Spatial Microscopy and Computational Pathology in 2026
Perspective for the 10x Genomics Spatial Oncology Project
Executive perspective
Spatial biology and computational pathology are increasingly difficult to regard as separate fields. One begins by measuring more information from tissue—RNA, protein, cell identity, molecular state, and spatial relationships at increasingly high plex and resolution. The other begins with ordinary pathology images and attempts to compute more information from what is already visible, using artificial intelligence to quantify morphology, identify cellular neighborhoods, predict molecular states, and increasingly infer signals that previously required molecular assays.
By 2026, those trajectories are converging.
This convergence is particularly relevant to the 10x Genomics project evaluating a multiplex slide-based protein/spatial oncology assay. The important question is no longer simply whether 20, 50, or 100 proteins can be visualized on a slide. The more consequential question is whether high-dimensional spatial measurement identifies a clinically meaningful tissue phenotype—a pattern of target expression, cell states, stromal architecture, immune context, or cellular neighborhoods—that predicts something important enough to change patient management. The current 10x project appropriately frames CSP around that clinical and economic problem rather than around multiplexing for its own sake.
The field now appears to be developing a three-layer architecture:
High-dimensional spatial assays create biological ground truth and discover phenotypes. Computational methods identify the features within those data that actually matter. Clinical products then either retain the high-dimensional assay, compress the discovery into a much smaller practical assay, or attempt to infer the phenotype directly from routine H&E.
That last possibility is both an opportunity and a competitive threat for high-plex spatial platforms.
1. Where the field stands in 2026
The term spatial microscopy covers several overlapping technologies rather than a single method. Conventional immunohistochemistry is already spatial biology: it identifies a molecular target and its location within tissue. Multiplex immunofluorescence extends the same principle to several proteins simultaneously. Cyclic fluorescence, imaging mass cytometry, highly multiplexed antibody imaging, digital spatial profiling, spatial transcriptomics, and related methods extend the observable molecular space from several analytes to dozens, hundreds, or thousands.
A useful 2026 status map is:
| Technology layer | Current status |
|---|---|
| Routine H&E microscopy | Universal clinical infrastructure |
| Digital whole-slide imaging | Clinically established and increasingly interoperable |
| Conventional IHC | Routine standard of care |
| Multiplex IF / modest-plex spatial protein assays | Clinically feasible; established in selected laboratory-developed tests and translational programs |
| High-plex spatial proteomics | Strong research and pharmaceutical-development tool; limited routine diagnostic deployment |
| Spatial transcriptomics | Rapidly maturing discovery and translational platform; still uncommon as a routine patient-level diagnostic |
| AI detection/classification from H&E | Real clinical products and regulatory precedents exist |
| AI prediction of molecular or spatial phenotypes from H&E | Rapidly advancing; striking research results, but mostly pre-routine clinical use |
| Multimodal pathology foundation models | Very rapid development; clinical-grade performance shown experimentally, but validation and deployment remain major issues |
Reviews by Liu, Dai, and Wang and by Bollhagen and Bodenmiller describe a field that has moved beyond proving that tissue contains useful spatial information. The emphasis has shifted toward identifying functional niches, cellular neighborhoods, treatment-response patterns, and clinically translatable signatures. (Liu et al., PubMed; Bollhagen & Bodenmiller, PubMed). (PubMed)
At the same time, Pentimalli and colleagues emphasize that high-resolution spatial transcriptomics still faces substantial translational problems: tissue handling, cost, standardization, computational burden, multimodal data integration, prospective validation, and the need to reduce a research-scale readout to something usable in clinical workflows. (PubMed; DOI). (Annual Reviews)
Thus, "spatial biology is clinically here" and "spatial omics remains largely translational" can both be true. The answer depends on the plex, workflow, intended use, and clinical claim.
2. Spatial biology has moved from making maps to finding mechanisms
The first generation of spatial studies often produced visually remarkable maps demonstrating that tumors were heterogeneous. That observation was biologically important but rarely constituted a diagnostic product.
The newer literature asks more consequential questions: Does the spatial arrangement explain treatment response? Does it identify resistance? Can it separate clinically important populations? Does the spatial phenotype predict recurrence?
Several 2026 studies illustrate that shift unusually well.
Triple-negative breast cancer: the unit of biology becomes the cellular community
Yan and colleagues analyzed 427,857 cells from 101 patients with treatment-naive triple-negative breast cancer and spatial transcriptomic data from 44 patients. Rather than simply cataloging cell types, the investigators identified cancer-cell programs, 49 immune and stromal states, and eight spatially organized "ecotypes." Macrophage states, cancer-cell interferon signaling, HLA expression, and cell-cycle programs were associated with chemotherapy response. (PubMed; DOI). (PubMed)
The conceptual advance is important. The candidate biomarker is not necessarily "protein X is positive." It may be something closer to:
cell type A in state B, adjacent to cell type C, in a particular stromal and immune configuration.
That is a fundamentally different diagnostic object from traditional one-analyte pathology.
Colorectal metastasis: spatial biology finds clinically relevant microscopic ecosystems
Liu and colleagues combined Visium and Visium HD spatial transcriptomics, whole-genome sequencing, and high-plex protein imaging in 49 primary and metastatic colorectal tumors from 19 patients. Liver micrometastases showed early clonal divergence, stem-like/quiescent tumor states, immunosuppressive niches, and T-cell exhaustion. A six-gene micrometastasis signature was associated with ctDNA-defined MRD, disease-free survival, and chemotherapy resistance. (PubMed; DOI). (PubMed)
Again, the value of spatial measurement is not simply better visualization. The assay is revealing a biological state whose relevance would be diluted or lost by homogenizing tissue.
