Multidimensional mutual information methods for the analysis of covariation in multiple sequence alignments.

Area of Science:

• Bioinformatics
• Computational Biology
• Structural Biology

Background:

• Multiple sequence alignments (MSAs) offer methods to detect covarying positions.
• Covariation analysis can predict protein folds from MSAs with many sequences.
• For known protein structures, covariation data aids in understanding protein mechanisms and dynamics.

Purpose of the Study:

• To evaluate a multivariate extension of mutual information (MI) for studying residue covariation.
• To assess the performance of two multidimensional MI (mdMI) methods against existing statistical models.

Main Methods:

• Tested two multidimensional MI (mdMI) methods on 9 MSAs (<400 sequences each).
• Compared mdMI performance with methods based on maximum entropy/pseudolikelyhood statistical models.
• Analyzed covarying residue pairs within 8 Å in X-ray structures.

Main Results:

• mdMI methods showed performance comparable to state-of-the-art statistical models.
• All tested methods identified similar numbers of covarying residue pairs within 8 Å.
• Less than 65% overlap was observed between top-scoring pairs from different methods.

Conclusions:

• No single method may be universally optimal for detecting protein covariation.
• Combining results from diverse methods can yield richer insights for each protein family.
• This integrated approach is particularly useful for small MSAs or low-quality alignments.