Prediction of compound-protein interactions based on deep learning methods

Abstract

As the number of potential compound-protein interactions (CPIs) that could be assayed is essentially infinite, brute-force experimental screening for CPIs is highly wasteful. Attention has thus been given to CPI prediction models that can guide researchers to fast lanes for hit discovery. Existing CPI prediction models have mostly used a curated database of interactions for building a single fixed model, with the Support Vector Machine (SVM) often used for model construction. On a dataset of 100,000 CPIs, the SVM can train a model in less than one day. Yet the size of available datasets can be in the millions, and since SVMs require an exponential increase in resources, model construction on such datasets is infeasible. We investigated the ability of Deep Learning to handle large volumes of CPIs that cannot be processed by SVMs. Deep learning does not require learning on all input data at once as in the standard SVM, but rather the model is iteratively tuned over the course of data input. We evaluate the learning error rate as a function of the number of learning iterations, our method based on Deep learning outperformed the method based on SVM.