Brain tumor MRI can predict some molecular features of a glioma well and others not at all. Models that read a tumor's mutation in the IDH gene from the scan reach areas under the ROC curve near 0.9. Hundreds of models trying to read MGMT promoter methylation, a chemical tag on DNA that silences a repair gene, sit at chance. The usual explanation is that MRI only sees molecular events that change tissue physics: blood vessels, how densely cells are packed, dead tissue. Yet the field still trains imaging models against the cheapest molecular label available, which is a whole-tumor DNA call or messenger RNA (mRNA) measured from ground-up tissue. Proteins and cell populations are what build vessels and pack tissue. mRNA explains only part of protein abundance. And one ground-up sample blends many cell types and tumor regions. So the targets these imaging models chase may be one step removed from what the scanner physically records.
We propose to test this directly. We will collect a cohort of image-guided multi-region glioblastoma biopsies, with each biopsy site registered to the patient's preoperative scan. Every site gets three measurements: matched mRNA, protein, and a cell-type census taken from multiplexed protein imaging. We then ask whether the local MRI signal predicts protein abundance and cell-type fractions better than it predicts the matching mRNA. The test is gene by gene and corrected for how noisy each assay is. If the answer is yes, the "virtual biopsy" problem should be posed differently: MRI should be trained to estimate local protein and cell-composition maps, not bulk genotypes or transcript signatures. If the answer is no, a widely repeated motivation for proteomics-aware radiogenomics is false, and that is worth knowing too.
Core Claim
Our hypothesis is that, for genes in pathways that shape tissue physics, MRI supervoxels predict local protein abundance and protein-defined cell-type fractions better than they predict the matched mRNA. A supervoxel is a small, irregular cluster of neighboring image voxels with similar signal. The pathways we have in mind are angiogenesis and hypoxia, proliferation, and immune and myeloid infiltration. We further expect the advantage of protein over mRNA to be largest for genes whose mRNA and protein disagree most within the same tissue. For nuclear, physically silent features, MGMT among them, we expect neither layer to be predictable.
The claim rests on three findings in the literature. Kickingereder et al. showed that IDH is visible on perfusion MRI through an intermediate step: a metabolite made by the mutant enzyme accelerates breakdown of the HIF1A protein, which suppresses blood-vessel growth. Their molecular evidence, however, was bulk TCGA mRNA, and the imaging was a whole-tumor percentile. Adhikari reports that mRNA tracks protein abundance only 40–60% of the time. That leaves room for imaging to align better with protein than with transcript. Two large benchmarks on MGMT, Saeed et al. and the BraTS 2021 radiogenomics validation, found that networks attend to the right tumor region but find no usable signal. We read this as a warning, not a target. Mapping straight from the scan to a microscopic nuclear mark fails, and we do not try to rescue it.
What Is New Relative to Existing Work
Three groups of prior work each connect two of the relevant pieces, and none connects all three. The first group links imaging to molecular labels. FoundBioNet predicts IDH from multi-sequence MRI across 1,705 patients. The habitat study of IDH genotypes using DCE-MRI and DWI divides the tumor into physiologically distinct subregions to predict the same whole-tumor label. The UPenn-GBM cohort pairs subregion radiomics with bulk IDH and MGMT calls. All three supervise on one label per tumor taken from ground-up tissue. We supervise on many labels per tumor, each tied to a location in the scan, and at the protein and cell level.
The second group resolves the tumor in space but without molecules. Ziegenfeuter et al. split PET/MRI scans into supervoxels and classify true progression versus pseudoprogression, where pseudoprogression is treatment-related change that mimics tumor growth. Their labels come from follow-up scans, not from tissue. We borrow their supervoxel representation and replace radiographic labels with molecular ones.
The third group handles the mRNA–protein gap or the cell-mixture problem, but only in excised tissue. scChord translates single-cell RNA into protein distributions. POPPY maps bulk transcriptomes onto single-cell references. KRONOS learns cell phenotypes from multiplexed protein images. None of them touches imaging of the living patient. We use KRONOS-style phenotyping as a measurement tool for ground truth. We use scChord-style RNA-to-protein translation only as a competing route.
Research Program
Data. The first deliverable is a dataset that, as far as the material we reviewed shows, does not exist. We will enroll roughly 40 newly diagnosed glioblastoma patients undergoing resection. Each patient will have preoperative structural MRI, diffusion MRI, and dynamic susceptibility contrast perfusion MRI, which tracks a contrast bolus through the vessels; this matches the UPenn-GBM protocol. During surgery, 4–6 neuronavigated biopsies per patient will be taken from the enhancing rim, the non-enhancing infiltrated edge, and perinecrotic tissue, giving about 200 sites. Each site's coordinates will be recorded on the preoperative scan.
