When a surgeon removes a tumor, a pathologist checks the cut edge under the microscope. The edge is called "R0" if no cancer cells are seen there and "R1" if cancer cells are seen. That yes-or-no call explains recurrence poorly. In one retroperitoneal liposarcoma cohort, every recurrence came back at the original site, even though no visible tumor had been left behind (Gao et al.). One explanation is that the tissue around a tumor can look normal but already be biologically changed. A natural response is to measure the margin's molecules, and the cheapest molecule to measure is RNA. RNA is a weak stand-in for protein, though. Messenger RNA levels track protein levels only 40 to 60 percent of the time (Adhikari et al.). Where in a tissue that link breaks down, and whether it breaks down at the tumor border, has never been measured.
We propose to measure it, and then to test whether the breakdown can be read from the routine stained slide. On resection specimens profiled for both spatial RNA and multiplexed protein, we group genes into co-regulated modules using weighted gene co-expression network analysis. This method links genes whose expression rises and falls together and summarizes each group by a single score. We then ask how far each module's protein-level activity departs from what its RNA-level activity predicts, as a function of distance from the tumor edge. The claim is that this module-level RNA–protein gap is larger in peritumoral margin tissue than in the tumor core or in distant normal tissue. If so, a model trained on the slide stained with hematoxylin and eosin (H&E, the routine pink-and-purple stain) can predict the gap, and the predicted gap adds recurrence information beyond R0/R1 status and Ki67, a stain for dividing cells.
Core Claim
The new object is a spatial map of discordance at the level of modules, anchored to the surgical margin. The method works in three steps. First, we fit, within the tumor core of each patient cohort, a mapping from each module's RNA summary score (its "eigengene," the first principal component of its member genes) to its protein summary score, built from the module members that appear on the protein panel. Second, we apply that mapping at every tissue location. Third, we define the discordance $D_m(x)$ for module $m$ at location $x$ as the observed protein score minus the predicted one.
The hypothesis is that the size of $D_m$, averaged over modules involved in stress, matrix remodeling and immune activation, peaks in a band roughly within a few hundred micrometers outside the tumor boundary. A second hypothesis is that the H&E-predicted version of this quantity, evaluated at the inked margin, stratifies local recurrence among R0 and R1 patients.
We work at the module level rather than gene by gene for two reasons. Spot-level RNA is sparse and noisy, and averaging over co-regulated genes is the standard way to recover a stable signal. The module also explicitly encodes the network structure that, per the scChord authors, gene-by-gene RNA-to-protein models lose as panels grow (scChord).
Relation to Prior Work
Two lines of work supply the image-to-molecule machinery. Neither looks at RNA–protein discordance or at margins. HistoPlexer generates an 11-protein imaging mass cytometry map from H&E in melanoma, using multiscale and contrastive losses to tolerate misalignment between adjacent sections (HistoPlexer). HEX predicts a 40-protein panel from H&E in lung cancer on top of a pathology foundation model (HEX/MICA). On the RNA side, STimage-1K4M pairs H&E tiles with spatial transcriptomes. We reuse these models as image encoders and as baselines. We change the prediction target from single markers to a cross-modal residual, and we change the evaluation region to the margin.
Three studies motivate the hypothesis without testing it in tissue. Breen et al. ran this method on proteomics and found that RNA–protein correlation weakens with age (r = −0.56) and differs by module, but in ground-up bulk brain tissue with no spatial information. HEIST combines gene co-expression with spatial cell neighborhoods. It does so on coordinate tables without images, and it handles RNA and protein separately rather than modeling the gap between them. Shen and Cooper correlated co-expression modules with whole-slide image features. They also warn that cohorts of 100 to 800 patients make module discovery fragile. This warning constrains our design: modules are built across tens of thousands of tissue spots within patients and must replicate across patients before use.
Clinically, Niu et al. document frequent positive posterior margins in distal pancreatectomy for pancreatic ductal adenocarcinoma (PDAC). That makes PDAC our primary setting, with liposarcoma as a second tissue.
Research Program
Discovery cohort. We assemble a discovery cohort of PDAC resection specimens from which sections can be cut that span tumor core, invasive front, inked margin and distant parenchyma. On these sections we run spatial transcriptomics and a 40–50 marker multiplexed protein assay. We use the same section where the platform allows. Otherwise we use adjacent sections registered to the H&E image, with alignment error measured on landmark nuclei.
Panel design. The protein panel limits which modules can be scored. We therefore design the panel after an initial round of module building on public PDAC transcriptomes. The goal is to place at least three proteins in each of the roughly ten modules we expect to be most important. Modules with fewer measured proteins are reported as unscorable rather than imputed.
Building and tracking modules. Modules are built on the spatial RNA using standard soft-thresholded co-expression networks. We keep only modules that are preserved in a held-out set of patients. For each kept module, we compute $D_m(x)$ and model it against signed distance to the annotated tumor boundary, with a random effect per patient. We also report module-preservation statistics between the RNA and protein layers within distance bins, which asks whether the network itself rewires, not only whether the levels shift.
