Contextual gene states support gene-expression reconstruction, while CLS-conditioned reconstruction and contrastive learning provide complementary supervision for the cell-level representation.
SpaAxFormer first contextualizes genes within each cell and then exchanges corresponding-gene states across spatial neighbors, while a separate CLS representation provides a cell-level summary.
This stacking strategy lets the ensemble leverage the complementary strengths of different integration and embedding methods (e.g., Harmony, Scanorama, scVI, scGPT, scANVI).