---
pretty_name: NPPI-Net reproducibility resources
license: other
task_categories:
  - tabular-regression
tags:
  - natural-products
  - protein-ligand-interaction
  - molecular-embeddings
  - protein-embeddings
size_categories:
  - 10K<n<100K
---

# NPPI-Net reproducibility resources

This dataset repository contains the exact large inputs used by the retained
NPPI-Net S1-S4 result chain. Source code, fixed split manifests, paper
configurations, raw metrics, and verification scripts are published separately
in
[`Cai-z-us/natural-product-protein-interact`](https://github.com/Cai-z-us/natural-product-protein-interact).
Paper tables and images are not duplicated here.

## Files

| Path | Purpose | Bytes | SHA256 |
| --- | --- | ---: | --- |
| `data/dataset_all.csv` | canonical headerless interaction dataset | 23,281,585 | `0bef27914ed06e5fe0af9ad64ecd16f27d007369b738f65240795505708d8017` |
| `data/external/uniprot_human_reviewed.fasta` | exact historical reviewed-human UniProtKB snapshot | 13,663,585 | `89b398077ec571c34bd67fd10fc1640eeae37d1421d6f84c3a8a94e1db88dac3` |
| `data/mol/unimol_v2_1.1B.pth` | pooled Uni-Mol2 molecule representations | 1,460,616,917 | `f87c3673e911e1cc21022e4c4ddcb40cfb2b9ee4e355899ba02f60ae21ad83da` |
| `data/prot/proteinglm_xtmlm_1b.pth` | pooled ProteinGLM protein representations | 9,691,794,119 | `1abb24b598db851733b7345618284aab78ef0ce4d431b454878fa63451d20cb4` |

The CSV has 32,802 rows, 8,959 unique molecule identifiers, 1,860 unique
protein identifiers, and no header. Its ten columns are, in order:
`mol_id`, `smiles`, `uniprot_id`, `sequence`, `pchembl`,
`standard_relation`, `standard_value_nM`, `standard_units`, `standard_type`,
and `confidence_score`.

The retained model consumes pooled `float32` vectors only: 1,536 dimensions
for molecules and 2,048 for proteins. Token inputs are disabled in the paper
configuration.

## Download and verify

From the root of a clone of the source repository:

```bash
hf download Caizus/natural-product-protein-interact \
  --repo-type dataset \
  --revision d67ea7b4f75b35f1f362caebcd3d5565cedc16dc \
  --local-dir .
sha256sum --check SHA256SUMS
uv run --frozen python scripts/verify_resources.py --deep
```

`scripts/verify_resources.py` checks byte size, SHA256, schema, key alignment,
pooled dimensions, and dtypes against
`artifacts/resources/manifest.yaml`. PyTorch files are opened with the
restricted `weights_only` loader.

## Provenance

- The canonical interaction dataset was built from ChEMBL 34 and a reviewed
  human UniProtKB snapshot. The exact FASTA is published above because the
  historical UniProt release response header was not retained; its byte
  identity, rather than an inferred release number, is authoritative.
- The molecule file was generated with `unimol-tools==0.1.5` and Uni-Mol2
  1.1B. The package wheel and final output are content-addressed in the source
  repository.
- The protein file was generated from
  `biomap-research/proteinglm-1b-mlm` at revision
  `676c1bcb9f737c166a25587f83fd580610acb1ad`.

See
[`ARTIFACTS.md`](https://github.com/Cai-z-us/natural-product-protein-interact/blob/main/ARTIFACTS.md),
[`DATASET.md`](https://github.com/Cai-z-us/natural-product-protein-interact/blob/main/DATASET.md),
and
[`THIRD_PARTY_NOTICES.md`](https://github.com/Cai-z-us/natural-product-protein-interact/blob/main/THIRD_PARTY_NOTICES.md)
for the full build and attribution record.

## Terms

This repository contains artifacts with different governing terms; the
project's future source-code license does not replace them.

- ChEMBL-derived dataset content is subject to ChEMBL 34's CC BY-SA 3.0 terms
  and required attribution.
- UniProt-derived content is subject to CC BY 4.0 attribution.
- Uni-Mol tooling/model attribution is retained; applicable source-data terms
  continue to apply to the generated molecule representations.
- ProteinGLM is CC BY-NC 4.0. The generated protein representations must be
  treated as noncommercial material and must not be advertised as unrestricted
  commercial assets.

No additional artifact-wide license is asserted beyond those file-specific
terms while the authors finalize the paper release metadata.