The most seductive story in biotechnology right now is also, according to Daphne Koller, the most dangerous one. Koller, founder and CEO of insitro, opened the debut issue of Deep Phenotype1 with a pointed, two-part manifesto that takes direct aim at what she calls the AI Magic Wand narratives reshaping capital flows and scientific strategy across drug discovery. The argument is as precise as it is sobering: powerful AI aimed at poorly understood biology will not cure disease. It will simply accelerate failure.
The Bottleneck No One Is Talking About
Koller frames drug discovery in three stages: identifying a biological mechanism underlying disease, designing a molecule to hit that mechanism, and taking the drug through clinical development to patients. The overwhelming majority of AI investment and excitement has concentrated on stage two, the design of molecules. That focus, she argues, is profoundly misplaced.
"We are doing a pretty good job at manufacturing keys," Koller writes, "but they are often for the wrong locks. Even if AI lets us make better keys at an accelerating pace, that won't improve our ability to identify the right locks."
The numbers bear out the severity of the problem. Over 90% of drugs that enter clinical trials fail, a statistic that has barely moved in decades. In the large majority of cases, the molecule performed exactly as designed. The mechanism it targeted was simply wrong. Meanwhile, the industry has responded not by expanding into new biological territory, but by concentrating capital on already-validated targets. There are currently 37 targets with over 50 programs each competing against them. Novel target advances fell from roughly 100 per year in 2015 to approximately 30 in 2024.
What AI Can and Cannot Reason Through
Koller is equally direct in her critique of LLM-driven approaches to disease understanding, the proposition that sufficiently powerful reasoning across published scientific literature will surface new mechanistic hypotheses. The fatal assumption embedded in that thesis, she argues, is that the necessary data already exists.
"Human biology is incredibly complex," Koller notes, "spanning multiple interconnected biological layers — DNA, protein, cells, multi-cellular environments, entire organisms." Biology was not engineered. It is the product of billions of years of stochastic evolution, producing variation too vast and too idiosyncratic to be reasoned about from existing published data alone. It must be measured. And the measurements, at the scale required, largely do not yet exist.
The largest cell atlases assembled to date span hundreds of millions of cells and remain orders of magnitude too small to cover the relevant biological space. Virtual cell efforts, while promising, sample only a vanishing fraction of possible perturbations. And critically, they do not address the deeper challenge: connecting biological mechanisms to human clinical outcomes.
The Agentic Loop Problem
Even the most sophisticated argument for AI in drug discovery, the closed-loop agentic laboratory, runs into a structural wall that Koller identifies with clarity. Agentic systems thrive when feedback is fast, cheap, and accurate. Drug development offers the opposite. "The ultimate scorecard," she writes, "cannot be captured well by computational models or high-throughput assays. The only true ground truth is a human clinical trial."
Accelerating the wrong objective function does not get medicine closer to patients. It scales the failure pipeline.
The One Pathway That Actually Works
Koller is not dismissive of AI in drug discovery broadly. She acknowledges real and meaningful gains available in clinical operations, toxicology prediction, patient identification, and regulatory filing efficiency. But she is precise about where the genuine transformation lies: in biological mechanism. A deep mechanistic understanding of disease enables the identification of clinical readouts that select responsive patients, confirm target engagement, and detect early biological signals of efficacy. "Better trials," as she puts it, "are downstream of better biology."
Five Key Takeaways for Advisors and Investors
- The mechanism problem, not the molecule problem, is what drives clinical failure. Investment narratives centered on AI-accelerated molecular design are addressing a secondary bottleneck, not the primary one.
- Novel target attrition is a systemic risk. The collapse in novel target advances from 100 to 30 per year signals that the industry is recycling conviction rather than generating new biological insight. That is a long-term pipeline problem.
- Agentic AI in drug discovery requires the right objective function first. Without a validated proxy for clinical benefit, automation accelerates the wrong outcomes. Evaluate AI drug discovery platforms on whether they are solving for mechanism or molecule.
- Human-specific diseases represent the hardest and most underfunded frontier. Rodents do not get Alzheimer's. Primates do not recapitulate ALS. The diseases with the least scientific progress are precisely those most resistant to model-organism and computational shortcuts.
- The companies building causal human biology datasets are building durable moats. Koller's thesis implies that proprietary, large-scale perturbational data in human-relevant cellular contexts will be a foundational and scarce asset in the next generation of drug discovery.
Footnote:
1 Koller, Daphne. "Drug Discovery Has No Magic Wands." Deep Phenotype, insitro, 3 Aug. 2026, https://deepphenotype.substack.com/p/drug-discovery-has-no-magic-wands.
