Newsletter · · Ashutosh Agarwal

Anthropic's Claude Finds an Enzyme in 21 Hours as Pharma Weighs the Frenemy Risk - AI Drug Discovery Weekly - Week of September 24, 2026

AI Drug Discovery Weekly podcast synthesis for the week of September 17–24, 2026. Anthropic opened a San Francisco wet lab and reported that Claude turned up a previously unknown enzyme system within 21 hours, a Novartis CEO joined Anthropic's board and drug makers voiced the 'frenemy' risk, while among the public names Recursion licensed its TxFM RNA foundation model to Tempus and Eli Lilly stayed busy on non-AI catalysts.

AI Drug Discovery Weekly

Week of September 24, 2026: Anthropic's Claude Finds an Enzyme in 21 Hours as Pharma Weighs the Frenemy Risk

Issue #13, Week of September 17–24, 2026.

TL;DR

  • The story of the week wasn't a public stock, it was Anthropic. The company that makes Claude quietly opened a "wet lab" (a real, physical biology lab) in San Francisco, pointed Claude at it, and reported that within 21 hours it had turned up a "previously unknown enzyme system hidden in the DNA of bacteriophages" that CEO Dario Amodei said "could represent a new gene editing mechanism." Multiple podcasts dug into what it means, and what it doesn't. AI Breakdown
  • The "frenemy" problem went mainstream. Reuters reported the CEO of Novartis has joined Anthropic's board, and that Anthropic is partnering with Novo Nordisk, even as drugmakers openly worry that the AI company they pay could learn from their data and eventually out-compete them. Reuters World News
  • A big-pharma insider drew the clearest line yet between AI hype and AI reality. Bristol Myers Squibb's chief research officer Robert Plenge said AI-designed molecules are working today, including one sickle-cell program that "wouldn't have happened without AI," but that reading the deepest layers of human biology is years away and shouldn't be over-funded. Biotech 2050 Podcast
  • Among our public names, the only genuinely AI-related move was quiet and easy to miss: Recursion (RXRX) licensed its RNA foundation model ("TxFM") to Tempus AI and extended their data partnership to 2029, a rare, real, dollars-attached monetization of an AI platform. Eli Lilly (LLY) had a busy week of drug approvals and a $3.25B discovery deal, but nothing new on its TuneLab AI platform. Schrödinger (SDGR) hit a fresh 52-week high and then gave a little back on no news.
  • Prices (Sep 24 close): RXRX $3.89 (+9.9% on the week), SDGR $29.45 (−2.6%), LLY $1,183.86 (+2.8%).

What's new

1. Anthropic Put Claude in a Real Lab, and It Found Something in Under a Day

This was the dominant new development of the week, and it reframes the whole "AI in drug discovery" conversation.

Here's what happened, in plain terms. Anthropic, the company behind the Claude AI models, and a firm that until now sold AI tools to drugmakers rather than doing drug discovery itself, set up a physical biology lab in the San Francisco Bay Area. They connected Claude to it, gave it a research direction, and let AI "agents" plan and interpret experiments while human scientists ran the bench work. On AI Update (Sep 18) the setup was described as Claude "directing laboratory robots with very limited human intervention," focused on "rare and undruggable diseases," and, tellingly, "deliberately avoiding direct competition with all of the big pharmaceutical companies that are paying them... millions of dollars for their technology."

Then came the result. As AI Breakdown (Sep 23) recounted, quoting Anthropic's own announcement: "Claude has discovered a previously unknown enzyme system hidden in the DNA of bacteriophages" (bacteriophages being viruses that infect bacteria) and "beside the enzyme's gene sits a long array of repeated DNA, a structure that looks somewhat similar to CRISPR." (CRISPR is the gene-editing tool that won a Nobel Prize; a "new" system that resembles it is, potentially, a big deal.) Amodei's own words, relayed on the show: "we suspect this could represent a new gene editing mechanism. Its precise function, biotechnological utility, if any, or level of significance is not yet clear."

