Deducing high-accuracy protein contact-maps from a triplet of coevolutionary matrices through deep residual convolutional networks
Fig 1
The architecture of TripletRes, which formulates the contact-map prediction as a pixel-level labeling problem, where a pixel in the image represents a pair of residue positions in the contact-map of the query protein.
Starting from the MSA generated for the query sequence, three L×L×441 feature matrices (also called tensors) are computed for the three sets of coevolutionary features (PRE, PLM, COV). Here, L is the length of the query sequence while 441 = 21×21 is the combination of all 21 amino acid types (including the gap) for two positions in the MSA. Each tensor is input to a separate ResNet, where the first layer reduces the number of feature channels from 441 to 64, followed by instance normalization and 24 consecutive residual blocks to get an L×L×64 tensor. Details of a residual block are shown on the right-hand side inset. The three tensors from the three ResNets are concatenated into an L×L×192 tensor to feed into a final ResNet. In this ResNet, the first layer again reduces the feature channels from 192 to 64, followed by instance normalization, and 24 residual blocks to get an L×L×64 tensor, which is further reduced to L×L×12. Finally, a softmax layer is used to scale the values in the tensor between 0 and 1 and to make the sum of all values for each pixel (i.e. residue pair) equal to one. Since a protein contact/distance map is symmetric, TripletRes averages the corresponding softmax output of residue pair (i,j) and (j,i) to get the final L×L×12 distance-map prediction, where 12 stands of the number of distance bins. The contact-map is obtained by summing up the first 4 distance bin.