against a net loss of US$119 million. The stock market fills the gap. In October 2026 it sold another US$200 million of shares, which the company says will keep it funded into 2029.
7. The shares trade at less than a quarter of the IPO price. Recursion went public in 2021 at US$18 a share, raising US$501.8 million including the underwriters’ extra shares. On 3 November 2026, after the latest REC-4881 data, the shares barely moved, trading around US$3.84.
8. What to watch. Recursion says it will give more details on the registrational trial later in 2026. The questions are which endpoint the FDA will accept and whether the trial needs a control group. Three things are due in 2027: the Sanofi candidate entering Phase 1, the REC-1245 endometrial cancer data, and whether the REC-4881 registrational trial starts on schedule.

This is the third time around

Twice before, someone has sold this industry a technology that made the front end of discovery much cheaper, with the pitch that drug discovery was now solved: genome sequencing in the 1990s, and Watson from IBM (IBM) in the 2010s. Each time, a company’s valuation rested on the idea that the front end was the whole business, and to hold up that valuation, the company had to produce a drug. Human Genome Sciences became a biotech to get one and ran straight into the cost of trials and the risk of the target. IBM’s health unit was eventually sold for far less than had been put into it. Both times the technology stayed on as a tool, bought at tool prices by companies already in the industry, and the valuation was gone. BenevolentAI, the best-known name of the current wave, has already been through it. There is more money this time, but the pitch and the business are the same.

In every row, the technology is still in use, while the company valued on it was sold, shrank or was delisted.

What each wave left behind

1. Human Genome Sciences was not hype, and the science held up. The genome, the patents and the drug were all real: belimumab was the first new lupus drug in fifty years. It just took twenty years to produce a single drug, because in 1990s drug discovery, sequencing was the cheap half, as molecule design is now. The company that patented genes was eventually bought by a company that could run trials, and its management was gone within fifteen days.
2. BenevolentAI went down the same road, only faster. It called itself a machine brain for bioscience, raised money at a valuation of two billion dollars and went public through Europe’s largest SPAC deal (a SPAC is a listed shell company that merges with a private one to take it public). When investors asked where the drugs were, it built a pipeline. When the pipeline burned through the cash, it cut the pipeline. It laid off staff round after round, announced it would go back to selling the platform, and finally delisted. The whole cycle took twelve years.
3. The technologies all survived. Genomics is now the basis for target validation (confirming that a target really is linked to the disease in people) across the industry, and nobody would start a program without it. Natural-language systems that read the literature are in use at every major drugmaker. The molecule in the next section was designed with free-energy calculations.
4. They misjudged who gets paid for a tool. The toolmaker does not own what the tool finds: Illumina (ILMN) sells sequencers, but it does not own the drugs discovered with them. If buyers can build the tool themselves, as with chemistry software, the toolmaker gets even less, because a tool cannot sell for more than it would cost the buyer to make one. Schrödinger sells FEP software (free-energy perturbation, the free-energy calculation mentioned in the previous item), yet the money it made from zasocitinib (a US$111 million distribution) came from its stake in Nimbus, not from the software. Until its first clinical readout, a discovery platform is just a tool valued as if it were a drug company.
5. So far, the only way out anyone has found is to sell the asset. Human Genome Sciences did it by getting acquired, and Nimbus by selling a subsidiary. BenevolentAI did not manage it, because by the time it wanted to sell, it had nothing anyone wanted to buy.
6. Biotech stocks have badly lagged tech. Using adjusted monthly closes, from the end of 2011 to the end of October 2026, XBI, an index fund of small and mid-sized biotech stocks, rose about 632%, while QQQ, which tracks the Nasdaq-100 and its large tech companies, rose about 1,304%, a gap of nearly 700 percentage points. Before the pandemic the two ran close: from the end of 2011 to the end of 2019, XBI rose about 343%, slightly ahead of QQQ’s 315%. Almost all of the gap opened after 2020, the years when AI advanced fastest and AI drug discovery raised the most money. Index returns reflect many things, including interest rates, acquisitions and drug-pricing policy, so not all of this can be pinned on AI, but there is no sign that AI has lifted biotech as a whole.