Two things are worth pulling out.

First, the honesty about how it worked. This was not Claude alone. In Amodei's telling: "The work was done mostly, though not entirely, by Claude. Our life sciences team suggested a broad area of research, Claude read through the literature and a bunch of genome data, and discovered something interesting. Then Claude proposed experiments to verify the discovery, and our team carried them out." In other words, humans aimed it, Claude read and reasoned across a mountain of genomic data faster than any person could, proposed what to test, and humans did the testing. The host's analogy: "It's just like you or I going into a domain that we understand deeply and asking Claude to help us do some research so we can figure stuff out."

Second, the "21 hours" framing and why Amodei thinks it matters beyond this one enzyme. The discovery reportedly came after only 21 hours of the lab being online. Amodei's argument for why you shouldn't dismiss it as a fluke, quoted on the show: "It's easy to dismiss this as a one-off or curiosity, but we've repeatedly seen a pattern where AI performance in new intellectual domains goes from weak to superhuman in a matter of a few years." He drew the parallel to mathematics (models went from high-school-level math in 2023 to solving serious open problems by late 2026) and concluded: "we believe AI for biology is on a similar exponential trend." That is the real claim. Not "we found one enzyme," but "biology is the next field to be swept up in the same curve."

A note on how loud this got: the same story was picked up almost word-for-word by roughly ten AI-news podcasts in a single week (AI Chat, The AI Podcast, Practical News, and others). When a niche biology result becomes a general-tech headline, that tells you the narrative has escaped the lab even if the science is early.

2. The Uncomfortable Part: Your AI Vendor Might Become Your Rival

The more interesting business angle came from Reuters World News (Sep 21), which added detail the tech-news shows glossed over. Anthropic is "partnering with drug makers including Novartis, whose CEO joined Anthropic's board, to use Claude to automate lab equipment and tackle rare diseases." So a sitting big-pharma chief executive is now on the AI company's board.

But Reuters also named the tension out loud, and it's the thing pharma strategy teams should be chewing on: drug companies fear that "as those drug companies use Claude for their own drug programs... Anthropic will learn from that and then out-compete those companies in the drug industry." That is the "frenemy" problem in one sentence. You pay Anthropic to accelerate your pipeline; in doing so you show it exactly how drug discovery works; and it has just proven it can run a lab of its own. Anthropic's stated answer, deliberately working on "undruggable" diseases the big companies aren't chasing, is a promise, not a wall.

There's a related governance thread. Moonshots with Peter Diamandis (Sep 17) noted an "Anthropic and Novo Nordisk... sprawling collaboration for biotech" landing in the same week Anthropic flagged five cases of people trying to misuse Claude for bioweapons work, with one panelist wryly summarizing the double standard as "advanced biotech, for me, but not for thee." The same power that finds a new enzyme in 21 hours is the power regulators worry about.

3. The Anthropic Protein Story From Last Week Now Has a Public-Company Receipt

Two weeks ago we tracked a rumor that Anthropic had designed protein binders that worked about half the time. This week it got a paper trail. On Chip Stock Investor (Sep 17), the hosts explained: "Anthropic published some results from a protein design campaign against 15 targets. And then Twist filed an 8-K noting that it was one of the independent external evaluators, meaning it manufactured the AI-designed proteins and then tested them for binding. Anthropic reported that that was a success for some of those protein binders."

Why this matters for investors: an 8-K is an official regulatory filing, and Twist Bioscience (a company that makes synthetic DNA) validating that AI-designed proteins actually bound their targets is independent, on-the-record confirmation, not a press release. The hosts also flagged that Twist is reporting "triple-digit percent growth for AI-driven drug discovery in fiscal year 2026 over fiscal year 2025." That is the "picks and shovels" read: whoever wins the AI-protein race, the companies that physically make and test the designed molecules get paid either way. (The hosts still weren't buying the stock on valuation, hence the title, but the data point stands.)

