A family of feed-forward models for protein sequence classification

Abstract

Advances in sequencing have greatly outpaced experimental methods for determining a protein's structure and function. As a result, biologists increasingly rely on computational techniques to infer these properties of proteins from sequence information alone. We present a sequence classification framework that differs from the common SVM/kernel-based approach. We introduce a type of artificial neural network which we term the Subsequence Network (SN) that incorporates structural models over sequences in its lowest layer. These structural models, which we call Sequence Scoring Models (SSM), are similar to Hidden Markov Models and act as a mechanism to extract relevant features from sequences. In contrast to SVM/kernel methods, which only allow learning of linear discrimination weights, our feed-forward structure allows linear weights to be learned in conjunction with sequence-level features using standard optimization techniques. © 2012 Springer-Verlag.