Interpretable machine learning classifies β-lactamase subclasses and prioritizes boundary variants for antimicrobial...

Antimicrobial resistance is a major global health threat, and β-lactamase-mediated inactivation of β-lactam antibiotics remains a leading cause of treatment failure. Rapid identification of emerging β-lactamase variants and improved understanding of their functional diversity are therefore essential for resistance surveillance and therapeutic development.

Here, we developed an interpretable machine learning framework to classify major Class A β-lactamase subclasses (TEM, SHV, and CTX-M) and distinguish them from non-β-lactamase hydrolases using multidimensional protein descriptors. Descriptors based on sequence composition, physicochemical properties, residue accessibility, and structural topology were evaluated to identify a minimal panel of 30 descriptors while preserving biological interpretability.

The selected descriptors provided strong discriminatory power and were linked to steric environment, hydrophobicity, flexibility, solvent exposure, and residue interaction patterns, consistent with roles in substrate recognition and catalysis. Projection into reduced descriptor space identified transitional variants near subclass boundaries, including CTX-M/SHV convergence zones and non-β-lactamase hydrolases with partial β-lactamase-like signatures. Representative case studies indicated that some apparent misclassifications may reflect annotation inconsistencies, scaffold convergence, or early adaptive trajectories, highlighting candidates for experimental follow-up. External validation using 271 independent proteins annotated as non-β-lactamase hydrolases demonstrated 96.95% specificity, with only limited false-positive assignments. Most incorrect classifications showed weak statistical support, indicating that confidence filtering could further improve screening performance.

Overall, this study demonstrates that compact descriptor-based models can provide accurate and interpretable classification of β-lactamase-related proteins while identifying emerging boundary variants relevant to antimicrobial resistance surveillance. The framework offers a scalable approach for protein annotation, monitoring resistance evolution, and supporting future inhibitor development.