Using deep learning to annotate the protein universe.

Area of Science:

• * Computational biology
• * Bioinformatics
• * Structural biology

Background:

• * Predicting protein function from amino acid sequences is crucial for biological and translational research.
• * Current alignment-based methods fail to annotate approximately one-third of microbial protein sequences, limiting data utilization.
• * Gaps in functional annotation hinder the exploitation of genomic data from diverse organisms.

Purpose of the Study:

• * To develop and evaluate deep learning models for accurate functional annotation of unaligned amino acid sequences.
• * To assess the models' ability to infer evolutionary patterns and cluster novel protein families.
• * To integrate deep learning with existing methods for enhanced remote homology detection.

Main Methods:

• * Training deep learning models on a comprehensive dataset from the Protein Families database (Pfam).
• * Rigorous benchmark assessments using 17,929 Pfam families to evaluate model performance.
• * Combining deep learning predictions with established alignment-based techniques.

Main Results:

• * Deep learning models accurately predict functional annotations for unaligned sequences, inferring evolutionary substitution patterns.
• * Models successfully cluster sequences from previously unseen protein families.
• * Integration with existing methods significantly improves remote homology detection, indicating complementary information capture.
• * Pfam coverage is extended by over 9.5%, surpassing additions from the past decade.
• * Functional annotations were predicted for 360 human reference proteome proteins lacking prior Pfam annotation.

Conclusions:

• * Deep learning offers a powerful approach to overcome limitations of alignment-based protein function prediction.
• * These models enhance protein annotation tools, significantly expanding the functional knowledge of protein databases like Pfam.
• * The findings suggest deep learning will be integral to future protein annotation strategies and biological discovery.