EGA-DTA

EGA-DTA: An Energetic-Geometric Augmented Graph Neural Network with Target-Conditional Gating for DTA Prediction

• BDEGNN.py: Core model definition including EnhancedGATNet, TCGI, and Multi-Strategy Fusion.
• DTA_train_test.py: Main training script with Cosine Annealing scheduler and MSE loss.
• precompute_ESM2_3B.py: Script to extract 2560-dim protein features using ESM-2 3B.
• drug_graph_construct.py: Generates molecular graphs with BDE/BL attributes (requires alfabet for BDE prediction：https://github.com/ NREL/alfabet ).
• convert_pytorch_data.py: Prepares PyTorch Geometric datasets.
• utils.py: Metrics (CI, MSE, $r_m^2$) and data loading utilities.

Step 1: Drug Graph Construction Note: The BDE calculation relies on the TensorFlow-based alfabet library. To save time, we provide the pre-computed graph dictionary: smile_graph.pkl: Contains graph structures, BDE, and Bond Lengths for the datasets. Place this file in the root directory. python drug_graph_construct.py

Step 2: Protein Feature Extraction (ESM-2 3B) Extract the 2560-dimensional embeddings for protein sequences. This requires a GPU with approx. 16GB+ VRAM (or reduce batch size). python precompute_ESM2_3B.py davis 4

Step 3:Convert to PyTorch Format Combine the drug graphs and protein features into .pt files for training. python convert_pytorch_data.py --use-precomputed-esm2

Since the ALFABET model is based on TensorFlow, we have placed all preprocessed data (smile_graph.pkl, esm2_features_.pkl, <datsetname_task>.pt) and trained model weights at the following link: https://drive.google.com/drive/folders/1eQPO3dr4NfqUja_oW_bSAq6DUVAIyGWl?usp=drive_link