What computation has actually achieved

There are now two kinds of evidence. One is the aggregate: a 2024 count covering every molecule that came from an AI-native biotech and completed a clinical phase. The other is the strongest single molecule that computational design has produced. In aggregate, AI-designed molecules pass Phase 1 far more often than the historical rate, and pass Phase 2 at about the historical rate. Phase 1 asks an engineering question: is the molecule safe at the target dose, and does it behave in the body the way the chemistry predicted? Phase 2 asks a scientific question: does this target actually matter in this disease? The strongest single molecule, zasocitinib, used computation to solve a hard chemistry problem, but against a target someone else had already validated. So far, the computation has helped at the stage where programs rarely die.

The numbers behind the evidence

1. The aggregate. In 2024, a team at the consultancy BCG went through every molecule from an AI-native biotech that had completed a clinical phase. Of the 24 that finished Phase 1, 21 passed, about 88% (the report gives 80 to 90%), against a historical 40 to 65%. Of the 10 that finished Phase 2, 4 passed, about 40%, against a historical 30 to 40%. The authors’ own reading is that AI is good at producing molecules with drug-like properties (absorption, metabolism, safety and the other properties a drug needs). Comparing the midpoints of the ranges, AI-designed molecules beat the historical pass rate on the engineering question by roughly twenty to forty percentage points. On the scientific question, ten molecules are far too few to prove anything, and their pass rate looks like the historical one. The authors also list well-validated targets as one possible reason for the high Phase 1 pass rate, though they argue it is not the whole explanation.
2. What a doubled hit rate is worth to a buyer. The arithmetic is simple. If the buyer’s own screen succeeds ten percent of the time and a vendor’s succeeds twenty percent of the time, the buyer can screen twice. It only needs one hit (one compound that binds the target), and it can afford to screen, so doubling the success rate only halves its work. It is like a coupon from a food-delivery app: people on a tight budget use it, and people with money do not even look at it. That value is real, just far smaller than the business plan says. There are two exceptions. One is time: if a patent filing is waiting on the screening results, speed is worth money. The other is a target that cannot be screened, and only there does a model open up something new. A model earns its keep by getting a hit where there was none, not by raising ten percent to twenty, and nobody has shown that yet.
3. The best case: zasocitinib. In 2016, a joint team from Nimbus and Schrödinger set out to make a selective TYK2 inhibitor. TYK2 belongs to the JAK family of enzymes, and the binding sites of TYK2, JAK1, JAK2 and JAK3 are almost identical, so selectivity (hitting TYK2 without touching the others) was the whole challenge. They started with JH1, the catalytic domain (the part of the enzyme that does the actual work), using free-energy calculations to score about four thousand designs, pick 46 to synthesize, and find nine that met the project’s criteria. They then turned to JH2, a regulatory domain away from the catalytic one, and screened 2.4 million commercially available compounds computationally against a set of protein structure models. That turned up a new kind of starting molecule, which they optimized into zasocitinib. Binding the regulatory domain instead of the catalytic one is what made selectivity possible: according to Takeda, the molecule’s selectivity over the other JAK enzymes is more than a million-fold. It was a hard, expensive chemistry problem, and the computation solved it.
4. Up close, the story is less impressive. It was Bristol Myers Squibb (BMY) that validated the target and found the allosteric pocket (a binding site away from the enzyme’s active site). Nimbus came later, working on a target whose risk someone else had already removed, and making a second drug. Its method was physics-based computation, the FEP described earlier, not a generative model trained on chemistry text. Nimbus’s own president of R&D has said plainly that he would not call a compound identified back in 2020 the first AI-approved drug. And the company that supplied the software was Schrödinger, which now runs its own pipeline alongside its software business. What works here needs three things: a target someone else has already validated, a question narrow enough to be solved as a physics problem, and a wet lab (a lab that really synthesizes and tests compounds) that can afford to make and test dozens of molecules in every round.
5. What FEP calculates, and what it does not. FEP estimates the difference in binding free energy between two similar molecules binding to the same protein structure. It simulates one molecule gradually turning into the other, accounting along the way for the energy of the water, the protein and the ligand (the candidate molecule). Done well (with enough sampling and a reliable structure), it is accurate to about one kilocalorie per mole, which corresponds to roughly a fivefold difference in affinity (how tightly a molecule binds). That is enough to rank similar molecules and decide what to make next. It does not tell you whether you picked the right target. Nor does it predict whether the compound will be metabolized within an hour, block the hERG channel (a heart ion channel whose blockade carries a risk of arrhythmia), build up in the liver, or work in mice. It assumes you already know how the molecule sits in the pocket, and it quickly becomes unreliable once the chemical change is large, the protein moves, or the structure itself is wrong. It answers one narrow question well. Every other question in the cycle of designing, making and testing compounds still needs an experiment.

