MorphoProt: RNA-Conditioned Integration of Multiscale Histology for Spatial Protein Inference
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…
MorphoProt is a novel multimodal framework designed to predict spatial protein abundance from RNA expression, multiscale histology, and tissue neighborhoods. By using RNA-guided cross-attention over five histological fields of view, the model learns the morphological context for each spot based on its molecular state. This fused representation is then refined through gated aggregation of RNA-derived information from both physical and transcriptomic neighborhoods before joint protein prediction.
In evaluations using held-out sections of human tonsil, mouse spleen, and human liver, MorphoProt demonstrates superior performance compared to RNA-only and graph-based baselines, as well as histology-based HIPI and multimodal SPRINT approaches. The framework achieves mean proteinwise Spearman correlations of 0.652, 0.604, and 0.372, respectively, across the three tissues.
The key advantage of MorphoProt lies in its ability to combine fields of view rather than relying on a single one. This multiscale model outperforms every single-field model in spleen and liver tissue, and its learned scale preference effectively separates white-pulp from red-pulp compartments. Furthermore, the predicted proteins maintain protein-defined tissue domains and their respective marker identities, accurately recovering pathologist-annotated germinal centers as well as measured proteins.
Rewritten by urgent.news; may contain errors — check the original.
Source: biorxiv.org