Newsletter · 2026-09-10 · Ashutosh Agarwal

AI Writes a Working Virus From Scratch as Biologists Push Back - AI Drug Discovery Weekly - Week of September 10, 2026

AI Drug Discovery Weekly for the week of September 3 to 10, 2026. Podcast synthesis on scientists using AI to design entire working bacteriophage genomes, an Anthropic model designing lab-verified protein binders at a roughly 50 percent hit rate, DeepMind's map of every possible human mutation, and a physicist, a bioengineer, a genetics chief and Eli Lilly's CEO all arguing that AI still does not understand the language of biology.

AI Drug Discovery Weekly

Week of September 10, 2026: AI Writes a Working Virus From Scratch as Biologists Push Back

TL;DR

• After last week's total silence (not a single on-theme podcast in seven days), the microphone came roaring back this week. But it came back on the science and the big AI labs, not on our public companies. Recursion, Schrödinger and Eli Lilly were essentially absent from the podcasts. The action was upstream, in the research.
• The landmark: scientists at Stanford and the Broad Institute used AI to design entire working bacteriophage genomes (viruses that infect bacteria) that don't exist in nature. It made the front page of The New York Times. Roughly 1 in 20 of the AI's designs actually came to life in the lab. Big for medicine (think programmable viruses that kill drug-resistant bacteria); genuinely scary for biosecurity (This Week in Virology).
• The other landmark: an Anthropic AI model designed protein "binders" (molecules that latch onto a target protein and change what it does) that labs then verified worked, with a roughly 50% hit rate on 12 targets, versus the usual 10–15% (Last Week in AI).
• But the skeptics came out in force, and their argument sharpened. Last week's worry was whether AI can verify that a drug works in a body. This week the worry went deeper: does AI even understand biology, or is it just pattern-matching the surface? A physicist-turned-drug-hunter, a Stanford bioengineer, the head of Regeneron's genetics engine, and Eli Lilly's own CEO (quoted second-hand) all landed on versions of the same line: "AI models speak English. They don't speak biology, because we don't speak biology."
• One real corporate item for our names: Schrödinger helped launch a brand-new drug company, Tectora Therapeutics, which raised a $55 million first round. Schrödinger put in two early-stage drug programs in exchange for a stake plus future milestones and royalties (MT Newswires, Sep 9).
• Prices: a rough week for the whole group. Recursion -8.1%, Schrödinger -9.2%, Eli Lilly -3.1% on the week. Recursion is now sitting near its 52-week low.

What's new

An AI designed a working virus, and it made the front page

This is the story of the week, and it deserves the space. On the long-running science podcast This Week in Virology (Sep 6), the hosts walked through a new paper in the journal Science from Stanford University and the Broad Institute of MIT and Harvard, titled "Generative design of bacteriophages with genome language models." It was front-page, above-the-fold news in The New York Times on August 7 under the headline "In lab, AI designs viruses not found in nature."

Here is what actually happened, because the reality is more interesting, and more reassuring, than the headline. A bacteriophage (or "phage") is a virus that infects bacteria, not people. The team took a tiny, famous phage called phiX174 (its genome was fully sequenced back in 1977) as a template. Then they used two AI models called Evo 1 and Evo 2. These are "genome language models": think of them as the biology equivalent of ChatGPT, except instead of learning the patterns of English from books, they learned the patterns of DNA from vast amounts of genetic sequence. The four billion years in the episode's title is a nod to the idea that evolution itself generated the training data.

The crucial leap: earlier work had used AI to design individual genes, or small circuits of genes. This is the first time anyone has had AI generate whole genomes (the complete instruction set for an organism) from scratch.

The numbers matter, and they cut against the hype:

• The researchers fine-tuned the models on about 15,000 sequences from the phage's family, then had the AI generate roughly 3,800 candidate genomes that passed their filters.
• They curated that down to 302 designs (nicknamed "EvoPhi"), and managed to physically build 285 of them (the other 17 were too complex for the DNA-synthesis machines).
• Of those 285, exactly 16 produced a working, self-replicating virus that could kill its target bacteria. That is a success rate of about 5%, 5.3% for Evo 1, 6.9% for the newer Evo 2.

