Deep and Self-Taught Learning for Protein Accessible Surface Area Prediction

ASA captures the degree of burial or surface accessibility of a protein residue. It is a very important indicator of the behavior of amino acids within a protein as well. It can be used to find protein interactions, interfaces, folding states, etc. Calculation of the ASA requires the presence of the structure of the protein. However, structure determination for proteins is expensive and requires significant technical effort. As a consequence, the prediction of ASA is a very important and fundamental problem in Bioinformatics and Proteomics. In this work, we have investigated self-taught machine learning methods along with deep neural network to predict the residue level accessible surface area (ASA) of a protein. We have found that deep learning neural networks can predict the ASA of the residues in a protein accurately. Furthermore, the proposed deep learning based method does not require the use of computationally demanding features such as the position specific scoring matrix (PSSM) which have been used in previous works. A simple Blosum62 matrix based position dependent representation of amino acids in a sequence window gives comparable performance. This is particularly attractive for proteome wide prediction of ASA. We have used various self-taught learning schemes for obtaining an optimal feature representation from unlabeled data. These include a sparse and regularized autoencoder neural network and a dictionary based learning scheme. We have used unlabeled data from the protein universe in an attempt to improve the feature representation. We have also evaluated the performance of a stochastic gradient based predictor of accessible surface area for different feature representations.