2 Interpretation of Coupling Matrices

Contact prediction methods learning a Potts model for the MSA of a protein familiy, map the inferred 20 x 20 dimensional coupling matrices \(w_{ij}\) onto scalar values to obtain contact scores for each residue pair as outlined in section 1.2.4.6. As a result, the full information contained in coupling matrices is lost, such as the contribution of individual couplings \(\wijab\), whether a coupling is positive or negative, higher order dependencies between couplings or possibly biological meaningful signals. The following sections give some intuition for the information contained in coupling matrices.