RAANMF: An adaptive sequence feature representation method for predictions of protein thermostability, PPI, and drug–target interaction

Abstract

The effective representation of sequence is essential for analyzing protein structure and function. Sequence representation based on reduced amino acids plays an important part in protein research, as it preserves key sequence features while simplifying feature processing. However, it is a challenge to select an appropriate reduced amino acid method for various downstream analysis tasks. Developing reduced amino acid methods that can adapt to various downstream tasks is essential to promote protein-related researches. In this paper, we propose a novel reduced amino acid method based on non-negative matrix factorization (NMF) named RAANMF, which can adaptively generate the reduced amino acid schemes for different tasks. Through validating the effectiveness and universality of RAANMF on three mainstream tasks including protein thermostability prediction, protein–protein interaction prediction, and drug–target interaction prediction, the results demonstrate that the reconstructed models using RAANMF to characterize amino acid sequences can achieve comparable or superior predictive performance with greatly reduced feature dimensions compared to the original models. Moreover, the interpretability of RAANMF which is analyzed from the perspective of the non-negative matrix clustering principle helps us understand the biological significance and enhances its credibility and utility in practical applications. As a method developed from NMF, RAANMF offers a straightforward and interpretable approach for extracting latent features, and it is expected to help study the relation of protein sequence, structure and function.