Reading tumor ecosystems from routine histologyDownload PDF Download PDF Research HighlightPublished: 10 August 2026Ludvig Bergenstråhle1 &Joakim Lundeberg ORCID: orcid.org/0000-0003-4313-16012 Cell Research (2026) Cite this articleSave articleView saved researchSubjectsCancer imagingProteomicsLi et al. introduce CANVAS, an AI framework for translating hematoxylin and eosin images into spatial maps of tumor habitats. By predicting stable, biologically anchored labels with foundation-model image analysis, CANVAS extends habitat mapping to archival pathology samples and suggests a path toward more accessible precision oncology.Tumors are complex ecosystems comprising a multitude of interdependent cell populations. By measuring proteins and transcripts in situ using multiplex spatial proteomics and transcriptomics, researchers have characterized the spatial organization of tumors across a wide range of cancers, revealing recurrent cellular neighborhoods associated with disease progression and clinical outcomes.1,2 Yet, the high cost and technical complexity of these technologies have confined them to a research setting and impeded their adoption in routine clinical practice. Thus, the potential of spatially resolved molecular profiling to inform diagnosis, prognosis, and treatment decisions remains largely untapped.By contrast, hematoxylin and eosin (H&E) staining is inexpensive, ubiquitous, and routine in diagnostic histopathology but captures morphology rather than molecular state. Researchers have therefore sought to predict the expression of individual molecules directly from H&E images.3,4 However, these efforts have been hampered by the limited availability of experimental data and considerable technical variability between samples, making the learned mappings fragile and difficult to generalize.5 In their recent study, Li et al.6 reframe the problem: instead of predicting noisy molecular markers, they predict the higher-order cellular neighborhoods that make up the tumor ecosystem. First, they measured 41 proteins in more than 18 million cells from 457 patients with non-small-cell lung cancer (NSCLC) and distilled the data into ten reproducible neighborhoods spanning tumor, immune, and stromal compartments. The distilled data encoded clinically meaningful biology, from prognostic associations to fibroblast barriers that can wall off cytotoxic T cells from the tumor cells they target.7 They next fine-tuned a pretrained pathology foundation model to predict patch-level cellular neighborhoods (habitats) from co-registered H&E images (Fig. 1a).8 Despite being fine-tuned on only eight whole-slide images, the resulting model, CANVAS, achieved a macro-averaged F1 score of 0.79 and an overall classification accuracy of 0.88 in an independent test cohort.Fig. 1: From spatial depth to histological scale.Full size imagea 41-plex CODEX imaging defines recurrent cellular neighborhoods, exemplified by tumor core, B-cell niche, and stromal-fibrotic domains, and provides a reference habitat map. Co-registered H&E images are used to train CANVAS, which combines the MUSK image encoder (partially frozen) with a multilayer perceptron (trainable) to infer habitats from routine histology. b The trained model enables virtual habitat mapping in large H&E cohorts, supporting analyses of prognosis, patient-level spatial ecotypes, and immune checkpoint blockade outcomes. MLP multilayer perceptron, ICB immune checkpoint blockade.The translational promise of CANVAS comes from applying it beyond the spatial proteomics data the model was trained on. To this end, the authors used their method on archival H&E slides from more than 5000 patients in public clinical datasets spanning nine cancer types (Fig. 1b). In NSCLC, the composition of the inferred habitat maps recapitulated the prognostic signal of the spatial proteomics reference. Specifically, T- and B-cell habitats were associated with more favorable outcomes; tumor core, fibrotic, and neutrophil-rich habitats with poorer ones. At the patient level, the maps further grouped tumors into recurrent spatial ecotypes linked to distinct genomic alterations. The inferred habitats remained prognostic in other cancer types, albeit in a type-dependent manner.One of the most striking results concerned immunotherapy. In a smaller cohort of 149 NSCLC patients treated with immune checkpoint blockade, CANVAS-derived features from pretreatment H&E images were used to model