Comparative performance of structural aligners in functional domain annotation.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Protein Informatics

Background:

• Accurate protein domain annotation is crucial for understanding protein function.
• Databases like Pfam use sequence-derived signatures, but structure-based methods offer higher sensitivity for distantly related proteins.
• AlphaFold provides accurate protein structure predictions, enabling new structure-based annotation approaches.

Purpose of the Study:

• To systematically evaluate structure-based versus sequence-based methods for Pfam domain annotation.
• To benchmark structural aligners (Reseek, Foldseek, TM-align) against sequence-based methods (MMseqs, HMMER).
• To assess the impact of profile-derived information on structural annotation performance.

Main Methods:

• Benchmarking of three structural aligners and two sequence-based methods using all-against-all and split-family searches.
• Evaluation of residue-level alignment accuracy against Pfam multiple sequence alignments.
• Investigation of profile-informed rescoring for structural hit ranking.

Main Results:

• Reseek demonstrated the highest sensitivity in all-against-all searches (AUC=0.85), outperforming other structural and sequence aligners.
• HMMER remained superior in split-family evaluation (max F1=0.991), confirming the strength of sequence-profile methods.
• Structural aligners provided more accurate residue-level mappings than pairwise sequence methods.
• Profile-informed rescoring improved performance for short domains.

Conclusions:

• Structure-based methods, particularly Reseek, show high sensitivity for protein domain annotation, especially at low sequence identities.
• Sequence-profile methods like HMMER are still optimal for family-level annotation.
• Combining structural information with sequence profiles offers a promising direction for enhanced domain annotation.