Optimizing skin sensitization prediction across activity cliffs: a comparative analysis of K-nearest neighbours vs random Forest.

Computational models for skin sensitization prediction face critical challenges in handling activity cliffs, where structurally similar compounds exhibit different biological activities, limiting their regulatory applicability. This study systematically compared Random Forest (RF) and K-Nearest Neighbors (KNN) models using five molecular fingerprint approaches integrated with structural alerts and physicochemical properties. Models were developed using 1174 chemicals and evaluated across progressive feature integration levels: fingerprints alone, fingerprints with structural alerts, and fully integrated models. Performance was assessed using standard classification metrics and chemical similarity analysis for compounds with ≥70% Tanimoto similarity but discordant experimental outcomes. RF consistently outperformed KNN across all fingerprint approaches, achieving 81% balanced accuracy compared to (74%) in fully integrated models on the test set. Critically, RF demonstrated superior handling of activity cliffs compared to KNN. Substructure-based fingerprints (Avalon, PubChem and MACCS) consistently outperformed hash-based approaches (Morgan, Atom Pair), with Avalon showing optimal performance across metrics. SHAP interpretability analysis identified vapor pressure (VP) as the most consistently important physicochemical predictor and identified key reactive structural features aligning with known sensitization mechanisms. These findings provide evidence-based guidance for model selection when developing tools for skin sensitization and establish a systematic methodology for evaluating activity cliff performance that can be applied across other toxicological endpoints.