Newsletter · · Ashutosh Agarwal

A Generative AI Drug Company With 32 Candidates and GPT-5 Running the Lab - AI Drug Discovery - Week of August 13–20, 2026

AI drug discovery weekly for the week of August 13–20, 2026. Insilico Medicine details a 32-candidate generative-AI pipeline public in Hong Kong, Ginkgo lets GPT-5 beat a Stanford lab by 40%, and skeptics question whether a virtual cell is real yet.

AI Drug Discovery Weekly

Week of August 13–20, 2026: A Generative AI Drug Company With 32 Candidates and GPT-5 Running the Lab

TL;DR

  • The theme this week was "prediction to generation." For years, AI in drug discovery mostly predicted things, will this protein fold this way, will this molecule be toxic. Several podcasts this week described the field crossing over into generating brand-new drugs from scratch. The clearest proof point is a company most investors have never heard of: Insilico Medicine, which is generative-AI-first, clinical-stage, and quietly public in Hong Kong (ticker 3696). Its founder laid out a pipeline of 32 drug candidates, with 8 in Phase 1, 3 in Phase 2, and 1 in Phase 3 on The Big Unlock.
  • AI is starting to run the lab, not just read the data. Ginkgo Bioworks' CEO described letting OpenAI's GPT-5 design 30,000 experiments a week for six weeks, with humans simply carrying out whatever it asked, and it beat a benchmark Stanford lab by 40%.
  • A useful splash of cold water. The team at Turbine, which builds "virtual cells," argued the field is "where AI was 10 or 20 years ago," we have narrow, one-trick systems, "not the LLMs of biology." A general-purpose virtual cell, they said, "is not possible now."
  • A brand-new name to watch: NOETIK. Its founder (an ex-Recursion scientist) described an AI agent doing "reverse drug discovery," handed a group of patients who respond to a drug, it figured out the correct biological target on its own, from scratch.
  • A genuinely new tool entered the story: quantum computing. Cleveland Clinic and IBM said they simulated a 12,600-atom enzyme, up from just 10 atoms 18 months earlier, and argued quantum is especially good for rare diseases, exactly where AI runs short of data.
  • Stocks: a quiet week for the pure-plays. Schrödinger (SDGR) rose about 6% on the week; Recursion (RXRX) was roughly flat; Eli Lilly (LLY) climbed to a fresh high before slipping on news that rival Novo Nordisk is testing lower doses of its weight-loss pill.

What's New

A public, generative-AI drug company you probably weren't tracking

The single richest conversation this week was Alex Zhavoronkov, founder and co-CEO of Insilico Medicine, on The Big Unlock (Aug 20), recorded at the AI4 conference. Two things make this worth your attention.

First, Insilico is not a private science project or a pure software vendor. As Zhavoronkov put it, it's "a generative AI company that is clinical stage, we actually have real drugs in the clinical trials," and it is "publicly traded... in Hong Kong under the ticker symbol 3696" (it listed there last year). That makes it one of the very few ways a public-market investor can own a company whose whole identity is generating drugs with AI, rather than predicting things about them.

Second, the pipeline is real and deep. Zhavoronkov rattled off the numbers: 32 developmental candidates, 8 in Phase 1, 3 in Phase 2, and 1 in Phase 3. He also described a pain drug that started life as an anti-aging target and, through sheer volume of experiments, turned out to work, in animals, better than morphine and better, taken orally, than the standard of care for a pain pathway called NAV1.8.

How does he do it? He described a "Lego system" of many specialized AI models, some generate molecules, some generate synthetic biological data, some predict clinical-trial outcomes, all "orchestrated" by frontier models. "Now we've got over 1,200 tasks in drug discovery," he said. "It basically can do prompt to drug." The phrase captures the whole "prediction to generation" shift in three words.

A bit of history explains why this isn't hype. Insilico published its first paper on using generative AI (then called GANs) to invent molecules back in 2016, the idea being that instead of hunting for "a needle in a haystack," you "generate perfect needles." By a 2018 Nature Biotechnology paper, the team could design, synthesize and test molecules all the way into mice in 46 days.

AI is starting to run experiments

If Insilico shows AI inventing molecules, Ginkgo Bioworks (DNA), one of the names in our coverage universe, showed AI running the lab. On Inside the ICE House (Aug 17), CEO Jason Kelly described a collaboration with OpenAI in which "we let GPT-5 design 30,000 experiments a week for six weeks." The goal was to beat a benchmark set by a Stanford lab (Mike Jewett's) for making a certain amount of protein per dollar. "After the fourth round it had beat it," Kelly said, "and after the sixth round we'd beat it by 40%." Crucially, the humans weren't steering: "We weren't controlling. We were not designing the experiments... If it had said, hey, run water... we would have done it."

Kelly framed the prize in dollars: pharma spends roughly $80 billion a year on research (the labs and the people in them, not clinical trials), and the U.S. National Institutes of Health another $40 billion, "so we're talking like $120 billion that's being spent largely on manual labs. That's my market." His bet is on "autonomous labs", you can already go to cloud.ginkgo.bio, place an experiment through a plug-in to coding assistants like Cloud Code or Codex, and have robots run it with no person in the room.