4. A Big-Pharma R&D Chief Drew the Sharpest Real-vs-Hype Line of the Week

If you read one debate episode this week, make it Biotech 2050's interview with Robert Plenge, EVP and Chief Research Officer at Bristol Myers Squibb (Sep 24). Plenge is not a hype man (he spent years as an academic being unimpressed by AI) and that makes his current conviction more useful. He splits AI in drug R&D into three buckets, and he is refreshingly specific about which are real now and which aren't. See "The debate" below for the full breakdown; the headline is that he thinks AI-designed molecules are delivering today, while AI reading the deepest layers of human biology is a much longer road that companies risk over-funding.

The debate

The question this week: now that an AI has made a real, if early, biology discovery on its own, how close are we actually to AI doing the hard part of drug discovery? The most grounded answers came from people inside the industry, and they converged on a nuanced middle.

Robert Plenge (BMS Chief Research Officer), the three tiers. Plenge's framework is the cleanest we've heard. He divides AI's usefulness into three layers:

  1. Designing better molecules, faster: real, right now. "Our very first investment was really around this idea of predictive molecule invention, especially in the lead optimization space... We are there today." His concrete proof: "we have one program that I think wouldn't be in patients today... that wouldn't have happened without AI," a "targeted protein degradation molecular glue" that switches fetal hemoglobin back on in sickle cell patients, "offering the potential for functional cures." The reason AI works here, in his words, is that molecule design "is a rapid cycle that's verifiable," you can quickly make the molecule and measure whether it's better. Fast feedback is what AI thrives on.
  2. Reading "causal human biology": real eventually, but don't over-invest. This is the deep question of why a disease happens in a human body. Plenge splits it into three tiers of knowledge: biology that's known; biology that's "hidden, latent somewhere in datasets" and could be extracted with AI; and biology that is simply "not known today," where "you could do all the AI that you want and you're never going to be able to extract it." His judgment: "the majority of causal human biology is within that third layer, that it's going to require direct experimentation." His investing conclusion is the money quote for portfolio thinking: "you don't want to underinvest, but you also don't want to overinvest there."
  3. Modeling clinical data to predict success: the emerging frontier. As a drug moves from research toward the clinic, Plenge says AI agents can run "mechanistic models" against early trial data to see whether a program is "trending favorable or unfavorable," and, crucially, the same agents can "scan the external environment to understand the competitive landscape." So AI becomes a tool not just for making molecules but for deciding, earlier and with more evidence, whether to keep spending on them.

The trials skeptics: the drug usually isn't the problem. A complementary and slightly contrarian view came from AI For Pharma Growth E236 (Sep 22), an episode built around the fact that ~90% of clinical trials fail. The guest's provocative claim: when trials fail, "in most cases, it's not [because] the drug is not working." The real culprits are operational and statistical: wrong protocol design, wrong patient population, poor site selection, weak monitoring, and under-use of the "totality of data." Their pitch is "causal ML" (machine learning that can explain why, not just predict) run over a data lake of "millions and millions of patients," and their proof point is "rescuing a previously failed diabetes trial by identifying a subgroup of patients that actually responded well." The read-through: the near-term money in AI may be less in inventing molecules and more in not botching the trials that test them.

The grounding skeptics: the bottleneck is data, not the model. Data in Biotech (Sep 17), interviewing DrugBank's CEO, made the case that raw AI models are dangerously incomplete without a reliable data foundation. The vivid example: "if you ask a general model how many approved drugs hit PD-L1, it will tell you three. The real, complete answer is six." She cited an Anthropic benchmark where, on a biology task, "the best frontier models returned the right answer as little as 17% of the time. When those same models were given a deterministic tool, the data accuracy went north of 99%. So their own conclusion was the bottleneck was never the model, it's the data infrastructure." That is a direct, sobering counterweight to the "21 hours to a discovery" excitement: the same models can be confidently wrong 83% of the time without the right data plumbing.

