Perceiver CPI

Perceiver CPI predicts compound–protein interactions (CPI) by integrating molecular and protein data using a nested cross-attention architecture to support drug discovery and interaction prediction.

Key Features:

• Nested cross-attention network architecture: Aggregates information from compounds and proteins via a nested cross-attention structure to model inter-modality relationships.
• Cross-attention mechanism: Replaces simple concatenation to capture complex interactions between drug compounds and target proteins for improved representation learning.
• Extended-Connectivity Fingerprints (ECFPs): Uses ECFPs to enrich molecular representations and encode detailed compound structural information.
• Benchmark evaluation: Demonstrates improved performance compared with state-of-the-art methods on the Davis, KIBA, and Metz datasets.

Scientific Applications:

• Drug discovery: Predicts compound–protein interactions to aid identification of potential therapeutic compounds.
• Lead prioritization: Provides interaction predictions that can help prioritize candidate molecules for experimental validation.
• Alternative to docking: Offers predictive modeling that can reduce reliance on costly molecular docking simulations during early-stage screening.

Methodology:

Aggregates compound and protein information using a nested cross-attention network, represents molecules with Extended-Connectivity Fingerprints (ECFPs), and evaluates model performance on the Davis, KIBA, and Metz datasets.