Improve low-homology receptor generalization

Improve the generalization of the Dual-State Query (DSQ) model to GPCRs with low sequence homology to receptors represented in the training data, addressing its unresolved limitation on such cold-target receptors.

Background

The paper evaluates DSQ on test receptors stratified by their maximum sequence identity to training receptors. Although DSQ performs well on high-homology receptors, its performance deteriorates substantially on the low-homology subset, with near-zero or negative coefficients of determination for overall, agonist, and antagonist prediction. The authors identify this low-homology generalization issue as an unresolved limitation of the sequence-based formulation.

The limitation is especially relevant because DSQ relies on frozen ESM-2 and MolFormer representations and does not directly incorporate receptor-specific structural or conformational information. Extending the model to generalize reliably across large evolutionary distances would improve its utility for novel, orphan, or otherwise poorly represented GPCR targets.