#
#
#
Alternate Opening
Dawood et al. “Confounding factors and biases abound when predicting molecular biomarkers from histological images.” Nature Biomedical Engineering, 2026. DOI: 10.1038/s41551-026-01616-8.
#
#
#
Dawood as the cautionary frame
Dawood et al. is a warning shot against the common claim that deep learning can infer molecular biomarkers directly from routine H&E slides. The authors show that many biomarker labels are statistically interdependent with other biomarkers and clinicopathologic variables, especially grade, and that H&E models may learn the whole associated phenotype rather than the intended biomarker-specific morphology. They test multiple model families across TCGA, CPTAC, and breast-cancer validation data, then stratify model performance by co-dependent biomarkers and clinicopathologic variables. In breast cancer, ER/PR and mutation predictions can look strong globally but degrade substantially within relevant strata, implying Simpson’s-paradox-like inflation. Their conclusion is not “H&E contains no molecular signal,” but rather: current models are risky as substitutes for molecular testing; they may be useful for triage or complementary decision support; and stronger causal, stratified, dependency-aware validation is needed.
Reference: Dawood et al. “Confounding factors and biases abound when predicting molecular biomarkers from histological images.” Nature Biomedical Engineering, 2026. DOI: 10.1038/s41551-026-01616-8.
#
#
Couture et al. 2018 — H&E prediction of grade, ER, subtype, and recurrence score
Couture et al. is almost a pre-Dawood case study of the exact issue. The paper reported impressive proof-of-principle results from H&E image analysis: prediction of grade, ER status, basal-like subtype, ductal versus lobular histology, and ROR-PT recurrence score. Importantly, Couture already recognized one major Dawood concern: tumor grade is strongly associated with ER status, and the authors tried to reduce grade-related bias. But Dawood would ask for more: ER performance within PR, HER2, Ki-67, PAM50, TP53, PIK3CA, grade, and subtype strata; comparison with grade-only or subtype-only baselines; and larger external validation. The reported concentration of ER errors in Luminal B tumors also fits Dawood’s point: misclassification may occur where the usual morphologic-molecular correlations are less clean. So Couture remains a valuable early proof-of-principle, but its safest clinical framing is triage or resource-limited prioritization for IHC/genomic testing, not replacement.
Reference: Couture et al. “Image analysis with deep learning to predict breast cancer grade, ER status, histologic subtype, and intrinsic subtype.” npj Breast Cancer, 2018; 4:30. DOI: 10.1038/s41523-018-0079-1.
Zhen et al. / ecPICK — H&E prediction and localization of ecDNA
Zhen et al.’s ecPICK paper makes a bolder claim: that H&E slides can identify and spatially localize extrachromosomal DNA across cancers, with reported agreement against FISH and prognostic associations. The FISH validation is a major strength relative to many H&E-to-molecular-label studies, because it anchors the model to an orthogonal molecular/spatial assay. Still, Dawood’s concern remains highly relevant. ecDNA is biologically entangled with oncogene amplification, aggressive morphology, nuclear atypia, proliferation, hypoxia, treatment resistance, advanced stage, and tumor microenvironment changes. If ecPICK predicts “ecDNA-rich” regions that also look like high-grade, genomically unstable, aggressive tumor regions, it may be capturing a composite malignant phenotype rather than ecDNA-specific morphology. Dawood would not dismiss it, but would ask for tumor-type-specific and breast-subtype-specific validation, grade/stage/receptor/TMB/comutation stratification, and comparison to simple morphologic-aggressiveness baselines. Best current interpretation: powerful spatial hypothesis generator and possible triage tool, but not yet a stand-alone ecDNA assay.
Reference: Zhen et al. “ecPICK: A deep learning-enabled spatial diagnostic platform for direct ecDNA identification and clinical prognosis across pan-cancer histopathology.” Theranostics, 2026; 16(12):7067–7083. DOI: 10.7150/thno.134316.