Each site will be split three ways. One part goes to bulk RNA sequencing. One part goes to reverse-phase protein array profiling of total and phosphorylated proteins, in the style of the pathway-array medicine reviewed by Pal et al.. Phosphorylation is chemical switching of protein activity that sequencing cannot see. The third part goes to multiplexed protein immunofluorescence. From these images, a KRONOS encoder fine-tuned on a hand-labeled subset will give per-site fractions of tumor, endothelial, myeloid, lymphoid, and proliferating cells. A subset of sites will be assayed twice to measure test-retest reliability for every readout.
The protein panel is chosen so that each target has a matched transcript. It covers angiogenesis and hypoxia (VEGFA, HIF1A, CA9, and the vessel-lining marker CD31/PECAM1), proliferation (Ki-67/MKI67), myeloid and immune markers (CD68, CD163, CD3), and MGMT protein as an expected negative.
Measurement model. For each site, we take the supervoxel that contains the biopsy coordinate. We describe it with the 126-feature radiomic set of Ziegenfeuter et al., or its UPenn equivalent, plus a learned patch embedding from the FoundBioNet encoder. Brain shift during surgery and navigation error are typically a few millimeters. We therefore set supervoxel size to match measured registration error, and we report results as a function of a deliberate spatial blur.
Predictors are fitted with leave-one-patient-out cross-validation. For each target $y$ in gene $g$, we compute the held-out correlation $\hat\rho$ and correct it for assay noise:
$$\rho_{g,\ell} = \hat\rho_{g,\ell} / \sqrt{r_{g,\ell}}.$$
Here $\ell$ indexes the molecular layer (mRNA or protein) and $r_{g,\ell}$ is that readout's test-retest reliability. The correction is the standard attenuation formula for noise in the target only. Our primary statistic is the per-gene gap $\Delta_g = \rho_{g,\text{protein}} - \rho_{g,\text{mRNA}}$. We test whether the average $\Delta_g$ over physics-coupled genes is positive. We also test whether $\Delta_g$ grows as the within-cohort mRNA–protein correlation for gene $g$ shrinks. Both tests use a patient-level bootstrap.
To rule out the possibility that the signal is just IDH in disguise, all primary analyses are restricted to IDH-wildtype tumors. Patient random effects absorb overall differences between tumors, so the claim is about variation within a tumor.
Spatial model. If the measurement study succeeds, we fit a model that estimates protein and cell-fraction maps over whole tumors. It is a graph attention network over the supervoxel adjacency graph, in the spirit of PerfGAT. Its output head predicts cell-type fractions constrained to sum to one, plus protein levels. In effect, it deconvolves an image voxel into cell types the way POPPY deconvolves bulk RNA. Neighboring supervoxels are included because tissue signaling is local. We then apply the model to UPenn-GBM to produce maps for 630 patients that have no biopsy labels.
Baselines
The naive reference predicts each site's value from its subregion label alone (enhancing, necrotic, or edema), plus a patient intercept. The subregion labels come from UPenn-style segmentation. This tests whether supervoxel imaging adds anything beyond the coarse anatomy a radiologist already draws. Any proposed model must beat it on held-out correlation for physics-coupled targets, by a margin outside the patient-bootstrap interval.
The whole-tumor histogram baseline follows Kickingereder et al. It assigns every site in a tumor the same 90th-percentile rCBV and diffusion summaries. On IDH this approach reached 88% accuracy in their 73-patient cohort. It represents the view that imaging sees only tumor-level state, and it should fail by construction on within-tumor variation. If it does not fail, our premise about spatial heterogeneity is wrong.
The supervoxel radiomics baseline is the Ziegenfeuter et al. pipeline: SLIC supervoxels, 126 radiomic features, and a random forest. It is the strongest existing method for local tissue state from multimodal imaging, reporting AUC 0.875 and 80.0% accuracy for supervoxel progression versus pseudoprogression. No result for predicting site-level protein or cell fractions from MRI was found on Lacuna. This baseline is run on exactly our sites, folds, and feature budget.
The deep encoder baseline fine-tunes the FoundBioNet SWIN-UNETR encoder end to end on biopsy-centered patches. It gets the same hyperparameter search budget as our graph model, measured in trials and GPU-hours. For orientation, FoundBioNet reports IDH AUC of 90.58% on EGD and 80.31% on UPenn-GBM.
The cascade baseline predicts site mRNA from imaging first, then maps mRNA to protein with a model learned on the same training sites. This is the image-to-transcriptome-to-protein route that Gap 1 of the literature implies. Bulk RNA cannot feed scChord, which needs single-cell input, so we use a regularized regression together with a distributional variant.
