Germline Signal Lost to Subjective Tumor Grading

Pathologists grading the same tumor slide often disagree, because they are cutting a continuous spectrum of tissue appearance into a few discrete bins. In genetics, a noisy or poorly defined trait is one well-known reason that DNA variants explain less of a trait's variation than family studies predict, the "missing heritability" gap. This proposal asks how much inherited (germline) signal in tumor phenotypes is lost to subjective grading, and whether it can be recovered. We replace pathologist grades with continuous phenotypes defined by gene co-expression modules: groups of genes whose expression rises and falls together across patients, each summarized by a single score. We predict these module scores from routine microscope slides, then run genetic association with a correction that keeps the image model from inventing genetic signal.

In two sentences: tumor grade is a subjective, unreliable measurement, and its genetic signal is diluted. We test whether molecularly defined continuous phenotypes, measured directly from RNA or predicted from slides with statistically valid correction, carry more recoverable germline heritability than grade, and we separate how much of the gain comes from better reliability and how much from choosing a better biological axis.

Core Claim

Weighted gene co-expression network analysis (WGCNA) clusters genes into co-regulated modules. It summarizes each module by its "eigengene," the first principal component of the module's expression across samples. Prior work treats eigengenes as things to correlate with clinical traits, as in Lian et al., who linked modules to TNM stage, and Breen et al., who tracked developmental protein modules. We instead treat eigengenes as the phenotype itself: a continuous, molecularly anchored replacement for a subjective grade.

The claim has two parts. First, some immune- and stroma-related module eigengenes show detectable SNP-heritability across a pan-cancer cohort, while pathologist grade shows little. Second, this gap decomposes into two pieces. One is measurement unreliability, which a reliability correction alone would fix. The other is a phenotype that sits on the wrong biological axis, which only redefining the trait fixes. The second part is what makes this more than a reliability study.

The decomposition rests on the classical attenuation relation. If measurement error is independent of genotype, then

$$h^2_{\text{obs}} = \rho \, h^2_{\text{true}},$$

where $h^2_{\text{obs}}$ is the SNP-heritability estimated on the measured trait, $h^2_{\text{true}}$ is the heritability of the underlying error-free trait, and $\rho$ is the trait's reliability. Reliability is the fraction of its variance that repeats across independent raters or independent slides of the same tumor. We measure $\rho$ directly for both grade and the module phenotypes. Any heritability gap that survives division by $\rho$ is attributable to the choice of axis.

How This Differs From the Closest Work

Four papers set up the problem, and none of them does this.

Sun, Kranzler and Bi motivate the idea. They reweight 13 addiction symptoms to maximize chip heritability, reaching a cross-validated 0.87 against a best single symptom of about 0.6. Their approach searches for whatever combination is most heritable, works only on questionnaire items, and operates in a cohort of 1,752 people where heritability estimates are noisy. We do not search for the most heritable combination of features. We fix the phenotype axes in advance from transcriptome structure alone, without looking at genotypes, which removes the winner's-curse inflation that heritability maximization invites.

Miao et al. constrain the design. They show that GWAS on machine-learning-imputed traits had an 81% replication failure rate for type 2 diabetes because the predictor leaks the genetics of its input features. Their correction, POP-GWAS, cut that to 0% and found 116% more loci than labeled data alone. We reuse POP-GWAS as the inference engine for image-predicted phenotypes. Its first application to histopathology is part of what we test.

TANGLE supplies the image method. It aligns slide embeddings with bulk RNA through contrastive learning, which trains the model to place each slide near its own patient's RNA profile and away from other patients'. It feeds the full expression vector into an unstructured network and never touches germline genetics. We use its encoder but regress onto module eigengenes.

MPRS-NetWAS is the closest network-plus-genetics precedent. It aggregates Alzheimer's GWAS weights over protein-interaction modules mediated by MRI amygdala density. Its modules come from a fixed interaction graph, its intermediate phenotype is organ-scale MRI rather than tissue, and it does not address rater disagreement.

Research Program

The testbed is TCGA, which offers H&E slides, bulk RNA-seq on 20,530 genes, clinical grade, and controlled-access germline arrays across 33 cancer types and more than 11,000 cases. Power dictates the design. For unrelated individuals, the sampling error of SNP-heritability shrinks roughly in proportion to 1/N. Per-cancer cohorts of about a thousand patients therefore give uninterpretable estimates. We pool across cancers, with cancer type, sex, ancestry principal components and estimated tumor purity as covariates. We restrict primary analyses to modules that replicate across cancer types. We expect heritability estimates and genetic correlations to be the main readouts, not new genome-wide significant loci.

The build has four steps.

First, we build WGCNA modules on a held-out half of the RNA data, with soft thresholding and topological-overlap clustering, and record module stability across perturbations of the soft-threshold power and the tree-cut settings. A known weakness of WGCNA is that bulk eigengenes may track cell-type mixture, as the file notes. We do not treat this as a nuisance. Composition-driven immune and stromal modules are exactly the tissue traits pathologists eyeball, and we label each module by its correlation with purity.

