The AI Drug Discovery Race: How Silicon Valley Is Trying to Cure Every Disease
In February 2024, the US Food and Drug Administration approved a drug called Rentosertib for the treatment of idiopathic pulmonary fibrosis: a devastating lung disease with few effective treatments. The approval was notable not primarily for the drug itself but for how it was found. Rentosertib was identified by an AI system developed by Insilico Medicine, a drug discovery company that had used machine learning to predict that the molecule would be effective, design its structure computationally, and suggest a synthesis pathway: all steps that traditionally required years of laboratory chemistry and biological screening.
From initial AI-assisted identification to FDA approval took approximately seven years: still a long time by ordinary standards, but dramatically faster than the typical drug discovery timeline of twelve to fifteen years. And the pipeline behind it is substantial: dozens of AI-discovered or AI-optimized drug candidates are in clinical trials across therapeutic areas ranging from cancer to Alzheimer's to rare genetic diseases. The FDA approval of Rentosertib is not an endpoint. It is evidence that a fundamentally new approach to drug discovery is working.
Traditional drug discovery is a process of searching an almost incomprehensibly large space. The number of possible small molecules with drug-like properties: what chemists call chemical space: is estimated at ten to the sixtieth power, or roughly ten trillion trillion trillion trillion trillion molecules. Laboratory screening can test perhaps a few million per year against a target of interest, with heroic effort. The odds of finding a molecule that is both effective and safe in this space by anything resembling random search are astronomically bad.
The process that has historically worked is expert-guided search: medicinal chemists with deep knowledge of how molecular structure relates to biological activity making educated guesses about which directions in chemical space to explore, testing those guesses, and iterating based on results. This process is effective but slow: it typically takes two to five years to identify a candidate molecule worth advancing to clinical trials, and only one in ten candidates that enter clinical trials ultimately receives approval. The total time from initial research to approved drug averages twelve to fifteen years, and the total cost averages over a billion dollars per approved drug.
AI approaches to drug discovery are fundamentally different from this expert-guided random walk. Machine learning models trained on databases of known molecular structures and their measured biological activities can predict, with useful accuracy, how novel molecules will behave: their binding affinity for specific proteins, their toxicity profile, their solubility, their metabolic stability: before any laboratory testing occurs. This transforms the search from a primarily experimental process to a primarily computational one, with laboratory work focused on the most promising candidates rather than distributed across a broader exploratory space.
AlphaFold, DeepMind protein structure prediction system released in 2021, was the breakthrough that made much of this possible. Proteins are the molecular machines of biology: enzymes, receptors, ion channels: and drug molecules work by interacting with specific proteins in specific ways. Predicting what a drug-protein interaction will do requires knowing the three-dimensional structure of the target protein. For decades, determining protein structure experimentally was the rate-limiting step in structural biology: extraordinarily expensive, slow, and limited to proteins amenable to experimental structure determination.
Please enable JavaScript to read the full article.