Drug Discovery and AI

Before a drug reaches a pharmacy shelf, it passes through distinct stages that together typically take a decade or more, and the vast majority of candidates that enter the pipeline never make it out the other end.

Roughly nine in ten candidates that make it as far as human trials still fail somewhere along the way — most often in the phase built specifically to test whether the drug actually works, rather than the earlier phase that just checks whether it is safe. Nothing about a faster start changes how unforgiving that filter is.

Candidate discovery is where AI has made the clearest dent. Searching the space of chemically possible molecules by hand or by brute-force lab testing is slow — a model that has learned which molecular shapes tend to bind to a given target can narrow millions of theoretical candidates down to a shortlist to synthesise and test, in a fraction of the time a purely manual search would take.

Target identification benefits too, in a smaller way — models that scan research literature and biological data can surface plausible targets a human researcher might not have connected on their own.

Both are real, and both are also two of the narrowest stages in the pipeline above — which is exactly why the improvement doesn't show up as a shorter overall timeline yet. Be precise about which tool did which job before crediting AI with the whole pipeline.

DeepMind's AlphaFold is a well-known example of AI in this space, and be exact about what it did, because the popular version of the story credits it with more than it delivered. Predicting the three-dimensional shape a protein folds into, from nothing but the sequence of amino acids that make it up, was an open problem biologists had chased for decades. Solving it well enough to compete with a physical lab measurement was a scientific breakthrough, and the resulting database now holds predicted structures for well over 200 million proteins, most of which had never been measured directly at all.

Knowing a target protein's shape is useful for candidate discovery — it lets a model reason about which molecules might physically fit against it, rather than searching blind. It is also a smaller piece of the pipeline than the headline suggests.

Trial design is a third category, distinct from finding a target and generating candidate molecules, and it is where a newer wave of AI use is concentrated. Matching eligible patients to a trial used to mean a coordinator manually checking medical records against a long list of inclusion and exclusion criteria — a model that scans records at that scale can surface eligible patients in days instead of months, which matters enormously for a trial that cannot start recruiting until enough patients are found.

Adaptive trial designs go a step further, using interim results to adjust which dosing arms keep enrolling patients while the trial is still running, rather than waiting for a fixed design set in advance to finish before learning anything. Both speed up parts of running a trial. Neither shortens the part that actually takes the years — following enough patients for long enough to know whether a drug works and stays safe, which is a biological constraint no amount of faster matching or smarter enrolment gets around.

Nothing about a faster shortlist changes what has to happen after it. A candidate a model is confident about still has to survive lab testing, animal testing, and three separate phases of human trials, because none of those stages exist to find a good molecule — they exist to prove a specific molecule is safe and effective in an actual human body, which a model trained on existing data cannot certify on its own.