Trained Model Weights for OmniBind: Pan–Pharmacological Drug–Target Interaction Prediction with 3D–Informed Protein Encoding at Scale

### OmniBind: Pre-trained model checkpoints for compound-protein interaction prediction This repository contains the pre-trained model weights for **OmniBind**, a unified framework for pan-pharmacological prediction of drug-target interactions (DTIs) using 3D-structure-informed adaptive gated fusion. OmniBind simultaneously predicts four key binding affinity metrics (**Ki, Kd, IC50, and EC50**) from:- Compound SMILES- Protein amino acid sequences- Foldseek 3Di structural alphabet sequences #### Included CheckpointsThe provided weights are organized to ensure both scientific reproducibility and practical utility: - **benchmark/seed[42, 123, 369, 777, 2024].pth**: These five checkpoints correspond to the performance evaluation results (mean ± std) reported in the associated paper. Each represents the best test-set RMSE model for its respective random seed.- **application.pth**: The specific model selected for practical application analyses, such as drug repositioning and off-target screening, due to its robust generalization to novel compound-target pairs. #### UsageTo use these weights, download the files and place them into the `checkpoints/` directory of the OmniBind repository. For installation, environment setup, and detailed usage instructions, please visit our GitHub repository:https://github.com/Shimizu-team/OmniBind #### Associated Publication**"Pan–Pharmacological Drug–Target Interaction Prediction with 3D–Informed Protein Encoding at Scale"**