Behavior of ConGLUDe on predicted protein structures and highly divergent proteins

Determine the performance and generalization characteristics of Contrastive Geometric Learning for Unified Computational Drug Design (ConGLUDe) when applied to predicted protein structures (such as AlphaFold-derived models) and to proteins that are highly divergent from known structural templates, including whether its virtual screening, target fishing, and pocket prediction capabilities remain reliable in these settings.

Background

ConGLUDe is introduced as a unified contrastive geometric learning framework that couples a VN-EGNN-based protein encoder with a ligand encoder to jointly leverage structure-based and ligand-based data. It demonstrates strong results across multiple tasks, including virtual screening, target fishing, and ligand-conditioned pocket selection, primarily evaluated on proteins with experimentally resolved structures.

In the Limitations section, the authors note uncertainty regarding the model’s behavior on predicted structures and on proteins that deviate substantially from known templates. Since many real-world targets lack high-resolution experimental structures and rely on predictions (e.g., AlphaFold), characterizing ConGLUDe’s robustness in such scenarios is crucial for broader applicability.