TALE: Transformer-based protein function Annotation with joint sequence-Label Embedding.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in genomics

Background:

• High-throughput sequencing generates vast data, but functional insights lag.
• Existing computational protein function annotation methods face limitations in applicability and generalizability.
• Reliance on non-sequence data or poor generalization to novel sequences/functions hinders progress.

Purpose of the Study:

• To develop a novel deep learning model for protein function annotation using solely sequence information.
• To enhance generalizability to novel sequences, species, and functions.
• To provide a high-throughput, accurate, and broadly applicable alternative to experimental methods.

Main Methods:

• Proposed Transformer-based protein function Annotation through joint sequence-Label Embedding (TALE) model.
• Utilized self-attention-based transformers to capture global sequence patterns for generalizability.
• Employed joint embedding of protein function labels (hierarchical GO terms) and sequence features for unseen functions.
• Combined TALE with a sequence similarity method (TALE+) for enhanced performance.

Main Results:

• TALE+ outperformed competing methods using only sequence input.
• TALE+ surpassed a state-of-the-art network-based method in two out of three Gene Ontology (GO) categories.
• Demonstrated superior generalizability to proteins with low sequence similarity, new species, and rare functions.
• Ablation studies confirmed the contributions of algorithmic components to accuracy and generalizability.

Conclusions:

• TALE offers a powerful, sequence-only approach for accurate protein function prediction.
• The model exhibits significant improvements in generalizability, addressing key limitations of current methods.
• TALE provides deep insights into protein sequence-function relationships and advances computational annotation capabilities.