Hybrid Dual-Context Prompted Cross-Attention Framework with Language Model Guidance for Multi-Label Prediction of...

Area of Science:

• Computational chemistry and drug discovery
• Bioinformatics and computational biology
• Artificial intelligence in pharmaceutical research

Background:

• Accurate drug off-target identification is crucial for pharmaceutical development, yet current deep learning methods struggle to integrate diverse data types.
• Predicting ligand binding to proteins is essential for reducing drug toxicity and enhancing drug discovery success rates.
• Existing models often fail to effectively fuse chemical structure, protein biology, and multi-target information.

Purpose of the Study:

• To introduce HDPC-LGT (Hybrid Dual-Prompt Cross-Attention Ligand-Protein Graph Transformer), a novel framework for predicting ligand binding to proteins.
• To develop a model capable of integrating chemical structure, protein language model embeddings, and structural priors for robust interaction prediction.
• To predict ligand binding across sixteen human translation-related proteins associated with antibiotic toxicity.

Main Methods:

• HDPC-LGT combines graph-based chemical representation with protein language model embeddings and structural information.
• The framework was trained on over 216,000 experimentally validated ligand-protein pairs from ChEMBL and BindingDB.
• Model performance was evaluated using rigorous scaffold-level, protein-level, and combined holdout strategies, alongside external datasets like Papyrus, PDBbind, and Yamanishi.

Main Results:

• HDPC-LGT achieved a macro ROC-AUC of 0.996 and a micro F1-score of 0.989, significantly outperforming existing state-of-the-art models (DeepDTA, GraphDTA, MolTrans, CAT-DTI, HGT-DTA) by 3-7%.
• External validation datasets confirmed the model's strong generalization capabilities to novel chemical structures and proteins.
• Interpretability methods (cross-attention maps, IG, Grad-CAM) highlighted key interactions and residues, aligning with known biochemical mechanisms.

Conclusions:

• HDPC-LGT effectively integrates multimodal biochemical data for accurate ligand-protein interaction prediction and off-target toxicity assessment.
• The framework offers biologically interpretable insights, aiding in structure-based drug design and lead optimization.
• HDPC-LGT presents a valuable tool for antibiotic development, safety profiling, and polypharmacology research.