Caizus/natural-product-protein-interact

NPPI-Net retained paper checkpoints

This model repository is reserved for the retained NPPI-Net final_greedy_chain checkpoints: scenarios S1-S4 and seeds 42-46. Exploratory, rejected, and intermediate checkpoints are intentionally excluded.

Publication status

No checkpoint binary is currently published. A read-only audit of the complete experiment account found none of the 20 paths required by the historical final_greedy_chain runner and no adjacent run manifests. It hashed all 80 recovered metrics.json files; none matched the 20 retained evaluator-output identities in the source repository.

The archive does contain 20 main checkpoints and 19 single-seed ablation checkpoints, but they have no content-addressed run manifests and the main metrics match 0/20 retained final-chain metric identities. They are therefore not uploaded or relabeled as paper checkpoints. This metadata-only repository records that absence instead of presenting an unverifiable model as the retained paper chain.

The implementation, exact paper configuration, fixed split manifests, raw metrics, and verification code are published in Cai-z-us/natural-product-protein-interact. The exact input dataset and embeddings are published in datasets/Caizus/natural-product-protein-interact. Paper tables and images are not duplicated here.

Checkpoint contract

Each published checkpoint must be bound to:

• configs/paper/nppinet_final.yaml;
• one of the checked-in S1-S4, seed 42-46 split manifests;
• dataset SHA256 0bef27914ed06e5fe0af9ad64ecd16f27d007369b738f65240795505708d8017;
• molecule embedding SHA256 f87c3673e911e1cc21022e4c4ddcb40cfb2b9ee4e355899ba02f60ae21ad83da;
• protein embedding SHA256 1abb24b598db851733b7345618284aab78ef0ce4d431b454878fa63451d20cb4;
• the public source commit recorded in the repository metadata; and
• its exact byte size and SHA256 in SHA256SUMS.

If eligible checkpoint bytes are recovered later, the required layout is:

checkpoints/
  s1/seed42.pt ... seed46.pt
  s2/seed42.pt ... seed46.pt
  s3/seed42.pt ... seed46.pt
  s4/seed42.pt ... seed46.pt

Only files that can be matched to the retained chain by scenario, seed, configuration, and recovered run metadata are eligible for this repository. The source publication does not infer a checkpoint identity from a filename alone.

Scope and limitations

The retained evidence supports the S1-S4 metrics documented in the source repository. S2-S4 are distribution-shift evaluations; the weak S3 result is reported rather than hidden. These checkpoints are research artifacts, not a clinical, safety-critical, or universal protein-ligand predictor.

The historical evaluator JSON did not record checkpoint digests. Any later recovery must record the bytes as a new content-addressed publication identity without claiming that an absent historical digest was recovered.

Terms

These checkpoints depend on ProteinGLM-derived representations. ProteinGLM is published under CC BY-NC 4.0, so the checkpoints must be treated as noncommercial material and must not be advertised as unrestricted commercial assets. ChEMBL and UniProt attribution/data terms also continue to apply.

No project-wide checkpoint license is asserted while the authors finalize the paper and source-code license. See THIRD_PARTY_NOTICES.md for the upstream attribution record.