Generalizable compound protein interaction prediction with a model incorporating protein structure aware and compound...

Area of Science:

• Computational chemistry
• Bioinformatics
• Artificial intelligence in drug discovery

Background:

• Compound-protein interaction (CPI) prediction is vital for drug discovery, but current deep learning models face limitations due to reliance on sequence data and insufficient labeled datasets.
• These limitations hinder the accuracy and broad applicability of existing models in identifying binding affinities between small molecules and proteins.

Purpose of the Study:

• To develop a novel deep learning model, GenSPARC, for more accurate and generalizable CPI prediction.
• To overcome the data scarcity and representation limitations of current sequence-based deep learning approaches.

Main Methods:

• GenSPARC integrates structure-aware protein representations from AlphaFold2 predictions and FoldSeek's 3D interaction alphabet.
• Compound features are extracted using graph convolutional networks and a pre-trained chemical language model for multimodal representation.
• An attention mechanism is employed to capture complex binding patterns and enhance interaction modeling.

Main Results:

• GenSPARC demonstrated strong generalizability across diverse CPI benchmark datasets, including challenging data splits.
• The model achieved competitive performance in virtual screening tasks, indicating its effectiveness in practical drug discovery applications.
• Validation confirmed the model's ability to accurately predict binding affinities and interactions.

Conclusions:

• GenSPARC represents a significant advancement in AI-driven drug discovery by improving CPI prediction accuracy and generalizability.
• The model's multimodal approach, leveraging structural and chemical information, addresses key limitations of previous methods.
• GenSPARC is poised to accelerate the identification of potential drug candidates by enhancing the understanding of molecular interactions.