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

Aug 4, 2026

Explainable and interpretable unsupervised machine learning clarifies the underlying structure of amino-acid sequence data. An ensemble of restricted Boltzmann machine models consolidates interpretation and improves robustness across different initializations. The approach compresses information into two or three bits, revealing how five-amino-acid contexts at α-helix and β-sheet boundaries encode amphiphilic patterns and unexpected roles for specific residues.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

AccessFree to download

Abstract

Introduction

Interpretable machine learning for proteins

Amphiphilic patterns in secondary structures

Statistical ensemble approach with RBMs

Results overview and information compression

This article is licensed under CC-BY 4.0 [pubs.acs.org/JCTC](pubs.acs.org/JCTC) Article Interpretable Machine Learning of Amino Acid Patterns in Proteins: A Statistical Ensemble Approach Anna Braghetto, Enzo Orlandini, and Marco Baiesi* Cite This: J. Chem. Theory Comput. 2023, 19, 6011−6022 Read Online | ACCESS | Metrics & More | | Article Recommendations | | *sı Supporting Information | | --- | --- | --- | --- | --- | --- | ABSTRACT: Explainable and interpretable unsupervised machine learning helps one to understand the underlying structure of data. We introduce an ensemble analysis of machine learning models to consolidate their interpretation. Its application shows that restricted Boltzmann machines compress consistently into a few bits the information stored in a sequence of five amino acids at the start or end of α-helices or β-sheets. The weights learned by the machines reveal unexpected properties of the amino acids and the secondary structure of proteins: (i) His and Thr have a negligible contribution to the amphiphilic pattern of α-helices; (ii) there is a class of α -helices particularly rich in Ala at their end; (iii) Pro occupies most often slots otherwise occupied by polar or charged amino acids, and its presence at the start of helices is relevant; (iv) Glu and especially Asp on one side and Val, Leu, Iso, and Phe on the other display the strongest tendency to mark amphiphilic patterns, i.e., extreme values of an effective hydrophobicity, though they are not the most powerful (non)hydrophobic amino acids. 1. INTRODUCTION Various machine learning (ML) methods are applied to proteins. 1−25 For example, outstanding advancements have shown how ML can boost the prediction of protein native states13−15 and complexes 16, 17 based only on amino acid sequences. However, the aim of several approaches is not to achieve a reliable (black box) tool for protein structure prediction but to get informative knowledge from the big data available for protein sequences and structures. Interpretable ML26,27 focuses on understanding the cause of a model’s decision and enhancing human capability to consistently predict the model’s result. Interpretable ML versions are more complex and informative than standard statistical analysis and can improve our understanding of proteins.1,2,4−10 In particular, they might detect patterns notemerging naturally from studying the abundance and correlations of amino acids in secondary structures. Among the well-known patterns, for instance, there is the amphiphilic structure of several α-helices and β-sheets,28,29 which are mostly (charged or) polar () on one side d nonpolar () on the other side. In an α-helix, with pitch of 3.6 residues, the typical (non)polarity switch occurs every two residues. On the data statistics encoded in local biases. Conveniently, the weights and biases learned by RBMs can be visualized and easily interpreted. This established approach has already revealed correlated amino acids within protein families, 1 drug−target interactions,2 and correlations within DNA 41,42 sequences. A novelty of our work is a statistical ensemble approach to unsupervised ML, which improves the robustness of the findings. By training RBMs of the same size but with different weight initializations, we checked whether they all converge to the same final set of learned weights. The maximally complex RBMs preserving this ensemble coherence are optimal, as they perform encoding of the correlations within data samples while providing stable and transparent information on the data. We show that our optimal RBMs perform extreme information compression to two or three bits, encoding the essential correlations between amino acids at the beginning or end of α-helices and β-sheets. In addition to recovering the expected amphiphilic structures, this approach (i) discovers more subtle yet relevant amino acid patterns in each portion of the secondary structure and (ii) provi

FAQs

What problem does the work address in protein analysis using machine learning?

It targets extracting informative, interpretable knowledge from protein sequence and structure data, rather than building a black-box predictor.

How does the authors’ statistical ensemble improve interpretability and robustness?

It trains restricted Boltzmann machines of the same size with different weight initializations and checks whether learned weights converge to a consistent representation.

What kinds of amino-acid patterns are identified, and where in protein secondary structure?

The method recovers amphiphilic patterns linked to α-helices and β-sheets, especially using information stored at the start or end of these secondary-structure elements.

Aug 4, 2026

Upload by Lucas Martin