Trost et al. / SPARK — agentic, interpretable pathology-feature discovery
Trost et al.’s SPARK paper is, philosophically, one of the better answers to Dawood. Instead of asking a black box to jump directly from H&E to ER, MSI, PD-L1, prognosis, or therapy response, SPARK generates biologically motivated tissue concepts, converts them into computable parameters, filters and decorrelates them, and then tests them across many cohorts and cancer types, including breast cancer. In breast cancer, SPARK parameters correlate with grade, subtype, ER/PR, and other known variables, which is useful but also exactly where Dawood would raise an eyebrow. The advantage is that SPARK makes the intermediate features visible: lymphocyte architecture, tumor-stroma relationships, necrosis, spatial organization, and cell-type patterns can be inspected, stress-tested, and compared to known pathology. Dawood would still require conditional validation: do SPARK features add independent value within grade, stage, receptor subtype, treatment, and cohort strata? But compared with opaque H&E-to-biomarker prediction, SPARK moves the field toward a Dawood-compliant model: interpretable intermediate biology first, prediction second.
Reference: Trost et al. “An agentic framework for autonomous scientific discovery in cancer pathology.” Nature Medicine, 2026. DOI: 10.1038/s41591-026-04357-y.
Liang et al. / PathPrism — semantic spatial biomarkers and transparent modeling
Liang et al.’s PathPrism is another “post-Dawood” style paper. It converts whole-slide images into interpretable spatial biomarker spectra and uses transparent modeling for prognosis, molecular alteration prediction, treatment-response exploration, and in silico perturbation through VirtualWSI. This is attractive because it does not merely say, “the neural network predicts outcome”; it says, in effect, “these spatial tissue categories and relationships are associated with outcome or molecular state.” That makes it easier to ask Dawood-style questions. But the same caution applies: spatial patterns tied to mutation, prognosis, chemotherapy benefit, immunotherapy benefit, or breast-cancer treatment response may be proxies for broader biology — mucinous morphology, lymphocyte density, necrosis, stroma, subtype, grade, treatment selection, or line of therapy. The transparent feature layer reduces black-box risk, but it does not eliminate confounding. For breast cancer especially, response or molecular-prediction claims would need validation within ER status, HER2 status, TNBC versus HR-positive disease, grade, stage, prior therapy, and cohort source.
Reference: Liang et al. “Spatial biomarker discovery via interpretable semantic learning in histopathology.” Cancer Cell, 2026; 44:1–18. DOI: 10.1016/j.ccell.2026.05.014.
Frascarelli et al. — task-oriented AI models for molecular prediction
Frascarelli et al.’s npj Breast Cancer review is broadly aligned with Dawood, even before getting to Dawood’s specific statistical tests. It argues against naïve “bigger model, more data, better medicine” enthusiasm and favors smaller, task-oriented, clinically deployable AI systems for HR, HER2, Ki-67, BRCA-related status, somatic mutations, and related breast-cancer endpoints. It also stresses curation, harmonization, explainability, workflow fit, and multi-institutional validation. Dawood adds a sharper statistical requirement to that framework: for every molecular-prediction claim, ask whether the model is truly detecting the target or simply exploiting grade, subtype, proliferation, stromal pattern, receptor status, and co-mutation structure. So this review’s preferred future — focused, explainable, modular, clinically anchored models — is compatible with Dawood, but it should explicitly add dependency-aware stratified validation as a core requirement.
Reference: Frascarelli et al. “Computational pathology in breast cancer: optimizing molecular prediction through task-oriented AI models.” npj Breast Cancer, 2025; 11:141. DOI: 10.1038/s41523-025-00857-1.
Chua et al. — AI across breast cancer management
Chua et al.’s Communications Medicine review surveys AI across detection, diagnosis, prognosis, treatment, and recovery, including pathology examples such as molecular subtype prediction, invasive-cancer detection, H&E prediction of grade/ER/subtype/ROR, 3D H&E biomarker prediction, and HER2 image-analysis tools. As a broad review, it appropriately emphasizes generalizability, reproducibility, transparency, bias, regulation, cost, and prospective validation. Dawood functions here as a filter on the pathology/prognosis sections: reported accuracies across heterogeneous breast-cancer AI studies should not be read as equivalent to clinical validity. A model that predicts subtype, recurrence, or treatment response may be learning the background clinical distribution of the dataset. Dawood would push the review’s caution one step further: future reviews should classify studies not only by accuracy and dataset size, but by whether they prove incremental value over grade, stage, receptor status, subtype, treatment, and simple clinical/pathology baselines.
