Daphne Koller on Machine Learning, Biology and Drug Discovery

Daphne Koller tells Joel Hellermark that boredom at school led her to begin university at thirteen and a half while still attending high school. She was drawn to early personal computers because programming made a versatile machine respond to her instructions.

At Stanford she studied multi-agent reasoning and then uncertainty more generally. She came to believe that decisions under uncertainty require a model of the world, but that constructing such models by hand was too difficult.

That problem brought her to machine learning before it became central to AI. She jokes that the question is not when she started doing AI, but when AI started doing what she was already doing.

Koller began working with biological data in the late 1990s, initially because it posed more interesting technical questions than the standard datasets available to machine-learning researchers. Early projects included tuberculosis transmission and gene-regulatory networks.

Over time biology and medicine became interests in their own right. Her lab pursued both core machine learning and biological research, an unusual combination that sometimes surprised colleagues in either field.

Koller compares the potential role of AI in biology to calculus in physics: a mathematical foundation for prediction. She hopes better models will eventually predict the effects of interventions such as medicines, while acknowledging that biology remains difficult to predict reliably.

She sees two converging forces: progress in machine learning and growing biological and clinical datasets. Advances in cell biology, bioengineering, imaging and health records make it possible to measure systems at greater scale and resolution.

Koller uses the metaphor of an exponential curve to explain why progress can look slow before becoming visible. Her view that biological AI is approaching an inflection point is a forecast, not proof that specific treatments are ready.

Hellermark asks why insitro emphasizes data quality. Koller recalls finding published datasets unavailable, incompletely annotated or missing important experimental assumptions.

Even datasets that appear similar may use different assays and experimental conditions. Models can latch onto those differences instead of biology, so insitro combines generating biological data with analyzing it computationally.

Koller says laboratory automation can generate large single-cell datasets. Her team also selects a limited number of external datasets worth the effort to clean, harmonize and curate, rather than assuming every available dataset can be combined.

Hellermark asks about an experiment using pathology scores. Koller says the scores were high quality, but they represented what a trained human had been taught to notice.

When the model was no longer anchored to those scores, it found different patterns that might also be valuable. Koller compares human knowledge to illuminated areas under streetlights: useful, but not necessarily the only places to look.

The challenge is to distinguish a genuinely new signal from an artifact. Her team seeks confidence through explainability, replication in diverse cohorts and checks against independent data sources the algorithm has not seen.

On foundation models, Koller says task-specific refinement remains important. Models trained on ordinary web images provided a starting point for cell images, but fine-tuning on cells improved performance.

Domain assumptions can help further: cells have no fixed upright orientation, unlike many pictures of cars. She sees value in incorporating such biological properties into model design.

insitro collects multiple data modalities in both human and cellular studies. Koller says different measurements can reveal different aspects of the same system, making their combination more informative.

She describes bridging biological scales: MRI, tissue pathology and cellular measurements can be connected through approaches such as spatial biology, which measures proteins in their tissue context.

The ultimate output must be actionable, such as identifying a target to investigate for a potential drug. Koller distinguishes having a useful target from fully understanding every mechanism behind it.

She describes using machine learning to infer patterns of gene activity from pathology images. These synthetic layers can help researchers form hypotheses about pathways, but the inferred signals still require validation.

Building a shared language is a central organizational challenge. insitro brings together biologists, machine-learning scientists, automation engineers and drug-discovery experts, and expects open, constructive and respectful discussion.

Some colleagues bridge disciplines from the outset; others learn over time. Koller describes biologists learning programming and engineers learning enough cell biology to hold meaningful conversations.

Asked whether AI could make biologists unnecessary, she says no: biological data must be produced through carefully designed assays, protocols and instruments. Experts are also needed to ask the right questions and connect evidence to interventions.

Koller warns against technological hubris: entering a field and treating AI as a universal fix without respecting its difficulties and accumulated knowledge. Respect does not require accepting every existing practice without question.

They discuss Eroom’s law, the observed long-term decline in drug-development efficiency. Koller points to incomplete biological understanding and high failure rates as important contributors to cost.

Better predictions might help select targets and identify patients more likely to respond in a trial. Koller says this could improve efficiency, but disease progression and clinical development impose time limits that cannot simply be removed.

Drug development takes years, so she expects a dramatic overall change to take well over a decade. She cites earlier genetic-association evidence as an example of rigorous data analysis helping prioritize drug targets.

For her own work, Koller says the primary bottleneck is data rather than compute. She admires the UK Biobank’s long-term follow-up and wants a much larger, more diverse cohort with multiple measurements over time.

She argues that people who want to contribute their health data to research need better ways to opt in. A useful system would require consent, infrastructure and sustained coordination, not just a checkbox.

To attract machine-learning researchers, she emphasizes both patient impact and demanding technical work. The problems can be intellectually challenging even when they are not centered on the largest language models.

At insitro, experimental biology and AI can be interleaved: a researcher may combine cell editing, microscopy and computational analysis in one workflow. Koller describes this as working with two sets of technical building blocks.

Koller says building a company brings scientific, funding and personnel challenges, but she is motivated by the possibility of giving patients more healthy years. She mentions ALS and acknowledges that development timelines may be too long for some patients she has met.

Looking decades ahead, she hopes diseases can be defined more precisely by underlying biology, with earlier detection and treatments tailored to patient groups. She presents this as a hoped-for future, not a claim that all disease can be eliminated.