The research team, led by corresponding authors Liping Sun of China Medical University and Huakang Tu of Zhejiang University School of Medicine, set out to determine whether digital pathology slides could serve as a window into the tumor microbiome. Current sequencing-based approaches for detecting intratumoral microbes lack spatial histopathological context, meaning they can confirm a microbe’s presence but cannot reveal where in the tissue it resides or what morphological features accompany its colonization. By contrast, the deep learning approach developed here preserves that spatial dimension while requiring only the slides that pathologists already produce every day.

To test their hypothesis, the investigators assembled an unusually large and diverse patient population: 1,301 colorectal cancer patients drawn from four independent cohorts. These included 608 patients from the TCGA-CRC dataset, 102 from CPTAC-COAD, 298 from the Ningbo Clinical Pathology Diagnosis Center (NCPDC), and 293 from China Medical University (CMU). The primary microbial labels were derived from metagenomic sequencing, with additional label-sensitivity analyses performed using relative abundance thresholds of 0.001% and 0.01% to ensure the findings were not artifacts of a single labeling convention.

On the computational side, the team evaluated multiple pathology foundation models, including CONCH, Gigapath, UNI, UNI2, and Virchow_v2, for predicting the Fusobacterium genus and five representative species from H&E whole-slide images. These foundation models, pretrained on vast collections of pathology images, encode subtle visual patterns that human observers may not consciously register. Among single models, UNI and Virchow_v2 demonstrated the most robust performance, with the highest area under the curve (AUC) reaching 83.0% for Fusobacterium massiliense, one of the five species examined.

The researchers did not stop at single models. They constructed ensemble models using an averaging fusion strategy, combining the outputs of multiple foundation models to improve reliability. The ensembles showed consistently robust performance and high inter-model agreement across different Fusobacterium species, particularly for Fusobacterium nucleatum, where agreement reached 99.2%. Model performance and ensemble robustness were evaluated using classification metrics, agreement analysis, and Shapley value attribution, a technique borrowed from game theory that quantifies each model’s contribution to the final prediction.

Beyond raw predictive accuracy, the study revealed biologically meaningful patterns. Predicted Fusobacterium colonization was consistently enriched in mucinous adenocarcinoma and microsatellite instability (MSI) tumors across multi-cohort association analyses, suggesting that the models were capturing genuine histomorphological niches rather than spurious correlations. These associations align with known clinicopathological features of Fusobacterium-positive colorectal cancers reported in prior sequencing-based studies, lending external validity to the deep learning predictions.

Perhaps the most striking technical contribution came from Grad-CAM analysis, a visualization method that highlights which image regions drive a model’s predictions. The analysis revealed that microbial-associated signals were mainly concentrated in tumor gland lumens and tumor-stroma interface regions, suggesting distinct histomorphological contexts associated with microbial colonization. Multiple-instance learning attention heatmaps and representative pathological review further indicated that high-attention regions frequently involved tumor-stroma interfaces, inflammatory cell-rich areas, mucinous components, and luminal debris or necrosis-like material, painting a recurrent morphological picture of where Fusobacterium-related labels reside in the tissue.

The implications extend well beyond colorectal cancer. The findings demonstrate that routine histopathology contains latent spatial signals associated with intratumoral microbial ecology, highlighting the potential of foundation model-based computational pathology for microbiome characterization and biomarker discovery. Because H&E slides are produced universally and archived in hospitals worldwide, the approach could in principle be applied retrospectively to millions of existing specimens, enabling large-scale studies of tumor-microbe interactions without new tissue collection or sequencing costs.

For clinical practice, the ability to predict intratumoral Fusobacterium from digital slides could support risk stratification and treatment planning. Fusobacterium-positive colorectal cancers have been linked in prior research to distinct molecular features and potentially different responses to therapy, and a slide-based proxy for microbial status could complement existing biomarkers. The authors note that the models were trained and validated across four independent cohorts, providing a degree of generalizability that single-center studies often lack.

The study complies with the Declaration of Helsinki, with ethics approval from Zhejiang University School of Public Health, the Ningbo Clinical Pathology Diagnosis Center, and the First Affiliated Hospital of China Medical University. Written informed consent was obtained from all participants, and analyses involving TCGA and CPTAC datasets used publicly available, de-identified data. The work was supported by the National Natural Science Foundation of China, the National Key R&D Program of China, and the Healthy Zhejiang One Million People Cohort. As foundation models continue to mature, the ability to read microbial ecology from stained tissue may become a standard tool in the computational pathologist’s arsenal, transforming what was once invisible into actionable insight.

Subject of Research: Deep learning prediction of intratumoral Fusobacterium from digital colorectal cancer H&E slides

Article Title: Deep learning prediction of intratumoral Fusobacterium from digital colorectal cancer H&E slides

Article References: Xiang, H., Hu, L., Xu, J., Cai, H., Lu, K., You, Z., Shang, B., Huang, Q., Su, G., Zhang, H., Dai, W., Chang, X., Zhang, Z., Sun, L., & Tu, H. (2026). Deep learning prediction of intratumoral Fusobacterium from digital colorectal cancer H&E slides. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02696-3

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02696-3

Keywords: Deep, learning, prediction, intratumoral, Fusobacterium, digital, colorectal, cancer, slides, scientific research

The research, led by Guangchuan Deng of The Second Affiliated Hospital of Chongqing Medical University together with collaborators at Shandong Cancer Hospital and Institute, The Second Qilu Hospital of Shandong University, Qilu Hospital of Shandong University, and Neusoft Medical Systems, tackles the question of intratumoral and peritumoral heterogeneity. Heterogeneity, in the imaging sense, means that a tumor is not a uniform blob. Different voxels within the enhancing mass show different signal intensities on contrast-enhanced T1-weighted imaging, T2-weighted imaging, and diffusion-weighted imaging, reflecting differences in cell density, necrosis, edema, and microvascular architecture. Around the tumor, in the peritumoral zone, similar variation exists in the brain parenchyma that appears normal or merely edematous on routine reads but may harbor infiltrating tumor cells and treatment-resistant clones.

