K-means-based feature learning for protein sequence classification

Protein sequence classification has been a major challenge in bioinformatics and related fields for some time and remains so today. Due to the complexity and volume of protein data, algorithmic techniques such as sequence alignment are often unsuitable due to time and memory constraints. Heuristic methods based on machine learning are the dominant technique for classifying large sets of protein data. In recent years, unsupervised deep learning techniques have garnered significant attention in various domains of classification tasks, but especially for image data. In this study, we adapt a k-means-based deep learning approach that was originally developed for image classification to classify protein sequence data. We use this unsupervised learning method to preprocess the data and create new feature vectors to be classified by a traditional supervised learning algorithm such as SVM. We find the performance of this technique to be superior to that of the spectrum kernel and empirical kernel map, and comparable to that of slower distance matrix-based approaches.