The knowledge that matters has never been written down

Ask a veteran medicinal chemist, an old hand, what he does in a round of lead optimization, and the answer sounds routine: look at the activity data (test results showing how strongly the molecules act on the target), pick the next batch of analogs (molecules with small changes to their structure), send them off to be made, wait three or four weeks, and look again. But ask what separates an old hand from a novice, and the answer is no longer routine.

The difference lies in a few judgments: what in the data is signal and what is noise, which paths were tried and abandoned before, which few molecules to make next, and when to stop. None of this is in the literature, because the literature records what worked. And the published record is not reliable enough to serve as training data anyway. When Amgen (AMGN) and Bayer (BAYN.DE) tried to reproduce published results before building programs on them, most of those results did not fully hold up. The science benchmarks reported in model launches do not measure these judgments either. They test knowledge recall and machine-learning engineering (MLE), as the items below explain.

The first three rows are what model launches report as scientific ability, and every one of them is graded against an answer that existed in advance. The last row has no answer until a clinical trial reads out.

What the record leaves out

1. What an old hand actually decides. He has to judge which differences in a structure–activity table (a table of how changes to a molecule’s structure changed its activity) are signal and which are assay noise. Test the same compound twice and the results can differ twofold, so a twofold improvement proves nothing. Knowing when a change is real is a knack nobody has written down. He remembers that a scaffold (the core structure of a molecule) was dropped at his previous company because of a toxicity that was never published. Working within a synthesis budget, he picks thirty analogs out of a thousand plausible ones, knowing he will see results only once a month. And he has to decide when a series (a family of molecules built on the same scaffold) is dead. Of everything he does, this decision costs the most and rests on the least evidence.
2. Failures are not published, and the published record is unreliable. Failures stay in internal company databases, the most tightly guarded assets in the industry, so a model trained on the public record learns mostly from attempts that worked. Amgen’s hematology and oncology team spent years trying to reproduce 53 landmark cancer papers before betting programs on them, and confirmed six. A Bayer team reviewed 67 of its own in-house projects and found that the published data fully matched its own results in only about 20 to 25% of them. A third replication effort, the Reproducibility Project: Cancer Biology, repeated 50 experiments from 23 high-impact papers. For results that were positive in the original papers, the replicated effect sizes (how big a change the experiment measured) were 85% smaller at the median. Of the effects it could evaluate, fewer than half (46%) met most of its criteria for success. Part of the reason is the incentive: every paper needs a gene nobody has claimed yet, so search almost any gene together with the word tumor and you will find a paper linking the two. So a large share of a big drugmaker’s internal research is filtering: before a program is bet on, its scientists repeat the published findings and throw out the ones that do not hold.
3. Who sets the test. Drug discovery cannot pick its own judge. First the human body judges, then the FDA. Nobody can define a benchmark, win on it and declare the problem solved, the way computer vision and language modeling do every day. Outside grading is slow and costly, and it also makes it hard for the field to fool itself. To train the kind of judgment described above, you would need boring, specific records from inside real programs: the situation at the time, the decision made and what happened next. For each series, such a record would show what was known, which thirty compounds were picked, what data came back and where the series died. It would include the failures (which are most of it) and carry timestamps, so that a model is scored only on what could have been known at the time. Data like this exists only in fragments inside a few dozen companies, and it will not be pooled, because the record of failures is itself a competitive asset. This is an institutional problem, and more computing power does nothing about it.
4. The science benchmarks in model launches are Kaggle competitions. A Kaggle competition gives everyone the same fixed dataset and a question, and the organizer holds the answer key. You submit, you get a score right away, and you can submit again. The science benchmarks reported in model launches come in roughly three kinds (see the table above), and each has the same parts: a fixed dataset, a question, and an answer written in advance. One widely used agent benchmark consists of 75 Kaggle competitions, scored against Kaggle’s own leaderboards. Some newer tests use real data and real research workflows, which is progress, but every question still comes from a result someone already has. A high score shows a model can find answers that already exist, and a drug program pays only for the ones that do not.
5. Critics inside the field agree. The leaderboards most used within drug discovery are built the same way. A researcher who has spent many years in industry cheminformatics (the use of computers to work with chemical data) argues that the most widely used leaderboards are too flawed to draw conclusions from. In December 2025, a paper with 56 authors argued that mainstream science benchmarks test knowledge out of context and overlook what drives discovery: step-by-step reasoning, proposing hypotheses and interpreting observations.
6. What a real AI-for-science result looks like. A real result is a prediction about something nobody knows yet, checked afterwards by someone else. In drug discovery, that means predicting that a molecule will bind before it has been made, or predicting that a target matters before the clinical trial, and then letting a lab or a trial grade the prediction. Mathematics already has results like this: a model raised the known lower bound on the kissing number in 11 dimensions (the largest number of equal-sized spheres that can touch a central sphere of the same size at once) from 592 to 593. Nobody had found that arrangement before, and anyone can check it. A result in mathematics can be checked by other mathematicians. A result in drug discovery can only be checked in a lab or in patients, which takes months or even years and real money. So model companies report tests that can be scored the same day.

