Jake Lawrence · AI Systems / Structural Biology / STS · AI Systems theme
AlphaFold made the whole protein universe legible, and it ships a per-residue confidence that tells you exactly where to trust it. Rotate a real prediction of p53 and watch its certain core give way to tails the model honestly cannot place.
An interactive essay on AlphaFold, DeepMind's protein-structure prediction system, read through the site's legibility lens. For fifty years a protein's amino-acid sequence was easy to read and its folded shape was not, so shapes were measured one at a time; decades of that work left on the order of 200,000 experimental structures against hundreds of millions of known sequences. In 2020 AlphaFold2 reached a median 92.4 GDT at the CASP14 blind test, near experimental accuracy, and the open AlphaFold Protein Structure Database now covers over 214 million sequences, free and CC-BY. The signature interaction is a real AlphaFold prediction of human p53, rotatable and colored by the model's own per-residue confidence (pLDDT): a compact high-confidence core flanked by disordered tails the model correctly marks as unplaceable, not failed. The reading, in the series' idiom: AlphaFold is the rare instrument that renders something newly legible and, in the same breath, publishes the borders of its own legibility. Content-as-code figures re-derive live and in CI; the structure parses from AlphaFold DB entry AF-P04637-F1. Covers AlphaFold 3 and the 2024 Nobel Prize in Chemistry, with the standing caveat that a prediction is not an experiment.
Classification as Infrastructure names the pattern where a category system does the quiet work of deciding what is visible while staying invisible itself. AlphaFold inverts it: a legibility instrument that ships a per-residue confidence, drawing the borders of what it can and cannot see onto the map itself instead of hiding them.
Seeing Like an AI Company reads a frontier lab's proposal to classify the AI industry from the inside and finds the problems it will not name. AlphaFold is the constructive counterpoint from the same industry: an AI system that renders something newly legible and publishes, residue by residue, exactly where not to trust it.
Two readings of legibility from opposite directions. The Legibility Gap measures what a wartime state cannot see about its own damaged homes; AlphaFold measures the protein universe into view and then marks, in low confidence, the regions where no single structure exists to be seen.
Back to the release table · See it on the network · Machine-readable catalog
I design and ship AI tools, full-stack apps, and data pipelines — end to end, to production. Tell me the problem in a sentence; I'll give you an honest read on fit within a day.
Work with me →