Molecular Rank Disagreement as an Anchor for Pathologist Disagreement

When several expert pathologists grade the same prostate biopsy, they often disagree. Current models of that disagreement learn it from images and labels alone. They can tell us that a case is ambiguous, but not why. The disagreement could come from a staining artifact, from human inconsistency, or from tissue that is a genuine mixture of more and less aggressive cell populations. This proposal tests whether a molecular signal can separate these causes. The signal is a model that predicts, from a hematoxylin-and-eosin (H&E) image, which of two genes is more highly expressed at each location. It is trained on spatial transcriptomics, which measures gene expression at many small spots across a tissue section.

Predicting relative orderings rather than absolute amounts matters for two reasons. Orderings survive the scale distortions that differ between laboratories, so the signal should travel across hospitals better than raw image features do. And where a biopsy's predicted orderings disagree from region to region, that disagreement is a direct, measurable sign of mixed tissue. The central hypothesis is that within-biopsy disagreement of predicted gene orderings predicts between-pathologist disagreement. It should also add information beyond what a purely visual "mixed patterns" score already captures, and it should keep doing so when a model trained at one site is applied at another.

Core Claim

Inter-observer variability in Gleason grading has a component explained by heterogeneity of relative gene expression. Two quantities expose it. The first is how uncertain the predicted orderings are at each patch. The second is how much those orderings disagree across patches of the same core.

Because this signal comes from scale-invariant orderings, it should keep its relationship to pathologist entropy across sites and datasets. Visual uncertainty, by contrast, should degrade under stain and scanner shift. Using the signal to condition a model of the full distribution of pathologist labels should therefore improve that model mainly out of distribution.

The claim is falsifiable in a specific way. If the rank-heterogeneity score adds nothing once a visual patch-grade dispersion score is controlled for, the molecular anchor is redundant and the proposal fails.

Where this sits in prior work

Three papers supply the method. STRank replaces regression on spot counts with pairwise Binomial and listwise Multinomial ranking losses on top of a frozen CONCH image encoder. It reports a top average Spearman correlation of 0.345 across seven HEST-1k cancer cohorts, and 0.828 versus 0.583 for squared-error loss on synthetic batch-shifted data. We reuse its pairwise loss essentially unchanged. STRank, however, only evaluates gene-trend prediction. It never asks whether the predicted orderings mean anything for diagnosis, and it treats each spot independently rather than measuring disagreement across a tissue sample. STAMP and PAST both use rank-based tokenization of genes to resist cross-lab noise when pretraining image–gene encoders. Both evaluate against hard labels or a single expert reference; PAST, for example, reports 0.91 correlation of virtual H-scores with one pathologist. These motivate rank inputs as a batch-robust interface, but neither models a panel's disagreement.

On the disagreement side, the Latent Doctor Model is the method we build on and compete with. It is a conditional variational autoencoder with a per-pathologist style vector, and it predicts the panel's label distribution. On Gleason2019 it reaches a KL divergence of 0.461, and on a 20-pathologist ISUP set its predicted entropy reaches a Spearman correlation of up to 0.728 with panel entropy. It is image-only, and it was not tested under site shift.

That gap is exactly what Gustafsson and Rantalainen motivate. Using UNI and CONCH features on PANDA, they show quadratic weighted kappa falling from about 0.88 in distribution to about 0.24 when a Radboud-trained model is applied to Karolinska. Adding more source-site data made this worse, because the classifiers overfit site artifacts.

Two papers act as constraints. The batch-agnostic encoder of Nishimura et al. shows that raw Visium profiles cluster by patient rather than tissue class. Their scVI-corrected guidance still did not beat an ImageNet ResNet50 (accuracy 0.601 vs. 0.677). Molecular guidance is therefore not automatically useful for image tasks, and we must show added value over strong vision baselines rather than assume it. Count Bridges and POPPY deconvolve mixed or bulk measurements into cell-level states. They show that mixture structure is recoverable, but both work on absolute or normalized counts, have no image branch, and so cannot score H&E slides at other hospitals.

Research Program

Stage one: train a rank predictor. Using HEST-1k prostate (PRAD) sections, with the other HEST-1k cohorts as optional additional pretraining, we train a pairwise rank model. Its output $p_{ij}(x)$ is the predicted probability that gene $i$ is more highly expressed than gene $j$ in image patch $x$. It runs on a frozen pathology encoder and uses STRank's count-weighted Binomial loss.

The gene pairs are chosen without diagnostic labels. We keep pairs whose ordering is bimodal across PRAD spots, meaning it flips in a sizable fraction of spots while staying consistent within contiguous regions. We cap the panel at a few hundred pairs to avoid the quadratic blow-up in the number of pairs. This label-free selection keeps the pathologist annotations used in later stages out of the molecular target.

Stage two: define two summary statistics per biopsy core. The first is local ordering uncertainty, the average binary entropy of $p_{ij}$ over pairs and patches. The second is cross-patch ordering heterogeneity, defined as

$$H_{\text{rank}}(C) = \frac{1}{|P|}\sum_{(i,j)\in P} \mathrm{Var}_{x \in C}\, p_{ij}(x),$$

where $C$ is the set of patches in a core and $P$ is the selected pair panel. A core made of uniformly ordered tissue scores low. A core in which some regions order a pair one way and others the opposite way scores high.

