docs/data_card.md

MolAS active-acquisition dataset card

This card documents the Hugging Face release's data schema and provenance. The author confirmed applicable redistribution terms; see the confirmation record. Exact end-to-end historical reproduction has not been newly verified. Original acquisition inputs have been recovered and hash-verified; numerical values are preserved. Existing feature arithmetic was independently repeated only as an integrity check. No docking, active-learning runs, or learned-model training was performed.

Scope

The benchmark comprises 3,179 MOAD-curated protein–ligand complex identifiers and seven docking algorithms: surfdock, unimol, diffdock, diffdock_l, gnina, autodock, and smina. Each of the three raw tables contains 22,253 measurements in the same row/algorithm order. These are historical outcomes used as an offline oracle for cell-wise acquisition. Their public availability must not be confused with permission for an acquisition policy to access unrevealed outcomes or held-out labels.

The package supports inspecting the existing outcome/runtime matrix, auditing saved acquisition/downstream results, and preparing reproducible inputs. It does not establish performance on other benchmarks, hardware, portfolios, or prospective docking campaigns.

Data dictionary

File/field	Meaning
All raw tables: PDB_CCD_ID	Original compound complex identifier, e.g. 19gs_BSP_A_210. Preserve it as a string; it includes PDB, ligand/CCD, chain and residue components.
raw/moad_rmsd_trimmed.csv: <algorithm>_rmsd	Historical docking RMSD. The manuscript interprets RMSD in Angstrom. Values are preserved exactly, including extreme finite values.
raw/moad_pb_ratio_trimmed.csv: <algorithm>_pb_ratio	Historical PoseBusters validity ratio. PB validity for scoring is the exact rule ratio == 1.0, not a tolerance or rounding rule.
raw/moad_runtime_trimmed.csv: <algorithm>_runtime	Historical per-cell runtime, treated as seconds by the experiment code (summed seconds / 3600 for hours). Hardware and precise timing boundaries still require original acquisition documentation.
targets/full_strict_score_matrix.csv: <algorithm>_strict_score	Original restart target; deterministic score derived from RMSD and PB validity, higher is better.
targets/strict_score_matrix.csv: <algorithm>	Earlier target table retained unchanged, with slightly different decimal serialization and column names.
targets/moad_score_native_reconstructed.csv: <algorithm>	Reconstructed native downstream evaluation score using the recovered MolAS source's RMSD >5 or <0 zeroing rule. Distinct from the acquisition targets above.
ordered_complex_ids.csv: PDB_CCD_ID	The raw-table row sequence only. Do not use it as unverified feature provenance.
splits/moad_5fold_split_seed42.csv: row_index, protein, ligand, instance_id, test_fold	Original split; zero-based row index, original compound ID, and test-fold label 0–4. protein + '_' + ligand == instance_id.
splits/recovered_heldout_folds_from_downstream.csv: PDB_CCD_ID, fold	Recovered existing test-set membership, with folds 0–4; a row is held out when its fold matches the evaluated outer fold.
descriptors/moad_physical_descriptors_filtered.csv	PDB_CCD_ID plus ligand_heavy_atoms, ligand_rotatable_bonds, protein_residues, and pocket_heavy_atoms_r6. These are the original four numeric descriptor columns; exact generation conventions should be taken from the upstream preprocessing documentation.
features/complex_features_ligand_protein_d256_seed0.npy	Original 3,179 × 256 float32 cache, unchanged.
features/complex_feature_ids.csv: PDB_CCD_ID	Cache row order verified against the source transformation for all 3,179 rows.
features/ligand_chemberta_moad3179.pt	Original PyTorch ZIP archive with ligand_ids and a 3,179 × 384 float embedding tensor.
features/ligand_feature_ids.csv: PDB_CCD_ID	The ligand archive's verified embedded ID sequence. It equals the raw table order.
provenance/protein_graph_inventory.csv	One filename/complex, logical bytes, and cloud-placeholder status for the historical restart copy; no graph values or hashes are inferred from filenames.

The continuous acquisition score is

strict_score = ((1 + exp(-6)) / (1 + exp(3 * (rmsd - 2)))) * (pb_ratio == 1.0).

