Embeddings from deep learning transfer GO annotations beyond homology.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics and Proteomics

Background:

• Experimental protein function annotation is limited, with fewer than 0.5% of known proteins experimentally annotated.
• Existing computational methods for bridging the sequence-annotation gap rely on homology-based transfer or evolutionary information.
• A significant need exists for advanced computational strategies to accurately annotate protein functions.

Purpose of the Study:

• To develop and evaluate a novel method for predicting protein functions using protein sequence embeddings.
• To assess the performance of annotation transfer based on proximity in SeqVec embeddings versus sequence similarity.
• To explore the potential of deep learning language models for enhancing protein function prediction.

Main Methods:

• Utilized SeqVec, a deep learning language model, to generate protein sequence embeddings.
• Predicted Gene Ontology (GO) terms through annotation transfer based on protein proximity in the SeqVec embedding space.
• Replicated the conditions of the Critical Assessment of Function Annotations (CAFA3) challenge for performance evaluation.

Main Results:

• Achieved Fmax scores of 37±2% (BPO), 50±3% (MFO), and 57±2% (CCO) in CAFA3 conditions, comparable to top-performing methods.
• Demonstrated superior performance over naïve sequence-based transfer, even when restricting transfer to proteins with <20% sequence identity.
• Preliminary CAFA4 results suggest consistent performance and confirm the method's potential.

Conclusions:

• Annotation transfer based on SeqVec embeddings offers a powerful alternative to traditional homology-based methods.
• This approach shows particular promise for annotating proteins in smaller families or those with intrinsically disordered regions.
• The findings suggest a paradigm shift in protein annotation strategies, leveraging deep learning for functional prediction.