To test this, we encode ~24,000 CASP structural models using different representations - contact maps (N(N-1)/2 pairwise binary labels) and core identity (N binary labels) for example - and ask: how well does each predict the accuracy of the backbone (LDDT)? 2/8
