Newsletter · 2026-10-01 · Ashutosh Agarwal

Scientists Push Back on Anthropic's AI Discovery as Podcasts Say Data Beats the Model - AI Drug Discovery Weekly - Week of October 1, 2026

AI Drug Discovery Weekly podcast synthesis for the week of September 24 to October 1, 2026. Scientists questioned Anthropic's claimed enzyme discovery as its cost and IPO context emerged, Deep Genomics showed an RNAi model cutting a 2,000-molecule screen to 20, Ginkgo pitched fully automated labs, and the podcasts converged on proprietary data being worth more than the model, while SDGR hit a new high and LLY slipped despite strong obesity data.

AI Drug Discovery Weekly

Week of October 1, 2026: Scientists Push Back on Anthropic's AI Discovery as Podcasts Say Data Beats the Model

Issue #14, Week of September 24 – October 1, 2026.

TL;DR

• The Anthropic biology story got its reality check. Last week Anthropic said its Claude AI found a new CRISPR-like enzyme system in 21 hours. This week the podcasts filled in the details and the doubts. About 1,000 Claude agents used roughly 210 million tokens, which is about $10,000 of computing. Working scientists pointed out that finding an odd gene cluster is "often the easy part." Working out what it actually does is the hard part, and Anthropic has not done that yet. (The AI Daily Brief, Sep 24)
• The defense came from All-In. David Friedberg said the lab is a low-risk BSL-1/BSL-2 benchtop lab, built to check whether the AI's predictions are right. He also said Anthropic has "the best life sciences models." (All-In, Sep 26)
• Deep Genomics had the clearest AI result of the week. Its DeepRNAi model cut a screen from 2,000 candidate molecules to 20. It did this on a gene it had never seen, so the model "has actually learned the rules." The COO was candid that this was a retrospective test, run on past data, and that the problem could still be brute-forced. (Between Two COO's, Sep 30)
• Data beats models. Ben Lamm of Colossal said "the data set is more valuable than the model" (Moonshots #297, Sep 30). Deep Genomics says its partners' data sets were dwarfed "by over 10x." A Telltales panelist suggested AI may simply let drug companies compete away their profits, unless they hold proprietary data like Lilly's on GLP-1 drugs.
• Ginkgo's Jason Kelly made the biggest claim on lab automation. Pharma labs run "at best 40 hours a week" out of 168. In one project, GPT-5 ran 30,000 experiments a week on Ginkgo's robot lab and beat the state of the art by 40%. (Grow Everything Biotech #200, Sep 25)
• Stocks: Schrodinger (SDGR) +3.4% to $30.46. It hit a new 52-week high of $32.45 on more than 3x normal volume, again with no company news. Recursion (RXRX) +4.6% to $4.07, also on no news. Eli Lilly (LLY) -2.7% to $1,151.42, despite a week of strong obesity-drug data. None of the three made AI-specific news.

What's new

1. Anthropic's enzyme "discovery": the numbers, and the doubters

Last week's lead story was Anthropic's claim that Claude found a new enzyme system in the DNA of bacteriophages (viruses that infect bacteria). This week the podcasts supplied the price tag and the pushback.

What it took. On The AI Daily Brief (Sep 24), host Nathaniel Whittemore reported that Anthropic says the discovery took "a thousand Claude agents, 21 hours, and around 210 million tokens, so somewhere on the order of $10,000 of compute." (Tokens are the chunks of text an AI model reads and writes, and are how its usage is billed.) Tech Brew Ride Home (Sep 24) put the count at nearly 950 agents. It added that Anthropic itself admits the practical uses are unclear and the announcement is early.

What Anthropic is claiming. The system looks like CRISPR, the bacterial gene-editing tool, because it contains a repeated array of genetic sequences. Dario Amodei, Anthropic's CEO, who has a PhD in biophysics, framed it as part of a trend. He pointed to math: in 2023 AI models "struggled to do math at the level of an average high school student," and by late 2026 they are "beginning to solve the top few open problems in all of mathematics. We believe AI for biology is on a similar exponential trend." He also said Claude cannot yet run lab equipment on its own, though that is the goal.

