Supplementary data: Interpretable Machine Learning for the Rational Design and In Vitro Validation of SET-M33 Dendrimeric Antimicrobial Peptides

This repository contains the datasets, trained models, archived predictions, experimental measurements, and numerical analysis outputs supporting the study “Interpretable machine learning identifies target-dependent predictive patterns for a dendrimeric antimicrobial peptide.” The study develops an interpretable machine-learning framework for the design and prioritization of SET-M33 antimicrobial peptide analogs against Escherichia coli, Pseudomonas aeruginosa, and Klebsiella pneumoniae, followed by experimental MIC validation of selected candidates. The workflow includes Random Forest MIC-prediction models built from physicochemical descriptor pools selected with WEKA CfsSubsetEval/BestFirst; generation and prediction of N-terminally extended cationic SET-M33 analogs; organization of candidates into High-Similarity Physicochemical Networks (HSPNs); and post-prediction interpretation using surrogate Random Forest models, SHAP analysis, grouped-SHAP UMAP projections, and robustness analyses. This version additionally provides a leakage-controlled validation of the feature-selection workflow. The historical global-selection strategy is compared with a matched nested procedure in which descriptor selection is repeated within each outer training fold across 11 target–charge datasets, 10 random seeds, and 5 folds (1,100 Random Forest fits). The repository also includes surrogate-model fidelity analyses, SHAP seed and bootstrap stability analyses, HSPN robustness results, the primary grouped-SHAP UMAP outputs, and UMAP sensitivity analyses. The deposited material is organized into thematic packages containing training data and WEKA models, generated candidate libraries and archived predictions, experimental MIC validation data, leakage-controlled validation outputs, post-WEKA SHAP/UMAP/HSPN robustness analyses, and network-analysis outputs. SHA-256 manifests are provided to support file-integrity verification. Analysis and reproducibility code are maintained separately on GitHub: Ismaelcasku/set-m33-interpretable-ml (release v1.0.0, commit 7127b8e). This Zenodo record is therefore intended as the persistent data and numerical-results archive associated with the study, while the GitHub repository provides the corresponding computational workflows.