Tissue proteins without antibodies: MorphoProt lets RNA pick the H&E scale, beating SPRINT on held-out tonsil, spleen and liver

Spatial protein abundance can now be inferred from the H&E images routinely collected alongside spatial transcriptomics plus RNA expression, covering the whole protein panel rather than the few proteins an antibody panel selects. Earlier spatial protein inference relied mainly on transcriptomic profiles and spatial adjacency, or predicted proteins from histology alone; the multimodal SPRINT combined histology with spatial transcriptomics but did not let each spot's molecular state decide which morphological scale to read. On held-out sections of human tonsil, mouse spleen and human liver, MorphoProt reached mean protein-wise Spearman correlations of 0.652, 0.604 and 0.372, the highest among the compared methods in all three tissues; it exceeded the matched RNA-only network by 0.038, 0.040 and 0.065, and the strongest multimodal baseline SPRINT by 0.018, 0.068 and 0.018.

Which histological scale a spot reads is set by that spot's own RNA state, so regions in different molecular states within one section can favor different fields of view. A single fixed field or one uniform weighting across spots cannot differentiate multiscale information by spot, and concatenating the five fields into a Transformer performed worse than RNA-guided cross-attention. In mouse spleen, white-pulp spots placed the largest attention weight on the finest field (0.29 versus 0.19 on the coarsest), while red-pulp spots placed it on the coarsest (0.27 versus 0.13 on the finest), with the same ordering across all four seeds; removing the physical graph lowered Spearman correlation by 0.010 to 0.014, more than removing the transcriptomic graph.

The predicted proteins preserve protein-defined tissue domains and germinal-center structure, not just per-spot numerical agreement. Correlation alone does not show whether predictions recover tissue architecture; the comparison methods agreed less with measured-protein domains under the same clustering. Clustering measured and predicted proteins independently with k-means, MorphoProt reached an adjusted Rand index of 0.703 in tonsil and 0.334 in spleen at k=5, versus 0.513 and 0.195 for SPRINT and 0.656 and 0.280 for RNA-only; tonsil predicted domains correlated 0.95 to 0.98 per domain with measured marker profiles, and 86.8% of spots fell in the same domain.

Evaluation is limited to one held-out section partition each for human tonsil, mouse spleen and human liver, with protein panels of 29, 21 and 31, and training requires protein-measured reference sections paired with spatial transcriptomics. The intended reader is a researcher working on spatial transcriptomics and tissue-image analysis; generalization across donors, staining protocols, platforms and unseen panels still needs independent validation.

The histology gain in tonsil is small, with the validation-selected single-field model matching the five-field model (0.651 versus 0.652) and the five tonsil field embeddings the most similar; in liver the paired interval against SPRINT includes zero, and the two methods perform comparably on germinal-center identification; per-protein comparisons were not adjusted for multiple testing, and the confidence intervals do not capture variability across independent donors or sections.