RatKpContext Benchmark (RKpC Benchmark): Aggregate Evaluation Outputs and Reproducibility Metadata for Tissue-Context-Aware Rat Kp Prediction

Abstract

The RatKpContext Benchmark (RKpC Benchmark) is the data companion to the manuscript “Dissecting Tissue-Context Effects in Graph Neural Networks for Rat Kp Prediction: Additive Shifts, Fusion Position, and Extrapolation.” It contains aggregate evaluation outputs for four graph neural network architectures under repeated parent-group and scaffold splits and leave-one-tissue-out analyses. The package also includes tissue-only and additive controls, preprocessing-robustness analyses, fold-error summaries, tissue-volume-weighted distribution-volume-index results, run-level metadata, index-only split assignments, and SHA-256 provenance records. The package intentionally excludes the harmonized model-development Kp_Data records, molecular structures, record-level model predictions, record-level Poulin–Theil and Rodgers–Rowland transcriptions, source workbooks, article PDFs, and model checkpoints. The aggregate Poulin–Theil and Rodgers–Rowland outputs do not redistribute source-paper values. Software and reproducibility workflows are available from the RatKpContext GitHub repository: https://github.com/Pluswick/RKpC-Benchmark.