Predicting and designing proteins (3): How structure prediction became end to end

By the late 2010s, protein structure prediction was a chain of strong but distinct procedures. A sequence search found relatives. An alignment summarized their variation. A neural network predicted contacts, distances, and orientations. An optimizer then searched for coordinates that satisfied those predictions. Each stage had been improved, but information moved mainly in one direction. The coordinate search could not easily tell the alignment representation what it had learned, and a local error in one stage could propagate through the rest. End-to-end learning changed this organization. “End to end” means that the system is trained so that its internal representations jointly serve the final objective—in this case, accurate three-dimensional coordinates—rather than optimizing every intermediate stage independently. AlphaFold2 and RoseTTAFold did not eliminate alignments, pairwise geometry, or physical constraints. They put these sources of information into architectures where each…