Second, we measure reliability. For module phenotypes, we use TCGA patients with multiple slides (frozen and diagnostic) to compute cross-slide agreement of image-predicted eigengenes. For grade, we commission blinded re-grading of a subset of TCGA-BRCA diagnostic slides by three pathologists using the Nottingham system. This gives an empirical $\rho$ instead of an assumed one, and it directly quantifies the rater disagreement that meningioma grading work attributes to arbitrary bins on a continuous spectrum.

Third, we train the slide-to-eigengene predictor: a TANGLE-pretrained attention-based multiple-instance model, which pools thousands of patch embeddings into one slide vector, regressed onto eigengenes.

Fourth, we run genetics. We estimate GREML heritability on grade, measured eigengenes, and image-predicted eigengenes, with POP-GWAS for the imputed traits.

Baselines

The naive reference point is GREML on pathologist grade as recorded in TCGA, the trait a clinician would hand a geneticist. Its strength is that it is the current practice. We report it both raw and divided by the measured rater reliability. Beating it means a module phenotype whose heritability, after reliability correction on both sides, exceeds corrected grade's heritability, with non-overlapping confidence intervals on the pooled cohort. The file contains no reported heritability for any histopathology trait; there is no reported result found on Lacuna for SNP-heritability of tumor grade or image-derived tumor phenotypes. This baseline is therefore established in-study.

The strongest existing phenotype-construction method is Sun et al.'s heritability-maximizing reweighting, which reported cross-validated chip heritability of 0.87 on SSADDA cocaine dependence. We apply it to the same inputs our modules see, namely the top transcriptome principal components or image embedding dimensions. We use the same covariates and the same genotyped samples, and we fit the weights on one genotype split and estimate heritability on a disjoint split. If held-out heritability matches our pre-specified module eigengenes, the WGCNA step is not adding value. We count our approach as winning only if it matches held-out heritability while showing smaller train-to-test shrinkage.

For the image predictor, the strongest competitor is TANGLE itself with its unstructured full-transcriptome head. Its reported gain is +11.0% macro-AUC over supervised ABMIL in 10-shot TCGA-BRCA classification. It has no reported eigengene-regression result. We compare it against supervised ABMIL trained directly on eigengenes, with identical patch features, splits and tuning budget. The comparison metric is held-out eigengene correlation and downstream POP-GWAS heritability agreement.

The ablations isolating the key ingredient replace WGCNA modules with three alternatives, all at matched dimensionality: plain principal components of the transcriptome, curated gene-set mean scores, and hard-threshold correlation-network modules. A further ablation replaces POP-GWAS with naive GWAS on image predictions.

Why not just…?

Why not just run genetics on the measured RNA eigengenes and skip images? In TCGA, that is exactly what we do, and it serves as the gold standard; the image arm cannot beat it there. Images matter only for the much larger H&E archives with genotypes but no RNA. We demonstrate their validity by masking RNA for a random 70% of TCGA and checking whether POP-GWAS on image predictions plus the 30% labeled subset recovers the full-RNA estimates. If no RNA-free deployment cohort materializes, the image arm stays a validated tool rather than a discovery, and we will say so.

Why not just correct grade for rater unreliability? Attenuation correction fixes noise but not a trait that mixes biologically distinct tumors under one label, which is the heterogeneity Sun et al. identify as the larger cause of diluted heritability. The decomposition above tests this directly. If reliability-corrected grade matches the eigengenes, the honest conclusion is that subjectivity was the whole story, and the result is still useful.

Why not just maximize heritability over image features directly? With high-dimensional features and roughly ten thousand genotyped patients, heritability maximization overfits genotype noise. With an image model in the loop, it also invites the proxy leakage Miao et al. documented. It is our strongest baseline and is run with held-out heritability estimation, so the shrinkage comparison settles the question.

Why not just use principal components or curated pathway scores instead of WGCNA? This is a real threat, because PCA is simpler and the eigengene is itself a principal component. The matched-dimensionality ablation decides it. If principal components or gene sets give equal heritability and reliability, we drop the WGCNA claim and keep the rest.

Decisive Checks

Three results decide the proposal.

First, the masking experiment must show that POP-GWAS heritability and genetic correlations from image-predicted eigengenes agree with full-RNA estimates. Naive image-predicted GWAS should show inflation. We test for this inflation with permuted-genotype nulls and with negative-control genetic correlations against sex- and ancestry-linked signals, which should be zero after correction.

Second, at least one replicated module eigengene must exceed reliability-corrected grade in heritability on the pooled cohort.

Third, cross-slide reliability of image-predicted eigengenes must exceed the measured rater agreement on grade. If it does not, the premise that continuous molecular anchors remove subjectivity fails.

GasHisSDB reports that classification of gastric histology patches reaches 96.47% with VGG16. That high accuracy reproduces existing labels and makes no claim about genetic validity, which is exactly why these three checks, rather than accuracy, are the endpoints.