The pushback was specific, and worth reading. The AI Daily Brief quoted two scientists:

• Ravid Schwartz-Ziv was blunt. If a PhD student told his advisor "we found an interesting system, but we still don't know what it does" and said he was ready to graduate, "Bill would have kicked him out of the room. But somehow, when the IPO is around the corner, this becomes AI is starting to drive biological discovery."
• Lucas Harrington did the same kind of genome-mining work for his PhD. His view: "Finding a weird cluster of genes and repeats is often the easy part. The hard part, and where the real discoveries come from, is figuring out what the system actually does." Anthropic has shown the array produces RNA molecules, but not what those RNAs do or whether the system can be programmed the way CRISPR can. He said he is "genuinely rooting for" frontier AI labs in biology, but wants them to "set the bar high now."

The host's summary: "It's very cool and people should be excited. But it's also very, very, very preliminary."

The defense. On All-In (Sep 26), David Friedberg explained what the lab is for. It is BSL-1/BSL-2, the lowest biosafety levels, used for harmless material, and "there's hundreds of labs" like it in the Bay Area. Most of the work is "taking DNA, putting it in a bacteria to make a protein, and then measuring what that protein does." The point is "experimental proof" that the software's predictions hold up, the same way AlphaFold's predicted protein shapes had to be checked against real proteins. His overall verdict: Anthropic has "the best life sciences models," and this is "the fast track to basically seeing if these models can be used by therapeutic companies, by pharma companies."

Why the IPO matters here. Schwartz-Ziv's IPO point is not just a jab. Much of the week's AI podcasting was about Anthropic's leaked IPO paperwork. Tech Brew Ride Home (Sep 29) reported that it shows a $42 billion net loss on $4.6 billion of revenue, with 25% of revenue from two customers. Reuters World News (Sep 29) reported the listing is planned for after the November midterms. With a listing that size on the horizon, readers should treat any Anthropic science claim the way they would treat a biotech press release before a fundraise: interesting, and still to be proven.

Another data point on Anthropic and pharma. On House Calls (Sep 30), Cain Brothers managing director Jason Moran said Anthropic bought a stealth biotech startup, Coefficient Bio, for $400 million in April. He called it Amodei having "put his money where his mouth is." We have not independently confirmed the deal terms, so treat them as the speaker's claim.

2. Deep Genomics: the most concrete AI result of the week

The best episode for anyone who wants to know whether AI drug discovery works was Between Two COO's with Tom Masterson, COO of Deep Genomics (Sep 30).

The problem. Deep Genomics works on siRNA (small interfering RNA) drugs, the class made famous by Alnylam. An siRNA is a fragment "a couple dozen base pairs long" that sticks to a faulty RNA in the cell and gets it destroyed. Masterson said hitting the target is "actually the easy part." The hard part is that the fragment "can be hitting all sorts of different RNAs all over the cell," which causes side effects that often only show up late in development.

The result. The company's DeepRNAi model predicts these off-target hits. Instead of screening 2,000 candidate molecules, "just screen 20 of them, put them into rats or mice or whatever and start your pipeline there." The key detail: "the model had never seen that gene... yet it was still able to make the predictions. What that tells us is that the model has actually learned the rules, not that it's memorized the answers to some test." Two separate pharma companies set the challenge because, before AI, the problem was "just intractable."

The honest caveats. Masterson made them himself:

• It was a retrospective study, a test run on past data. Scientists will say "that's cool, but you'll never convince them that something retrospective is as good as it discovered this and it went out in the world and succeeded."
• The problem could technically be brute-forced: "You can just run 10 times as many molecules and you'll find good ones eventually." The really exciting moment, he said, "is when we start doing things that genuinely nobody else can do, even with brute force." He doesn't think they are there yet, but "I wouldn't bet against us."
• The host noted that, to his knowledge, there is still no FDA-approved drug discovered by AI.

The strategy shift. Deep Genomics once had about 40 task-specific AI models. It scrapped that approach for one biological foundation model, a general model of biology that works like "an AI textbook for biology," with smaller application models built on top. Masterson said the returns from the 40 models "weren't staggering."