As one host put it, when you read the news coverage "it sort of seemed like, oh, they just told AI make me a virus and AI made me a virus. It's a little more complicated than that." A 95% failure rate is a very long way from "AI designs life on demand."

Two things are worth holding onto. First, the safety design was deliberate and thoughtful: the researchers trained the models only on phage sequences and specifically excluded any viruses known to infect humans, animals or plants, the logic being that if the model never sees a human pathogen, it can't generate one. (The Arc Institute funded the original model training, reportedly "tens of millions of dollars.") Second, the medical promise is real: engineered phages are one of the most promising routes to killing antibiotic-resistant bacteria, an area where the traditional drug industry has largely given up. But the dual-use shadow is unmistakable, and the hosts did not shy from it.

An Anthropic model designed protein binders that actually worked

On the AI-industry podcast Last Week in AI (Sep 8), the hosts flagged a bio result buried inside Anthropic's latest model release. In plain terms: a large part of drug discovery is designing protein binders, molecules that stick to a specific protein and, by grabbing onto it, change its shape and therefore its behavior. That is how a great many drugs work.

The claim: Anthropic's newest model designed protein binders that were then verified in real labs, with "roughly a 50% success rate" across "a sample size of 12" different targets. The host's reference point: "10 to 15% is the norm here. You're hitting 50%. This is industry changing stuff if it gets deployed." The same model also rewrote the low-level computer code ("GPU kernels") for seven open-source biology models, producing 1.4x to 2.5x speed-ups and cutting computing costs 30–60%.

Treat the specific numbers with care: this is a podcast relaying a company's own release, not a peer-reviewed result, and the sample of 12 is tiny. But the direction is the same as the phage story: the generation step, the "design me a molecule that binds here," is getting genuinely, measurably good. The hosts also sounded the biosecurity alarm again, predicting that "at some point there will be an AI-designed biopathogen," the flip side of the same capability.

Google DeepMind maps every possible mutation in the human genome

On The Daily AI Show (Sep 9), the hosts covered DeepMind's new AlphaGenome Atlas, a predictive map of what happens if you change any single letter of human DNA. It boils each change down to a single "variant impact" (AVI) score: the higher the number, the more likely that mutation is to cause disease. DeepMind says it has pre-computed this for all ~9 billion possible single-letter changes in the genome, and made it free for researchers. As one host noted, checking every variant by hand "for one second" each would take on the order of tens of years.

The same episode flagged a headline-grabbing AI milestone with a bioscience read-through: OpenAI reportedly pointed 10,000 AI agents at a hard math problem for 88 hours and produced a claimed proof toward one of the famous Clay Millennium Prize problems (only one of the seven, the Poincaré conjecture, has ever been solved). The hosts' bigger point was a bet on the 2030s: "we solve math first, we bridge to bioscience." The counterpoint, that math has clean right-and-wrong answers and biology does not, became the theme of the week.

The debate: does AI understand biology, or just imitate it?

This is where the week got genuinely interesting. For the past few issues the skeptics' complaint was about validation: sure, AI can design a molecule, but can it tell you whether the thing will work inside a messy human body? This week a different and deeper complaint took over: even before you get to the body, does AI actually understand biology, or is it just very good at mimicking the surface patterns? Four credible voices, coming from completely different directions, converged on the same answer.

The physicist: "precision and complexity are mutually exclusive." On AI For Pharma Growth (Sep 8), Dr. Jacek Macik, an ex-aerospace and automotive engineer (Airbus, BMW, Silicon Graphics) who now co-runs a company called BioDynLab, made the contrarian case for physics over machine learning. His core idea, which he calls the "principle of incompatibility," is that "high precision is incompatible with high complexity, you can't have both at the same time." His memorable framing: "in highly complex situations, precise statements are irrelevant and relevant statements are imprecise."

