MorphoProt: RNA-Conditioned Integration of Multiscale Histology for Spatial Protein Inference

Abstract

Spatially resolved protein abundance reveals cellular states and tissue organization that RNA expression cannot fully capture, yet direct spatial protein profiling remains costly and limited in coverage. Histology images routinely accompany spatial transcriptomics and offer complementary morphological information for protein inference. Here, we introduce MorphoProt, a multimodal framework that jointly predicts spatial protein abundance from RNA expression, multiscale histology and tissue neighborhoods. MorphoProt uses RNA-guided cross-attention over five histological fields of view, so that the morphological context read for each spot is conditioned on its molecular state. The fused representation is refined by gated aggregation of RNA-derived information from physical and transcriptomic neighborhoods before joint protein prediction. On held-out sections of human tonsil, mouse spleen and human liver, MorphoProt achieves mean proteinwise Spearman correlations of 0.652, 0.604 and 0.372, outperforming matched RNA-only and graph-based baselines and the histology-based HIPI and multimodal SPRINT baselines in all three tissues. Combining fields of view, rather than any single one, carries the histology gain: in spleen and liver, the multiscale model outperforms every single-field model, and its learned scale preference separates white-pulp from red-pulp compartments. Predicted proteins preserve protein-defined tissue domains and their marker identity, and recover pathologist-annotated germinal centers as accurately as measured proteins.