AI drug discovery: Biology, not just code

The tech world is captivated by the idea that a superintelligence could cure cancer and other diseases. This premise, while seductive, rests on a flawed assumption: that we already understand human biology well enough for AI to find the cures. Daphne Koller, founder and CEO of Insitro, argues in a recent essay that while AI will transform health, it's not a magic wand. The real challenge lies in our fundamental knowledge gaps about biology. The original analysis can be found on the a16z Blog.

Drug discovery is typically broken into three stages: identifying a disease-to-mechanism link, designing a mechanism-to-drug molecule, and then navigating the drug-to-patient clinical trial process. Most AI efforts have focused on stage two, spurred by successes like AlphaFold in protein structure prediction. AI can indeed design novel molecules and therapies at unprecedented speed. However, Koller points out that the biggest hurdles aren't in molecule design but in the first stage: understanding which biological mechanisms are actually responsible for disease and amenable to intervention.

The Undruggable Target Fallacy

While AI excels at designing molecules for known targets, many diseases remain untreatable because the underlying biological mechanisms are unknown. We have a surplus of potential drug candidates, but they often target the wrong biological pathways. This is why over 90% of drugs entering clinical trials fail, a rate that has stagnated for decades. The industry’s response has been to double down on a few well-understood targets, leading to a decline in the number of novel targets pursued annually. This misallocation of resources delays progress for millions suffering from diseases with no effective treatments.

Bridging the Data Chasm

The optimism around AI reasoning over vast biological literature overlooks a critical point: the data itself is insufficient. Human biology is an immensely complex, multi-layered system shaped by billions of years of evolution. Understanding it requires not just abstract reasoning but extensive, causal, perturbational measurements. Current efforts, even those building massive cell atlases or 'virtual cells,' sample only a fraction of the biological space. Furthermore, many human-specific diseases, like Alzheimer's or ALS, are poorly recapitulated in animal models, making data collection expensive, scarce, and ethically complex.

StartupHub.ai data indicates that while platforms like Substack have a score of 10/100 for their current utility in scientific communication, other platforms like Rillet (54/100) and Ghost (53/100) offer more robust features for creators. This highlights a broader trend: the tools that facilitate scientific discourse and data sharing are still evolving, mirroring the challenges in drug discovery itself.

The Slow Feedback Loop of Clinical Trials

The idea that AI agents can rapidly iterate in a closed loop with lab automation, similar to coding or molecular design, is appealing. However, drug development's ultimate feedback loop, human clinical trials, takes years and millions of dollars. This slow, ethically bound process cannot be fundamentally accelerated by compute alone. Agentic AI excels where there is a fast, accurate, and cheap scorecard. In drug discovery, the true scorecard is patient benefit, a metric that is years away from any given candidate.

AI can indeed help in clinical development by improving predictive toxicology or patient identification. But these are operational efficiencies. The real acceleration comes from AI enabling a deeper mechanistic understanding of disease. This allows for the identification of novel clinical readouts that can select better-responding patients, confirm target engagement, and detect efficacy signals earlier. This capability is inseparable from solving the disease-understanding problem. Better trials are downstream of better biology.

This focus on deep biological understanding is reminiscent of the work being done at companies like Altos Labs, which are exploring AI in biology from the cellular level upwards. The challenges Koller outlines are central to the entire field of AI in biology, where translating complex biological data into actionable insights remains a formidable task.

Why This Matters

Koller’s analysis reframes the narrative around AI in drug discovery. It shifts the focus from AI’s potential to solve problems we already understand to its role in helping us understand problems we don't. For startups, this means the most impactful AI companies in biotech will be those that can generate high-quality, causal biological data and build models that truly capture disease complexity. For investors, it underscores the need to back companies with deep biological expertise, not just AI talent. The ultimate impact of AI on human health hinges not on the cleverness of algorithms, but on our ability to measure and interpret the intricate workings of life itself.