Interpretable Machine Learning of Amino Acid Patterns in Proteins: A Statistical Ensemble Approach

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in protein science

Background:

• Understanding protein secondary structure (alpha-helices and beta-sheets) is crucial for deciphering protein function.
• Unsupervised machine learning offers powerful tools for uncovering hidden patterns in complex biological data.
• Interpretable models are essential for deriving meaningful biological insights from machine learning analyses.

Purpose of the Study:

• To develop and apply an ensemble analysis of machine learning models for enhanced interpretation of protein sequence data.
• To investigate the information content and patterns within amino acid sequences at the boundaries of protein secondary structures.
• To identify novel properties of amino acids and their contributions to protein secondary structure formation using machine learning.

Main Methods:

• Utilized restricted Boltzmann machines (RBMs) as a core component of the machine learning ensemble.
• Applied ensemble analysis to consolidate interpretations from multiple machine learning models.
• Analyzed the learned weights of the RBMs to identify significant amino acid patterns and properties.

Main Results:

• Restricted Boltzmann machines effectively compress information from five-amino acid sequences at helix/sheet termini.
• Identified specific amino acid propensities and their roles in amphipathic patterns of alpha-helices.
• Discovered that His and Thr have minimal impact on alpha-helix amphipathicity, while Ala-rich helices exist.
• Proline's position and role in initiating helices were highlighted, often replacing polar/charged residues.
• Observed strong amphipathicity markers in Glu, Asp, Val, Leu, Ile, and Phe, related to effective hydrophobicity.

Conclusions:

• Ensemble machine learning provides a robust framework for interpreting complex biological sequence data.
• The study reveals nuanced, previously unrecognized roles of specific amino acids in protein secondary structure formation.
• Findings contribute to a deeper understanding of protein folding and structure-function relationships through interpretable AI.