The bull megaphone: Cathie Wood. For the maximalist case, Cathie Wood on Equity Mates (Sep 21) put numbers on it: today "to discover and develop one new drug costs $2.4 billion and takes 13 years. With AI and sequencing technologies, we believe that in the next five years the cost will drop from $2.4 billion to $600 to $700 million, and the time will drop from 13 years to eight years or fewer." She also made a market-structure argument worth remembering: healthcare is "the most inefficiently priced part of the market" for AI, because tech analysts and healthcare analysts have different "DNA" and neither fully prices the crossover. Take the specific numbers as an advocate's, not gospel, but the direction is the whole bull thesis in one breath.

Where the debate nets out this week: the sober insiders (Plenge, the trials and data skeptics) and the loudest bull (Wood) actually agree on the shape: AI is real and compounding in molecule design and trial operations, and slower in the deep biology. The disagreement is only about speed and how much to pay for it today. Anthropic's 21-hour enzyme is a data point for the fast camp; the "17% right without the right data" benchmark is a data point for the careful one. Both are true.

Stocks in play

Eli Lilly (LLY), $1,183.86, +2.8% on the Week

The busiest of our three names, but notably not on AI-platform news. The week's items:

  • Two FDA approvals. Full approval of Inluriyo (an oral SERD) in combination with Verzenio for a common form of advanced breast cancer, based on the Phase 3 EMBER-3 trial (Sep 18); and approval of Onswik, a once-weekly basal insulin for type 2 diabetes, non-inferior on blood-sugar control to standard daily insulins (Sep 24). (Source: thefly / MT Newswires, Sep 18 and Sep 24.)
  • The oral GLP-1 ramp. CEO Dave Ricks said Lilly's oral GLP-1 pill, Foundayo, is now roughly one-third of new patients starting a GLP-1 pill, with share growing week over week, and the company broke ground on a $6.5 billion manufacturing plant in Houston to make it (operational by 2030). (Source: MT Newswires / CNBC, Sep 21.)
  • A $3.25 billion discovery deal, the closest thing to an AI story. On Sep 24, Lilly announced a collaboration with China's InnoCare Pharma worth up to $3.25 billion in milestones (plus up to $100M upfront and single-digit royalties), in which InnoCare uses "its drug discovery platform" to find compounds across up to five targets. (Source: MT Newswires, Sep 24.) Important caveat: this is a discovery-platform partnership, but it is not branded as a TuneLab deal, so don't file it under the AI-flywheel story without more detail.
  • Analyst move: Guggenheim raised its price target to $1,284 from $1,273 (Buy), citing prescription trends and Zepbound's reinstatement on the CVS formulary from Oct 1. FactSet's consensus mean target sits around $1,359. (Source: thefly / MT Newswires, Sep 18.)
  • The overhang to watch: research and lawsuits linking the whole GLP-1 class to a rare vision-loss condition (NAION), naming both Lilly and Novo Nordisk (WSJ via thefly, Sep 20). And Lilly is a defendant in a jury trial with Nektar Therapeutics over the Rezpeg program (Sep 23).

The AI angle: quiet this week. After last week's TuneLab momentum (the Twist Bioscience and Ginkgo Bioworks data partnerships and the AtaiBeckley close), there were no new TuneLab ecosystem partners or disclosed economics. The stock is being driven by its core obesity/diabetes and oncology engine, not the AI platform, a reminder that TuneLab is still strategic optionality, not yet a visible revenue line.

Recursion Pharmaceuticals (RXRX), $3.89, +9.9% on the Week

After roughly seven quiet news weeks, RXRX finally produced a genuinely on-theme, AI-platform data point, and it's arguably the most interesting public-name item of the week.

