FlexBind: Multi-Scale Complex Learning for Imbalance-Robust IDR and Binding Residue Prediction

A deep learning framework for residue-level prediction of protein properties, including:

• Intrinsically Disordered Regions (IDRs)
• Protein-binding residues
• RNA-binding residues
• DNA-binding residues

---

📌 Reproducibility

To fully satisfy requirements for generalization, experimental validity, and methodological rigor, this repository explicitly provides:

• Exact Preprocessing & Split Commands: Provided in data/preprocess.sh
• Configuration Files & Random Seeds: Defined in configs/default_config.yaml
• Pretrained Weights: Trained FlexBind models are provided in the weights/ directory. The base ProtT5 weights can be downloaded via weights/download_weights.sh.
• Baseline Instructions: Detailed guide located in baselines/baseline_instructions.md
• Environment Versions: Exact dependencies are exported in environment.yml

---

Installation

conda env create -f environment.yml
conda activate flexbind

Pretrained Weights

The trained FlexBind models (flexbind_dp81.pt, flexbind_dp93.pt, flexbind_dp94.pt) are already included in the weights/ directory.

Before running the model, you only need to download the base ProtT5 encoder weights by running the following script:

cd weights
bash download_weights.sh
cd ..

Usage

Run model training and evaluation:

python scripts/train.py --npz_path data/processed/DP81/train_val.npz --test_npz_path data/processed/DP81/test.npz