How AI Is Accelerating Drug Discovery and Where the Hype Ends

Drug discovery is slow by design. The process of identifying a promising compound, testing it in cells, testing it in animals, moving through three phases of human trials, and finally getting regulatory approval takes roughly twelve years and costs somewhere between one and two billion dollars. The vast majority of drug candidates fail somewhere along the way. The system is built to catch failure before it reaches patients, and that caution comes at enormous expense.

AI is beginning to change this, though probably not in the way most headlines suggest.

The moment that shifted the field's understanding of what machine learning could do came in 2020, when DeepMind released AlphaFold2. For fifty years, biologists had been trying to solve the protein folding problem: given a protein's amino acid sequence, predict the three-dimensional shape it folds into. Shape determines function. If you know a protein's shape, you can start designing molecules that interact with it. AlphaFold solved this problem with extraordinary accuracy across almost every known protein structure, releasing predictions for over 200 million proteins into a public database.

That was genuinely revolutionary. Drug discovery has historically required crystallography or cryo-electron microscopy to determine protein structures, processes that take months and significant expertise. AlphaFold made the structural prediction step nearly instantaneous.

But structural prediction is only one step in a process that has many. Knowing the shape of a protein does not tell you how to design a molecule that binds to it safely and effectively. It does not predict how that molecule will behave in a living system, how it will be metabolized, whether it will cross the blood-brain barrier, or whether it will cause side effects that only appear after years of use.

Companies like Insilico Medicine, Recursion Pharmaceuticals, and Exscientia have been racing to automate more of the pipeline. Insilico made headlines in 2023 when it moved a drug candidate from AI-generated hypothesis to Phase 1 clinical trials in under 18 months, a fraction of the normal timeline. Recursion uses computer vision to analyze how cells respond to drug candidates at massive scale, generating data that would take traditional labs years to collect.

The honest picture is that AI has meaningfully shortened the early stages of drug discovery. It is helping researchers generate better candidates faster and screen out obvious failures earlier. What it has not done is change the fundamental biology of human clinical trials, which still require time, large populations, and careful monitoring that no software can accelerate.

There is also a concern that the enthusiasm for AI drug discovery is outpacing the evidence. Several high-profile AI-designed compounds have entered trials and failed, which is normal for drug development but raises questions about whether AI screening is actually better at predicting clinical success or just better at predicting early-stage lab results.

My thoughts: AI already looks useful early in the process, especially for narrowing down targets and molecules. The bigger claim, that it can reliably turn early discovery into approved drugs faster, still needs more proof. Drug development has a way of humbling every shortcut.