Annotating the microbial dark matter with HiFi-NN.

Area of Science:

• Bioinformatics
• Computational Biology
• Enzymology

Background:

• Accurate computational annotation of protein sequences with enzymatic function is a significant challenge.
• Existing methods struggle with precision and recall, particularly for proteins with distant sequence homology.

Purpose of the Study:

• To introduce HiFi-NN (Hierarchically-Finetuned Nearest Neighbor search) for precise protein functional annotation.
• To evaluate HiFi-NN's performance against state-of-the-art methods, including deep learning and BLASTp.
• To demonstrate HiFi-NN's utility in correcting database errors and annotating uncharacterized proteins.

Main Methods:

• Developed HiFi-NN, a novel Hierarchically-Finetuned Nearest Neighbor search algorithm.
• Benchmarked HiFi-NN against existing deep learning methods and BLASTp for Enzyme Commission (EC) number annotation.
• Investigated the impact of lookup set diversity on HiFi-NN performance.
• Applied HiFi-NN to correct mis-annotations in the BRENDA database and annotate NMPFamDB sequences.

Main Results:

• HiFi-NN achieved higher precision and recall than state-of-the-art deep learning methods for EC number annotation.
• HiFi-NN correctly identified EC numbers at lower sequence identities compared to BLASTp.
• Increased diversity in the lookup set significantly improved HiFi-NN's performance.
• Successfully corrected specific mis-annotations in the BRENDA database.
• Annotated functional dark-matter protein sequences from NMPFamDB.

Conclusions:

• HiFi-NN offers a more accurate and robust approach for *in silico* protein functional annotation.
• The method is particularly effective for proteins from distant sequence space and aids in refining existing enzyme databases.
• HiFi-NN advances the field of bioinformatics by enabling better understanding of protein function.