Three ablations isolate the key ingredients. First, removing the perfusion sequences tests whether protein predictability runs through blood-vessel physiology, as the HIF1A mechanism predicts. Second, swapping protein targets for mRNA targets under an otherwise identical pipeline is the core comparison itself. Third, removing graph neighbors tests whether spatial context matters beyond the single supervoxel.
Why not just…?
Why not just train an end-to-end network from MRI to the clinical biomarker? For physically silent markers, the evidence says this hits a ceiling. Saeed et al. report MGMT AUCs of 0.57–0.62, with the training loss stuck near the coin-flip value of 0.69. The BraTS validation found 80.2% of 420 models indistinguishable from chance in accuracy, and the challenge winner fell to 56.2% AUROC on the external SNUH cohort. For markers that do leave a physical footprint, such as IDH, end-to-end models work well, and we do not compete with them on that task. Our question concerns local, within-tumor variation, which a whole-tumor label cannot supervise. The deep encoder baseline above is the honest version of this alternative applied to our targets.
Why not just predict mRNA signatures from imaging and translate them to protein afterward? This is the most serious threat to the proposal. If mRNA predicts protein well within glioblastoma tissue, the cascade will match direct protein supervision, and the protein assays are not worth their cost. The cascade baseline and the per-gene $\Delta_g$ analysis settle this directly. We will also report the within-cohort mRNA–protein correlation for every gene. A finding that the cascade wins on concordant genes and loses on discordant ones would itself support the claim, because it would show discordance is the deciding variable.
Why not just deconvolve biopsy RNA with POPPY or a reference atlas? That method still needs a biopsy, which cannot be repeated over time. One needle core also misses spatial mixtures: Ziegenfeuter et al. found mixed progression and pseudoprogression within the same tumor in 26% of cases. The method also stays at the transcript layer. We do run reference-based RNA deconvolution on our sites, comparing its cell fractions against the protein-imaging census, to measure how far RNA-based composition drifts from protein-defined composition.
Why not just use whole-tumor radiomics with IDH as a covariate? Restricting the analysis to IDH-wildtype tumors and fitting patient random effects removes exactly the signal that whole-tumor features capture. The histogram baseline tests directly whether anything remains beyond it.
Proposed Experiments and Decisive Checks
The first decisive check is the sign and size of the average $\Delta_g$ over physics-coupled genes in IDH-wildtype sites, after reliability correction. A positive gap with a bootstrap interval excluding zero supports the claim. A gap of zero or less refutes it, and we would report that as the main result. Within that check, we test whether $\Delta_g$ correlates negatively with each gene's mRNA–protein concordance. We also test whether the gap disappears when perfusion sequences are removed, which would tie it to blood-vessel physiology.
As a negative control, MGMT protein and MGMT mRNA should both be unpredictable. If either becomes predictable, we would suspect leakage through tumor location or patient identity. We will test for that by permuting the site–label pairing within each patient.
The second check concerns transfer. In UPenn-GBM, we summarize the predicted maps per patient, for example the myeloid fraction in the enhancing rim, and ask whether they add prognostic value for overall survival. The comparison is against the cohort's 145 subregion radiomic features plus IDH and MGMT status, measured by the concordance index under nested cross-validation. Gaining nothing here would mean the maps are biologically faithful but clinically redundant, which would limit the scope of the claim without undoing the measurement result.
A third, optional check applies the maps to an interval-scan cohort like Ziegenfeuter's. It asks whether predicted proliferative and myeloid fractions separate regions of true progression from pseudoprogression beyond their 0.875 AUC. Because that cohort is not public, this check depends on access.
The main risk is biopsy-to-scan registration error. We address it with the blur sweep. If predictive correlations collapse at the measured navigation error, the study cannot distinguish the two molecular layers. In that case we will say so rather than report the gap.
References
- IDH mutation status is associated with a distinct hypoxia/angiogenesis transcriptome signature which is non-invasively predictable with rCBV imaging in human glioma
- A high-stringency blueprint of the human proteome
- MGMT promoter methylation status prediction using MRI scans? An extensive experimental evaluation of deep learning models
- Validation of MRI-Based Models to Predict MGMT Promoter Methylation in Gliomas: BraTS 2021 Radiogenomics Challenge
- FoundBioNet: A Foundation-Based Model for IDH Genotyping of Glioma from Multi-parametric MRI
- Preoperative prediction of IDH genotypes and prognosis in adult-type diffuse gliomas: intratumor heterogeneity habitat analysis using dynamic contrast-enhanced MRI and diffusion-weighted imaging
- The University of Pennsylvania glioblastoma (UPenn-GBM) cohort: advanced MRI, clinical, genomics, & radiomics
- Resolving spatial response heterogeneity in glioblastoma
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- A Foundation Model for Spatial Proteomics
- Proteomics, pathway array and signaling network-based medicine in cancer
- Spatiotemporal Graph Neural Network Modelling Perfusion MRI