Dataset and Pre-Trained Weights for BioX-DTI: Structure-Aware Drug–Target Interaction Prediction via Pretrained Language Models and Cross-Attention Fusion

This archive contains the reproducibility materials for the manuscript "Structure-Aware Drug–Target Interaction Prediction via Pretrained Language Models and Cross-Attention Fusion" Contents of this archive: Datasets (data/): The fully preprocessed Yamanishi enzyme benchmark CSVs, including the balanced (1:1) and imbalanced (1:9) splits with SMILES strings and FASTA sequences. Feature Caches (cache/features/): The pre-extracted multimodal embeddings to bypass heavy computations. Includes protein_graphs.pkl (ESM-2 650M residue embeddings and AlphaFold-derived 3D spatial proximity graphs) and drug_emb_cache.pkl (ChemBERTa-2 substructural semantics). Model Weights (checkpoints/): The pre-trained PyTorch weights (biox_dti_1to1_bce_best.pt) for the BioX-DTI model, allowing instant zero-training evaluation.