Sources and verification record

Reviewed 2026-10-01 against current official documentation, PyPI metadata, and a freshly fetched, SHA-256-verified PyTDC 1.1.15 source distribution. Current upstream main was also inspected at commit c310c35f27e3f506411018ac43d97b8ba23ca652. Executed checks used the installed PyTDC runtime with small local synthetic files and blocked HTTP calls, except separate explicitly bounded public Dataverse probes. No full dataset, checkpoint, remote oracle, docking, or authenticated scientific workflow was run.

Release and package metadata

1. PyTDC on PyPI
2. Stable release: 1.1.15
3. Uploaded: 2025-03-31
4. Distribution: source tarball only, 154,168 bytes
5. SHA-256: cd6164859af7b9b6f60e0c6d6e50679eacaffd09cfdea1acfc8bb7360e8e2205
6. License metadata: MIT
7. No Requires-Python or Python classifiers
8. PyPI JSON for 1.1.15
9. Used to verify exact Requires-Dist, artifact metadata, and absence of Requires-Python.
10. Official setup.py
11. Package name pytdc; version loaded from tdc/version.py; dependencies loaded from requirements.txt; no python_requires.
12. Official requirements.txt
13. 1.1.15 pins/constrains a large dependency graph, including cellxgene-census==1.15.0, NumPy <2, RDKit <2024.3.1, Hugging Face packages, and TileDB-SOMA.
14. cellxgene-census 1.15.0 JSON
15. Declares Requires-Python: >=3.8,<3.12.
16. RDKit 2023.9.6 JSON
17. Provides CPython 3.8–3.12 wheels for common platforms, including macOS ARM64, but no CPython 3.13 wheel.
18. Setuptools release history
19. pkg_resources was deprecated long before this snapshot and removed in setuptools 82.0.0 (2026-02-08).
20. PyTDC 1.1.15 still imports it at runtime. The isolated smoke test therefore pins the verified compatibility release setuptools==80.9.0.

The pinned smoke environment uses CPython 3.11. PyTDC itself does not publish a supported Python range, so this skill describes Python 3.11 plus setuptools 80.9.0 as the verified target rather than claiming broader upstream support. On the current tested macOS ARM64 resolver, the test environment contained 128 packages (including pytest) and included large Torch, RDKit, TileDB, Arrow, and scientific-Python artifacts.

Official source used for API verification

1. Package metadata registry
2. Task/dataset names, evaluator names, oracle categories, benchmark names, benchmark metrics, and split metadata.
3. Top-level public imports
4. Evaluator, Oracle, and deprecated generic BenchmarkGroup.
5. Single-prediction imports
6. Multi-prediction imports
7. Generation imports
8. Base loader
9. Single-prediction loader
10. Pair-prediction loader
11. General multi-prediction loader
12. Generation loader
13. Split implementations
14. Evaluator implementation
15. Oracle implementation
16. Download/load implementation
17. Metadata retrieval helpers
18. Specialized BenchmarkGroup base
19. Deprecated generic BenchmarkGroup

The stable PyPI source distribution was inspected directly rather than assuming that main or old generated documentation exactly matched 1.1.15.

Official user documentation

1. TDC quick start
2. Problem/task/dataset hierarchy and constructor/get-data/get-split workflow.
3. Dataset splits
4. Random defaults, documented scaffold scope, cold_split plus column_name, and combination split.
5. Model evaluation
6. Exact evaluator examples, input types, thresholds, and metric definitions.
7. Benchmark/leaderboard guide
8. get, get_train_valid_split, evaluate, evaluate_many, fixed test set, and at least five independent runs.
9. ADMET benchmark group
10. Dataset-specific benchmark metrics and scaffold protocol.
11. Oracle documentation
12. Local, checkpoint, synthesis-service, and docking examples/requirements.
13. Molecule generation task
14. Stable MolGen names and random split examples.
15. ADME task
16. Dataset-specific descriptions, splits, citations, and heterogeneous license labels.
17. DTI task
18. DTI datasets, cold-drug/protein intent, and per-dataset licenses.
19. Trial outcome task
20. Evidence that some TDC datasets carry non-commercial terms.
21. TDC 0.4.1 ReadTheDocs
22. Generated API signatures and source links used only as a cross-check. Its displayed release is behind PyPI 1.1.15.
23. Harvard Dataverse TDC collection
24. Persistent collection identifier linked by the official README. The landing page is not treated as a universal license for individual datasets.
25. Dataverse Data Access API
26. GET file bytes by numeric ID, public/restricted access distinction, and optional format behavior; PyTDC supplies no auth or pagination.
27. PrimeKG resource source
28. Grouped graph mapping, node_type parameter, feature files and lossy undirected display-name graph.
29. Chemical evaluator source
30. Actual KL/FCD transformations and SciPy compatibility failure.

Primary TDC papers

1. Huang, K., Fu, T., Gao, W. et al. (2021). Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development. NeurIPS Datasets and Benchmarks 2021. Published 2021-12-06.
2. Defines the original TDC task/dataset/benchmark/data-function scope.
3. Huang, K., Fu, T., Gao, W. et al. (2022). Artificial intelligence foundation for therapeutic science. Nature Chemical Biology 18, 1033–1036. Published 2022-09-21.
4. Describes the Commons as infrastructure for AI-ready tasks, datasets, and benchmarks across therapeutic science.

