Unraveling protein secrets: machine learning unveils novel biologically significant associations among amino acids

Abstract

Accurate classification of amino acids is fundamental to protein engineering and structural biology. Traditional schemes often rely on single-dimensional properties, potentially overlooking complex structure–function relationships. In this study, we present an automated, AI-driven hierarchical clustering pipeline utilizing 22 novel graph-theoretic molecular descriptors to uncover high-dimensional biochemical associations. Using average linkage with Manhattan distance, our approach achieved a high cophenetic correlation (0.847), significantly outperforming K-means and DBSCAN in both cluster quality (Silhouette score: 0.573 vs. 0.548 and 0.412, respectively) and biological interpretability. The analysis revealed two dominant clusters: one comprising aromatic (Trp, Phe, Tyr) and positively charged residues (Arg, His, Lys), and a second encompassing aliphatic, polar, and acidic residues. Notably, we identified a robust, high-stability association between Arginine and aromatic residues (consensus > 0.85), suggesting a functional basis in cation–π interactions often missed by classical hydrophobicity scales. Conversely, Glycine and Proline emerged as distinct structural outliers with low co-clustering probabilities (< 0.3). Sensitivity analysis demonstrated remarkable robustness, with core clusters persisting across 98.6% of descriptor subsets. These findings validate graph-theoretic descriptors as a powerful tool for refining amino acid alphabets, offering a robust framework for predicting mutation effects and guiding de novo protein design.