Molecular Prototype Construction

H&E-only clustering for prototypes recovers only half the performance gain of full cross-modal prototypes, confirming that molecular discovery rather than extra projection capacity drives the core improvement. Performance degrades consistently at K=64 prototypes, suggesting that only a compact set of molecular concepts is necessary to span the molecular axes accessible from H&E. The key architectural choice is that clustering is performed in gene expression space to discover molecular states, but each prototype resides in morphological feature space to directly condition H&E patch embeddings. Prototypes are built per organ because molecular concepts differ substantially across tissue types. The prototype bank is computed once per organ and stored offline, making it reusable without recomputation.