Both target files are unchanged copies of their recorded sources: the original restart ground_truth/full_strict_score_matrix.csv and the earlier MolAS_AL/data/processed/pilot1a/strict_score_matrix.csv. Both contain all 22,253 values and are checked against the raw-data formula. Their column suffixes and decimal serialization differ; they are not claimed to be byte-identical. Runtime does not enter this target formula.

Native evaluation target is separate

The recovered downstream source, MolAS_AL-main/data/dataset.py, defines rmsd_score as zero when RMSD is greater than 5 or less than 0; otherwise it uses the same normalized sigmoid. pb_score again requires the ratio to equal 1 exactly. An RMSD of exactly 5 retains its sigmoid value. This native evaluation definition differs from the continuous acquisition target and must not be silently interchanged with it.

The historical analysis notebook explicitly separates GROUND_TRUTH_PATH (server-side moad_score_trimmed.csv) from ACQUISITION_GROUND_TRUTH_PATH (a full acquisition score matrix). The original moad_score_trimmed.csv was not found in the relevant local source/data folders. data/targets/moad_score_native_reconstructed.csv therefore has an explicit reconstruction name and provenance: it was deterministically calculated from the unchanged raw RMSD/PB tables using the recovered native source rules. It is not represented as the recovered original CSV.

All 22,253 reconstructed native engine scores were independently compared with the seven engine-score columns in the five saved cost_tiers_5/seed0/b001_pct downstream descriptive tables. These tables jointly cover all 3,179 complexes exactly once. The maximum absolute difference is 1.11e-16, with zero mismatches at a 1e-12 tolerance. The continuous acquisition and native definitions differ by up to approximately 0.00012097 in these data; source tables and existing scientific results were not modified. Source-function excerpts, file hashes, and validation references are in data/provenance/native_evaluation_target.json and native_score_functions.py.txt.

scripts/reconstruct_native_evaluation_target.py --check-only verifies this derived file. Without --check-only, the utility writes only the explicitly named reconstructed target. This is deterministic data preparation, not a docking or model-training experiment. Supply native and acquisition paths separately to analysis tooling; do not silently alias either to the other.

Sources and recovery

The three raw matrices were copied byte-for-byte from the resident MolAS_AL/data/raw directory. All three are now confirmed hash-identical to the corresponding recovered MolAS_AL_restart/AL_MolAS/data/raw inputs. The three raw matrices, original split, original descriptors, and combined cache also match all six hashes in the preserved original input_audit.json and experiment_manifest.json. data/provenance/raw_sources.json and hydrated_original_inputs.json record exact sources. Initial bounded-read failures remain recorded as history and no longer describe current input availability.

While the original split was initially unavailable, existing test membership was recovered from the five cost_tiers_5/seed0/b001_pct/test_composite/default/fold_*/..._descriptive.csv files. Their protein and ligand fields reconstruct the original complex IDs. The resulting sets are disjoint, cover all 3,179 raw IDs, and contain 636/636/636/636/635 rows. Ordered IDs also agree with ten additional resident checkpoint files. The original split was subsequently hydrated: every original instance_id/test_fold pair matches the independent recovery exactly, with no label offset required. Original row_index is 0–3178 and follows raw-table order. Both files are retained. The recovery did not generate a new split. Evidence is in data/provenance/recovered_split.json and hydrated_original_inputs.json.

The ligand archive was copied unchanged from the restart structural-data folder. ZIP entries and pickle opcodes were inspected without unpickling or executing objects. Its explicit 3,179 ligand_ids match the raw ID order, and the single float storage contains 1,220,736 finite values consistent with the recorded 3,179 × 384 shape. The combined projected feature cache was checked separately as described below.