He also had a memorable way of explaining why biology is so much harder than software or math: cells are "basically alien technology... not invented by humans," like a computer "landed from Mars." You can insert your own code, "but you don't know what half the parts do." And he flagged a policy tailwind, a new White House science report (from technology-policy director Michael Kratzios) arguing that U.S. science, run largely the same way since the 1940s, can no longer beat China on cheap labor, and should shift to AI and automated labs instead.

A genuinely new tool: quantum computing

A fresh angle this week came from Cleveland Clinic and IBM on What's Your Problem? (Aug 20). Cleveland Clinic's research lead described their "Discovery Accelerator," which even put a quantum computer in the hospital cafeteria to shift how 3,000 researchers think.

The useful, non-hype lesson: quantum is not a faster version of a normal computer. They tried using it on huge electronic-health-record datasets and it "failed miserably." Where it shines is small, well-defined problems that hinge on simulation, like predicting whether a drug molecule (the "key") will fit a protein (the "lock"). The reason this matters for drug discovery is subtle but important: quantum models that fit purely "based on the physical characteristics" of a molecule, "does not rely on the previous data." That makes it well suited to rare diseases, where there simply isn't enough past data to train ordinary AI. For an Alzheimer's target, they said, the quantum-based prediction was "much better... than the AI generated one."

And a striking measure of progress: in October 2024, the largest biological molecule anyone could simulate on quantum was 10 atoms. Last month (April 2026), Cleveland Clinic, IBM and Japan's RIKEN published a simulation of a 12,600-atom enzyme (trypsin). The breakthrough came from treating quantum and AI as teammates rather than rivals, what they call "quantum-centric supercomputing."

The debate: is a "virtual cell" real yet?

The most valuable counterweight this week came from Turbine, whose CTO Kristóf Szalay and scientist Gerold Csendes appeared on The AI in Business Podcast (Aug 17). A "virtual cell" is the dream of a computer model so good at simulating a living cell that you can test drugs on it before ever touching a lab.

Szalay was refreshingly honest about where that dream actually stands. Biology AI, he said, is "somewhere like where AI was 10 or 20 years ago." We don't yet have "the LLMs of biology... generally applicable problem solvers." What we have are "the Deep Blue or AlphaGo kind of machines", brilliant at one narrow task, given the right data, and useless outside it. A general virtual cell, he said plainly, "is an AGI for biology... It is not now. The reality is that it's not possible now."

What does work today he calls "virtual assays": start with a small real experiment (say, immune-cell responses in 10 patient donors) and use AI to reliably extrapolate to 100 more donors, but only within the exact boundaries of that experiment. The payoff is speed: an experiment-and-analysis loop that normally takes "easily half a year" inside a big pharma company (because of lab queues and waiting on the bioinformatics team) can be compressed "ideally to a day."

Csendes added a warning that ties directly to a running theme of this newsletter: we can't even properly grade these models yet. There's "no unified evaluation framework." Last year's "Virtual Cell Challenge... brought more questions than answers," and one honest conclusion was that "virtual AI cells don't work, at least on the problem that they proposed." With roughly 20,000 gene readings per cell and mostly faint signals, "it's really easy to fool yourself."

That skepticism was echoed, from a different chair, by Mohammed AlQuraishi of Columbia (creator of the open-source OpenFold) in part two of his interview on Artificial Intelligence and You (Aug 17). He put AlphaFold's arrival in perspective, it "compressed something like 10 years of progress in like two years," jumping the field "like 30 points in two years" after decades of inching forward 3–4 points a year. But on whether AI can now design its own breakthrough biology models, he was measured: "there's not been an AI that has developed an AlphaFold... but I don't think it's very far away," predicting that "in 6 to 12 months we'll probably see models designed by AI that are actually tackling serious questions in molecular biology." He also underscored how cheap prediction has become, a single protein takes about a minute on a GPU, "a 60th of a dollar," versus the slow, costly crystallography it replaces, which is precisely why it now sits at the front of pharma's discovery funnel.

The honest read across these three: generating molecules is real and improving fast; simulating whole living systems is not there yet. The value today is in narrow, well-defined tasks, not a magic cell-in-a-box.

Stocks in play

A quiet week for company news, so the podcasts carried the issue. Here's where the three names we track most closely stand.

Eli Lilly (LLY): $1,245.69. Lilly spent the week climbing toward a fresh high (its 52-week high is now $1,292.65) before pulling back 2.7% on the final day. The likely trigger: Reuters reported (Aug 19) that rival Novo Nordisk has begun a study (OASIS-5) testing lower doses of its oral Wegovy weight-loss pill to find the minimum effective dose, a move to improve tolerability and defend against Lilly in the obesity market (Reuters). Two other items: Lilly paid $10 million upfront to AlzeCure for global rights to an Alzheimer's candidate (Alzstatin/ACD680), a deal that could exceed $1 billion in total milestones excluding royalties; and China Merchants raised its price target to $1,252 from $964 (Buy). Net on the week: up about 3%.

