ACon: predicting protein contact-map using additive neural networks and subgroups balancing method

Proteins’ contact-map is a binary matrix which shows contact points between residues in a protein sequence that is very essential in defining tertiary structure of proteins. Although contact-maps have been studied vastly over the last decade, predicting contact-maps from protein’s sequence is a difficult task to do. Many algorithms are developed to predict contact-maps from protein sequences by using evolutionary information and physical constraints but the accuracy of almost all of them are not promising[1]. In this paper, we present a neural network based method called ACon. This method uses a multi-level additive neural networks to predict proteins residue-residue contact-maps. Inputs of each level of the network are provided by the output of previous levels. For training the unbalanced data of contact-map, we used subgroups method which shows better results over prevalent sampling methods. In this method, we keep all samples instead of randomly selecting a portion of them. We have devided our samples to seven groups. Each of these groups have trained separately. This method, divides these groups to L subgroups. Where L is the ratio of the number of non-contacts to the number of contacts in our contact map matrix. Each of these subgroups have trained separately and the results of each L subgroups have summed up by votting. ACon reaches the accuracy of 77 percent with TPR and SPC equals 74 and 78 percent respectively.

کلیدواژه ها:

protein structure prediction ، contact-map ، additive neural networks ، unbalanced datasets