• Invent an unsupervised method to extract structure and coevolutionary information from protein language model (e.g., ESM2)
• Mutate each residue with 20 AAs, compute logits using ESM2, and subtract the WT logit to get an [L,20] logit matrix for each residue
• Iterate over all residues to compile a [L,20,L,20] Jacobian matrix. Apply Average Product Correction to get the final pairwise weight matrix for contact predictions