ProSST-Evotuned
ProSST-Evotuned is a fine-tuned ProtSSN graph neural network (GNN) checkpoint used in PET-Flow for structure-based single-site mutation scoring in PET hydrolase engineering.
The checkpoint uses the ProtSSN egnn / k20_h512 architecture. Despite its release name, this file contains ProtSSN GNN weights and must be loaded with ProtSSN-compatible code.
Model details
| Item | Value |
|---|---|
| Checkpoint | ProSST-Evotuned.pt |
| Format | Native PyTorch GNN state_dict |
| GNN | ProtSSN EGNN, 6 layers, hidden channels 512 |
| Graph neighborhood | k = 20 |
| Protein language model | facebook/esm2_t33_650M_UR50D (frozen, loaded separately) |
| Input | Wild-type protein PDB structure |
| Output | Scores for all 19 nonself substitutions at each retained residue |
| File size | 592,418,469 bytes (about 565 MiB) |
The file contains the GNN parameters only. ESM2 weights, graph-normalization statistics, optimizer state, and training metadata are not bundled in this checkpoint. Inference requires the upstream ProtSSN repository, including src/ and norm/, and a CUDA GPU with the current PET-Flow inference script.
Download
pip install "huggingface-hub>=1.0"
hf download GinnM/ProSST-Evotuned ProSST-Evotuned.pt --local-dir weights
Or download from Python:
from huggingface_hub import hf_hub_download
checkpoint = hf_hub_download(
repo_id="GinnM/ProSST-Evotuned",
filename="ProSST-Evotuned.pt",
)
Run mutation scoring with PET-Flow
Install the dependencies described in the PET-Flow README and upstream ProtSSN instructions. Use the updated PET-Flow prot_ssn.py with --gnn_checkpoint support.
From the PET-Flow directory, after the CLI download above:
python prot_ssn.py \
--pdb_file protein.pdb \
--out_file protein_saturation.csv \
--protssn_root /path/to/ProtSSN \
--gnn_model_name k20_h512 \
--gnn_checkpoint weights/ProSST-Evotuned.pt
The updated script also uses this Hugging Face checkpoint by default: omit --gnn_checkpoint to download and cache it automatically. ESM2 downloads separately; --plm /path/to/local/esm2 can point to a local copy of the same model.
The output CSV contains mutant and ProtSSN_k20_h512. Mutation names such as A1V use 1-based positions in the retained structure-derived sequence, which may differ from original PDB residue numbers when residues are missing or excluded.
For each substitution, the score is:
score = log P(mutant amino acid | WT sequence and structure)
- log P(WT amino acid | WT sequence and structure)
Higher scores indicate greater relative preference under the model. They are not predicted activity percentages, melting temperatures, or calibrated probabilities of experimental improvement. The script emits 19 substitutions per retained residue and excludes self substitutions.
Training and evaluation scope
The released checkpoint was initialized from a ProtSSN checkpoint previously fine-tuned with masked language modeling on protein structures. Its subsequent fine-tuning used:
- 20 measured PETase single-mutant activity labels (
Activity at 50 ยฐC (%)). - Auxiliary masked language modeling over 2,308 eligible protein structures.
- A combined objective containing activity Pearson loss, masked-token cross-entropy, and explicit ranking supervision for the 20 selected mutation identities.
- A frozen ESM2 encoder and training of all EGNN parameters, without a separate regression head.
- Five epochs and 185 optimizer updates; the selected checkpoint was from epoch 5.
Training and evaluation used the same experimental mutation table. Candidate identities were explicitly included in the ranking objective. Reproduction of those candidates and correlations on that table measure training-data fit; they do not establish held-out generalization or prospective experimental improvement.
Inference scores mutations directly from the wild-type sequence and structure, without reading activity labels or adjusting scores from a reference candidate list. The checkpoint can be used to reproduce the associated PET-Flow scoring workflow. Predictions for new mutations, proteins, or combinations require independent evaluation; the single-site workflow does not directly score combinatorial variants.
Integrity
SHA-256 of ProSST-Evotuned.pt:
4e1285ef39cb41a00e063be615caa5e0d5c75083e2c40746cbf16090f15821f9
sha256sum weights/ProSST-Evotuned.pt