This archive pins KGCNN commit ceaaba2d45e2273f12d46f2d0ff9a167b8e1df8e, version 2.1.0, from September 2022. Its recorded final-epoch results round to the MEGAN paper's Appendix C / Figure 6: MAE 0.44 ± 0.03 and RMSE 0.60 ± 0.05, over random five-fold ESOL cross-validation.
original_hyper.json and original_score.yaml are unmodified upstream files. provenance.json contains their upstream paths, every bundled source hash and the archive hash. source.tar.gz bundles the original kgcnn/ package, training script, configuration, results, requirements and license. The runner verifies all hashes before extracting it into ignored .runtime/kgcnn-esol-paper/.
The original final MAEs are 0.4779, 0.4190, 0.4073, 0.4732 and 0.4071 logS; mean 0.4369 and population SD 0.0319. Original final RMSE is 0.6048 ± 0.0461. These are historical upstream results, not new Drugzello measurements. A January 2023 upstream rerun has different scores, so using that newer source would not target the exact table quoted in the paper.
The reproduction imports the original model and chemistry preprocessing without source edits. Drugzello supplies a local-CSV adapter, explicit training seeds, compact result export, resumable fold execution and checkpoint replay. TensorFlow 2.12 / Python 3.11 compatibility and CPU execution are recorded differences: the upstream artifact does not record complete runtime versions, hardware or training RNG seeds. Split seed 42 is recorded upstream.
The original model configuration uses three channels, concatenated attention heads, layer widths [60, 50, 40, 30], MLP widths [50, 30, 10, 1], 41 atom features, no bond features, dropout 0.3, and zero explanation/sparsity factors. It trains for 400 epochs with Adam and a fixed linear learning-rate schedule. It is different from Drugzello's maintained PyTorch implementation.
KGCNN source is MIT licensed; its original license and attribution are preserved. See the reproduction report for commands, measured results and limitations.