TEMPROT: protein function annotation using transformers embeddings and homology search.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• High-throughput sequencing generates vast protein data, necessitating computational methods for function analysis.
• Existing sequence-based protein classification methods have limitations, particularly regarding input protein length.
• Accurate protein function prediction is crucial for understanding biological processes and disease mechanisms.

Purpose of the Study:

• To introduce TEMPROT, a novel computational method for protein function prediction.
• To present TEMPROT+, an ensemble method combining TEMPROT with BLASTp for enhanced prediction accuracy.
• To evaluate the performance of TEMPROT and TEMPROT+ against state-of-the-art methods.

Main Methods:

• TEMPROT utilizes fine-tuning and embedding extraction from pre-trained protein sequence architectures.
• TEMPROT+ is an ensemble of TEMPROT and BLASTp, a local sequence alignment tool.
• Model performance was evaluated on a dataset derived from the CAFA3 challenge database.

Main Results:

• TEMPROT and TEMPROT+ demonstrated competitive performance across Biological Process (BP), Cellular Component (CC), and Molecular Function (MF) ontologies.
• Key metrics including F-max, AuPRC, and IAuPRC showed strong results, with F-max scores of 0.581 (BP), 0.692 (CC), and 0.662 (MF).
• The proposed methods achieved state-of-the-art results compared to existing literature approaches.

Conclusions:

• TEMPROT and TEMPROT+ offer competitive protein function prediction by effectively analyzing amino acid sequence patterns and homology.
• The developed models overcome the protein length input restrictions prevalent in many current literature methods.
• These advancements contribute to more efficient and accurate protein function annotation in bioinformatics.