MODEL-FREE INFERENCE FOR CHARACTERIZING PROTEIN MUTATIONS THROUGH A COEVOLUTIONARY LENS

Area of Science:

• Bioinformatics
• Computational Biology
• Structural Biology

Background:

• Multiple sequence alignment (MSA) data are vital for studying protein mutations.
• Protein contact prediction is a key application of MSA, but current methods lack statistical inference for uncertainty quantification.

Purpose of the Study:

• To develop a novel statistical framework for protein contact prediction.
• To incorporate statistical inference for quantifying prediction uncertainty.
• To identify amino acid combinations contributing to correlations in protein contacts.

Main Methods:

• Transformed contact prediction into a statistical testing problem.
• Utilized one-hot encoding of MSA data to construct a partial correlation graph for categorical variables.
• Introduced a novel spectrum-based test statistic to assess partial correlations between protein positions.

Main Results:

• The proposed method controls Type I errors and demonstrates statistical power.
• Successfully identified contacts based on partial correlations in MSA data.
• Enabled the discovery of amino acid combinations associated with protein contacts and mutations.

Conclusions:

• The novel framework provides a statistically rigorous approach to protein contact prediction.
• It enhances the analysis of coevolution and mutation patterns in proteins.
• The method offers practical utility for understanding protein structure-function relationships.