Anca Petre
In 1964, two chemists sat down with a stack of paper and tried to write an equation for why one molecule works and another one does not. Corwin Hansch and Toshio Fujita had no computer worth the name. What they had was a conviction that the behaviour of a molecule could be reduced to numbers, and that those numbers could be compared.
That conviction is the whole story. Everything that followed, the expert systems of the seventies, molecular docking in the eighties, deep learning in 2012, AlphaFold in 2021, is a longer and better funded version of the same bet: that chemistry can be turned into data a machine can reason about.
The first attempt was a filing system
DENDRAL, built at Stanford from 1965 onwards, was not designed to invent anything. It was designed to read mass spectrometry output and propose which molecular structures were consistent with it. It worked, narrowly, and it established the pattern that held for the next thirty years: computers were good at sorting through what already existed, and useless at proposing what did not.
Every generation believed it had crossed from sorting to creating. Every generation was one layer of abstraction away. Tech Anatomy · episode 134
Docking software arrived at UCSF in the early eighties and turned the lock and key metaphor into geometry. It also inherited the metaphor's flaw: proteins are not locks, they move. The correction, induced fit, took another decade to become standard, and it is the reason a generation of predictions aged badly.
Between 1950 and 2010, the number of new drugs approved per billion dollars of research spending halved roughly every nine years. Eroom's law, Moore's law backwards, is the pressure behind every computational promise made since.
2012, and the part everyone remembers wrong
The Merck Kaggle challenge is usually told as the moment deep learning arrived in pharmaceutical research. It is more accurate to say it was the moment the industry noticed. The winning margin was real but modest, and the models were trained on datasets that would embarrass a modern team. What changed was not accuracy. It was who started paying attention.
What followed, generative models proposing molecules that had never existed, is the first genuine break in the pattern. For the first time the machine was not filtering a catalogue. It was writing one.
What to watch now
Rentosertib, the first drug with both an AI discovered target and an AI designed molecule to report phase two results, is the honest test case. Not because one trial settles anything, but because it is the first time the full claim can be checked end to end rather than argued about.
Sources
- Hansch C, Fujita T. J Am Chem Soc. 1964;86:1616-1626.
- Scannell JW et al. Nat Rev Drug Discov. 2012;11:191-200.
- Jumper J et al. Nature. 2021;596:583-589.
