Coevolution Contact Map: How AI Reads Structure from Sequence

Sequence Alignment

MSA depth (sequences) 150

Mutation noise 35%

Top-K contacts used 24

Show true structure (ghost)

Live Readouts

Contact precision–

Distance RMSD to truth–

Mean MI of top-K (bits)–

Backbone (folding model) Predicted contact (true positive) Predicted contact (false positive) Ghost = true 3D fold

How it works

Real proteins that fold correctly show coevolution: if residue i mutates, a spatially contacting residue j often mutates too, to keep the pair chemically compatible. Across a deep multiple sequence alignment (MSA) of homologous sequences, this leaves a statistical footprint even though i and j can be far apart along the chain.

Mutual information between columns i, j:
MI(i,j) = Σ_a Σ_b p(a,b) · log₂[ p(a,b) / (p(a)·p(b)) ]

a, b = residue states observed at columns i, j
p(a,b) = joint frequency across the M sequences
High MI ⇒ i and j are statistically coupled ⇒ likely in 3D contact

This simulator generates a toy helix-turn-helix protein (32 residues) with a known 3D fold, then synthesizes an MSA where a subset of true contact pairs carry a coevolution signal, corrupted by the noise slider. It ranks all distant residue pairs (|i−j| ≥ 4) by MI and keeps the top-K as predicted contacts — exactly the pairwise "distogram" signal that AlphaFold's Evoformer extracts from real MSAs.

Those predicted contacts are then fed into a live physics relaxation: backbone springs hold consecutive residues at the correct bond length, contact springs pull predicted-contact pairs together, and a soft repulsion keeps the chain from collapsing — turning a 2D contact map back into a 3D structure, the same job AlphaFold's structure module performs from its predicted distances.

MSA depth — more sequences average out sampling noise, so MI estimates converge to the true coupling and precision rises (just as real structure prediction needs deep alignments).
Mutation noise — degrades both the background signal and the coevolutionary coupling strength; push it high enough and even a deep MSA can't recover the fold.
Top-K — how many of the highest-MI pairs are trusted as contacts; too few underconstrains the fold, too many admits false positives that distort it.

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.