Hypernetwork-guided fusion with intra-class MixUp for breast cancer subtyping

Accurate breast cancer subtyping guides treatment selection, yet histopathology captures morphology without molecular state, while genomic profiling captures molecular signatures without spatial context. Existing fusion methods rely on concatenation, or on attention applied only after each modality is encoded independently. This work identifies a scale-dependent asymmetry in the direction of cross-modal conditioning: the direction that performs best under limited samples is not the one that holds at scale, and the reversal is traced to the capacity of the modulation pathway rather than to the fusion principle. The comparison is carried out within a hypernetwork-guided framework in which an auxiliary network maps one modality to conditioning parameters that modulate the other's feature representation, shaping features at the parametric level rather than the decision stage; modulation is patient-specific rather than patch-specific. Both directions are instantiated-gene-to-image (HyperG2I) and image-to-gene (HyperI2G) - and trained under a label-aware MixUp strategy that interpolates within-class samples across both modalities, preserving the hard binary labels clinical decisions require. The framework is evaluated on two paired TCGA-BRCA cohorts-one limited-sample, one independently assembled at scale-under a single protocol spanning two whole-slide representations, multiple visual backbones, and both conditioning directions. On the limited-sample cohort, gene-to-image conditioning at its optimal augmentation setting exceeds early fusion and both unimodal baselines, giving the highest recall on the aggressive Basal/HER2 class of any configuration evaluated, and an ablation favours intra-class over inter-class mixing. At scale this ordering does not hold: image-to-gene conditioning sustains its performance whereas gene-to-image does not, recovering only partially under the full tissue bag and isolating the capacity of the modulation pathway as the binding constraint. Direction and capacity of cross-modal conditioning, rather than fusion depth alone, therefore govern how such frameworks scale.