Schrödinger (SDGR): $19.58. The best performer of the three, up roughly 6% on the week and holding its post-earnings strength. The only fresh item was an analyst move: UBS assumed coverage at Neutral with a $19 target (up from $13), praising the software growth outlook but citing "limited near-term catalysts and the slow adoption of computational drug discovery" as reasons to wait. Notably, its AI co-scientist platform (Bunsen) and its Bristol Myers Squibb tie-up drew no podcast discussion at all this week, a fourth straight week of silence on that story.

Recursion (RXRX): $3.34. Essentially flat on the week and quiet on news; shares slipped 4.6% on the final day but on nothing company-specific. The one item was a small analyst trim: Morgan Stanley cut its target to $5.30 from $5.50, keeping Equal Weight. Like Schrödinger, Recursion got no coverage by name in this week's podcasts.

The contrast worth sitting with: the two U.S.-listed pure-plays we track are trading quietly at roughly $1.5 billion (SDGR) and $1.8 billion (RXRX) market caps and got zero podcast attention this week, while the deepest, most concrete AI-drug-discovery conversation was about a Hong Kong-listed company (Insilico) that trades entirely off most U.S. investors' radar.

Read-throughs

  • The "data, not the model" theme keeps deepening, and now points to two escape hatches. For months the running lesson has been that whoever owns the best proprietary data wins, because the models themselves are becoming commodities. This week both NOETIK and the Cleveland Clinic/IBM quantum story pushed on the flip side: what do you do when the data doesn't exist? NOETIK's answer is to generate its own high-quality human-patient data from scratch (tumors profiled across 19,000 genes) rather than rely on messy public datasets, its founder was blunt that public data like TCGA's pathology images are often "terrible quality." Quantum's answer is to skip data entirely for the hardest cases, modeling drug-target fit from physics alone. Both are direct responses to the same bottleneck.
  • Watch NOETIK as a new name. On Once a Scientist (Aug 18), founder Ron Alfa (ex-Recursion) described building "biology foundation models" aimed at the clinical end, predicting which patients will respond to a drug, "the bottleneck... at the clinical side." His most striking anecdote: an AI agent ran "reverse drug discovery", given patients who respond to a first-in-class drug, it "literally discovered the target" on its own, "zero-shot," from the underlying biology. If that generalizes, it points at target discovery being automated, which is exactly the task Insilico's Zhavoronkov said is already "completely demonetized."
  • The competitive threat to the specialists is the big AI labs, not each other. Zhavoronkov was explicit that the companies he worries about are "Anthropic, OpenAI, Google... many Chinese players like Lanssen, Alibaba" building foundation models that reason in biology, not fellow biotech startups. His survival strategy is telling: build tools that help the frontier labs train, then buy their models, and keep the durable moats, "the drugs... they're like diamonds... AI comes and goes every six months," plus physical infrastructure (labs, robotics, a trusted network of testing partners). It's a useful lens for any AI-drug-discovery holding: which layer are they defending?
  • A pricing myth, debunked in plain terms. The Chat GPT Podcast episode "How AI reverses Eroom's Law" (Aug 20), which independently cited Insilico's 18-month idiopathic-pulmonary-fibrosis candidate, made the case that AI making drug discovery cheaper will not make drugs cheaper. Because so many drugs fail late (a new drug still costs ~$2.6 billion with a ~90% failure rate), the payoff from AI is mostly accuracy, screening out losers early. And by the "Jevons paradox," companies will plow the savings into chasing more targets and more trials, not into lower prices. The likely result is a much bigger research pipeline and a shot at more rare-disease cures, good for research volume, neutral-to-negative for the "AI will crush drug prices" narrative.

What changed vs last week

  • Last week the marquee was Chai Discovery's $400M raise and the AlphaFold reality check; this week the center of gravity moved to "prediction to generation" and biology foundation models. Insilico Medicine (public in Hong Kong, 32 candidates) was the standout, appearing in two independent podcasts, a real signal, not one feed double-counted.
  • AlQuraishi returned (part two of the same interview thread), this installment focused on OpenFold and his "6 to 12 months" timeline for AI-designed models, rather than last week's "data not the model" argument, though the cheap-prediction economics reinforced it.
  • New names added to the watch list: NOETIK (patient-data biology foundation models, ex-Recursion founder) and the quantum-computing angle (Cleveland Clinic/IBM). Ginkgo (DNA), in our coverage universe, resurfaced with a concrete GPT-5 result.
  • Still silent, week over week: no follow-up on Chai's $400M round details or a fifth pharma partner; nothing new on federated learning / Aferis (last week's theme); still no confirmation of Anthropic hiring John Jumper or of Anthropic's preclinical programs; no coverage of Nabla Bio (the de novo-antibody name flagged last week); and a fourth straight week of zero podcast engagement with Schrödinger's Bunsen / Bristol Myers Squibb story.
  • Prices: SDGR +6% and LLY +3% on the week; RXRX roughly flat. LLY hit a fresh high before Novo's oral-Wegovy dosing study pressured it late.