Every technology on my beat eventually has to leave the demo and enter the exam room. This year it happens to the most hyped promise in science: drugs designed by artificial intelligence are now moving into human clinical trials, including the first candidates from the famous protein-folding lineage — the AlphaFold family of models that cracked biology's most famous prediction problem and collected a chemistry Nobel for it.
Since this desk exists to sort results from demos, let me set up the scoreboard properly.
The demo era, first, and credit where it is earned: the protein-structure breakthrough was real. For fifty years, "given this gene sequence, what shape does the protein fold into?" was a grand-challenge problem — shape is function in biology, and finding one structure experimentally could consume a PhD. The models collapsed that to minutes, then published predicted structures for essentially every protein science knows. That is one of the great computational results of the century, full stop. No calibration needed.
But a structure is not a drug. A drug is a molecule that binds the right protein, avoids ten thousand wrong ones, survives a liver hell-bent on shredding it, and does all this inside a seventy-kilogram chemical reactor with opinions. The old, brutal statistic of pharmaceutical R&D is that roughly nine in ten candidates that enter human trials fail — usually not because the chemistry was wrong but because the biology was more tangled than anyone's model of it. The industry's whole cost problem lives in those failures.
So the claim under examination this year is specific: that models which understand molecular shape can design candidates that fail less often. Not faster paperwork — better guesses. The AI-native labs have now pushed their first oncology and immunology candidates into early trials, and the money has followed at a scale that tells you how seriously the industry takes the question; the flagship AI-drug lab raised one of the largest private rounds in biotech history this winter to run the experiment.
Now the part every reader should tattoo somewhere: early trials cannot prove the thesis. Phase 1 measures safety in small groups; the drug-quality verdict — efficacy, the nine-in-ten gauntlet — arrives years out, in the phases with control arms. The honest scoreboard for 2026 reads: molecules in the exam room, no grades posted. Anyone declaring victory this year is selling something; anyone declaring failure is guessing.
What would move my scoreboard, in either direction? Three markers, posted in advance so December can grade me instead of the other way around. Up: trial candidates hitting targets the field previously considered "undruggable" — that is the shape-understanding advantage cashing out, and a couple of the current candidates aim at exactly such targets. Up: portfolio-level attrition visibly below the industry's historical nine-in-ten, once enough candidates accumulate to make the ratio mean something. Down: AI-designed candidates failing in the same old ways at the same old rates, which would file the whole revolution under "faster paperwork" — worth billions, but not the promise on the slide.
The refrigerator-decades lesson from my Nobel piece in the fall applies here with the sign flipped. Quantum circuits took forty years to go from curiosity to industry, and the patience paid. Drug discovery is attempting the reverse trick — compressing decades of trial-and-error into model weights — and the compression is precisely the claim the next few years will grade. Biology, unlike a benchmark, cannot be overfit. That's what makes this the best experiment on my beat: for once, the hype has agreed to be falsifiable.
Molecules in the room. Pencils down in a few years. This desk will post the grades as they come — both columns.
