Sequence-Derived and Molecular Descriptors for Interpretable Modeling of Molecular Systems: Insights from Peptide Hemolysis

Understanding how molecular representations encode structure-property relationships is a central challenge in chemoinformatics, particularly for complex biomolecular systems such as antimicrobial peptides (AMPs). Although numerous computational models have been developed to predict peptide hemolysis, less attention has been given to how different descriptor representations influence both predictive robustness and mechanistic interpretability. Here, we present a comparative computational analysis of sequence-derived and structure-based molecular descriptors to identify the physicochemical properties governing AMP-induced hemolysis. Our analysis identifies a reduced set of key descriptors that preserve the predictive performance of the process. It shows that toxicity is primarily associated with hydrophobic clustering, amphipathic polarity patterning, solvent accessibility, and specific dipeptide motifs, whereas reduced toxicity correlates with higher aggregation propensity and earlier accumulation of polarizable residues. Complementary molecular descriptors suggest that periodic organization of electronic and aromatic properties and localized charge distributions contribute to membrane-disruptive behavior. These findings demonstrate how the representation choice might provide mechanistic insights and guiding principles for descriptor-based analysis and rational design of selective antimicrobial peptides.