Prediction from H&E. We train an H&E encoder, initialized from the HEX architecture, with three output heads that predict RNA module scores, protein module scores and $D_m$ directly. We use HistoPlexer-style multiscale losses to absorb misalignment between sections.
Clinical test. The trained model is applied to archival margin slides from a retrospective PDAC cohort with local-recurrence follow-up.
Baselines
Naive reference point. The do-nothing reference is the current clinical picture: a Cox model for local-recurrence-free survival using R0/R1 status, tumor grade and manual Ki67. Ki67 and malignancy grade were the only independent risk factors in the liposarcoma cohort (Gao et al.). No C-index for this kind of model is reported on Lacuna, so we compute it on our own cohort.
Strongest existing method. The strongest competitor is HEX-style virtual proteomics applied to the margin, with predicted markers pooled into a Cox model. HEX reports an average Pearson r of 0.790 across 40 proteins in cross-validation (HEX/MICA). HistoPlexer reports a 3.18% C-index gain from adding virtual protein features to H&E features (HistoPlexer). We retrain HEX on our margin sections using the same encoder, the same training spots, the same patient-level splits and the same hyperparameter search budget as our model. Its survival head sees the same margin regions.
RNA-only counterpart. A STimage-style H&E-to-RNA model predicts module eigengenes from RNA alone, with no protein term.
Ablation of the key ingredient. The ablation that isolates our contribution keeps everything fixed but replaces module-level discordance with per-gene residuals between RNA and protein over the same panel genes.
What counts as winning. Beating the baselines requires a statistically significant C-index improvement over both the clinical model and the retrained HEX model on held-out patients. It also requires that the improvement shrink substantially when the discordance head is replaced by the protein-only or RNA-only heads.
Why not just…?
Why not just predict proteins from the margin H&E and skip RNA entirely? This is the real threat. If protein abundance is what drives recurrence, the RNA comparison adds nothing. The experiment settles it directly: the protein-only head versus the discordance head, with an identical encoder and identical survival model. If protein-only matches or beats discordance, we report that. The spatial discordance map then remains a biological result, but the clinical claim reduces to "apply HEX at margins," which is a smaller and still untested contribution.
Why not just use spatial RNA, which covers the whole transcriptome? The motivation is that RNA can miss functional state. Han et al. needed mass cytometry to see shifts in immune activation after neoadjuvant therapy that transcriptomics missed. Whether this matters at margins is exactly what the discovery phase measures. If $D_m$ is flat across distance from the tumor, RNA suffices and the proposal stops at that negative result.
Why not just correlate each gene's RNA with its protein instead of building modules? Single spots carry too few RNA counts for stable per-gene residuals. A 40-marker panel also gives only a few dozen such pairs, each noisy. The per-gene residual ablation tests whether the averaging over modules actually helps.
Why not just have pathologists or a tumor-cell detector look harder at the margin? Detecting cancer cells at the margin is already what R0/R1 status measures, and it misses the recurrences observed after macroscopically complete resection. We include automated tumor-cell density at the margin as a covariate, so any gain from discordance must hold up after accounting for it.
Decisive Checks
The first decisive check comes in the discovery phase. A positive answer requires the patient-adjusted mean absolute $D_m$ in the margin band to exceed both the core and distant-tissue levels for a pre-registered set of modules. It must also survive two controls: a check in which RNA and protein locations are deliberately shifted relative to each other, which tests whether the signal is just misregistration, and a check in which module membership is shuffled, which tests whether the network structure matters.
The second check is predictability. The H&E-predicted $D_m$ must correlate with the measured $D_m$ on held-out patients. We report this correlation next to the HEX r of 0.790 for raw markers, expecting it to be lower, because a residual between two predicted quantities is a harder target than either alone.
The third check is the recurrence analysis among R0 patients alone. This is where the margin call is least informative, and it is where a margin biomarker would matter most.
Liposarcoma specimens serve as a second tissue. They test whether the margin band of high discordance is specific to PDAC stroma or a general feature of peritumoral tissue.
References
- Analysis of recurrence and metastasis patterns and prognosis after complete resection of retroperitoneal liposarcoma
- A high-stringency blueprint of the human proteome
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- Histopathology-based protein multiplex generation using deep learning
- AI-enabled virtual spatial proteomics from histopathology for interpretable biomarker discovery in lung cancer
- STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics
- Temporal proteomic profiling of postnatal human cortical development
- HEIST: A Graph Foundation Model for Spatial Transcriptomics and Proteomics Data
- Session Introduction
- Clinical outcome comparison of laparoscopic radical antegrade modular pancreatosplenectomy vs. laparoscopic distal pancreatosplenectomy for left-sided pancreatic ductal adenocarcinoma surgical resection
- High-dimensional single-cell proteomics analysis of esophageal squamous cell carcinoma reveals dynamic alterations of the tumor immune microenvironment after neoadjuvant therapy