Feature learning using stacked autoencoders to predict the activity of antimicrobial peptides

Abstract

In recent years, pattern recognition methods have been applied to determine the activity of biological molecules, including the prediction of antimicrobial activity of synthetic and natural peptides where Quantitative Structure-Activity Relationship methodologies are widely used. Traditionally, works focused on designing descriptors for sequences to yield better correlations with the biological activity and improve predictors performance. Albeit there have been remarkable results, the small size of available datasets leave large room for improvement. In this work, rather than hand-crafting new descriptors, our approach consists in automatically learning them from existing ones. We use stacked autoencoders (a class of unsupervised neural networks), and the descriptors learnt are fed to a support vector regression task to predict biological activity. This method improves results in existing literature by roughly 12% simultaneously in different metrics, providing interesting insights into the nature of descriptors learnt and suggesting its applicability in other areas in protein properties prediction.