Predicting Lifespan-Extending Compounds Using Supervised Machine Learning

Identifying lifespan-extending compounds is an important objective in ageing research, but experimental screening remains costly and time-consuming. This study evaluates supervised machine learning for predicting lifespan-extending compounds in Caenorhabditis elegans using compound–protein interaction features derived from the STITCH database. Three feature representations were compared: continuous interaction probabilities, binary features generated using a fixed threshold of 0.5, and binary features generated using an optimised threshold selected within cross-validation. Random Forest and XGBoost classifiers were evaluated using stratified 10-fold cross-validation, with performance assessed by accuracy, area under the receiver operating characteristic curve (AUROC), and area under the precision– recall curve (AUPRC). Model interpretation was examined using Mean Decrease in Impurity (MDI), SHAP, TreeSHAP, and LIME, with Jaccard similarity used to compare the top-ranked protein features. Random Forest achieved its highest accuracy (78.84%), AUROC (0.7391), and AUPRC (0.5202) using the fixed-threshold binary representation. XGBoost achieved its highest accuracy (80.36%) using the optimised-threshold representation, whereas continuous features yielded its highest AUROC (0.7605) and AUPRC (0.5332). SHAP and TreeSHAP showed the strongest agreement in protein rankings, producing identical top-10 feature sets across all three XGBoost representations. These findings demonstrate that feature representation influences both predictive performance and model interpretation, highlighting the importance of evaluating classification metrics alongside explainability when prioritising candidate compounds for further ageing research.