Machine Learning Model Predicts Clinical Adverse Events of Small Molecule Kinase Inhibitors in Cancer Patients Using On‐/Off‐Target Engagement and Tissue Selectivity

Adverse events (AEs) of small molecule kinase inhibitors (SMKIs) at therapeutic doses in cancer patients are largely unpredictable in phase I-III studies and clinical use, despite extensive preclinical toxicity testing under good laboratory practice conditions. To address this gap, we developed a machine learning (ML) framework to predict the occurrence and time to onset of clinical AEs caused by SMKIs using on-/off-target engagement and tissue/cell selectivity. The analysis included 1939 unique AEs from 3,433 patients treated with 16 SMKIs. On-/off-target engagement was evaluated by linking the inhibition (Ki) constants and expression of 442 kinase targets to SMKI exposure, expressed as dose-normalized AUC across plasma and 36 tissues. For each AE, we constructed random survival forest models composed of ensembles of binary decision trees, evaluated predictive accuracy using the concordance index, and applied variable importance (VIMP) measures to identify kinase targets potentially responsible for tissue-specific AEs. The final models successfully predict the most common AEs (rash, nausea, fatigue, headache) along with the most severe hematological AEs (neutropenia, leukopenia, lymphopenia, thrombocytopenia, anemia). VIMP analyses highlighted previously unrecognized kinases potentially involved in tissue-specific AE profiles. External validation using data from a Phase II neratinib monotherapy trial demonstrated strong model performance, yielding Pearson correlation coefficients (PCCs) ≥ 0.87 between predicted and observed AE incidences. These findings show that integrating exposure, on-/off-target engagement, and tissue-specific selectivity enables robust prediction of the likelihood of SMKI-associated AEs for both blood- and organ-related toxicities, offering a scalable approach at both the patient- and population-level.