Perceiver CPI: a nested cross-attention network for compound-protein interaction prediction.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery

Background:

• Compound-protein interaction (CPI) prediction is crucial for drug discovery but traditionally relies on costly molecular docking.
• Existing AI models, including graph convolutional neural networks and descriptor-based neural networks, show promise but have limitations in effectively integrating compound and protein information.
• Current methods often use simple concatenation, failing to fully capture complex inter-molecular interactions.

Purpose of the Study:

• To introduce the Perceiver CPI network, an AI model designed to improve the prediction of compound-protein interactions.
• To enhance the representation learning of drug-target interactions by employing a cross-attention mechanism.
• To leverage extended-connectivity fingerprints for superior performance in CPI prediction.

Main Methods:

• Developed the Perceiver CPI network, incorporating a cross-attention mechanism.
• Utilized extended-connectivity fingerprints as input features for compounds.
• Evaluated the model on the Davis, KIBA, and Metz datasets.

Main Results:

• The Perceiver CPI network achieved satisfactory performance across all tested datasets.
• Demonstrated significant improvements compared to existing state-of-the-art methods for CPI prediction.
• The cross-attention mechanism effectively captured complex interactions between compounds and proteins.

Conclusions:

• Perceiver CPI offers a superior approach to predicting compound-protein interactions.
• The model's ability to learn rich representations through cross-attention advances AI applications in drug discovery.
• The proposed method provides a more effective way to integrate molecular information for improved CPI prediction accuracy.