AI Is Making Drug Pipelines Bigger. It Might Also Be Making Them Worse.

Daphne Koller did not mince the diagnosis. In a LinkedIn post this week, the founder of insitro and Stanford professor agreed with venture capitalist Bruce Booth’s uncomfortable prediction: AI will make the pharma industry’s R&D productivity worse before it makes it better. Her explanation was precise. “AI for molecular design is trained on known chemistry against known targets, so it produces more of that.” More molecules. More candidates. More of the same failures, arriving faster and cheaper than ever before. The pipeline fills up, the Phase II clock still runs two to three years, and the 90% clinical failure rate holds exactly where it has for decades.

That number is worth sitting with. Nine in ten drugs fail in the clinic. Faster chemistry does not move it. And yet the industry’s current AI investment thesis is concentrated almost entirely on making chemistry faster.

The Abundance Trap

The analogy Koller reached for is precise and damning. She cited the 2001 “Fruits of Genomics” report, which described target abundance without validation. The genomics era handed researchers thousands of potential targets and very little understanding of which ones actually drove disease. Two decades later, the industry is repeating the structure of that mistake at the molecular layer. Models like Chai-2, IsoDDE, and BoltzGen, cited by Cameron Sellers of Creandum in his announcement of an $11 million seed round in Atlas Bio, are “transforming what’s possible in pre-clinical R&D.” That is true. What Sellers also acknowledged, almost parenthetically, is that the trial itself is unchanged: a Phase II still takes two to three years, costs $20 million or more, and fails roughly 70% of the time.

Here is the counterintuitive read that the bullish AI-in-drug-discovery coverage consistently misses. The bottleneck in pharma R&D was never chemistry. Medicinal chemists were not the reason 90% of drugs failed. The reason is that most programs advance into humans testing a mechanism that does not actually drive the disease in the patients being enrolled. Speed up the part of the process that was not broken, and you accelerate directly into the part that is.

Koller puts the structural problem cleanly: “Thirty programs against one target, all enrolling all-comers, compete for the same patients, and most will not beat the standard of care.” That is not a hypothetical. Look at any crowded oncology indication right now, any PD-1 combination strategy, any KRAS program. The competitive crowding that Booth worries about is already visible in enrollment competition across sites. Now layer on a generation of AI-enabled candidate generation tools, and the crowding compounds before any of those programs read out.

Where the Real Gap Lives

The fix, in Koller’s framing, requires building a causal map: which mechanism drives which disease in which patients. That map is simultaneously a biology problem and an AI problem, and the data required to train it largely does not exist at the scale models need. This is where the industry’s AI investment is genuinely thin. The capital is flowing toward generative molecular design. The gap is in causal human biology, in understanding patient response at a mechanistic level before a single vial is administered.

insitro’s own MASH program illustrates what that alternative looks like in practice. The target, IRS1, came out of human genetics linking the mechanism to both liver fat and liver fibrosis, read at scale using AI. The molecule itself, a liver-targeted siRNA, is fairly standard design. Novelty in the biology, not the chemistry. Koller also pointed to David Fajgenbaum, who put his own Castleman disease into remission with sirolimus, a drug approved decades earlier for transplant patients. The repurposing insight was entirely biological. The chemistry was four decades old.

Simon Romanski, co-founder of Atlas Bio, which emerged from stealth this week with $13.2 million in combined funding from Creandum, First Momentum Ventures, and LEA Ventures, is building toward exactly this gap from a different angle. In his announcement post, Romanski argued that “the trial shouldn’t be where you find out if a drug is working on a patient.” Atlas Bio’s foundation model runs in silico clinical trials, predicting patient responses across treatments and indications before first-in-human work begins. Per Romanski’s post, the models already generate insights “uncovering effects that people later discovered in the clinic,” with immunology partnerships already informing clinical decisions for partners and customers.

That framing matters because it locates the value upstream of the trial, not inside it. The standard AI-in-trials pitch, which Sasha Sinclair called out with surgical precision in her post this week, is optimizing the operational shell around a broken core. “45% of conversations I’ve had with founders all claim” to reduce site burden with AI, she wrote, and each one arrives with the same slide about faster enrollment. Faster enrollment into a trial testing the wrong mechanism in the wrong patients is not a productivity gain. It is a more efficient path to a null result.

Counting the Wrong Things

The metric problem runs deeper than any single company’s pitch deck. Koller’s sharpest line in the post is a direct challenge to how the industry defines progress: “The field gets the kind of abundance it measures.” Right now, the field measures candidates per year, molecules generated, programs initiated. Press releases celebrate pipeline expansion. Investors count shots on goal. None of those metrics distinguish between a novel validated target in a defined patient population and the thirty-first inhibitor of a mechanism that has already failed to move overall survival in an all-comers trial.

Koller proposes three alternative metrics: novel validated targets, patient populations newly defined by mechanism, and first effective molecules against targets that have resisted every prior attempt. Those are harder to generate, harder to count, and structurally underincentivized in a capital environment that rewards velocity. A seed round announcement tied to “results in weeks rather than years,” as Sellers described Atlas Bio’s in silico output, is a compelling story. A five-year investment in causal human biology infrastructure, with no molecule at the end of it, is a harder fund to raise.

That misalignment between what the industry measures and what actually determines whether a drug reaches a patient is not a new observation. But AI acceleration makes it urgent in a way it was not before. When generating a new candidate took years of medicinal chemistry, the pipeline had a natural rate limiter. The queue moved slowly enough that bad ideas fell out before consuming too much capital. Strip that limiter out, and the queue floods. A Phase II that still takes two to three years and costs $20 million or more does not compress because the molecule was generated in weeks. The translational gap does not close. The clinical failure rate holds. What changes is volume and, with it, the competitive noise that makes every signal harder to read.

The sponsors who will come out ahead are not the ones running the most programs. They are the ones who invest now in the biology infrastructure that tells them which programs to kill before the IND is filed. The industry built faster guns. It still needs better targets.

References

1. Daphne Koller, LinkedIn post on AI drug abundance and R&D productivity, 2025
2. Simon Romanski, LinkedIn post announcing Atlas Bio emergence from stealth, 2025
3. Cameron Sellers, LinkedIn post announcing Creandum’s $11M seed investment in Atlas Bio, 2025
4. Sasha Sinclair, LinkedIn post on undifferentiated AI clinical trial pitches, 2025