AIP-DualBranch model files

This repository contains the released model files for AIP-DualBranch, a research model for anti-inflammatory peptide classification. The source code, reproducibility workflow, local predictor, and full usage documentation are maintained in the GitHub repository.

Repository contents

Fine-tuned PLM checkpoints

The checkpoints/ directory contains five outer-fold model weights:

fold1_best_last4_cls_l2_finetune.pt
fold2_best_last4_cls_l2_finetune.pt
fold3_best_last4_cls_l2_finetune.pt
fold4_best_last4_cls_l2_finetune.pt
fold5_best_last4_cls_l2_finetune.pt

Each checkpoint is approximately 2.6 GB before Hugging Face Xet deduplication. Inference extracts one 1,280-dimensional representation per checkpoint and averages the five representations.

Fitted downstream artifacts

The artifacts/ directory contains:

• plm_scaler.joblib: PLM-branch feature scaler;
• plm_grande.pkl: fitted GRANDE PLM-branch classifier;
• dde_scaler.joblib: DDE-branch feature scaler;
• dde_tabicl.pkl: fitted TabICLv2 DDE-branch classifier;
• meta_lr.joblib: fitted logistic-regression fusion model;
• manifest.json: model metadata, independent-test metrics, and SHA-256 checksums.

Download

Install the Hugging Face CLI and download these released files into a clone of the source repository:

python -m pip install -U huggingface_hub
hf download moyeyeye/AIP-DualBranch \
  --include "checkpoints/*" "artifacts/*" \
  --local-dir .

The public base PLM is maintained by its original authors and is not duplicated in this repository:

hf download danliu1226/PLM-interact-650M-humanV12 \
  --local-dir models/plm_base

Use with the source code

git clone https://github.com/YKRhs/AIP-DualBranch.git
cd AIP-DualBranch
python -m pip install -r requirements.txt

hf download moyeyeye/AIP-DualBranch \
  --include "checkpoints/*" "artifacts/*" \
  --local-dir .
hf download danliu1226/PLM-interact-650M-humanV12 \
  --local-dir models/plm_base

python scripts/check_installation.py
python app/web_app.py

Detailed instructions are available in the inference guide and reproduction guide.

Reported independent-test result

On the fixed 419-sequence independent test set, the finalized dual-branch deployment model recorded AUC 0.8559, accuracy 0.8091, MCC 0.6095, sensitivity 0.7667, specificity 0.8410, precision 0.7841, and F1 0.7753.

Important notes

• This is a research release and not a clinical decision system.
• The model was developed for peptide sequences using the 20 standard amino acids. The observed training and independent-test sequences were at most 25 residues long.
• Pickle and Joblib files can execute code during deserialization. Load these artifacts only from a trusted source and use the dependency versions documented by the source repository.