Evidence for paper review

The paper’s central methodological improvement is to make the patient—not a cohort-wide ranked gene list—the unit of evidence. Across 285 TCGA-BRCA patients, five-fold patient-grouped cross-validation, nested preprocessing, and 10,000 matched label permutations support genuine out-of-fold prediction from frozen embeddings: Spearman ρ ranged from 0.25 to 0.56 across four programmes and 11 reported backbone conditions; all 44 adjusted permutation tests reached p=0.0044. This is persuasive evidence for predictive association, not proof that specific visual structures causally encode molecular biology. Held-out molecular prediction

What actually carries the signal?

• Composition is a serious competing explanation. For basal, tissue compartment fractions achieved ρ=0.469 versus ρ=0.493 for embeddings; the difference was only +0.024 with paired p=0.77. The paper appropriately rejects any claim that embeddings add detectable basal-specific information beyond composition. For ER/luminal, proliferation, and immune, embeddings exceeded composition by +0.280, +0.284, and +0.479, respectively, although 54 interpretable cell-count features came within 0.043–0.085 of the embedding model. Composition and cell-count controls
• Geometry contributes little under the tested implementation. The graph first selected neighbours with Euclidean distance and only then reweighted those edges, so the claimed geometric topology was Euclidean by construction. The as-implemented Riemannian-minus-Euclidean difference was +0.0010, 95% CI −0.0007 to +0.0029. When neighbour selection was made metric-aware, performance was lower by −0.0117, with CI −0.0229 to −0.0004. Ridge regression also exceeded the graph decoder by +0.097, CI +0.069 to +0.127, across 24 of 27 cells. This supports “no demonstrated benefit here,” not “Riemannian geometry can never help.” Geometry ablation

The atlas-discovery claims are weaker than the benchmark

The paper correctly demotes driver recovery: 91.8% of random six-gene panels recovered at least five of six canonical drivers because 98.6% of tested genes passed FDR. Consequently, claims based on 5/6 or 6/6 driver recovery, multi-resolution differences, and pooling variants should be read as descriptive rather than predictive. The atlas’s qualitative tile grounding is also limited because pathologist-reviewed labels were unavailable. Driver-count metric critique

Generalizability and decisive next test

CPTAC-BRCA provides molecular replication only, because it lacks paired whole-slide images; the segmentation resource overlaps 284 discovery patients and supplies composition features rather than embeddings. Thus the paper contains no independent paired image–RNA replication of the complete pipeline. Survival analysis is exploratory: only 33 overall-survival events were available, and atlas coordinates added no significant information beyond PAM50. Confidence is therefore high for the within-cohort methodological conclusions, but moderate-to-low for external biological generalization. The most decisive follow-up is a fully independent, multi-institutional cohort with paired WSIs and RNA-seq, richer composition controls, and a genuinely metric-aware graph whose neighbourhoods are selected using the learned metric.

Code and reproducibility repository · Full preprint