Addressing inaccuracies in BLOSUM computation improves homology search performance

Area of Science:

• Bioinformatics
• Computational Biology
• Structural Bioinformatics

Background:

• BLOSUM matrices are standard for protein homology search and sequence alignment since 1992.
• Previous work identified miscalculations in BLOSUM matrix computation, leading to RBLOSUM matrices.
• RBLOSUM matrices, despite corrections, showed statistically significant performance degradation compared to BLOSUM62.

Purpose of the Study:

• To introduce a further corrected version of the BLOSUM code, termed CorBLOSUM.
• To conduct a comprehensive performance analysis of BLOSUM, RBLOSUM, and CorBLOSUM matrices.
• To evaluate homology search performance across various BLOCKS databases and ASTRAL subsets.

Main Methods:

• Derived and analyzed BLOSUM, RBLOSUM, and CorBLOSUM matrices.
• Assessed homology search performance using three BLOCKS databases.
• Benchmarked matrices on all versions of ASTRAL20, ASTRAL40, and ASTRAL70 subsets (51 benchmarks total).
• Focused analysis on BLOSUM50 and BLOSUM62.

Main Results:

• Corrected BLOSUM code yields improved substitution matrices beneficial for homology search.
• CorBLOSUM matrices matched or exceeded BLOSUM performance in ~75% of cases, outperforming them in >86% on current ASTRAL databases.
• RBLOSUM matrices outperformed corresponding BLOSUM matrices in most cases, unlike previous findings.
• CorBLOSUM matrices showed superior performance over RBLOSUM matrices on up-to-date ASTRAL databases (~74% of cases).

Conclusions:

• CorBLOSUM matrices demonstrate statistically significant performance improvements over BLOSUM matrices, particularly on recent ASTRAL databases.
• CorBLOSUM matrices align more closely with the original conceptual design of Henikoff and Henikoff.
• Recommends CorBLOSUM matrices for homology search tasks over (R)BLOSUM matrices.