Why its data is the advantage. Masterson's argument is that pharma's historical data is "very, very biased" for training AI. Pharma companies spent decades designing molecules meant to be safe, so their records hold very few examples of failure. Deep Genomics deliberately designs "test molecules that are intentionally toxic so that the model learns what toxic looks like." The data needs "probably more noise than good signals." He said partners' data sets were dwarfed by Deep Genomics' "by over 10x." He also described two different clocks: drug developers "think in 10 years," while machine-learning engineers think about "the simulation this afternoon."

3. Ginkgo bets on fully automated labs

On Grow Everything Biotech #200 (Sep 25), Ginkgo Bioworks CEO Jason Kelly described Ginkgo's turn toward "lights out," fully autonomous labs.

• Pharma's idle labs. Kelly asked heads of R&D at pharma companies how their lab spending splits between automated "work cells" and ordinary lab benches. The answer: "it's like 97% lab benches." Those benches run "at best 40 hours a week. We have 168 hours in a week." He estimated pharma companies spend "one to 3 billion a year" each on labs (not counting clinical trials), and the NIH spends $40 billion at universities on the same kind of work.
• The business model now. Kelly said Ginkgo used to be "about royalties and research partnerships." Now it is closer to Thermo Fisher: it sells and installs autonomous lab equipment. Most of its customers are now in pharma.
• AI plus robots. In one project with OpenAI, "GPT-5 got 30,000 experiments a week on our robot lab" and beat the "state of the art in cell-free protein synthesis by 40%." His summary: the AI model designs the experiment, the autonomous lab runs it, and "that's what science is."
• Government demand. Ginkgo announced it is building autonomous labs for MIT, Caltech, Northwestern and the University of Maryland through the NSF/White House "Genesis mission" program.

Moonshots #297 (Sep 30) linked the two stories. Ben Lamm of Colossal Biosciences mentioned Anthropic's announcement "a couple days ago" and said "you have Ginkgo and others that are now going to test it." His view: "You're still going to need to have a wet lab experiment to test and validate those," and AI's biggest near-term role will be "running simulation design across a myriad of different experiments and helping creatively come up with the next experiment" for "the next, you know, three to five years."

4. DeepMind builds a safety tag for AI-designed proteins

On Google DeepMind: The Podcast (Oct 1), DeepMind described SynthID Bio, a proof-of-concept system for watermarking AI-generated biology. It has two parts:

• BioStructure builds on AlphaFold3. It hides a watermark in a predicted protein structure by slightly shifting atom positions, changing "the distance between two atoms... or the angle between two atoms."
• BioSequence hides the watermark in a protein's amino-acid sequence by swapping individual amino acids without changing what the protein does.

Why it matters. DNA synthesis companies, which print DNA to order, screen orders against a database of known dangerous sequences. The worry is that AI can design something "very different than something that's present in that pathogen database but might fold into the thing that is of concern." A watermark lets a synthesis company tell whether a sequence came from a trusted AI system with safety guardrails built in.

Two other podcasts covered the same risk. A guest on The Cognitive Revolution (Sep 25) said "AI makes it easier for a bad actor to create a bioweapon or to synthesize a pandemic virus." New Books in Science, Technology, and Society (Sep 27) described AI acting as an "expert advisor" that could speed up pathogen development.

5. Smaller AI biotechs: real speed-ups and practical claims

• Congruence Therapeutics (Montreal). On Disruptors (Sep 29), CEO Clarissa Desjardins said machine learning and physics-based tools cut the time from target to clinic from "about six to seven years" to "three to four." The company now synthesizes "two to five hundred molecules" to find a good one, down from "two to five thousand." One drug is in the clinic and two more go in next year. She was skeptical of the big "digitize nature" startups: they cost "hundreds of millions of dollars" and "the jury's still out on whether they've been successful." Her own approach is narrower: "We're not trying to digitize biology yet."
• China is not ahead on AI. On Sinica (Sep 29), the host noted that Chinese companies signed about 186 out-licensing deals last year with a disclosed value of around $137 billion. That is roughly half of global out-licensing value and close to 10 times the 2021 figure. Lilly is among the buyers. Dr. Ruby Wang, author of China Cure, said: "China's AI-enabled drug factors aren't superior to the West... It's not doing it better so far." China's speed comes from people and process, not better AI.
• Data plumbing. Data in Biotech (Sep 30) covered a less glamorous problem. Flow cytometry, a common way of counting and sorting cells, still depends on scientists drawing "gates" by hand, which they described as "a time-consuming, subjective process." The guests from Dotmatics and a contract research lab want standard cell labels so AI can actually read the data.