Metastatic colonization: spatial information becomes temporal information
Sun and colleagues went one step further by constructing a spatiotemporal atlas of metastatic colonization. Integrating spatial transcriptomics, single-cell RNA sequencing, and chromatin accessibility, the investigators identified a transient quiescent PHGDH-high disseminated tumor-cell state followed by macrophage-mediated remodeling of the metastatic niche. Perturbing these processes restored immune surveillance and inhibited metastatic growth in experimental models. (PubMed; DOI). (PubMed)
The implication is that the biological object of interest may eventually be neither a gene nor a cell nor even a static neighborhood, but a trajectory of spatial ecosystems.
That is scientifically powerful, although considerably harder to turn into a reimbursable clinical test.
3. ADCs may be one of the strongest arguments for genuine spatial biomarkers
Antibody-drug conjugates provide an unusually persuasive use case because their pharmacology is intrinsically spatial.
An ADC does not merely require expression of a target. Efficacy may depend on target density and heterogeneity, accessibility of the target, vascular delivery, extracellular matrix, stromal barriers, internalization, payload release, diffusion into neighboring cells, macrophage uptake, and the geometry of target-positive and target-negative cells.
Wijerathna-Yapa and colleagues describe ADC pharmacology as operating across multiple physical scales: intracellular processing, local bystander diffusion, and whole-tumor antigen heterogeneity. Their proposed development pathway begins with high-dimensional spatial discovery and ultimately seeks to reduce the signature to a clinically deployable assay.
That framework is supported by direct patient data. Cai and colleagues analyzed 912,360 cells from 47 HER2-positive breast cancers treated with an ADC. Spatial proteomics identified higher tumor-cell H3K27ac as associated with efficacy and collagen-positive fibroblasts as associated with resistance. Importantly, the investigators did not stop with a discovery map: they constructed an ADC barrier-prediction model based on spatial cellular neighborhoods and proposed implementation by multiplex immunofluorescence. The article is available online ahead of its September 2026 issue. (PubMed; DOI). (PubMed)
This provides an especially useful model for the 10x CSP concept:
high-plex technology may be most valuable not because every measured protein becomes a separately reportable clinical result, but because high-plex measurement discovers a multivariate spatial phenotype that conventional IHC could not have specified in advance.
The eventual clinical assay might remain high plex. But it might also be a distilled 6-marker, 10-marker, or 20-marker panel plus an algorithm.
4. Computational pathology is approaching spatial biology from the opposite direction
While spatial platforms are making the tissue measurement richer, computational pathology is asking how much latent biological information can be recovered from the cheapest and most ubiquitous tissue measurement of all: the H&E slide.
This progression has moved rapidly from relatively straightforward applications—tumor detection, grading, segmentation, cell counting—to prediction of mutations, gene-expression programs, treatment response, prognosis, and tumor-microenvironment states.
The most provocative recent work suggests that the distinction between "molecular assay" and "image assay" is becoming less clean.
Path2Space: predicting spatial transcriptomics from H&E
Shulman and colleagues' Path2Space model was trained using spatial transcriptomic data but deployed on H&E images. It predicts spatial expression patterns for thousands of genes and was applied to 976 TCGA breast tumors. The resulting spatially inferred tumor-microenvironment maps defined patient subgroups and produced treatment-response predictions that, in the reported datasets, exceeded conventional bulk-sequencing biomarkers. (PubMed; DOI). (PubMed)
The conceptual consequence is striking. Spatial transcriptomics can act as teacher data even when the eventual clinical assay is simply a digital H&E slide.
That does not make spatial transcriptomics unnecessary. It may make it more valuable as the high-information reference modality used to discover and train scalable biomarkers.
Atlas H&E-TME: thousands of quantitative spatial features from routine tissue
Standvoss and colleagues describe Atlas H&E-TME, which generates more than 4,500 quantitative tissue and cell-level readouts from H&E. Its validation strategy is notable because the investigators explicitly confront a central problem in computational pathology: H&E itself is an imperfect ground truth for ambiguous cell identities.
They therefore used molecularly informed IHC consensus for "deep" validation while separately testing generalizability across more than 1,500 cases, eight cancers, more than 25 tissue sources, and more than eight scanner types. In the reported experiments, the system matched or exceeded H&E-only pathologist performance for several cell-classification tasks. The paper remains an arXiv preprint rather than a peer-reviewed clinical validation study. (arXiv).
This study makes an important conceptual point for spatial-platform companies: molecularly measured spatial data can become ground truth for AI systems whose eventual input is only morphology.
5. Foundation models are changing the scale of computational pathology
Earlier computational pathology systems were generally designed to perform one task. A prostate cancer detector detected prostate cancer. A tumor segmentation model segmented tumor. Each new task required new labels and substantial new development.
Foundation models aim to learn a broad representation of tissue once and then adapt that representation to many downstream tasks.
Vorontsov and colleagues' PRISM2 illustrates how far this has progressed. PRISM2 was trained on 2.3 million whole-slide images and approximately 14 million question-and-answer pairs derived from nearly 700,000 pathology reports. Rather than learning image morphology alone, it was supervised using clinical language so that tissue features were associated with pathologic reasoning. In research evaluations, prompt-based inference matched or exceeded several specialized clinical-grade products for selected cancer-detection tasks without task-specific retraining. (PubMed; DOI). (PubMed)
Yet the limitations are as informative as the accomplishments. The authors note hallucinations and omissions in report generation, difficulty with some grading and subtyping tasks, limited explicit global spatial reasoning, and the need for further robustness testing because the underlying training slides originated from a restricted scanning environment.
The lesson is that a spectacular model result is not equivalent to a deployable diagnostic.
6. The bottleneck is shifting from measurement to validation
There was a period when the principal obstacle to spatial biology was technical: it was difficult to measure many targets in intact tissue.
That problem has not disappeared, but it is no longer the only bottleneck. Modern platforms can generate extraordinarily rich data. The harder questions increasingly occur downstream:
What exactly is the reportable biological variable?
How reproducibly can it be measured across laboratories, scanners, lots, operators, fixation conditions, and specimen types?
Does the spatial feature add information beyond standard pathology, conventional IHC, sequencing, and known clinical variables?
