Much ado about nothing: modeling amino acid replacement with predicted protein structures

Area of Science:

• Bioinformatics
• Computational Biology
• Structural Biology

Background:

• Substitution matrices, such as BLOSUM62, are fundamental tools in protein sequence alignment, modeling amino acid replacement likelihoods.
• Despite the release of numerous matrices over three decades, none have leveraged large-scale 3D structure predictions from AlphaFold.

Purpose of the Study:

• To develop and evaluate the AlphaFold Substitution Matrix (AFSM), derived from extensive AlphaFold-predicted 3D structures.
• To compare AFSM's performance against established matrices like BLOSUM62 in key bioinformatics tasks.

Main Methods:

• Generation of AFSM using the BLOSUM methodology from over 20,000 predicted 3D protein structures.
• Benchmarking AFSM and 16 other matrices across five multiple sequence alignment (MSA) and protein homology search tasks.

Main Results:

• AFSM and BLOSUM62 showed similar performance to other matrices across tested tasks.
• Substitution matrices were most beneficial in scenarios with limited sequence data (sparse MSAs).
• Learned embeddings from vast sequence datasets outperformed substitution matrices in sparse data conditions.

Conclusions:

• 3D structural data, as used in AFSM, did not significantly enhance performance over BLOSUM62.
• Abundant sequence data diminishes the necessity for substitution matrices, rendering their extrapolation obsolete.
• BLOSUM62 remains valuable for its implicit physicochemical property representation, essential for sparse MSAs and pairwise sequence comparisons.