You are automating the cheap half

In practice, AI drug discovery now means a model that proposes molecules. That work falls within the nine percent of the cost that comes before the first human dose, and programs usually do not die there. The best existing version of it needed a target someone else had already validated, and the best-funded platform, Isomorphic, got just US$82.5 million in cash from its first two partnerships, while a single molecule that had been tested in people sold for US$4 billion.

By the industry’s own economics, the decisions that cost the most money are these: which validated target to follow, whether to buy an outside drug candidate and how much to pay at signing, how to design the trial, and above all when to stop, whether that means stopping a series or a whole program. Each of these decisions is made by a few senior people at a weekly meeting, yet nobody writes down what they knew and what they expected to happen before the answer comes in, so the call cannot be checked against the outcome later. These decisions are rare and expensive, and nobody checks them afterwards.

Neither drug people point to will settle the question soon. The company that designed zasocitinib will not call it the first AI-approved drug. And rentosertib from Insilico Medicine (a lung-fibrosis drug, and the molecule most often described as AI-discovered) has a Phase 3 trial run entirely in China, with results not due until December 2029.

The missing record, and three tests

1. Parts of these decisions have already been modeled and audited. The dose is set by exposure–response analysis (how the drug’s effect changes with the dose and its level in the blood). AstraZeneca has published a scorecard comparing its results before and after it overhauled how it decides which programs to advance and which to stop: the share of candidates that went from selection to the end of Phase 3 rose from 4% to 19%. But the reasoning behind individual decisions is not recorded anywhere. That is no knock on the people making them, whose experience is real and usually beats the alternatives.
2. Why the record does not exist. Hospital medicine explains this best, and portfolio meetings (the meetings where a company decides which programs to fund) work the same way. In the clinic, the real basis of a decision is liability: whoever pays if something happens to the patient gets to decide. And the decision rests on a judgment nobody has quantified, such as how bad this kidney is, how high this blood pressure is, and how much this patient’s age should adjust the formula. So the decision comes out as “the other center doses at 200, we dose at 150”, and the reasoning is never written down. This is street smart, not book smart, and it works. Portfolio meetings run on the same logic: whoever takes the blame if a program fails makes the call, and a written prediction could later be held against whoever wrote it, so it never gets written. Yet the moment someone writes a prediction down before the outcome, this kind of knowledge becomes testable.
3. These decisions suit a model well. They are rare and expensive, so being a little less wrong each time is worth a lot of money, and nobody needs an answer in milliseconds. The decisions themselves are already recorded in the minutes, because they are made at a meeting with a fixed agenda. What is not written down is what was known when each decision was made, and recording that only takes a change to the format of the minutes.
4. The hard part is scoring a decision to stop. You only ever see the outcome of the decision you actually made, and a series you stopped will never produce data. So a look back after the fact cannot score a decision to stop. Two things can be scored. One is predictions on the series that were continued, registered before the results came in. The other is the judgment of whichever side was followed when the model and the humans disagreed. Settling it takes forecasts written down and dated in advance, not a database.