Can the feature be reduced to a locked algorithm before validation?
Does the result predict an outcome in external patients?
Does acting on the result improve a decision?
Can the assay be performed reliably and economically at clinical scale?
The importance of technical generalizability is becoming increasingly obvious in computational pathology. Scanner, stain, laboratory, fixation, and image-processing differences can become unintended features learned by AI models rather than irrelevant nuisance variables. The breadth component of the Atlas H&E-TME validation—more than 25 sources and more than eight scanner models—reflects precisely this concern.
The same principle applies to high-plex microscopy. A test cannot be analytically validated only as a biological idea. It must demonstrate that the entire pipeline—antibody reagents, staining chemistry, imaging, segmentation, feature computation, and final algorithm—remains reproducible.
7. There are already clinical precedents, but they are revealing in their simplicity
Spatial pathology is not entirely waiting for the future.
TissueCypher for Barrett's esophagus is a particularly useful precedent. The assay applies multiplex immunofluorescence to nine protein biomarkers, extracts 15 quantitative tissue features, and uses a fixed algorithm to estimate a patient's risk of progression to high-grade dysplasia or esophageal adenocarcinoma. Multicenter studies have evaluated its prognostic performance. (Iyer et al., PubMed; DOI). (PubMed)
The commercial experience is no longer trivial. Castle Biosciences reported in June 2026 that TissueCypher had surpassed 100,000 cumulative clinical orders; the number is company-reported rather than an independently audited measure of clinical utility, but it demonstrates that a spatial, algorithmic tissue assay can operate at substantial clinical scale. (Castle Biosciences, June 2026). (Castle Biosciences, Inc.)
The striking feature of TissueCypher is not extreme plex. It is a locked clinical question, a manageable panel, quantitative spatial features, and a defined patient-level score.
That may be a more instructive clinical precedent for CSP than a research platform producing several thousand measurements.
8. Digital pathology itself is already past the proof-of-concept stage
The infrastructure necessary for computational pathology has also matured substantially.
FDA authorized the first whole-slide imaging system for primary diagnosis in 2017. (FDA). (U.S. Food and Drug Administration)
In 2021, FDA authorized Paige Prostate, software intended to assist pathologists in detecting suspicious prostate cancer foci in digitized H&E biopsies. It established a specific regulatory category for machine-learning software analyzing digital pathology slides for cancer detection. (FDA De Novo DEN200080). (FDA Access Data)
Therefore, several pieces of the future system already exist independently:
digital slides can be used for primary diagnosis;
AI can operate on those slides under regulated clinical conditions;
multiplex spatial assays can operate as clinical laboratory tests;
high-plex research assays can discover complex biological tissue states;
and AI can increasingly learn relationships between ordinary H&E morphology and those molecular states.
The unresolved problem is how those pieces will be combined into economically and clinically compelling products.
9. 10x sits unusually close to both sides of the convergence
The current 10x spatial portfolio already illustrates the merging of microscopy, molecular measurement, and computational analysis.
Visium provides whole-transcriptome spatial gene-expression analysis at single-cell scale. Xenium performs high-plex in situ RNA mapping at subcellular resolution and now supports same-section RNA/protein measurements; current Xenium offerings extend to thousands of RNA targets with optional protein measurement. (10x Visium; 10x Xenium). (10x Genomics)
That position creates several potential roles for a CSP-type oncology product.
At one extreme, CSP could itself become the clinical test: a patient sample is run on a standardized multiplex platform and the complete spatial profile generates the report.
At a second level, CSP could become the discovery and validation engine for finding lower-dimensional clinical biomarkers. A 50- or 100-marker assay might discover that six cellular states and three spatial relationships contain nearly all the predictive signal; a much simpler assay could then be commercialized.
At a third level, CSP could become the molecular truth set for computational pathology. Large numbers of spatially characterized slides could train an H&E algorithm to reproduce some clinically useful fraction of the CSP phenotype without performing CSP on every future patient.
These are not mutually exclusive business models.
10. The central competitive question: how much spatial molecular information is actually irreducible?
The development of Path2Space and Atlas H&E-TME raises a difficult strategic question for every high-dimensional spatial assay company.
If the important biology is visible indirectly in morphology, an AI system may eventually infer it from H&E for pennies relative to the cost of a complex molecular assay.
But morphology cannot reveal everything.
Protein isoforms, phosphorylation states, receptor activation, rare transcripts, subtle co-expression patterns, clonally restricted molecular events, and analytes with weak morphological correlates may remain inaccessible to H&E inference. An AI model can only predict molecular information to the extent that morphology contains a reproducible correlate of that information.
This produces a useful distinction:
Spatial measurement asks what is actually there. Computational pathology asks what can be inferred from what is visible.
The strongest long-term clinical role for high-plex spatial measurement may therefore lie in biological information that either cannot be reconstructed reliably from H&E or is important enough that direct measurement remains preferable.
The same high-plex assay may simultaneously serve as the training reference that makes H&E inference possible for other features.
11. The likely translational funnel
A realistic clinical translation pathway increasingly looks less like:
100 proteins → 100 clinical results
and more like:
100 proteins + spatial relationships → biological discovery → candidate phenotype → algorithmic feature selection → external validation → compact patient-level result.
The Wijerathna-Yapa ADC framework explicitly proposes such a process, ending high-dimensional discovery with a simplified clinical-grade multiplex assay.
The TissueCypher precedent demonstrates that the endpoint can be surprisingly compact: nine measured proteins, 15 quantified features, one patient-level risk result.
Path2Space suggests an additional endpoint:
high-dimensional spatial discovery → H&E-trained AI → inexpensive virtual spatial biomarker.
These possibilities should shape how prospective CSP clinical studies are designed. If clinical development merely demonstrates that the platform can enumerate many proteins, it will be vulnerable to both conventional multiplex IHC and computational pathology. If it demonstrates that a particular spatial architecture predicts response where existing diagnostics fail, it creates a clinical proposition rather than a technology demonstration.