The debate

This week's question: if every drug company gets the same AI, who actually wins?

Camp 1: the data owners win. This was the clear majority view.

• Ben Lamm: "I would make the argument that... the data set is more valuable than the model." (Moonshots #297, Sep 30)
• Tom Masterson (Deep Genomics): data that is "rigorously curated" and "causal in nature" is "the gasoline that fuels the engine." Pharma's own archives are biased toward molecules that worked, so they are poor training material. (Between Two COO's, Sep 30)
• This matches last week's DrugBank point that "the bottleneck was never the model."

Camp 2: AI gets copied, and the profits get competed away. On Telltales (Sep 30), one panelist made the cold-water case: "every pharmaceutical company will theoretically have access to this... everybody's cost of getting a drug to market will go down." Unless a company has "proprietary data or... something proprietary," the drug companies "just compete away the profits and the consumer wins." He thought that was more likely than one company crushing the rest. A co-host pushed back with Lilly: "you'd be hard-pressed to say that you have a better knowledge base to design molecules than Lilly if you're going to create a new diabetes controlling or weight loss drug." The same panel noted that language models mostly help with the paperwork, such as regulatory filings. The early science is "more of a physics engine," and the panel noted we have "gotten very good at simulating physics," while biology is a tougher problem.

Camp 3: watch the science, not the slogans. BioCentury This Week (Sep 30), recorded live at Grand Rounds Europe in Amsterdam, gave the measured view. On autonomous labs, a panelist said you can automate machines, but without building in "the knowledge of all your technicians on how to troubleshoot when it fails... you cannot really scale. You can't get to 24/7." The bigger missing piece is getting lab output into a form "a model can interpret and then make a decision about what experiment to do next and close that loop." The panel said that "is not here yet." The panelists also warned that even though AI is making discovery "much cheaper, much faster," the field risks sliding back into "high-throughput screening just with a different technology," meaning testing huge numbers of molecules by brute force instead of reasoning from biology. They pointed to David Baker's keynote as the better model of AI paired with a hypothesis.

Our read. The Anthropic episode and the Deep Genomics episode tell the same story. AI is very good at narrowing down options, whether that is 2,000 molecules down to 20 or a genome down to one odd gene cluster. It is not yet good at the step that creates value: proving what something does in living biology. So far, this week's evidence says the companies that win will own clean, causal data and a fast lab to test ideas. The companies that only own the model look less safe.

Stocks in play

Prices are FactSet closing prices for Oct 1, 2026, compared with the Sep 24 close in last week's issue.

Recursion (RXRX): up, quietly. There was no new company news in the 7-day news feed. The stock added to last week's bounce after the expanded Tempus AI license, which also covered Recursion's RNA foundation model, TxFM. It is still about 43% below its 52-week high. The next real event is the TUPELO trial data for REC-4881 in the inherited colon-polyp condition FAP, due Nov 2, 2026.

Schrodinger (SDGR): new high, still no news. SDGR traded as high as $32.45 intraday on Oct 1, a new 52-week high. Volume was about 5.1 million shares against a 1.5 million average, more than three times normal. The stock then closed down 1.9% on the day. We again found no company announcement, partnership or rating change in the newswires for the week. Last week we noted short interest of roughly 15% of shares available to trade. Momentum buying and short sellers covering their bets remain the most likely explanation, not a fundamental change. The fundamentals have not moved since Q2: annual contract value (ACV, the yearly value of software contracts) up 27% to $29.6M, and full-year 2026 ACV guidance of $218–228M.

