Beyond EEB · A paper from outside evolution and ecology
Combining two cell models is not always better
Richter et al., Nature Machine Intelligence (2026). Research article.
A foundation model trained on cell gene expression and another trained on tissue images each capture useful biology. But what happens when their outputs are combined? Researchers compared ten methods of integrating these data across spatial transcriptomics datasets, where measurements are tied to tissue locations. They introduced a Synergistic Information Score to ask whether the combination reveals useful information beyond either input alone. Some tasks, such as predicting cell composition, were already well served by one modality. Others, including certain tissue niche classifications, benefited from methods that learned cross-modal interactions. The lesson is practical: multimodal AI needs complementary evidence, not merely more inputs.