Combined-feature verification and structural inputs

The original four-descriptor CSV and combined 256-dimensional NPY cache were subsequently hydrated and copied unchanged. The cache's SHA-256 is 10b4a50f5cfc3de000ab06d0b8112146486e63571ad9db252455dd5af3976a91, matching the historical input audit. Its row order was independently verified, rather than inferred from filename or shape.

scripts/verify_feature_cache_alignment.py uses an allowlisted metadata decoder that replaces PyTorch/PyG classes with inert records and reads numeric tensor ZIP storages; it does not deserialize executable PyTorch/PyG objects. It reads the explicit ligand IDs, pools existing graph features in the original ID order, and repeats the existing producer's normalization and seed-0 Gaussian projection. The seven original producer functions are preserved in provenance/historical_acquisition_source/feature_builder_excerpt.py.txt, with original-file hashes and line ranges in feature_builder_provenance.json. All 813,824 cache values across all 3,179 rows agree within the predeclared combined tolerance abs(error) <= 2e-5 + 2e-5 * abs(reference). Maximum absolute difference is 2.19345e-5 and RMS difference is 2.32080e-6, consistent with float32 reduction differences between NumPy and the original PyTorch operations. The arrays are not bit-identical. The original cache was never rewritten. complex_feature_ids.csv was emitted only after this comparison passed; the full evidence and hashes are in feature_cache_comparison.json and combined_feature_alignment.json under data/provenance.

This check also identifies the historical cache preprocessing: normalization uses the fixed full set of 3,179 feature rows before projection. It must not be described as a scaler fitted separately within each training fold. No alternative representation or ligand-only cache was substituted.

Large protein graphs comprise 3,179 files and 14,985,036,224 logical bytes. The restart directory contains 2,414 cloud placeholders, but a separate fully resident source was located at /Users/wanweilin/Downloads/data/MolAS/protein_graphs_esmc_600m. The packaging process cloned that source locally into data/external/protein_graphs_esmc_600m: all filenames and logical sizes agree, and the local clone has zero cloud placeholders. These large assets are included in this Hugging Face upload payload; the corresponding GitHub code release excludes them. Refer to the parent package's graph hash manifest for payload verification. The original compressed structural archive is also outside this package. Graph availability and cache lineage are separate checks; the independent numerical cache comparison is documented above, and no replacement cache was substituted.

No physical descriptors were invented, no embedding model was run, and no existing feature cache was replaced. Deterministic feature arithmetic was repeated solely to compare against the existing cache. No docking results or random splits were generated. Check data/provenance/validation_report.json for current completeness rather than assuming a directory name means its input is available.

Data quality and interpretation

All three raw tables contain 3,179 unique IDs, aligned row order, and finite numeric values. Runtime is strictly positive and PB ratios lie in [0,1]. Raw RMSD includes extreme finite values, with a maximum approximately 1.5964e17 Angstrom. These were preserved; their possible failed-run or sentinel meaning has not been established. They must not be silently removed, clipped, or reinterpreted as validated physical poses.

Runtime is hardware/workflow dependent. The exact engine versions, compute allocation, startup/preprocessing inclusion, repeated-run aggregation, and failure/timeout conventions require source-author documentation. The five folds are existing complex-level memberships; this package does not certify protein-family or ligand-scaffold separation. The seed0 labels in downstream directories do not independently establish the MolAS training RNG seed.

Validation and reproducibility boundaries

python scripts/validate_data.py checks raw schemas and numeric domains, exact cross-table ID order, both acquisition score tables, the separate native-evaluation reconstruction, original/recovered fold agreement, descriptor/feature schemas, ligand internal IDs, cache-comparison evidence, six historical input hashes, and recorded checksums. It uses the Python standard library and never docks, trains, downloads, or deserializes PyTorch objects. verify_feature_cache_alignment.py is the separate NumPy-only source-to-cache comparison utility. Use --require-complete for a nonzero exit when acquisition inputs are missing. Acquisition inputs are complete. The numeric validator does not verify licenses; the author's redistribution confirmation is recorded separately from these integrity checks.

Attribution and publication

Upstream MolAS: code, processed data/checkpoints, and paper. Cite the camera-ready active-learning paper using its verified final bibliographic record and cite MolAS and the original dataset/feature sources. Do not invent a DOI or release URL.

Author-owned data contributions are licensed CC BY 4.0. Third-party assets retain their applicable terms; see the root LICENSE, LICENSE_DATA.md, and third_party.md. The mixed bundle uses custom component-specific Hub license metadata. For eventual Hugging Face publication, use explicit configurations separating raw outcomes, complex metadata, features, and derived results. The official dataset-card, file-configuration, and license guidance does not itself grant upstream redistribution rights.