MorphoProt links RNA signals to tissue morphology for spatial protein prediction
MorphoProt is built around a problem that has dogged spatial biology from the start: RNA can map where genes are active, but transcript counts do not always reflect the proteins that define cell state, tissue architecture, or functional compartments. In tissues where protein gradients separate immune niches, vascular zones, or parenchymal regions, the inability to measure protein directly at scale leaves an important gap between molecular mapping and pathology-grade interpretation.
The new framework addresses that gap by integrating RNA expression, multiscale histology, and tissue neighborhoods into one prediction pipeline for spatial protein abundance. Its central design choice is not generic image fusion. RNA guides cross-attention across five histological fields of view, so the morphology being read for each spot is conditioned on the molecular state of that spot rather than treated as a fixed visual template. In practice, that creates a more selective bridge between tissue appearance and expected protein output.
One sentence captures the clinical and computational consequence: protein inference improves when the model knows where to look and what molecular context to privilege.
RNA-conditioned attention improved correlation across tonsil, spleen, and liver
The held-out evaluation spanned human tonsil, mouse spleen, and human liver, three tissues with different structural logic and different demands on protein inference. MorphoProt reached mean proteinwise Spearman correlations of 0.652 in tonsil, 0.604 in spleen, and 0.372 in liver. Those values outperformed matched RNA-only baselines and graph-based baselines in all three tissues, while also beating the histology-only HIPI and multimodal SPRINT systems.
The tissue spread matters because it shows the model is not locked to a single morphological regime. Tonsil contains densely organized immune architecture, spleen alternates between white pulp and red pulp, and liver brings a different spatial organization altogether. A method that works across those settings has to capture both local texture and broader tissue context, then decide which scale carries the most relevant signal for each spot.
The liver result was lower than the others, but the framework still remained ahead of the comparison methods. In spatial proteomics, a lower absolute correlation in a structurally complex organ does not erase the value of a better rank-order mapping, especially when the goal is to recover tissue domains and protein-defined neighborhoods rather than to estimate single analytes in isolation.
Multiscale histology carried more signal than any single field of view
The most consequential design lesson comes from the multiscale component. Combining fields of view, rather than relying on any single image scale, produced the histology gain in spleen and liver. In those tissues, the multiscale model outperformed every single-field model, and the learned scale preference separated white-pulp from red-pulp compartments in spleen.
That separation is biologically plausible. White pulp and red pulp have distinct cellular compositions, vascular structures, and immune functions, so a model that can shift attention across histologic scales has a better chance of detecting the visual cues that align with protein expression. Small fields can capture cellular detail, while larger fields preserve architectural context. Protein expression sits somewhere between those layers, and MorphoProt appears to exploit that overlap more effectively than a fixed-scale approach.
The architecture does more than stitch modalities together. After RNA-conditioned cross-attention over histology, the fused representation is refined by gated aggregation of RNA-derived information from both physical neighborhoods and transcriptomic neighborhoods before protein prediction. In other words, the model does not treat a spot as an isolated patch. It uses adjacency, molecular similarity, and morphology together, which is exactly the sort of structure needed when proteins reflect local tissue organization as much as they reflect single-cell transcriptional state.
Protein-defined domains and germinal centers were recovered with greater fidelity
Beyond correlation scores, MorphoProt preserved protein-defined tissue domains and their marker identity. The predicted protein maps recovered pathologist-annotated germinal centers as accurately as the measured proteins. That is a meaningful result because germinal centers are not merely visual structures; they represent functionally distinct immune microenvironments whose boundaries are often clearer in protein space than in RNA space.
Here the biological logic is straightforward. Proteins act downstream of transcription, and in organized tissues they often track cellular specialization, cell-cell interaction, and microanatomic compartmentalization more faithfully than RNA counts alone. When a model can reconstruct those domains from histology plus RNA, it is not just interpolating missing values. It is learning how tissue organization encodes the molecular identity of a region.
This matters for spatial pathology because many diagnostic and research questions depend on recognizing compartment-specific protein patterns. Germinal centers, splenic pulp regions, and hepatic zones each carry distinct functional signatures. A method that can infer those signatures from routinely collected spatial transcriptomics slides could reduce the need for direct protein profiling across every sample, while extending the information value of existing datasets.
What this model can and cannot claim yet
The findings are strong for a preprint, but they remain bounded by the study design. The abstract describes held-out sections from three tissues and benchmark comparisons against several baselines, yet it does not provide broader external validation across additional organs, disease states, staining conditions, or platform types. It also does not report downstream clinical use cases, prospective deployment, or performance in sparse or damaged specimens.
Even so, the work clarifies a useful direction for spatial omics: protein inference improves when histology is read at multiple scales and when that reading is conditioned on the RNA state of the same spot. The model’s architecture reflects a more realistic view of tissue biology, where morphology, molecular state, and neighborhood effects are inseparable. For pathology-adjacent spatial analysis, that is a more credible route than single-modality prediction.
The next question is not whether morphology matters. The question is how far a multiscale, RNA-conditioned framework can be pushed across diseased tissues, heterogeneous sampling platforms, and proteins with weaker transcriptional coupling. Until those evaluations arrive, MorphoProt stands as a technically careful proof that spatial protein inference can gain resolution when histology is allowed to answer to RNA instead of standing apart from it.