PromisE

Minimal standalone PromisE inference package (no dependency on upper-level project code).

1. Repository Layout

• promisE/
  • Minimal model code (pl_trigo + mpnn + multitask head + wrapper)
  • Sequence embedder (ESM)
  • Inference entry (promisE/inference.py)
• sample_input.json: example input
• requirements.txt: minimal runtime dependencies
• params/: checkpoint location (params/params.pt)

2. Model Checkpoint (GitHub Release)

The model checkpoint is not stored in git history because it is large (~375MB).

Place the file at:

• params/params.pt

Recommended release asset name:

• params.pt

Example (replace v0.1.0 with your release tag):

curl -L -o params/params.pt \
  https://github.com/ChemBioHTP/PromisE/releases/download/v0.1.0/params.pt

3. Input Format

Input JSON fields:

• enzyme_sequence: protein sequence string
• substrate_panels: list of substrate iso-smiles (multiple entries)
• protein_structure: optional path string (accepted but not required by this minimal model)

Example (sample_input.json):

{
  "enzyme_sequence": "MNNRIT...",
  "substrate_panels": ["CC(=O)O", "CCO", "CC(C)O"],
  "protein_structure": "/path/to/optional_structure.pdb"
}

4. Install

pip install -r requirements.txt

5. Run Inference

python -m promisE.inference \
  --input sample_input.json \
  --output predictions.json

Optional flags:

• --device cuda or --device cpu
• --cache-dir ./esm_cache_msi_local
• --esm-model esm2_t30_150M_UR50D
• --checkpoint /custom/path/params.pt

6. Output

predictions.json includes:

• Panel-level summary (msi_mean/std, log_kcat_mean/std)
• Per-substrate predictions:
  • pred_msi
  • pred_log_kcat