progression-free survival (PFS). Here, the predictive signature captured the ecological organization of the tissue rather than habitat abundance alone. Co-localized T- and B-cell niches and higher habitat diversity were protective features, whereas fibrotic architecture and neutrophil accumulation were high-risk features. The resulting spatial signature outperformed established immunotherapy biomarkers (PD-L1 expression, tumor mutational burden, and tertiary lymphoid structures). PFS stratification was additionally validated in an independent 40-patient cohort.CANVAS highlights a shift in perspective on prediction target design. By mapping multicellular habitats instead of individual markers, CANVAS trades molecular granularity for a biologically structured target that may be less sensitive to technical variability. Although habitats compress the proteomics profiles they are based on, their interpretability and robustness could offer a valuable compromise in the low-signal, small-sample regime of contemporary translational oncology.The same design choice that makes CANVAS attractive also sets its limits: the model inherits the vocabulary of the habitats, a taxonomy of ten cellular neighborhoods that defines the boundaries of recognizable tissue ecology. Niches outside that vocabulary may stay invisible, even when observable from the underlying proteomics data. Separately, H&E-based inference is fundamentally bounded by the information encoded in tissue morphology, and the task-relevant signal that can reliably be extracted from it is not yet fully known.5 Finally, although habitat classification can aid interpretation, downstream associations are correlative. A habitat linked to a particular outcome may be mechanistically important, but it may also be a proxy for a different underlying process that mediates the effect. Causally connecting spatial ecology to clinical outcomes therefore remains an open challenge, and prospective, controlled validation will be essential before clinical deployment.Ultimately, the broader significance of CANVAS lies in its role as a template for supervising AI in digital pathology, demonstrating how information-rich assays can be distilled into habitats serving as discrete training targets. Researchers are now confronted with a productive tension: in today’s limited data setting, such targets can constrain and stabilize learning. As datasets grow, however, nuanced tissue patterns not captured by those same targets may become learnable. The next step for spatial diagnostics may therefore be a hybrid approach that pairs the predictive power of dense, continuous representations with the biological interpretability at the heart of CANVAS.ReferencesSchürch, C. M. et al. Cell 182, 1341–1359.e19 (2020).Article PubMed PubMed Central Google Scholar Elhanani, O., Ben-Uri, R. & Keren, L. Cancer Cell 41, 404–420 (2023).Article CAS PubMed Google Scholar He, B. et al. Nat. Biomed. Eng. 4, 827–834 (2020).Article CAS PubMed Google Scholar Li, Z. et al. Nat. Med. 32, 231–244 (2026).Article CAS PubMed PubMed Central Google Scholar Dawood, M., Branson, K., Tejpar, S., Rajpoot, N. & Minhas, F. U. A. A. Nat. Biomed. Eng. https://doi.org/10.1038/s41551-026-01616-8 (2026).Article PubMed Google Scholar Li, Y. et al. Cell 189, 4241–4259.e9 (2026).Article CAS PubMed PubMed Central Google Scholar Grout, J. A. et al. Cancer Discov. 12, 2606–2625 (2022).Article CAS PubMed PubMed Central Google Scholar Xiang, J. et al. Nature 638, 769–778 (2025).Article CAS PubMed PubMed Central Google Scholar Download referencesAuthor informationAuthors and AffiliationsDepartment of Bioengineering, Stanford University, Stanford, CA, USALudvig BergenstråhleDepartment of Gene Technology, KTH Royal Institute of Technology, Science for Life Laboratory, Stockholm, SwedenJoakim LundebergAuthorsLudvig BergenstråhleView author publicationsSearch author on:PubMed Google ScholarJoakim LundebergView author publicationsSearch author on:PubMed Google ScholarCorresponding authorsCorrespondence to Ludvig Bergenstråhle or Joakim Lundeberg.Ethics declarationsCompeting interestsJ.L. is a board member of Navinci AB that develops reagents to monitor proximity protein binding events. These reagents are not used in the study of Li et al. and do not impact the writing of this article.Additional informationPublisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.Rights and permissionsReprints and permissionsAbout this article