Applied to the AI gold rush, his argument is blunt: throwing more compute at biology hits a wall. "LLMs are already reaching a plateau phase where every additional 1% improvement in performance requires a 10x investment in training… there's not enough money in the universe to reach a 99% level of precision." He mocks the "give me five gigawatts and I'll find a cure for cancer" pitch: "there are hundreds of types of cancers." His alternative studies how every atom in a molecule moves and passes information, rather than predicting a static shape. His pointed critique of the technology everyone celebrates: "AlphaFold3 does a phenomenal job predicting protein structures, but they predict a static structure. They do not show you how the residues or even single atoms move." Notably, he says he isn't trying to replace AI: he wants physics and machine learning to be complementary. (One real data point: he says a physics-based method cut lab time 50–70% in one antimicrobial project, with no machine learning involved.)

The bioengineer: you can't design function you can't measure. On Mendelspod (Sep 3), Stanford physicist and bioengineer Polly Fordyce made a subtle, important point. We are great at reading DNA. We can now predict protein shapes and even invent new proteins. "But there's an odd gap in this extraordinary new world. We still have a hard time measuring what proteins actually do." Her verdict on AlphaFold: it solved "low energy structure prediction" (a 50-year grand challenge) but that is "just the very first step towards function." Real proteins "are wiggling around" and each does many different jobs.

Her fix is a wonderful analogy. Weather prediction became reliable because of three things: better computers, better models, and, crucially, a global network of stations measuring real physical quantities like temperature and wind speed. Biology has the first two but not the third. So her lab is building the measurement layer: standardized, physics-based measurements of protein function (like how tightly two molecules stick together), so that "next-generation AI algorithms capable of designing things that function" have real data to learn from. She won a $2.5 million Schmidt Sciences award to crowdsource these measurements using a cheap new bead-based device that "anyone can make… all they need is a pipette, a benchtop vortexer, and access to a [FACS] machine," feeding into a shared database she calls the "Functional Protein Observatory." Her framing of the whole AI question: AI is good at interpolating (filling in gaps between things it has seen); physics lets you extrapolate to conditions you've never measured. Biology needs the second.

The genetics chief: pure-AI target-hunting is "orders of magnitude off." On The BioCentury Show (Sep 4), Gonçalo Abecasis, who runs Regeneron's Genetics Center (a database of more than 3 million people's sequenced genomes paired with their health records), was refreshingly candid about where AI helps and where it doesn't. Where it clearly helps: making his software engineers far more productive, and "annotating images and extracting key features" from medical scans (how much fat is in a liver, how much muscle mass). Where it genuinely moves the needle: interpreting rare genetic variants by using an AI "trained on all the protein-coding sequences across the tree of life" (humans, bacteria, birds, fish) to judge whether a one-letter change looks harmless or damaging.

But asked whether you could just hand an AI all the genetic and health data and let it pick the best drug targets, he was clear: "I think we're at least orders of magnitude off from what would be required. I don't think we're there." Even 30 million genomes, he said, wouldn't be enough for AI to beat the plain statistical methods his team still relies on. Encouragingly, he sees no diminishing returns from more genetic data ("the number of discoveries grows a little faster than the data size") pointing to real examples like people missing a copy of the gene GPR75, who "really don't get as overweight as everybody else" and have better blood sugar, the kind of natural experiment that becomes a drug target.

The investor, quoting Lilly's CEO: "AI speaks English, not biology." Reacting to the DeepMind news on Motley Fool Hidden Gems (Sep 10), analyst Lou Whiteman was bracingly skeptical for anyone tempted to trade this stuff: AI drug discovery, he said, "is not investable" on any near-term horizon. He leaned on Eli Lilly CEO David Ricks, who "wasn't dismissive of AI but was definitely trying to hit the brakes on the hype." The line that stuck: "AI models speak English. They don't speak biology. And part of that is because humans don't speak biology. We still don't understand the language of biology, so we can't teach AI a language we don't understand." His reminder that biology is rarely simple: cystic fibrosis is caused by a single gene (chromosome 7) "and yet we still struggle with that"; most diseases are tangled interactions. His investing takeaway: this is "a decade or so" from producing marketed drugs, so "don't just load up on some biotech because they say they're licensing this."

