Datasets:
MolAS downstream code and saved analyses
historical_molas/ preserves the recovered Downloads/MolAS_AL-main source snapshot byte for byte: the completed-matrix training adapter, held-out evaluator, launcher, model/loss modules, dependency files, and analysis notebooks. Its original MIT license is included. historical_molas_manifest.json records all 27 source files and hashes; Python bytecode caches are excluded. The snapshot includes exploratory analyses beyond the paper. Its presence establishes available code, not execution of every notebook section or the exact configuration of every archived run.
historical_manuscript_scripts/ preserves two additional runtime-plotting scripts from the manuscript source folder. Their final-figure role is not established. upstream_molas/ is a separate, unchanged official MolAS reference snapshot at commit a6c417216ddb992f7dc513d511d0429aece4bd61, with MIT license and upstream_molas_manifest.json; it lacks the recovered completed-matrix adapter and is not the primary AI4DD workflow.
See ../../docs/downstream_reproduction.md for the recovered training protocol, source-to-figure evidence, saved CSV coverage, and remaining provenance limits. ../../provenance/figure_lineage.json maps all eight figures to preserved assets and relevant notebook cells.
Prepare portable paths without executing experiments
prepare_local_copy.py creates a new configured copy of the recovered sources. It never imports MolAS/PyTorch, executes a notebook, fits a model, or changes historical sources. Copy local_paths.example.json to your own JSON, supply authentic paths, and run:
python3 code/downstream/prepare_local_copy.py --config /absolute/path/to/your_paths.json --output /absolute/path/to/a_new_configured_molas_directory
Paths in the JSON are resolved relative to that JSON file. The example's relative paths assume it remains in code/downstream; update them if you move it. The acquisition path must contain expanded <method>/fold<fold>/seed0/strict_score_matrices/strict_score_NNNpct.csv directories. Use the acquisition artifact reader/extraction instructions to obtain those existing files; passing an archive directory alone is insufficient.
The utility configures dataset roots, raw/graph/ligand file links, and launcher paths. It preserves the recovered launcher's scientific settings and currently active two-method, four-checkpoint subset, rather than silently enabling a large new experiment. Review main.sh before any future execution. Activating its launcher would start new training and testing; preparing the copy does not.
Add --include-analysis only after resolving both target paths and the expanded downstream output root. molas_score_matrix and acquisition_score_matrix are deliberately separate required roles: the native MolAS evaluation score truncates RMSD above 5 Å to zero, while the acquisition strict score does not. The example leaves them unset to prevent an incorrect automatic alias. The native reconstruction in data/targets/moad_score_native_reconstructed.csv is independently checked against saved per-complex outcomes; see its provenance before selecting it. The primary notebook's paths are updated and its configured copy's old outputs are cleared. The original notebook, including saved outputs, remains unchanged. The separate runtime-probe notebook retains historical paths and is not automatically configured.
Saved-output packaging utility
catalog_saved_outputs.py was written for this release. It inventories and byte-copies existing CSVs; it is not the historical trainer/evaluator or plotting code.
python3 code/downstream/catalog_saved_outputs.py --source /absolute/path/to/port_as_results_260829 --output /absolute/path/to/MolAS_NIPS/results/downstream
python3 code/downstream/verify_saved_outputs.py --write-report --require-complete --original-archive-manifest provenance/downstream_original_archive_recovery.json
The verifier checks every archive/member hash, member CRC, and generated summary mirror. With the original archive manifest supplied, it also checks every member against the independently recovered original ZIP identities. It distinguishes integrity from completeness; --require-complete returns nonzero while original CSV payloads remain unavailable. Without --write-report, it does not modify the package.
The release contains all 4,508 saved CSVs for the six paper policies. The verified original port_as_results_260829.zip resolved earlier cloud-only gaps; archive/member evidence is recorded in provenance/downstream_original_archive_recovery.json. To rebuild, use a fully readable extraction of that historical result tree as --source.
Python's standard library is sufficient for these utilities. Source and archive output must be separate non-nested directories. iCloud placeholders are explicitly inventoried rather than treated as recovered payloads. No scientific experiments were executed during release packaging.