5. Nobody has published this test. Some companies sell pieces of it: trial-outcome prediction, digital control arms (using historical data in place of real control patients), and portfolio analytics. But none has published a time-stamped record that puts the model’s calls next to the humans’ calls on the same set of decisions, registered before the data read out. Meanwhile, the money and the talent have poured into molecule generators.
6. The two test cases will not settle it. Zasocitinib’s regulatory outcome arrives in the first quarter of 2027, but the company that designed it does not call it an AI drug. Rentosertib’s trial runs at 47 centers in China, and to win US approval it first has to get past the single-country data problem described earlier. That leaves rentosertib as the one real test of AI molecule design, and the result will not arrive until 2029. Even if it succeeds, it answers a question about the nine percent of the budget.
7. A 2026 Perspective reached the same conclusion. A Perspective published in Nature Reviews Drug Discovery in 2026, whose authors include the researcher who coined Eroom’s law (the observation that the cost of each new drug has kept rising for decades instead of falling), calls AI’s clinical impact “disappointingly limited”. AI-designed drugs still hit the same Phase 2 wall as everything else. An improvement of a given size is worth most when it raises the clinical success rate, above all in Phase 2. On this view, a model is worth roughly how much it raises the clinical success rate, multiplied by how many programs it can be used on. A molecule can bind its receptor tightly, kill cells in a dish and shrink tumors in mice, and still do nothing in people, or help only a small fraction of them. Only a molecule that shows a statistically significant effect in a Phase 2 trial has earned the right to be called a drug. But academia chases impact factors and grants, so important negative results rarely get published, and companies built around AI prefer a big story. The statistician John Tukey put it this way: an approximate answer to the right question is worth far more than an exact answer to the wrong one.
8. What would overturn this conclusion. First, a platform company reaches a Phase 3 readout on an asset it still owns, without turning itself into a conventional biotech along the way to raise the money. That would show the pipeline dilemma can be avoided and the structure described here is wrong. Second, a discovery platform with no clinical assets gets an upfront payment of US$500 million or more. An equity round does not count. It would show that buyers have started to price the method itself, and the fifty-to-one ratio above would be out of date. Third, someone publishes a time-stamped scorecard from inside real programs showing that, on the same budget and with failures counted, a model beats experienced people at the decisions listed above. It is the cheapest of the three: a company that already holds these meetings could register every prediction before the results come back and publish the scorecard when the program ends, at no extra experimental cost. Still, nobody has published one.

Papers are unreliable, and reliable data lives in the wet lab

Much of AI for science today means having a model read papers, patents and knowledge bases, then distill patterns and propose hypotheses from that text. The trouble is that the text itself is riddled with holes. In the life sciences, a large share of published results cannot be reproduced, a good deal is padding, and some is outright fraud. Most knowledge bases are in turn extracted from those same papers. However much a model reads and however well it summarizes, what it learns is that someone wrote this down, not that it is true. Reliable data lives in the wet lab: measurements made with one protocol and one batch of reagents, with the negative results kept and every result time-stamped. Data like that sits mostly with drugmakers and the few companies that run their own experiments, and that second group is what the next section says is worth backing.

