YudoX/iASAP-Fuse-weights

iASAP-Fuse Weights

Pre-trained weights for iASAP-Fuse, a deep-learning model that predicts the anti-skin-aging activity of peptides by fusing ProtBERT contextual embeddings with engineered Z-scale physicochemical descriptors.

The companion code, CLI and local web UI are released as the Python package iasapfuse on the project's GitHub repository.

Model summary

• Task: binary classification (anti-skin-aging peptide / non anti-skin-aging peptide)
• Backbone: Rostlab/prot_bert (frozen feature extractor)
• Head: fusion network combining ProtBERT [CLS] embeddings + Z-scale descriptors
• Training: 10-fold cross-validation, Stochastic Weight Averaging (SWA) per fold
• Ensemble: prediction-time average over 10 folds, calibrated with the saved stats.json

Files

.
├── stats.json              # ensemble normalisation / threshold metadata
├── fold_1/
│   ├── best_swa.pt         # SWA model weights
│   └── metrics_final.json  # held-out fold metrics
├── fold_2/
│   ├── best_swa.pt
│   └── metrics_final.json
├── ...
└── fold_10/
    ├── best_swa.pt
    └── metrics_final.json

All .pt files are PyTorch state dicts intended to be loaded by iasapfuse.inference.

How to use

Option 1 — via the iasapfuse CLI (recommended)

pip install iasapfuse  # or install from source
iasapfuse weights download \
  --repo-id YudoX/iASAP-Fuse-weights \
  --target-dir ./weights
iasapfuse predict examples/predict_sequences.csv --device cpu

Option 2 — via huggingface_hub

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="YudoX/iASAP-Fuse-weights",
    local_dir="./weights",
    allow_patterns=["fold_*/*.pt", "fold_*/*.json", "stats.json"],
)

License

The weights are released under CC BY-NC 4.0 (Attribution–NonCommercial 4.0 International). Academic and non-commercial research use is permitted with attribution. Commercial use requires separate permission from the authors.

Citation

A formal BibTeX entry will be added once the paper is published.

Disclaimer

These weights are intended for research only. They are not validated for clinical, diagnostic, cosmetic-product or any production use.