  • Tempus AI (TEM) extended and expanded its partnership with Recursion (Sep 21). The existing multi-year data license (first struck in 2023) was extended through November 2029, and Tempus entered a new license for Recursion's RNA foundation model, "TxFM" (non-exclusive, worldwide). (Source: thefly, Sep 21.) An "RNA foundation model" is an AI model trained on biological data to predict how genes behave, exactly the kind of "platform" asset AI-biotech bulls keep saying will one day be monetized. Here it actually is being licensed to a paying counterparty.
  • The economics have a twist worth understanding. The amended terms replace discretionary fees that could have totaled up to $84 million over two years with committed payments totaling $42 million over three years. So the headline ceiling comes down ($84M possible → $42M certain), but the cash becomes contracted and reliable rather than optional. Read it as de-risking, not a raise, a bird in the hand. (Source: thefly, Sep 21.)
  • Correction to last week's flag. Last week we flagged, from a podcast passing-mention, that founding CEO Chris Gibson may have "recently stepped aside" and marked it unverified. Now verified, and it is old news, not a September event. Gibson stepped down as CEO effective January 1, 2026 (succeeded by Najat Khan, formerly of J&J), declined re-election to the board, finished his term in June 2026, and moved to a strategic-advisor role. So there is no new governance shock here; the podcast was referring to a transition that happened months ago. We're closing this tracking item.
  • Still on the calendar: updated Phase 1b/2 data for REC-4881 in familial adenomatous polyposis (the TUPELO study) is scheduled for November 2, 2026 at the CGA-IGC annual meeting. Prior data showed a 43% median reduction in GI polyps at week 13, deepening to 53% at week 25, so this is a real, near-term catalyst.

The read: the +9.9% move is a bounce off near-lows (the stock is still well below its $7.18 52-week high), and the Tempus TxFM license is a small but real proof that Recursion's AI platform has outside value. Watch whether more model-licensing deals follow, that would be the thesis turning tangible.

Schrödinger (SDGR), $29.45, -2.6% on the Week

The two-week momentum story cooled. After exploding ~61% two weeks ago on the Sept 9 Tectora launch, SDGR pushed to a fresh 52-week (and two-year) intraday high of $31.59 on Sep 22, then gave a little back into the Sep 24 close.

  • No fresh fundamental catalyst this week. A careful check (widening the news window and running the catalyst feed) turned up nothing that explains the moves. The only company-specific items were routine: an insider sale by Karen Akinsanya (President, Head of Therapeutics R&D), who sold 31,599 shares for about $928,000 under a pre-set 10b5-1 plan adopted last November (routine and scheduled, not a signal), and small new-hire stock grants. (Source: MT Newswires / SEC Form 4, Sep 21; thefly, Sep 17.)
  • The move is momentum and technicals, not news. The healthcare sector was up only modestly on the week, so SDGR's swings are stock-specific. The stock had gotten technically stretched (overbought), and short interest is elevated at roughly 15% of the float, a setup that produces exactly this kind of sharp squeeze-then-fade action. The pullback looks like profit-taking, not a change in story.
  • Fundamentals recap (unchanged from Q2): software annual contract value grew ~27% year over year to $29.6M, with full-year ACV guidance reaffirmed at $218–228M; the Bunsen agentic-AI "co-scientist" and the Bristol Myers Squibb neurology collaboration are progressing. Street views are split post-rally: UBS is at Neutral with a $19 target (below the current price), while others are far more bullish.

The read: unchanged from last week, this is a momentum/short-squeeze move sitting on top of a real but early software-plus-pipeline story. Near a 52-week high with a stretched chart, the risk-reward is asymmetric to the downside if the momentum fades. Watch for any actual fundamental follow-through (software bookings, Tectora progress, the BMS collaboration) to justify the new level.