12. What would constitute a compelling CSP clinical use case?
The strongest use case would have five characteristics.
First, there would be a specific decision: choose drug A rather than drug B; identify patients likely to respond to an ADC; identify resistance despite nominal target positivity; or recognize patients requiring intensified therapy.
Second, conventional pathology or IHC would perform incompletely. If HER2 intensity alone already answers the question adequately, high-dimensional spatial profiling has limited incremental value.
Third, there would be a biologically plausible reason why spatial context matters. ADC penetration, bystander killing, stromal exclusion, immune neighborhoods, or target heterogeneity satisfy this criterion better than many purely cell-intrinsic biomarkers.
Fourth, the assay would produce a patient-level output, not a thousand-page description of the tumor.
Fifth, the spatial result would demonstrate incremental predictive or decision value over existing alternatives.
ADC selection therefore appears considerably more promising as a conceptual starting point than generic "cancer characterization." The Cai study is notable because it connects the full chain: molecular state → spatial microenvironment → treatment resistance → predictive model → lower-plex potential clinical implementation.
13. Perspective for the 10x project
The 2026 literature supports neither the view that high-plex spatial microscopy is merely an exotic research technology nor the view that it is about to replace routine pathology wholesale.
A more defensible interpretation is that the field is entering a translation-and-compression phase.
The ability to produce high-dimensional tissue maps is increasingly established. The frontier is determining which spatial features are sufficiently reproducible, predictive, actionable, and economically important to become clinical products.
Computational pathology is accelerating this transition because it can convert molecularly annotated tissue into scalable image biomarkers. In that sense, spatial microscopy and computational pathology are not simply competitors. They can form a productive loop:
spatial measurement discovers biology → computational pathology learns the phenotype → large H&E cohorts validate it → selected molecular measurements refine or confirm it → clinical assays are designed around the smallest reliable representation of the actionable biology.
For 10x, this suggests that the strategic unit of value should not automatically be the number of proteins measured. It should be the clinically actionable spatial phenotype that the platform can uniquely discover, quantify, validate, or deliver.
The most consequential question for CSP may therefore be:
What clinically important decision becomes possible because this assay knows not merely which proteins are present, but which cells express them, in what states, and in what spatial relationship to one another—and can that information be shown to outperform what pathology, IHC, sequencing, or H&E-based AI could otherwise provide?
If that question can be answered convincingly, the technological complexity becomes an asset. If it cannot, increasing plex alone is unlikely to create a clinical diagnostic category.
Bibliography
1. Liu Y, Dai Y, Wang L. 2026. Spatial omics at the forefront: emerging technologies, analytical innovations, and clinical applications. Cancer Cell. 44(1):24–49. DOI: 10.1016/j.ccell.2025.12.009. PubMed PMID 41478277. (PubMed)
2. Bollhagen A, Bodenmiller B. 2024. Highly Multiplexed Tissue Imaging in Precision Oncology and Translational Cancer Research. Cancer Discovery. 14(11):2071–2088. DOI: 10.1158/2159-8290.CD-23-1165. PubMed PMID 39485249. (PubMed)
3. Pentimalli TM, Karaiskos N, Rajewsky N. 2025. Challenges and Opportunities in the Clinical Translation of High-Resolution Spatial Transcriptomics. Annual Review of Pathology: Mechanisms of Disease. 20:405–432. DOI: 10.1146/annurev-pathmechdis-111523-023417. PubMed PMID 39476415. (Annual Reviews)
4. Yan Y, Lin Y, Kumar T, et al. 2026. Ecotypes of triple-negative breast cancer in response to chemotherapy. Nature. 654:1088–1097. DOI: 10.1038/s41586-026-10469-9. PubMed PMID 42129561. (PubMed)
5. Liu Y, Jadhav AS, Pan Y, et al. 2026. Spatial multi-omics landscape of colorectal cancer macro- and micrometastases. Cancer Cell. 44(8):1568–1586.e11. DOI: 10.1016/j.ccell.2026.06.009. PubMed PMID 42425074.
6. Cai Y-W, Jia S-H, Wang H, et al. 2026. Spatial proteogenomic profiling uncovers sensitization strategies for antibody-drug conjugate in HER2-positive breast cancer. Cell Reports Medicine. 7:102987. Published online ahead of the September 15, 2026 issue. DOI: 10.1016/j.xcrm.2026.102987. PubMed PMID 42594874.
7. Sun Y, Zhong Y, et al. 2026. Spatiotemporal multiomics uncover tumor ecosystem dynamics during metastatic colonization. Science. 393(6810):eadz7928. DOI: 10.1126/science.adz7928. PubMed PMID 42531396. (PubMed)
8. Shulman ED, Campagnolo EM, Lodha R, et al. 2026. AI-predicted spatial transcriptomics unlocks breast cancer biomarkers from pathology. Cell. 189(14):4225–4240.e25. DOI: 10.1016/j.cell.2026.04.023. PubMed PMID 42105763.
9. Vorontsov E, Shaikovski G, Casson A, et al. 2026. End-to-end multimodal pathology foundation model with clinical dialogue. Nature Medicine. Published online July 31, 2026. DOI: 10.1038/s41591-026-04521-4. PubMed PMID 42538427.
10. Standvoss K, Hägele M, Krupar R, et al. 2026. Atlas H&E-TME: Scalable AI-Based Tissue Profiling at Expert Pathologist-Level Accuracy. arXiv 2606.12346v2, July 14, 2026. arXiv. No PubMed PMID as of this writing.
11. Wijerathna-Yapa A, Mansfield J, Vasaturo A, Yeong JPS, Kulasinghe A. 2026. Spatial multiomics approaches for antibody-drug conjugate target discovery. Cell Press Blue. In press. DOI: 10.1016/j.cpblue.2026.100085. No PubMed record identified at the time of this review.
