Resolving Discrepancies in Kinase Activity Labels Using Machine Learning

Protein kinases regulate essential cellular processes by transitioning between active and inactive conformational states, primarily governed by structural rearrangements within the activation segment. Accurately distinguishing these states is critical for kinase research and drug discovery. In this study, we developed a classification framework that leverages 15 geometric descriptors derived from the activation segment to differentiate kinase conformations. We trained multiple machine learning models, including Support Vector Machines (SVM), Logistic Regression, Random Forest, and XGBoost, using only non-conflicting kinase activity labels, identifying over 300 discrepancies between resources. To resolve these inconsistencies, we applied Benchmarking, Randomized Search, Bayesian Optimization, and Coordinate Descent techniques. Additionally, we explored probabilistic models such as Kernel Density Estimation (KDE) and Gaussian Mixture Models (GMM) for kinase classification based on density estimation. Random Forest consistently achieved perfect classification performance, emerging as the most reliable model, with XGBoost as a strong alternative. By successfully distinguishing between active and inactive kinases, our classification scheme provides a robust tool for resolving conflicting labels and has important implications for structure-based drug discovery and guided drug design.