CCMpred — Supplementary Information

1 Markov Random Field Model The following section will outline the mathematical model in our contact prediction method that is essentially identical to the plmDCA (Ekeberg et al., 2013) and GREMLIN (Kamisetty et al., 2013) methods. We eliminate transitive interactions in the observed interaction network by learning a generative model of the MSA using a Markov Random Field (MRF). Assuming we have an MSA xi (n sequence index, i position index) with L columns and N sequences, we represent columns in the MSA as the vertices with single-residue emission potentials εi(a) (with i being a column index and a ∈ {1..20} representing the twenty possible amino acids) and covariation between columns as the edges with pairwise emission potentials εi,j(a, b) (with i and j column indices, a and b amino acid indices). The network of co-evolution can then be learned by maximizing the likelihood of observing the sequences in the input MSA, given the model parameters ε: