cofolding model

A cofolding model is an AI model that lets you preview, inside a computer, what 3D shape two molecules will form when they actually come together and bind. It's a bit like rotating two puzzle pieces on a screen to see if they fit, before you ever carve them out and try them by hand.

This matters especially in protein design. Before synthesizing anything in a lab, researchers use a cofolding model to screen how a newly designed small protein (a binder) might bind to its target protein—what shape the complex would take, and how tightly it might hold. Out of hundreds or thousands of candidate designs, only the ones predicted to bind well are sent on to actual experiments, which saves enormous time and cost.

But a cofolding model's prediction isn't the final answer. The structure drawn inside the computer doesn't always match how the protein actually behaves in a glass vial, so whether the prediction was right can only be confirmed by testing the binding directly in the lab.

The article notes: "Claude pulled together the internet, a collection of reference materials, Google Drive, Slack, Gmail, a bioRxiv connector, and GPUs to run structure design, sequence design, and cofolding models all on its own." A common misunderstanding here is thinking that a cofolding model gives a final, confirmed verdict on binding—but in reality it's only a prediction step used to narrow down candidates before experiments, with the final call made through actual binding assays in the lab.