Choosing What to Align: Predictability-Weighted RNA Targets Improve Histology--RNA Retrieval

Histology and spatial transcriptomics provide complementary views of tissue, but not all transcriptomic variation is reproducibly reflected in morphology. Nevertheless, RNA representations used for cross-modal learning are typically constructed from all measured genes or from transcriptomic variance alone, without considering whether the retained information is accessible from histology. We ask whether H&E--RNA alignment improves when the RNA space instead emphasizes genes whose expression is reproducibly predictable from H&E across patients. We test this in 31 lung patients comprising 28,362 paired H&E patches and 343-gene spatial transcriptomic profiles. Within each leave-one-patient-out fold, gene predictability is estimated exclusively from training patients and used to construct predictability-weighted RNA representations. Predictability weighting consistently improves exact-pair retrieval across held-out patients and in both retrieval directions. Improvements persist against permutation controls and are not fully explained by transcriptomic variance, while reversing the predictability weights substantially degrades retrieval performance. These results suggest that the transcriptomic target itself is an important design choice in multimodal alignment: RNA variation that is reproducibly accessible from morphology can provide a more effective target for cross-modal retrieval.