Daphne Koller Published an Essay on Using AI to Find the Causes of Disease Before Designing Drugs
Daphne Koller Published an Essay on Using AI to Find the Causes of Disease Before Designing Drugs
On 3 August, insitro founder and CEO Daphne Koller published an essay examining where AI is already accelerating drug development. She distinguishes the task of finding the cause of a disease from the tasks of developing a drug and testing it in patients.
In the essay, Koller describes three steps. First, researchers connect a disease to a mechanism. They look for a gene, protein, cellular process, or interaction between tissues whose alteration changes the course of the disease in humans. Chemists and computational models then develop a treatment directed at that target, such as a small molecule, antibody, RNA therapeutic, or gene therapy. Finally, the clinical team enrolls patients and measures the treatment's benefits and adverse effects.
In Koller's assessment, AI works especially quickly when both the target and a measurable outcome are known. A model proposes a molecular structure, the laboratory tests whether it binds to the protein, and the next cycle modifies the structure. An automated laboratory can similarly improve protein synthesis or an antibody against a known target because each step produces a result quickly.
Identifying a disease mechanism requires a different kind of work. Within the body, cells change state, tissues affect one another, and the effect of treatment depends on timing and the individual patient. Koller describes a clinical trial as the definitive test of whether a drug changes the disease in humans. This test takes years and costs millions of dollars.
A report by BIO, Informa Pharma Intelligence, and QLS Advisors analyzed 12 728 clinical program transitions during 2011–2020. Of the programs that entered Phase 1, 7,9 percent ultimately received FDA approval. In Phase 2, where clinical effects are tested, programs advanced in 28,9 percent of cases. Koller places the selection of the biological mechanism at the beginning of this sequence.
Koller also writes that 38 targets now have more than 50 programs each, while the number of new targets advanced by the industry each year fell from about one hundred in 2015 to thirty in 2024. She attributes this concentration to companies returning to areas of biology where they already know how to design experiments and select patients.
Koller therefore proposes using AI to map the causes of disease. Building such a map requires measurements from human cells and tissues, experiments that alter those cells and tissues, and data linking the resulting changes to patients' conditions. In Koller's model, this map can help researchers select a target, identify suitable participants for a trial, and determine earlier whether a treatment changes the biology of the disease.
For age-related diseases, this question must be addressed before a molecule is designed. These diseases affect many cell types and organs. Researchers must first determine which process in the human body the treatment needs to change.