Why it matters. Put the two halves of the week together and you get the sharpest picture yet of where AI in drug discovery actually stands. The generation engine is now genuinely powerful: it can spit out working viruses and high-hit-rate protein binders. But four independent experts, none of them Luddites, agree the understanding is still shallow: AI predicts static shapes, not the dance of atoms; it interpolates but can't extrapolate; it can't measure function; and it can't pick targets on its own. Our standing diagnostic holds and gets an upgrade: be excited about claims of faster design; be very skeptical of claims of compressed understanding or validation. This week, the design side delivered and the understanding side pushed back hard.

Stocks in play

A tough week for all three names, and, notably, the real news was corporate, not AI.

Week-over-week measured against last issue's Sep 3 closes (RXRX $3.44, SDGR $20.71, LLY $1,159.50). Prices via FactSet, Sep 10, 2026.

Schrödinger (SDGR): a real, concrete data point in an otherwise red week. Schrödinger, along with the venture firms New Enterprise Associates and RA Capital Management, helped establish a brand-new biotech called Tectora Therapeutics, focused on immunology and inflammation. Tectora simultaneously closed a $55 million Series A from those two firms. Schrödinger's role is the interesting part and fits its long-running strategy: rather than funding the programs with cash, it contributed two of its own early-stage small-molecule programs in exchange for an equity stake, and it will be entitled to future milestone payments and royalties while partnering with Tectora to push the programs toward clinical candidates (MT Newswires, Sep 9). This is Schrödinger's "use our computational platform to seed new companies and keep a piece of the upside" model in action, worth watching whether these launch-and-retain-a-stake deals become a more visible part of the story. Even so, the stock fell 9.2% on the week, its worst of the three; the move looks like broad risk-off in small-cap biotech rather than anything company-specific. And, tying back to this week's debate: Schrödinger is the purest public embodiment of the physics-based approach that both Macik and Fordyce argued is the missing complement to machine learning, though, to be clear, no podcast mentioned Schrödinger by name.

Eli Lilly (LLY): a bearish analyst raises his target anyway. The one news item was an odd one: HSBC raised its price target on Lilly to $940 from $850 while keeping a Reduce (sell-equivalent) rating (TheFly / MT Newswires, Sep 10). In other words, the analyst nudged the number up but still thinks the stock falls. The stated logic was housekeeping across the whole healthcare group ("a lower sector risk premium and pipeline updates"), with the blunt view that "sector multiples are elevated, and bottom-up stock picking is the only source of alpha." For context, the broader analyst crowd remains far more bullish: FactSet's polled consensus is an average "overweight" with a mean price target of $1,352.49, so HSBC is a genuine outlier on the bearish end. No AI-discovery news from Lilly this week; its TuneLab AI effort stayed quiet, and the stock drifted down 3.1%.

Recursion (RXRX): quiet again, and near the lows. No news and no podcast mentions for the sixth straight week. The stock fell 8.1% to $3.16, now uncomfortably close to its 52-week low of $2.77. The catalyst that matters remains unchanged: the Phase 2 data readout for its FAP program (TUPELO), expected in November, plus any milestone news from its Roche/Genentech neuroscience collaboration. Until then, the tape is drift, not thesis.

