ChIEDiff: trained weights, generated molecules, measurements, and evaluation receptors

Data and weights accompanying the ChIEDiff code release (https://github.com/DannyJPark/ChIEDiff, DOI 10.5281/zenodo.23000086). FILES chiediff_weights_iter1195000.pt (33 MB) Trained weights, iteration 1,195,000, selected by validation loss. Note that the benchmark split has an empty validation subset, so validation was computed on the test pockets: checkpoint selection is not independent of the evaluation set. chiediff_generated_molecules.pt (65 MB) The generated molecules with their docking scores. This file is the authoritative record of molecule IDENTITY. The sampler is not reproducible at the level of individual molecules: the affinity position gradient is a second backward pass over a graph the first retained, and re-traversing it accumulates in a different float order each time, which the 1000-step sampler amplifies. Two runs of the same code produce different molecules. A fresh run reproduces the reported statistics, not these molecules. chiediff_measurements.tar.gz (66 MB) Per-pocket measurement outputs for these molecules: exported SDFs, Vina, SMINA/Vinardo and GNINA docking, PoseCheck, ProLIF and PoseBusters. Unpacking this lets a reproducer start from aggregation instead of re-docking, which is days of compute. chiediff_receptors.tar.gz (48 MB) The CrossDocked2020 test receptors, and chain-remapped copies of them. The remapped copies matter: twelve receptors encode their chain copies only in the segID column, which Biopython ignores, so measuring against the originals silently uses a partial receptor for those pockets. SHA256SUMS Checksums for the four files above. LICENSING The weights, generated molecules and measurements are released CC-BY-4.0. The receptor files are derived from CrossDocked2020, which is dedicated to the public domain under CC0 1.0, and remain CC0. REPRODUCTION The code repository regenerates every table in the accompanying manuscript from aggregate CSVs it already contains, with no download required. This archive is needed only to re-measure from the molecules or to resume from the trained weights. See the repository's README section 3 and docs/reproduce.md.