Unconventional training for neural network predictions of inter-residue contacts

The ab-initio prediction of protein three-dimensional structures (protein folding problem) from protein residue sequences is a problem that has involved a great deal of scientific investigation in the last decades, but the results obtained so far are not yet satisfactory. One of the most interesting sub problem related to protein folding is the prediction of residue-residue contact maps: a complete or partial knowledge of residue-residue contacts would greatly improve the protein folding recognition. At the state of the art, the most efficient methods for contact prediction are neural network based. One of the most basic contact predictor is CORNET. Despite ten years of investigation on this problem, the performances of CORNET are still comparable to the state of the art in contact prediction. In this paper we investigate three different approaches to improve the performances of CORNET, and in general of neural networks based predictors. In our tests we train the CORNET predictor on different classes of examples in order to understand if there are classes of contacts that can be better predicted with respect to the standard training. In particular, we restrict the training set according to specific residue features, building specialized predictors. The features highlighted are solvent accessibility, sequence separation and secondary structure. For each of these features we compare the resulting predictor with performances of the original CORNET, giving hints for the construction of better neural network predictors.