HypSeek 18-seed affinity-ranking checkpoints

This repository contains the strict-comparability checkpoints used in the JACS-8/MERCK-8 HypSeek ensemble-scaling experiment.

Checkpoints

  • checkpoints/seed01.pt through checkpoints/seed18.pt: strict-comparability pool. Each is an average over epochs 41–50 from an independently trained HypSeek seed.
  • The epoch24–33 seed19 sensitivity checkpoint is deliberately not distributed in this repository.

Each checkpoint is approximately 1.45 GiB. Integrity hashes, repository paths, training windows, and byte sizes are recorded in checkpoint_manifest.csv.

Code and evaluation

  • Base code: https://github.com/jianhuiwemi/HypSeek
  • Reproduction data, portable analysis overlay, reference results, and a colleague quick-start guide are stored under reproduction/.
  • Model architecture: three_hybrid_model
  • Native single-model FEP score: max_pocket(pocket_embedding @ molecule_embedding.T) with alpha_poc=1 and alpha_prot=0.
  • Ensemble score: arithmetic mean of raw single-model scores.
  • Main benchmark: JACS-8 and MERCK-8, target-macro Spearman.

Strict seed1–18 result: JACS 0.7974, MERCK 0.5514, ALL 0.6744. The fixed historical seed1–5 result is JACS 0.7775, MERCK 0.5721, ALL 0.6748.

For native edge ACC, exhaustive random-subset expectations change from JACS/MERCK 0.7264/0.6483 at k=1 to 0.7346/0.6715 at k=5 and 0.7301/0.6696 for the unique k=18 ensemble. The ACC curve largely saturates by k=5–10; exact tables and subset intervals are included with the reference results.

Recipient workflow

  1. Obtain access to this private repository.
  2. Clone the public HypSeek code repository.
  3. Download this repository, extract reproduction/hypseek_ensemble_scaling_code.tar.gz over the clone, and extract reproduction/hypseek_jacs_merck_fep_data.tar.gz into a data directory.
  4. Follow reproduction/COLLEAGUE_QUICKSTART.md.

The public GitHub repository by itself is not a complete delivery because it does not contain the 18 checkpoints or the benchmark reproduction bundle.

Intended use and limitations

These files are research checkpoints for reproducing the reported strict ensemble experiment. Performance is benchmark- and target-dependent. The local seed19 sensitivity run is not training-stage matched and is not part of this release. Do not select a checkpoint subset on the reported test targets and present it as an unbiased result.

Before making this repository public, the owner should explicitly choose and document the weights license. The base repository currently advertises MIT in its README, but the checkpoint release terms should be confirmed separately.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support