Structure-Aware Multimodal Deep Learning for Drug-Protein Interaction Prediction.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Accurate prediction of drug-protein interactions (DPIs) is vital for drug discovery.
• Existing machine learning methods for DPI prediction often suffer from biased datasets and inadequate protein representation.
• There's a need for robust models that consider both molecular and protein features for enhanced virtual screening.

Purpose of the Study:

• To introduce STAMP-DPI, a novel structure-aware multimodal deep learning model for predicting drug-protein interactions (DPIs).
• To establish a high-quality, industry-scale benchmark dataset (GalaxyDB) for unbiased DPI prediction model training and evaluation.
• To improve the accuracy and interpretability of DPI prediction methods.

Main Methods:

• Developed STAMP-DPI, a deep learning model integrating structure-aware protein representations (using graph neural networks and contact maps) with pretrained molecular and protein embeddings.
• Curated and utilized GalaxyDB, a large-scale, high-quality benchmark dataset for training and validation, ensuring an unbiased training procedure.
• Employed a transformer-based interaction mechanism for model interpretability, enabling the identification of binding sites.

Main Results:

• STAMP-DPI demonstrated superior performance compared to state-of-the-art DPI prediction methods.
• Achieved a 7.00% reduction in mean square error (MSE) on the Davis dataset.
• Improved the area under the curve (AUC) by 8.89% on the GalaxyDB dataset.

Conclusions:

• The developed STAMP-DPI model offers a more robust and accurate approach to predicting drug-protein interactions.
• The structure-aware protein representation and unbiased dataset significantly enhance prediction capabilities.
• The model's interpretability allows for the identification of key binding sites, aiding in drug discovery efforts.