Eli Lilly (LLY): strong data, softer stock, and no AI news. It was a very busy week, all on drugs rather than AI (news reports, Sep 28 – Oct 1):

• Foundayo (oral GLP-1 pill) met its main goal in a Phase 3 heart-safety trial against insulin glargine. At 52 weeks, weight fell 8.8% on Foundayo against a 1.7% gain on insulin, with a 53% lower risk of cardiovascular death and a 57% lower risk of death from any cause. The results were published in The Lancet (Oct 1).
• EloraTZP (eloralintide plus tirzepatide) met all goals in a Phase 2b trial of 367 adults at 48 weeks. Phase 3 starts by the end of 2026 (Sep 30). UBS said the newer obesity drugs can deliver more than 20% weight loss and kept its Buy rating with a $1,425 target.
• Other news: an FDA approval for Olumiant in teenagers with severe alopecia areata (Sep 28), and positive Phase 3b data for Ebglyss in hand and foot eczema (Sep 30).
• JPMorgan raised its target to $1,500 from $1,400 ahead of Q3 results on Oct 29. FactSet's average target is $1,363.03.
• CEO David Ricks told Bloomberg Lilly plans larger deals in new areas such as infectious disease, women's health and psychiatry (Sep 28).
• Negatives: Lilly and Foghorn Therapeutics dropped their SMARCA2 programs, and Foghorn is cutting 40% of its staff (Oct 1). A jury awarded Nektar $90 million against Lilly, far below the roughly $1 billion Nektar had sought (Sep 27).
• TuneLab: no new partner or financial disclosure for a third straight week. On podcasts, Lilly's AI effort came up only in passing. House Calls mentioned the NVIDIA supercomputer, and Telltales described Lilly's GLP-1 data as its real moat.

Read-throughs

• For companies that supply lab tools and services: Kelly's "97% lab benches, 40 hours a week" point and the Genesis-mission university labs suggest automated-lab spending is still very early. The House Calls banker said pharma lab data typically runs across "20 or more systems." He named implementation firms such as Asterix and Slipstream as winners, and said biotech fundraising has been "more robust than it's been since '21," a tailwind for pharma services and deal-making.
• For Recursion and Tempus: this week's "data beats models" consensus supports both companies' pitch that their large, owned datasets are the asset. Recursion's TxFM license is the first real test of whether that data can be licensed for money.
• For Schrodinger: Congruence's "two to five thousand" down to "two to five hundred" molecules shows physics-based simulation plus machine learning working in practice. That supports Schrodinger's approach, but it also shows how many smaller companies now do the same thing.
• For gene-editing and RNA drug companies: if Anthropic's enzyme proves to be programmable, it is a possible new tool. That is still a big "if." Deep Genomics' off-target prediction for siRNA is directly relevant to the Alnylam-style RNA drug field.
• For biosecurity: DeepMind's watermarking work and AI-enabled bioweapon worries suggest DNA-synthesis screening rules may tighten. That would affect companies like Twist Bioscience, which also made last week's issue.
• For China-exposed big pharma: Wang's view that China's AI is not superior means the China out-licensing wave (about $137 billion last year) is about cost and speed, not technology. Lilly is a repeat buyer.

What changed vs last week

• Anthropic: from "breakthrough" to "prove it." Last week we reported the 21-hour discovery mostly in Anthropic's own words. This week brought the cost (about $10,000 of computing, roughly 1,000 agents), named scientific critics, and the IPO-timing angle. Still no peer review and no identified function.
• The debate moved from "how fast" to "who captures the value." Last week it was BMS's Robert Plenge on how much to invest. This week the podcasts converged on data ownership, plus a real counter-argument (Telltales) that AI could commoditize drug discovery.
• New voices, more concrete results: Deep Genomics (2,000 to 20 molecules) and Congruence (about 50% faster to the clinic) replaced last week's more abstract frameworks.
• Ginkgo arrived as the hardware story to go with Anthropic's software story. Twist was last week's "picks and shovels" name.
• Stocks reversed order. Last week RXRX led (+9.9%) and SDGR lagged (-2.6%). This week RXRX +4.6% and SDGR +3.4% with a new high, while LLY fell from +2.8% to -2.7% despite strong data.