The two statistics answer different questions. Ordering heterogeneity across patches is the quantity we tie to masked cellular heterogeneity. Local uncertainty is partly confounded by within-spot mixing, because a 55 µm Visium spot of 5 to 30 cells can show an intermediate ordering for biological reasons. We report the two separately rather than merging them.

Stage three: condition a label-distribution model on the rank statistics. We build a molecularly conditioned version of the Latent Doctor Model. The prior over the pathologist latent is conditioned on image features concatenated with the patch-level ordering vector $p(x)$, and on $H_{\text{rank}}$. We then ask whether this improves predicted label distributions, mainly when the image domain changes.

Baselines

The naive reference is a single classifier trained on majority-vote grades, with its softmax entropy used as the ambiguity score. This is what most deployed systems implicitly do. Against panel entropy it sets the floor any ambiguity estimate must clear.

The strongest existing method is the conditional Latent Doctor Model of the Latent Doctor Model paper. It reports a KL divergence of 0.461 on Gleason2019 and entropy Spearman up to 0.728 on the ISUP 20-pathologist set. We will re-run it on the same frozen encoder features as our model, with the same support-set sizes and the same hyperparameter search budget. We also report its published numbers so that our re-implementation can be checked against them.

Beating it means two things. In distribution, we must match it within noise, since it is not our claim that molecules help there. Under transfer from Gleason2019 to ISUP and across PANDA sites, we must improve KL or entropy correlation by a margin larger than the spread over five training seeds.

A deep ensemble is included at matched compute because the Latent Doctor paper found ensembles competitive on tumor budding.

The key ablations isolate each ingredient of the rank signal. The first ablation, which we consider the most important, replaces the rank targets with absolute targets: the same pipeline trained with squared-error or Pearson loss on log counts. This tests whether ordering invariance, rather than molecular supervision per se, carries the effect. The second replaces the selected pair panel with random pairs. The third shuffles spot-to-patch correspondence in HEST-1k, which gives a molecular-free control with an identical architecture.

Why not just…?

Why not just use the visual spread of patch-level grades? A patch classifier's predicted Gleason pattern, aggregated per core, directly measures "this core mixes pattern 3 and pattern 4." That may explain most pathologist disagreement without any transcriptomics, and it is the most serious threat to the proposal. We address it head on. The decisive analysis is the partial correlation between $H_{\text{rank}}$ and panel entropy after regressing out patch-grade dispersion. A second analysis is the incremental gain in the conditioned label-distribution model when the dispersion score is already included as an input. If the partial correlation is indistinguishable from zero on both Gleason2019 and ISUP, we report that molecular orderings add nothing to morphology for this task. That would itself be useful evidence about what pathologist disagreement is made of.

Why not just fix site shift with stain normalization or color augmentation? These are cheap, and they may recover much of the drop Gustafsson and Rantalainen reported. Their finding that more source data hurts suggests the overfitting is to fine-grained artifacts, not only to global color, but that is an inference rather than a test. We therefore apply strong stain augmentation to every baseline and to our model alike. The comparison then asks whether rank conditioning helps on top of augmentation, not instead of it.

Why not just fine-tune on a few target-site slides? Gustafsson and Rantalainen found that 100 to 1,000 local slides recover performance more than switching backbones does. For grade prediction this is likely the better practical fix. Our setting differs in that target-site multi-pathologist panels are expensive to collect. We test it anyway: we report how many target-site slides with labels are needed for the image-only Latent Doctor Model to match our zero-target-label model. If that number is small, the case for the molecular anchor narrows to interpretability.

Proposed Experiments

Molecular validity first. On held-out HEST-1k PRAD sections, we check that $H_{\text{rank}}$ computed from predicted orderings correlates with $H_{\text{rank}}$ computed from measured spot counts over matched regions. We also check that pair-ordering accuracy beats the shuffled-correspondence control. As a sanity check against the literature, the pipeline should approach STRank's reported 0.345 average Spearman correlation when run on the seven HEST-1k cohorts. We found no per-cohort PRAD number reported on Lacuna, so PRAD results are compared only internally.

Disagreement prediction. On Gleason2019 (244 cores, 6 pathologists) and the ISUP set (100 biopsies, 20 pathologists), we compute Spearman correlations of each ambiguity score with panel entropy. The scores compared are naive softmax entropy, Latent Doctor entropy, patch-grade dispersion, local ordering uncertainty, and $H_{\text{rank}}$. We also run the partial-correlation analysis described above. Train-on-Gleason2019, test-on-ISUP is our cross-dataset shift condition.

Site robustness. On PANDA, split into Radboud and Karolinska, we test whether each ambiguity score's distribution, conditional on consensus grade, differs between sites. We measure this with a site classifier's AUC on the scores; an invariant score sits near chance. We then run a referral experiment. A Radboud-trained grader defers its most ambiguous Karolinska cases, and we measure quadratic weighted kappa on the retained cases at matched deferral rates. A deferral score that targets genuine ambiguity, rather than site artifacts, should raise retained kappa more steeply.

Decisive outcomes. The proposal succeeds if three things hold together. First, $H_{\text{rank}}$ keeps a significant partial correlation with panel entropy after controlling for visual dispersion. Second, the rank-conditioned label model beats the Latent Doctor Model under transfer but not, or only marginally, in distribution. Third, the absolute-target ablation loses most of that transfer gain. If only the first holds, we have a scientific finding about the molecular basis of disagreement without a better tool. If none holds, pathologist disagreement in Gleason grading is, to the resolution of Visium-trained predictors, morphological, and the proposal will say so.