Reference: Chua et al. “Artificial intelligence for breast cancer management.” Communications Medicine, 2026; 6:79. DOI: 10.1038/s43856-025-01342-3.
Witowski et al. — multimodal AI for comprehensive breast-cancer prognostication
Witowski et al. is less directly vulnerable to Dawood than a pure H&E-to-biomarker paper because it is explicitly multimodal: it combines digital pathology with routine clinical variables to predict disease-free interval, rather than claiming that H&E replaces ER, HER2, or sequencing. The study is also large, with 8,161 patients, multiple cohorts, external evaluation, and comparison to a 21-gene assay. But Dawood still matters. Prognosis in breast cancer is deeply confounded by stage, grade, receptor status, treatment, cohort, follow-up time, histology, and selection for genomic testing. A high C-index or hazard ratio can still reflect learned cohort structure unless the model is well calibrated and demonstrably additive within clinically meaningful strata. The authors do report added prognostic information relative to Oncotype DX and grade, and they are careful that the study is prognostic rather than predictive of treatment benefit. Dawood would reinforce that distinction: this may become a useful risk-stratification tool, especially in TNBC where genomic assays are less established, but it does not yet prove who benefits from chemotherapy or any specific intervention.
Reference: Witowski et al. “Multi-modal AI for comprehensive breast cancer prognostication.” Nature Communications, 2026; 17:5879. DOI: 10.1038/s41467-026-73088-y.
Bottom line
Dawood changes the burden of proof. The older or more direct H&E-to-biomarker papers, like Couture and ecPICK, should be read cautiously unless they survive stratified, dependency-aware validation. The reviews by Frascarelli and Chua are directionally consistent with Dawood but should make confounding a more central evaluation criterion. The newer structured-representation papers — Trost/SPARK and Liang/PathPrism — look like the constructive path forward: decompose the slide into interpretable biology, then test whether those features add value beyond the variables pathologists and oncologists already know. Witowski is promising because it is multimodal and clinically anchored, but its claims should remain prognostic, not predictive, until treatment-benefit validation is shown.
#
Here are the same citations as plain bibliographic references, with one lead author only.
Dawood et al. “Confounding factors and biases abound when predicting molecular biomarkers from histological images.” Nature Biomedical Engineering, 2026. DOI: 10.1038/s41551-026-01616-8.
Couture et al. “Image analysis with deep learning to predict breast cancer grade, ER status, histologic subtype, and intrinsic subtype.” npj Breast Cancer, 2018; 4:30. DOI: 10.1038/s41523-018-0079-1.
Zhen et al. “ecPICK: A deep learning-enabled spatial diagnostic platform for direct ecDNA identification and clinical prognosis across pan-cancer histopathology.” Theranostics, 2026; 16(12):7067–7083. DOI: 10.7150/thno.134316.
Trost et al. “An agentic framework for autonomous scientific discovery in cancer pathology.” Nature Medicine, 2026. DOI: 10.1038/s41591-026-04357-y.
Liang et al. “Spatial biomarker discovery via interpretable semantic learning in histopathology.” Cancer Cell, 2026; 44:1–18. DOI: 10.1016/j.ccell.2026.05.014.
Frascarelli et al. “Computational pathology in breast cancer: optimizing molecular prediction through task-oriented AI models.” npj Breast Cancer, 2025; 11:141. DOI: 10.1038/s41523-025-00857-1.
Chua et al. “Artificial intelligence for breast cancer management.” Communications Medicine, 2026; 6:79. DOI: 10.1038/s43856-025-01342-3.
Witowski et al. “Multi-modal AI for comprehensive breast cancer prognostication.” Nature Communications, 2026; 17:5879. DOI: 10.1038/s41467-026-73088-y.
#
2018 NPJ BR CA Couture Image Analysis Predict Er status cf dawood.pdf
ZHEN Dawood ecDNA via H&E alone.pdf
2026 Nature Med TROST (also Dawood Liang) Agentic Autonomous Cancer Pathol.pdf
2026 Cancer Cell LIANG see also TROST DAWOOD Semantic layers Dig Pathol 26p.pdf
2026 Nat Biomed Eng Dawood ConFOundinG factors in H&E and Biomarkers Dig Path.pdf
2008 IJRO Dawood reply to wasserman.pdf