12. Iyer PG, Codipilly DC, Chandar AK, et al. 2022. Prediction of Progression in Barrett's Esophagus Using a Tissue Systems Pathology Test: A Pooled Analysis of International Multicenter Studies. Clinical Gastroenterology and Hepatology. 20(12):2772–2779.e8. DOI: 10.1016/j.cgh.2022.02.033. PubMed PMID 35217151. (PubMed)
13. Castle Biosciences. 2026. TissueCypher Test Surpasses 100,000 Orders, Advancing Precision Risk Assessment for Patients with Barrett's Esophagus. Corporate release, June 16, 2026. Castle Biosciences source. No PMID. (Castle Biosciences, Inc.)
14. U.S. Food and Drug Administration. 2017. FDA Allows Marketing of First Whole Slide Imaging System for Digital Pathology. April 12, 2017. FDA source. No PMID. (U.S. Food and Drug Administration)
15. U.S. Food and Drug Administration. 2021. De Novo Classification Request for Paige Prostate, DEN200080. September 21, 2021. FDA decision letter. No PMID. (FDA Access Data)
16. 10x Genomics. 2026. Xenium In Situ Platform and Visium Spatial Platform product and technical information. Xenium; Visium. No PMID. (10x Genomics)
Annotated Bibliography
1. Liu Y, Dai Y, Wang L. 2026. Spatial omics at the forefront: emerging technologies, analytical innovations, and clinical applications. Cancer Cell. 44(1):24–49.
DOI | PubMed PMID 41478277
This broad 2026 review is one of the most useful orientation documents for the field. It organizes the expanding spatial transcriptomic, proteomic, and related technology landscape while emphasizing the transition from descriptive maps toward functional niches, therapeutic response, multimodal integration, and clinically oriented study design. It is particularly useful for understanding spatial biology as an emerging clinical-information architecture rather than as a collection of microscopy platforms. (PubMed)
2. Bollhagen A, Bodenmiller B. 2024. Highly Multiplexed Tissue Imaging in Precision Oncology and Translational Cancer Research. Cancer Discovery. 14(11):2071–2088.
DOI | PubMed PMID 39485249
This review provides a strong overview of multiplex tissue-imaging technologies and their role in precision oncology. It explains why preserving tissue architecture adds information beyond dissociated-cell or bulk molecular assays and provides useful context for the capabilities and limitations of high-plex protein imaging. (PubMed)
3. Pentimalli TM, Karaiskos N, Rajewsky N. 2025. Challenges and Opportunities in the Clinical Translation of High-Resolution Spatial Transcriptomics. Annual Review of Pathology. 20:405–432.
DOI | PubMed PMID 39476415
Pentimalli and colleagues focus specifically on the gap between research performance and clinical translation. The review is important because it counters technological enthusiasm with the practical requirements of clinical pathology: reproducible specimens, standardized workflows, manageable data, validation, cost, and interpretability. (Annual Reviews)
4. Yan Y, Lin Y, Kumar T, et al. 2026. Ecotypes of triple-negative breast cancer in response to chemotherapy. Nature. 654:1088–1097.
DOI | PubMed PMID 42129561
This large study illustrates why spatial biology may create biomarkers qualitatively different from traditional single-analyte tests. The investigators identify cellular communities and spatial ecotypes associated with chemotherapy response, with macrophage and cancer-cell programs emerging as important features. It provides a strong example of the "multicellular phenotype" as the potential diagnostic unit.
5. Liu Y, Jadhav AS, Pan Y, et al. 2026. Spatial multi-omics landscape of colorectal cancer macro- and micrometastases. Cancer Cell. 44(8):1568–1586.e11.
DOI | PubMed PMID 42425074
This study integrates conventional pathology, spatial transcriptomics at two resolutions, genomics, and high-plex protein imaging. It demonstrates how microscopic metastatic deposits differ biologically and spatially from overt metastases and derives a six-gene signature linked to MRD, survival, and treatment resistance. It is a particularly good example of high-dimensional spatial discovery being distilled toward a much smaller biomarker.
6. Cai Y-W, Jia S-H, Wang H, et al. 2026. Spatial proteogenomic profiling uncovers sensitization strategies for antibody-drug conjugate in HER2-positive breast cancer. Cell Reports Medicine. 7:102987.
DOI | PubMed PMID 42594874
Cai and colleagues provide one of the most directly relevant examples for a spatial oncology protein strategy. High-dimensional imaging identified tumor-cell H3K27ac and collagen-positive fibroblasts as determinants of ADC response and supported a spatial neighborhood-based prediction model. Especially important for translation, the investigators explicitly propose implementation of the resulting model with multiplex immunofluorescence rather than requiring the full discovery platform.
7. Sun Y, Zhong Y, et al. 2026. Spatiotemporal multiomics uncover tumor ecosystem dynamics during metastatic colonization. Science. 393(6810):eadz7928.
DOI | PubMed PMID 42531396
This paper demonstrates that spatial biology can capture dynamic rather than merely static tumor states. The identification of a transient quiescent disseminated-tumor-cell state and subsequent macrophage-mediated niche remodeling shows how spatially resolved biology can identify mechanistic windows that conventional bulk measurements would obscure. (PubMed)
8. Shulman ED, Campagnolo EM, Lodha R, et al. 2026. AI-predicted spatial transcriptomics unlocks breast cancer biomarkers from pathology. Cell. 189(14):4225–4240.e25.
DOI | PubMed PMID 42105763
Path2Space is a key convergence paper. Spatial transcriptomics supplies the high-information training data, while routine H&E becomes the scalable input. The study therefore suggests a plausible future in which expensive spatial assays are crucial for biomarker discovery and model training even when the final clinical test is performed computationally on ordinary pathology images.
9. Vorontsov E, Shaikovski G, Casson A, et al. 2026. End-to-end multimodal pathology foundation model with clinical dialogue. Nature Medicine. Published online July 31, 2026.
