Coevolution Contact Map: How AI Reads Structure from Sequence (2D)

This is the statistical mechanism underneath AlphaFold, rendered as the contact-map heatmap itself: residues that touch in 3D tend to mutate together across a family of related sequences, so mining a multiple sequence alignment for mutual information between columns predicts which residue pairs are in contact — before any 3D coordinate exists. The simulator builds a hidden synthetic protein with a known helix-turn-helix fold (used only to define ground truth, never shown directly), generates an MSA carrying a coevolution signal (weakened by an adjustable mutation-noise rate), computes real mutual information for every eligible residue pair from the generated sequences alone, and renders the resulting NxN coupling heatmap live next to the hidden true contact map. Live readouts track the precision and recall of the top-K highest-scoring pairs against the known ground truth and the mean MI strength of the contacts actually selected — showing directly how MSA depth and noise determine whether the statistically inferred map can recover the hidden structure at all.