ENHANCING PROTEIN STRUCTURE AND FUNCTION PREDICTION THROUGH DEEP MULTIPLE SEQUENCE ALIGNMENTS

Abstract

This paper provides an overview of deep learning algorithms and discusses its potential future advancements. DeepMSA represents a transformative approach to constructing multiple sequence alignments (MSAs) by integrating deep learning techniques with iterative database searches. Leveraging extensive genomic and metagenomic datasets, DeepMSA refines traditional MSA methodologies, significantly enhancing alignment quality for remote homology and complex protein structures. The framework utilizes pre-trained sequence embeddings and neural network-based optimization to improve contact prediction, secondary structure inference, and fold recognition. Comparative benchmarks, such as CASP competitions, demonstrate DeepMSA's superiority over traditional methods like PSI-BLAST and Hblits, with improved SP scores and better tertiary structure modeling. The introduction of DeepMSA2 further advances this methodology by incorporating diverse databases (e.g., Uniclust30, MGnify) and hybrid MSAs for multimer proteins, achieving state-of-the-art performance in predicting both monomeric and complex structures. These results highlight DeepMSA's pivotal role in bridging MSA construction and downstream applications in computational biology, offering a robust platform for protein structure prediction, evolutionary studies, and functional annotation.