DOI | PubMed PMID 42538427
PRISM2 represents the rapid evolution from single-purpose pathology AI toward large generalist models. Its scale—millions of slides plus report-derived clinical language—and performance against specialized clinical algorithms show how quickly computational pathology is progressing. Equally important are the authors' stated limitations concerning hallucination, global spatial reasoning, and scanner/site generalizability.
10. Standvoss K, Hägele M, Krupar R, et al. 2026. Atlas H&E-TME: Scalable AI-Based Tissue Profiling at Expert Pathologist-Level Accuracy. arXiv 2606.12346v2.
arXiv
This preprint is useful less because of any single accuracy statistic than because of its validation architecture. The authors recognize that pathologists reading H&E are themselves an imperfect reference for some cell types, establish an IHC-informed molecular reference for deep validation, and separately test broad generalizability across many cancers, laboratories, and scanners. It illustrates how spatial molecular measurement and computational pathology can become mutually reinforcing rather than competing disciplines.
11. Wijerathna-Yapa A, Mansfield J, Vasaturo A, Yeong JPS, Kulasinghe A. 2026. Spatial multiomics approaches for antibody-drug conjugate target discovery. Cell Press Blue. In press.
DOI
This perspective provides a particularly useful framework for ADC applications, in which target density, microenvironment, stromal barriers, payload diffusion, and bystander effects occur at different spatial scales. Its proposed development funnel—high-dimensional discovery followed by biomarker selection and eventual clinically practical multiplex testing—is directly relevant to how a CSP-like technology might create value.
12. Iyer PG, Codipilly DC, Chandar AK, et al. 2022. Prediction of Progression in Barrett's Esophagus Using a Tissue Systems Pathology Test: A Pooled Analysis of International Multicenter Studies. Clinical Gastroenterology and Hepatology. 20(12):2772–2779.e8.
DOI | PubMed PMID 35217151
The pooled TissueCypher analysis provides a concrete example of a multiplex spatial pathology assay that has crossed into clinical laboratory use. Nine protein biomarkers and morphometric information are converted into quantitative image features and a patient-level risk result. It demonstrates that clinical success does not necessarily require extreme plex; it requires a validated output linked to a defined management problem. (PubMed)
13. Castle Biosciences. 2026. TissueCypher Test Surpasses 100,000 Orders. Corporate release, June 16, 2026.
Source
This company source provides contemporary commercial context rather than independent evidence of clinical validity. The reported 100,000 cumulative orders indicate that an algorithmic spatial tissue assay can be operationalized at meaningful scale, which makes TissueCypher a useful existence proof when considering clinical deployment models for multiplex spatial pathology. (Castle Biosciences, Inc.)
14. U.S. Food and Drug Administration. 2017. FDA Allows Marketing of First Whole Slide Imaging System for Digital Pathology.
FDA source
This FDA action marks an important infrastructure milestone: digital whole-slide images could be used for primary pathologic diagnosis rather than merely consultation or research. It created the regulated digital substrate upon which clinical computational pathology could subsequently develop. (U.S. Food and Drug Administration)
15. U.S. Food and Drug Administration. 2021. De Novo Classification Request for Paige Prostate, DEN200080.
FDA decision letter
Paige Prostate is a landmark regulatory precedent for machine-learning analysis of digitized H&E pathology. The authorization is narrower than the molecular and spatial inference envisioned by current foundation models, but it demonstrates that regulated AI-assisted interpretation of pathology images is already a clinical reality rather than a hypothetical future category. (FDA Access Data)
16. 10x Genomics. 2026. Xenium In Situ Platform and Visium Spatial Platform.
Xenium | Visium
The current 10x platforms illustrate the broader technological convergence described in the essay. Visium provides high-dimensional spatial transcriptomic measurement, while Xenium integrates subcellular in situ imaging, molecular detection, optional protein measurement, cell segmentation, and computational visualization. These capabilities place 10x near the intersection of high-dimensional tissue measurement and the computational extraction of clinically meaningful spatial phenotypes. (10x Genomics)
#
#
#
What is the five- to ten-year outlook for clinical testing with spatial biology? One expert says, "Unlikely!"
#
#
The statement is quite defensible—but much depends on what counts as a “real impactful clinical application.” If that means that by 2031–2036 pathology laboratories will routinely perform 50-, 100-, or 1,000-plex spatial assays on patients and use the results to select treatment, the skeptical position has considerable force. If it means that spatial measurement will materially influence biomarker discovery, companion diagnostics, trial enrollment, or a smaller number of specialized clinical tests, the prediction looks much less secure.
The case for skepticism
The strongest argument is historical. Molecular diagnostics is full of technologies that produced spectacular research papers years before producing useful clinical tests. Demonstrating that two groups of patients have different spatial ecosystems is comparatively easy; demonstrating that measuring that ecosystem prospectively improves a medical decision is much harder.
Much of the current literature still sits toward the discovery end of that spectrum. Spatial studies can now identify dozens of cell states, cellular neighborhoods, ligand-receptor relationships and microenvironmental structures. But the output is often a biological atlas rather than a clinical answer. A clinician cannot order “the 37-dimensional tumor ecosystem.” A test ultimately has to say something such as give this drug, avoid this drug, increase surveillance, or this patient has a 35% five-year risk. The necessary compression from thousands of measurements to a locked, validated patient-level result remains a major translational bottleneck.
Analytical validation is also unusually difficult. A spatial assay is not simply measuring whether analyte X is present. Its result may depend upon fixation, staining, antigen retrieval, image quality, segmentation, cell classification, neighborhood definitions and an algorithm. An industry review of clinical spatial workflows explicitly describes the historical problems of inconsistent staining, operator-dependent scanning, inter-instrument variation and computational complexity. These problems multiply as plex and algorithmic complexity increase.
There is also a statistical problem. High-dimensional spatial assays can generate enormous numbers of candidate features from relatively small clinical cohorts. A beautiful spatial signature discovered in 40 or 100 patients can disappear when taken to another institution, another scanner, another patient population or another treatment regimen. The ADC literature itself notes substantial attenuation between discovery and validation for some spatial signatures.
