Activity cliffs resist prediction within and across protein kinases: code and derived results for a leakage-controlled machine-learning analysis
Samuel S Agboola, Oluwaseun E. Agboola, et al
Code and derived results for a study of whether the chemical transformations thatgenerate activity cliffs on one protein kinase predict cliffs on another. Matched molecular pairs were constructed from measured Ki and Kd binding affinitiesretrieved from ChEMBL (release 37) for 20 human protein kinases, and activity-cliffprediction was evaluated by leave-one-kinase-out cross-validation against explicitprevalence and transformation-frequency baselines. Within-kinase performance wasadditionally assessed under five partitioning schemes of increasing stringency(random pair-level, and grouped by transformation, constant context, Murcko scaffold,and connected molecular component) to quantify the effect of information leakage. The repository contains the end-to-end pipeline, the figure and table generators, thethirteen additional analyses reported in the manuscript (assay-type separation,matched-pair geometry, leakage-controlled validation, correlation robustness, chemicaloverlap, feature ablation, clustered bootstrap, recurrence modelling, thresholdsensitivity, and transformation inventory), and the derived numerical results. All values derive from measured bioactivity data. No simulated, predicted, or imputedbioactivities are included. Analyses run on commodity hardware without a GPU.