How unreliable the literature is

1. Fraud has become an industry. More than 10,000 research papers were retracted in 2023, a record, and more than 8,000 of them came from journals of a single publisher, Hindawi. The retracted Hindawi papers had already been cited more than 35,000 times. A 2025 study in PNAS found that the output of suspected paper mills (operations that mass-produce fake papers and sell authorship) is doubling every year and a half, while the scientific literature as a whole doubles only every fifteen years.
2. Retractions catch only a small part of the problem. A screen of 20,621 biomedical papers found problematic duplicated images in 3.8% of them, at least half suggesting deliberate manipulation, and that study checked only one kind of error. An economic analysis estimated that more than half of preclinical research is irreproducible, costing about US$28 billion a year in the United States alone. The Amgen and Bayer replication efforts described earlier point the same way: most published findings do not fully hold up when someone tries to build a program on them.
3. One paper can lead a field astray for nearly two decades. A 2006 Nature paper reported that a specific amyloid assembly, Aβ*56, impairs memory, and it became one of the most cited papers in Alzheimer’s research, steering a great deal of later work and funding. In June 2024 it was retracted after investigators found manipulated images, making it the most cited paper ever retracted. Any model that learned the field from its literature learned that result as fact.
4. A model that reads the literature learns its errors too. Text cannot tell a model which results were fabricated, which were never replicated, and which experiments failed and were never written up. A knowledge graph built from papers turns a claim into an edge, and a claim cited a hundred times looks a hundred times more certain, even when every citation traces back to the same flawed experiment. More reading only adds more copies of the mistake.
5. Why wet-lab data can be trusted. Wet-lab data is not free of noise either: as noted earlier, the same compound tested twice can differ twofold. The difference is that the noise can be measured. The protocol is fixed, the controls run alongside, the failures are recorded, and any result can be run again. That is also why Lilly’s TuneLab and J&J’s antibody design challenge, both covered in the next section, insist on data measured in their own labs.

What is worth investing in

Drugmakers are reluctant to hand over much of their internal data. They would rather partner with others and build their own models on their own data. Since outsiders cannot get drugmakers’ data, the companies worth backing are those that can generate their own gold-standard data (the most trusted experimental data, the kind used as a benchmark for everything else) quickly and cheaply, in a field where drugs can actually be made.

That rules out a large part of the field. Companies selling AI as a software service, or building general-purpose biology models, will have a hard time, because frontier model companies such as Anthropic already work with big drugmakers, and drugmakers can get those services straight from them. Proprietary data counts for more than a proprietary model, because a company can iterate on its data again and again and eventually turn it into drugs of its own. Put another way, the company worth backing is one graded by its own wet lab and, in the end, by its own clinical trials. Lilly’s TuneLab, discussed below, is a drugmaker placing this bet itself.

