Model Framework
Architecture Overview
Feature Encoding: ESMC global and residue-level protein representations are combined with ECFP4, MACCS, and standardized RDKit2D molecular features.
Multiview Fusion: Learned gates weight complementary protein and molecular views before interaction modeling.
Interaction Decoder: Cross-modal compatibility matching, feature interaction, and bilinear coupling model compound–protein relationships.
Virus-Domain Adapter: Virus-aware adaptation refines the decoded interaction representation.
Prediction Heads: Separate released heads support classification and ranking; the two task scores are not interchangeable.
Classification probability ranges from 0 to 1. A larger value indicates a stronger predicted compound–protein interaction.Ranking score prioritizes compounds for one selected target. A larger value indicates a higher-priority candidate within the submitted library.