Read-throughs

• The diagnostic to keep using, now in two parts. When a company claims AI made discovery faster, ask which step. Faster generation of candidates (molecules, binders, even whole genomes) is now believable, the evidence is piling up. Compressed understanding of biology, or compressed clinical validation, remains extraordinary and should be met with skepticism. This week the generation side put real points on the board and the understanding side pushed back hard, from four directions.
• Biosecurity has moved from footnote to headline. AI-designed viruses (even carefully firewalled ones), explicit warnings of an eventual "AI-designed biopathogen," and DeepMind's map of every disease-causing mutation (with its embryo-screening implications) all landed in the same week. Expect regulators (the FDA/HHS AI roles we've flagged before) and the public to start treating "AI can design biology" as a safety question, not just a productivity story. That is a two-sided read for the sector: a moat for responsible, well-governed platforms; a headline risk for everyone.
• The physics-vs-machine-learning divide is becoming the sector's real fault line. Two separate experts this week argued that pure machine learning is hitting a ceiling and that physics, the actual laws governing how molecules move and bind, is the missing ingredient. That is conceptually favorable to physics-based platforms (Schrödinger being the obvious public example) and to "measure it, don't just predict it" approaches. It also reframes the AlphaFold achievement: spectacular, but a static snapshot, and only the first step toward the thing that actually matters, which is function.
• Capital is still flowing to AI/biology platforms even as public stocks sag. Infinimmune's up-to-$838 million Merck partnership and fresh $75 million Series A (Raising Biotech, Sep 3), Fordyce's $2.5 million Schmidt award, and Schrödinger's $55 million Tectora launch all closed while RXRX and SDGR fell ~8–9%. The private market's enthusiasm for AI-plus-antibody and AI-plus-function platforms hasn't cooled with the public tape. Infinimmune's pitch, read the human immune system directly and use an AI model ("Glimpse") to optimize the antibodies you find, rather than engineering them from synthetic libraries, is a clean example of "AI on top of better biology data," the exact recipe this week's skeptics were asking for.
• Clinical trials are the quieter, nearer-term AI opportunity. While the headlines chase molecule design, Dr. Anastasia Christianson (ex-Pfizer, ex-J&J) argued on Med Tech Gurus (Sep 9) that the biggest near-term wins are in trial design and patient recruitment: start from the business bottleneck, not the shiny tool. Her line on the buzziest ideas: digital twins and virtual trials are being used today to inform and accelerate trials, "not to replace trials with patients." This is the least glamorous and possibly most bankable slice of AI in drug development.
• Quantum computing for chemistry has a rough timeline now: ~2035. On Reuters Econ World (Sep 9), a quantum researcher said simulating nature (chemistry, biology, drug discovery, materials) is the single most exciting commercial category for quantum, but that "classical still wins today." His forecast: scientific breakthroughs by 2030, and by 2035 he "would not be surprised" to see drug-research and chemistry firms using quantum computers "as part of their overall process." A useful anchor: this is a real but multi-year-out tailwind, not a this-cycle catalyst.

What changed vs last week

• From silence to a flood, but on the science, not our stocks. Last week's issue was an honest zero: not one on-theme podcast surfaced in seven days, and the whole issue ran on Lilly's corporate news. This week the same net caught roughly ten substantive on-theme episodes, headlined by a genuine landmark (AI-designed working viruses) and a striking capability claim (Anthropic's ~50%-hit protein binders). The catch: our three public names were again absent from the podcasts, the conversation has moved to the underlying research and the big AI labs (DeepMind, Anthropic, Evo/Arc).
• The skeptics' argument evolved. Two weeks ago the frame was the "verification bottleneck": AI can't confirm a drug works in the body. This week it deepened into a "language bottleneck": AI may not understand biology at all, only its surface patterns. That is a more fundamental critique, and it came from a physicist, a bioengineer, a genetics chief and a CEO (relayed) all at once.
• Two long-standing watch items finally got concrete. Our open file has carried "Anthropic + protein design" and "what will Evo 2 actually do" for weeks. This week both got flagship results: Anthropic's lab-verified protein binders, and Evo 1/Evo 2 as the engines behind the first AI-designed whole viral genomes. (Note: the Anthropic figures come from a company release relayed on a podcast, not a peer-reviewed paper: treat as promising, not proven.)
• Prices turned from flat to broadly lower. Last week the three names were roughly flat. This week all three fell (Recursion -8.1%, Schrödinger -9.2%, Lilly -3.1%) with Recursion pressing toward its 52-week low. The one bright, concrete corporate data point was Schrödinger's Tectora launch.
• No double-counting. None of the previously covered episodes re-surfaced this week, and every episode dated September 3 is genuinely new (last week caught nothing on that boundary).