Papers support the Commons design and citation guidance; current Python signatures come from package source and official API documentation.

Confirmed migrations and removed stale guidance

• from tdc import BenchmarkGroup is implemented in benchmark_deprecated.py and prints a deprecation message. Prefer from tdc.benchmark_group import admet_group (or another specialized group).
• group.get(name) returns train_val, test, and normalized name; it is not indexed by seed.
• Multi-run input is a list of prediction dictionaries passed to evaluate_many, with at least five runs for non-docking groups.
• Generic cold split is method="cold_split", column_name=.... cold_drug_target is not implemented.
• Pair temporal split is method="time", time_column=...; temporal is not implemented.
• stratified=True is not a loader split argument.
• Registered Pearson correlation is PCC, not Pearson.
• Public generation imports are MolGen, Reaction, RetroSyn, and SBDD; PairMolGen is absent.

Unresolved upstream uncertainty

1. No changelog, tags, or GitHub Releases. PyPI release history establishes version/date, but upstream does not document a complete 0.4.x → 1.1.x migration.
2. Python support is undeclared. Requires-Python is absent, while transitive pins constrain viable interpreters/platforms. Re-run resolver/import smoke tests before changing Python or platform.
3. Legacy runtime dependency. PyTDC still imports deprecated pkg_resources; environments with setuptools 82+ fail unless upstream migrates or setuptools is pinned to a compatible release.
4. Unmarked backport dependency. PyTDC requires the dataclasses backport on modern Python without an environment marker even though Python 3.11 includes dataclasses in the standard library. Strict resolvers may handle that stale metadata differently.
5. ReadTheDocs lags PyPI. It identifies as 0.4.1 while PyPI is 1.1.15.
6. Website/source drift exists. Some website snippets use old names or output comments; stable source controls exact executable behavior.
7. Ambiguous dataset license labels. Some pages render “Not Specified” next to a Creative Commons link. Resolve terms from the original provider rather than inferring a license.
8. Evaluator metadata inconsistency. smina appears in the evaluator registry, but 1.1.15 does not bind it in Evaluator.assign_evaluator.
9. Oracle metadata understates side effects. logp and sa are grouped with trivial oracles, but their stable implementations call calculateScore, which downloads the fpscores artifact when it is absent.
10. Separate fork/package. pytdc-nextml is a distinct package/repository and was not treated as an upgrade or replacement for official PyPI PyTDC.

Current executed evidence and corrections

• CPython 3.11.11, PyTDC 1.1.15, setuptools 80.9.0, NumPy 1.26.4, pandas 2.3.3, scikit-learn 1.9.1, RDKit 2023.9.6. Unpinned setuptools caused an actual ModuleNotFoundError: pkg_resources; the pinned runtime imports successfully.
• All bundled task imports, registry keys, constructor names and generic split defaults were checked against the installed package. CLI help/discovery/plan commands and local ADME, label selection, random/scaffold/cold splits, QED, cached benchmark evaluation and local PrimeKG semantics were exercised.
• Public Evaluator forwards threshold 0.5 to pr@k and rp@k, despite upstream prose and the lower-level functions documenting 0.9. Regression arrays produce different results at the two thresholds. pr-auc is average precision.
• KL returns mean exp(-KL), higher for closer distributions. Unpatched native execution fails on modern SciPy's missing histogram export. A test-only NumPy histogram alias tested identical and changed molecular distributions; this is not evidence of an unpatched compatible workflow. The legacy histogram also can produce NaN when all sampled values fall outside reference bins.
• FCD TensorFlow transforms raw distance by exp(-0.2 * FCD); PyTorch returns raw distance. Both adapters were tested with controlled model outputs, with no ChemNet weights or real embedding inference. This is adapter testing only.
• Kabsch wrapper ignores translate; direct kabsch_rmsd(..., translate=True) aligns a translated tiny coordinate fixture. Benchmark scores are rounded before mean/population-SD aggregation. Zero-test-fraction random train/valid splits stay identical across requested seeds.
• The bundled molecular helper now marks invalid/zero-atom inputs with null scores and preserves positions. Benchmark input rejects normalized-name collisions, inconsistent run coverage/lengths, and mismatched fixed-test counts. The cache audit no longer reports a truncated scan complete at a directory boundary.
• Official release/tag APIs returned empty lists. ReadTheDocs still identifies 0.4.1. No separate fork was substituted for official PyTDC.

Public Dataverse verification used read-only GET /api/files/{file_id}/metadata for Caco2 4259569, Davis 5219748, BindingDB Patent 4724851, MOSES 4170962, PrimeKG 6180626, ADMET group 4426004, drug-combination group 4426002 and DTI DG group 4742443. Curl returned HTTP 200, expected artifact labels, and restricted: false for each. Python urllib requests received HTTP 403 in this environment, so these are curl/public endpoint checks, not proof that the package downloader works end to end. HTTP range GETs of Caco2 and PrimeKG returned 512 bytes each and the expected tabular headers; full downloads were not performed. The metadata response's id belongs to the file-metadata record; retain the requested datafile ID for download rather than replacing it with that value.

python tests/run_all.py --isolated pytdc passed 30 tests; spec validation and nine metadata/plan CLI invocations passed. Real local runtime checks are separate from the two controlled FCD adapters and the explicit KL compatibility alias.