Economics strengthens the skeptical argument. Routine H&E costs little because it is already performed. IHC is ubiquitous and clinically intelligible. Sequencing has established workflows, coding, quality systems and reimbursement. A new spatial assay therefore must show not merely that it provides additional information, but that the additional information is worth the incremental specimen handling, instrumentation, computation, professional interpretation and payment.
And computational pathology presents a particularly interesting threat. Path2Space was trained on spatial transcriptomic data but then inferred spatial gene-expression patterns from ordinary H&E. Its authors describe the resulting approach as scalable and inexpensive and report treatment-response predictions derived from routine histology. If that paradigm generalizes, some spatial technologies could perform the expensive discovery work only to teach an AI model that subsequently performs the clinical test from H&E.
Thus the skeptical expert could plausibly argue:
Spatial biology may become enormously important scientifically while the spatial assay itself never becomes a major clinical laboratory testing category.
That distinction has considerable precedent in laboratory medicine.
The case against the skeptical forecast
The strongest counterargument is that the predicted future has already begun in modest form.
TissueCypher is difficult to dismiss completely. It uses nine-protein multiplex immunofluorescence, quantitative digital microscopy and computational spatial features to generate a patient-level Barrett's esophagus progression-risk result. The supplied industry account reports more than 39,000 test reports in 2025. It is therefore incorrect in a literal sense to say that spatial biology has no clinically impactful application today.
The skeptic could reasonably respond that nine-marker multiplex fluorescence is quite different from the much more ambitious technologies now marketed as “spatial biology.” But that objection may reveal the likely translation mechanism rather than invalidate the field.
Clinical spatial biology may not arrive as a 100-marker clinical assay. It may arrive through a funnel:
high-plex spatial discovery → identification of the critical biology → algorithmic reduction → small multiplex panel → clinical score.
The new ADC work from Cai and colleagues is a striking example. Imaging mass cytometry on more than 900,000 cells identified spatial features associated with ADC response, including H3K27ac-positive tumor cells and collagen-positive fibroblasts. The investigators then constructed an ADC barrier score from relationships among epithelial cells, endothelial cells and fibroblasts. Crucially, they explicitly describe the resulting model as potentially implementable using multiplex immunofluorescence.
That is much closer to a plausible companion diagnostic than a 50-page spatial atlas.
ADCs also provide a biological reason to suspect that spatial information may eventually prove irreducible. ADC efficacy depends not merely upon whether HER2, TROP2 or another antigen is present. Antigen heterogeneity, membrane localization, stromal barriers, vascular access and bystander diffusion all occur at different spatial scales. A recent ADC perspective argues that these processes cannot simultaneously be represented by conventional IHC or bulk sequencing. If a spatial phenotype predicts a $100,000-plus cancer therapy substantially better than conventional IHC, the economics of testing change dramatically.
There is another way in which the skeptic could be wrong: clinical impact need not require routine spatial testing of every patient.
Spatial methods could become consequential in drug development and companion-diagnostic creation. A pharma company might use thousands of spatially characterized trial samples to determine why an ADC works, establish which patients should receive it, and then translate the finding into a simpler diagnostic. The spatial platform would have altered clinical medicine even though the FDA-approved companion test ultimately contained six stains rather than sixty.
Path2Space makes the same point from the opposite direction. If spatial transcriptomics supplies the molecular ground truth that enables an H&E algorithm to predict treatment response, spatial biology has had a major clinical impact even though the patient never receives a spatial transcriptomics assay.
Both sides may therefore be partly right
The disagreement becomes much smaller when “spatial biology” is divided into three different propositions.
The most doubtful proposition is that comprehensive high-plex spatial profiling becomes the next NGS—an ordinary diagnostic test ordered on large numbers of cancer patients, reimbursed at several thousand dollars, and routinely interpreted in clinical laboratories. Five years is very short for that transition, and even ten years may be optimistic.
A more plausible proposition is that several narrow clinical assays emerge in which a particular spatial phenotype has unusually strong utility: ADC response, immunotherapy response, progression risk, or another treatment-selection problem. Those tests may contain 5–20 markers rather than 100.
The most plausible proposition is that spatial biology becomes deeply embedded in biomarker discovery, pharmaceutical development and computational pathology, while remaining largely invisible to the practicing physician. High-dimensional spatial assays would discover the phenotype; smaller multiplex assays or H&E AI would deliver it clinically.
That last possibility makes the apparently bearish forecast surprisingly compatible with a bullish view of the underlying technology.
For the 10x project, this distinction is consequential. A convincing commercial thesis probably should not depend upon the assumption that oncologists will routinely order a massive spatial profile simply because such a profile can be generated. The stronger thesis is that high-dimensional spatial measurement can discover or directly measure a specific spatial biological state that changes an expensive and consequential treatment decision, and that 10x can own an important part of the path from discovery to clinically deployable result.
In that formulation, the skeptical pathologist's warning is less an argument against spatial biology than a warning against confusing technical capability with a clinical product. The 2026 evidence makes that warning look well founded. But it does not yet justify concluding that spatial biology will have little clinical impact over the coming decade. The more likely outcome is narrower, less spectacular, and perhaps commercially more interesting: a handful of important clinical applications, many more pharmaceutical applications, and substantial clinical impact delivered indirectly through compressed assays and computational pathology.
From H&E to Genes Standvoss vs Dawood
The two papers look, at first glance, as though they are on opposite sides of the same argument about how much biology can be extracted from H&E. On closer inspection, Dawood is much less damaging to Standvoss than to the H&E→gene-expression/mutation papers that Dawood was actually attacking. Standvoss has rather cleverly moved the target much closer to what is visibly present in the slide.