Read-throughs

  • "Picks and shovels" is the safest AI-biotech trade. Twist Bioscience's 8-K validation of Anthropic's proteins, plus its "triple-digit" growth in AI-driven drug-discovery revenue, is the clearest example: the companies that physically synthesize and test AI-designed molecules (DNA synthesis, antibody characterization, lab automation) get paid regardless of which AI model wins. This is the same logic behind Lilly's TuneLab partners (Twist, Ginkgo) last week. Chip Stock Investor
  • Data infrastructure is the under-priced layer. The DrugBank "17% right without a deterministic tool, 99% with one" benchmark says the value isn't only in the model, it's in the curated, verified data that keeps the model honest. Names that own proprietary or curated biological data (Tempus, Recursion's datasets, DrugBank-style layers) have a moat the foundation models can't easily replicate. Data in Biotech
  • The "frenemy" risk is now a real strategic variable for pharma. With a Novartis CEO on Anthropic's board and Anthropic running its own lab, every big drugmaker has to weigh how much of its process it exposes to an AI vendor that could compete. Expect more structured, walled-off partnerships, and possibly more in-housing of AI to avoid the dependency. Reuters World News
  • AI-native discovery companies are heading to public markets. BioSpace (Sep 23) noted that NVIDIA-backed Iambic Therapeutics filed to go public and, the same week, announced a partnership with AbbVie to discover small-molecule drugs in neurology, oncology and immunology. A fresh AI-drug-discovery IPO would give public investors a purer read on how the market values these platforms, worth tracking into the pricing.
  • Personalized cancer vaccines remain the AI-adjacent optionality. InvestTalk (Sep 22) revisited Moderna's Merck-partnered personalized cancer vaccine (which uses AI/mRNA to tailor a shot to a patient's tumor), noting Phase 3 data is coming but "that's still a long way from production... a bit of a risky play from here" with real upside if it works. Recurring theme; watch the data.
  • The clinical-operations layer is where near-term ROI is most defensible. Both the trials-failure episode and Plenge's "third bucket" point the same way: using AI to design better trials, pick the right patients, and kill failing programs earlier is lower-risk, faster-payback value than trying to invent biology from scratch. Favorable for tech-fluent CROs and trial-software players; a threat to those that don't adapt. AI For Pharma Growth E236

What changed vs last week

  • The narrative moved from public names to Anthropic. Last week the two stories were SDGR's squeeze and Lilly's TuneLab partners. This week the center of gravity shifted to a private company, Anthropic, whose 21-hour lab discovery, board addition (Novartis CEO) and Novo Nordisk deal made it the single biggest AI-drug-discovery story, and surfaced the "frenemy" question for the whole industry.
  • The Anthropic protein rumor became a filing. Last week's tracked rumor (Anthropic's protein binders working ~50% of the time) now has independent, on-the-record confirmation via Twist Bioscience's 8-K disclosure that it tested the AI-designed proteins against 15 targets.
  • The debate matured from "where AI helps" to "how fast, and how much to pay." Last week's frame was design-vs-clinical ("biology is biology"). This week Plenge sharpened it into a three-tier investment question, and the disagreement is now explicitly about speed and capital allocation, not whether AI works.
  • RXRX broke its ~7-week quiet streak with an actual AI-platform deal. The Tempus TxFM license is the first genuinely on-theme public-name AI monetization we've seen from Recursion in weeks. Contrast last week, when RXRX had no news at all.
  • The Chris Gibson departure flag is resolved and closed. Verified as a January 2026 transition (Najat Khan now CEO), not a September event, correcting last week's unverified podcast-sourced flag.
  • SDGR flipped from breakout to cool-down. Last week: at a 52-week high, best of the three. This week: new intraday high on Sep 22, then a pullback on no news; worst of the three on the week (−2.6%).
  • LLY stayed strong but for non-AI reasons. Two approvals, the Foundayo ramp, a $6.5B plant and a $3.25B InnoCare deal, but, unlike last week, nothing new on the TuneLab AI platform.

Still tracking (open items): more Recursion model-licensing deals (does TxFM licensing become a pattern?); Recursion's Nov 2 TUPELO data; whether Lilly adds TuneLab partners or discloses economics; the Iambic IPO pricing and AbbVie deal terms; whether SDGR shows real fundamental follow-through or fades; how pharma structures partnerships around the Anthropic "frenemy" risk; the Moderna/Merck cancer-vaccine Phase 3 data; and regulatory/governance moves around AI in biology (the bioweapon-misuse flags, the "trust" question).