What practice shows

1. AI can already help run experiments, but it cannot yet replace judgment. Some companies already use AI to design experiments and decide the order to run them in, and the boost is considerable. When it comes to decisions and judgment, though, AI still falls short. One company running a randomized Phase 2 trial worked closely with Anthropic in the months before database lock (the point after which trial data can no longer be changed).
2. The hardest part of biotech is translating academic findings, and translation is at heart a reproducibility problem. Most academic results cannot be carried straight into a drug program, which is the same problem as the failure to reproduce published results described earlier. One remedy is to work closely with academic founders who are willing to keep working on a finding. Their labs keep refining it long after the paper, which is closer to engineering, so what reaches the company is far more mature than what a typical lab hands over.
3. A company founded this way starts much further along. By the time the technology moves into a company founded by the professor who runs the lab or by one of the lab’s students, it is still pre-IND (the IND is the application to start human trials), but it already has months of research and several rounds of lead optimization behind it. It is no longer just the version in the paper, and building the company becomes much easier.
4. Lilly’s TuneLab is a drugmaker getting into the data business itself. In September 2025 Lilly launched TuneLab, with models built on its own drug disposition (how a drug is absorbed, distributed, metabolized and excreted), safety and preclinical data covering hundreds of thousands of molecules, backed by more than US$1 billion in research. Biotechs can use the models, and in return they contribute their own data to improve them through federated learning, which trains a model on each party’s data without moving or exposing it. Distribution has been quick: TuneLab is already plugged into Benchling, whose customers include more than 1,300 biotechs, and into CDD Vault, while Charles River (CRL) provides testing for participating companies. TuneLab itself also sits under Lilly’s Catalyze360 program, right next to Lilly’s venture arm. So Lilly keeps its own data, benefits from everyone else’s without ever seeing it, and gets an early view of which biotechs are worth investing in or buying. That last part may be worth more than the models themselves.
5. TuneLab still works on the cheap half. By its own account, the models help participating companies weed out candidate molecules faster. That is useful, and it is graded by real experiments, but it falls within the nine percent before the first human dose and does not touch the question of whether a target matters in people. There are two further risks. A biotech feeding its data to a company that may well end up acquiring it will sooner or later hold some of that data back. And there is no public, pre-registered evidence yet on how much federated learning gains across labs whose experimental methods differ. There is also a side effect: a drugmaker that opens models trained on a billion dollars of its own data to biotechs, asking for data instead of money, puts direct pricing pressure on companies that sell property-prediction models. Frontier model companies such as Anthropic complement TuneLab: the language model runs the workflow and the property model scores the molecules, but neither makes the decisions that cost the most money.
6. Johnson & Johnson shows what it looks like when the buyer sets the test. Investors now say they want companies that can run their own design, make and test loop: companies that build AI into their own systems and, above all, generate their own data. Companies that build foundation models on public datasets are everywhere and cannot set themselves apart. In October 2025, J&J (JNJ) launched a challenge in computational antibody design. According to people in the industry, each of the dozen or so antibody design companies taking part designed about 100 sequences for each of three targets, J&J is testing them in its own labs, and the winners will get the chance to work with J&J. The result is that J&J now holds head-to-head data on a dozen or so of the world’s leading AI protein design companies, on the same targets. The companies have never had a benchmark like this, and all of them want the data, which J&J keeps.
7. Data has to be distinctive, not just big. At a recent panel of AI biotech founders, a few practical points came up. First, a model that understands DNA does not understand disease. The chain from DNA to RNA, proteins, cells and finally patient outcomes is long, and the data that exists is not enough to predict it end to end, so it is more practical to model each layer and then connect them. In the same way, understanding the variants that cause disease is not the same as understanding the disease. Second, an unusual biological phenotype can teach more than a pile of samples. One company, for example, studies how the hearts of hibernating animals repair themselves after injury, looking for new leads in heart failure. Third, the goal for a model is to predict how people respond to treatment, and predicting sequences better is only a step toward that. Fourth, drugmakers may come to want drug assets more than pure technology platforms. AI should also explore chemistry that traditional methods cannot reach, against targets that already have clinical validation, instead of squeezing more efficiency out of known chemistry. Finally, some think the wave of protein language models has matured and the next opportunity may lie in small molecules, but in the end the value of any model or dataset has to be proven by commercial results.

Who is the hero, and who is the gimmick

There is one test running through all of this: who grades the results? A hero is graded from outside, by patients in a clinical trial, by a buyer who pays cash at signing, or by a wet lab that actually makes and tests the molecule. A gimmick grades itself, with a headline deal value, a pipeline slide in a business plan, an AI-discovered label, or a benchmark of its own choosing. By this test, what has worked in AI drug discovery is physics and chemistry computation aimed at specific problems and checked in the lab, and the value lies with the people and companies that pay for the trials and decide when to stop. Much of what has been sold under the name of AI drug discovery is just a label hung on the outside of that work.

Every hero in the table is graded by a clinical trial, a paying buyer or a lab, and every gimmick grades itself.

How do you spot a real hero? Look at just two things: real substance (data and molecules tested in a lab) and real money (cash paid at signing, or someone paying for the trial themselves). Ignore everything else.