Dawood's central objection is not that AI cannot extract useful information from H&E. It is that an impressive H&E prediction of, say, ER, TP53, BRAF, MSI, TMB, or another molecular variable does not establish that the network has recognized morphology caused specifically by that molecular variable. Grade, histologic subtype, other mutations, receptor status, TMB, etc. are correlated with one another. A model can therefore obtain a very good aggregate AUROC while mostly reading a correlated phenotype. Dawood shows exactly this by stratification: performance often falls substantially when those correlations are broken apart. For example, the paper finds that MSI prediction falls markedly after conditioning on hypermutation or other correlated variables, and ER/PR/mutation predictions weaken within grade or molecular strata.
Standvoss is doing something fundamentally different. Atlas H&E-TME does not claim to read TP53 mutations, RNA expression, or HER2 biology from an H&E slide. Its primary outputs are much nearer the optical phenotype itself:
tissue regions — carcinoma, epithelium, stroma, necrosis, blood, vessels;
cell classes — carcinoma cells, lymphocytes, plasma cells, macrophages, granulocytes, fibroblasts, endothelial cells, etc.;
quantities derived from those classifications — counts, densities, ratios, nuclear morphology and spatial neighborhoods.
Those become >4,500 numerical features, but the 4,500 should not be confused with 4,500 inferred molecular measurements. They are largely computed morphometric/spatial measurements of things the algorithm first identifies on H&E.
That distinction is crucial. Dawood asks, in effect:
Does this image really tell you the mutation, or is the AI merely recognizing grade/subtype/TME patterns correlated with the mutation?
Standvoss asks:
Can the AI accurately identify and quantify the grade/subtype/TME-type morphology itself?
The latter is much less vulnerable to Dawood.
In one important respect, Standvoss actually behaves as if it had absorbed the Dawood philosophy
Its most interesting methodological feature is the refusal to regard an H&E pathologist label automatically as ground truth. Standvoss explicitly acknowledges that macrophages, granulocytes, plasma cells, etc. can be morphologically ambiguous. The investigators stain the same physical section first with H&E and then, after bleaching, with a five-marker IHC panel, coregister the images, and have five pathologists reassess the cells with the molecular information available. That produces the IHC-informed consensus against which both AI and H&E-only pathologists are judged.
That is actually a quite strong answer to one family of "paper tiger" problems. They are not saying:
H&E AI agrees with H&E annotations; therefore it must be biologically right.
They ask whether the H&E inference survives confrontation with an orthogonal biological reference. And it does reasonably well: macro F1 0.738 for Atlas versus about 0.71 for the average H&E-only pathologist, with particularly high performance for carcinoma cells and lymphocytes and weaker performance for macrophages.
The paper also has an unusually extensive breadth test: >1,500 cases, >200,000 annotations, eight cancers, major metastatic sites, >25 sources and >8 scanners. That addresses another Dawood-adjacent concern—whether an apparently impressive result is really a peculiar property of one cohort.
But Dawood has not disappeared. He re-enters one step downstream.
This is where the wording in Standvoss becomes important. The authors say these 4,500 H&E-derived features can support "biomarker discovery or patient stratification" and describe Atlas as laying a foundation for a new generation of tissue-based biomarkers.
That is reasonable, but it is precisely at that second step that a Dawood analysis becomes necessary.
Suppose Atlas reports:
high macrophage density + particular tumor/stroma geometry + lymphocyte neighborhood pattern → response to pembrolizumab.
Or:
Atlas spatial signature X → BRAF mutation.
Or:
Atlas TME signature Y → HER2-low biology.
Now Dawood's questions apply almost verbatim. Does X independently contain information about BRAF, or is X merely tracking right-sided CRC, MSI, grade, histologic pattern and other variables correlated with BRAF? Does the putative immunotherapy-response signature work within PD-L1, MSI, grade, stage, histologic subtype and treatment strata? Does it add anything to information already available to the pathologist and oncologist?
So there are really two layers:
H&E → cells/tissue/spatial architecture:
Dawood has limited force. Standvoss validates this unusually well.
Cells/tissue/spatial architecture → genotype, expression, prognosis or treatment response:
Dawood comes roaring back. Every downstream "biomarker" created from the Atlas features would need dependency-aware, stratified validation.
There is even a little warning sign inside Standvoss itself. In the broad validation, macrophage classification is much weaker in certain environments—for example F1 0.54 at bone metastases and 0.65 in liver metastases—while carcinoma classification remains ≥0.96. Standvoss correctly attributes this to the difficulty of distinguishing morphologically similar resident phagocytic populations. This is not the Dawood molecular-confounding phenomenon, but it illustrates the underlying principle: an AI result can depend strongly on the biological context in which the visual feature occurs. Standvoss recognizes rather than conceals that limitation.
And, interestingly, Standvoss does not cite Dawood
I checked the Standvoss reference list and searched the paper specifically for Dawood. Dawood et al., Nature Biomedical Engineering 2026, is not among the 43 references. The related-work discussion includes H&E tissue profiling, foundation models, H&E annotation variability, IHC/mIF-derived ground truth, and newer benchmarks such as PhenoBench, but not Dawood's confounding paper.
That omission is understandable because the papers' immediate technical problems differ. But intellectually, I think Dawood deserved a citation, especially around Standvoss's claims about downstream biomarker discovery. The combination of the papers yields a very useful hierarchy:
H&E can be excellent at measuring morphology.
Morphology can certainly correlate with molecular biology.
But accurately measuring morphology does not prove that one has measured the molecular variable that correlates with it.
That last sentence is almost the reconciliation of Standvoss and Dawood.
Bottom line: Dawood does not knock down Atlas H&E-TME. In fact, Standvoss is one of the more defensible ways to push H&E AI forward after Dawood: extract explicit, interpretable morphology and validate it against IHC rather than making a black-box leap directly to invisible molecular states. But Atlas's impressive 4,500-feature output should be viewed as a very rich digital morphometry panel, not a virtual spatial-omics assay. The moment those features are claimed to stand in for gene expression, mutation status, drug response, or other molecular biomarkers, the full Dawood burden of proof applies again.
No comments:
Post a Comment
Note: Only a member of this blog may post a comment.