The team’s technical approach is what imaging scientists call habitat-based radiomics. Rather than treating the entire tumor as one region and extracting average features, the researchers segmented the index brain metastasis along with a 6-millimeter peritumoral shell, then performed voxel-wise habitat clustering. This means every voxel in the volume was assigned to one of several clusters, or habitats, based on its multiparametric MRI signature. Each habitat represents a distinct microenvironment, for example a region of high cellularity with restricted diffusion, or a region of necrosis with prolonged T2 relaxation. Radiomic features, quantitative descriptors of texture, shape, and intensity distribution, were then extracted separately within each habitat, so that the final signature captured how the tumor’s internal microenvironments were organized and how abundant each microenvironment was.

From these habitat-specific features, the investigators constructed three imaging signatures: an intratumoral heterogeneity (ITH) signature built from the enhancing mass alone, a peritumoral heterogeneity (PTH) signature built from the 6-millimeter shell, and a combined intratumoral-peritumoral habitat (PITH) signature integrating both. Feature selection reduced thousands of candidate descriptors to a compact set, and support vector machine classifiers were trained to predict the binary outcome: intracranial progression within 20 months of starting first-line third-generation EGFR-TKI monotherapy. Clinical variables were modeled separately using univariable and multivariable logistic regression, and the imaging and clinical models were compared head to head.

The study was retrospective and multicenter, drawing on patients diagnosed and treated at Shandong Cancer Hospital and Institute, The Second Qilu Hospital of Shandong University, and Qilu Hospital of Shandong University. A total of 141 patients with EGFR-mutant lung adenocarcinoma and de novo, treatment-naive brain metastases were analyzed, 90 in an internal training cohort and 51 in an external testing cohort. Of the 141 patients, 75 experienced intracranial progressive disease within the 20-month window and 66 did not. The external testing cohort is a critical design element, because it tests whether the model generalizes to patients from a different hospital rather than merely fitting the quirks of one dataset.

The results show a clear hierarchy of predictive power. The clinical model, relying on variables such as EGFR mutation subtype and tumor size, achieved an area under the receiver operating characteristic curve (AUC) of 0.662 in the testing cohort. The ITH model reached 0.715, the PTH model 0.724, and the combined PITH model 0.758. When imaging heterogeneity features were integrated with the independent clinical predictors, the final combined model delivered an AUC of 0.795, with 76.5 percent accuracy, 70.4 percent sensitivity, and 83.3 percent specificity. In plain terms, the model correctly identified roughly four out of five patients, and it was particularly good at reassuring the majority who would remain progression-free.

Two clinical variables emerged as independent predictors in multivariable logistic regression. EGFR mutation subtype carried an odds ratio of 2.390 (95 percent confidence interval 1.187 to 4.811, P = 0.015), meaning that certain mutation subtypes, such as exon 21 L858R substitutions compared with exon 19 deletions, roughly doubled the odds of early intracranial progression. The maximum diameter of the index brain metastasis carried an odds ratio of 2.048 (95 percent confidence interval 1.204 to 3.483, P = 0.008), so larger lesions predicted higher risk. These findings align with known biology, since EGFR mutation subtype is one of the strongest determinants of TKI benefit, and tumor burden influences the likelihood that resistant subclones already exist at treatment start.

Beyond the binary classification, the combined model stratified patients into risk groups with meaningfully different survival trajectories. Kaplan-Meier analysis showed that patients classified as high risk by the combined model had significantly shorter intracranial progression-free survival (iPFS) than those classified as low risk, with a log-rank P value of 0.019. Decision curve analysis and calibration plots further supported the model’s clinical utility, indicating that the risk scores were well matched to observed outcomes across the probability range rather than systematically overconfident or underconfident.

Why does the peritumoral zone matter so much? The answer lies in the biology of brain metastasis. Tumor cells do not respect the enhancing border seen on contrast-enhanced T1 imaging. They invade along perivascular spaces and white matter tracts into the surrounding parenchyma, where they interact with neurons, astrocytes, and microglia. This peritumoral niche is thought to be a reservoir of persister cells and a source of local recurrence after stereotactic radiosurgery. By quantifying heterogeneity in this 6-millimeter shell, the PTH signature effectively reads the tumor-brain interface, an area that conventional radiology reports describe only qualitatively as edema or mass effect. The fact that PTH alone outperformed ITH alone (0.724 versus 0.715) and that combining both improved performance further suggests that the tumor and its microenvironment carry complementary prognostic information.

The implications for clinical practice are substantial. If validated prospectively, a habitat-based MRI framework could be computed from scans that patients already receive before starting osimertinib or a similar third-generation TKI, adding no cost, no radiation, and no biopsy. Patients flagged as high risk could be considered for intensified surveillance MRI, enrollment in trials of combination strategies such as TKI plus stereotactic radiosurgery or TKI plus immunotherapy, or earlier escalation of systemic therapy. Patients flagged as low risk could avoid unnecessary interventions and maintain quality of life on monotherapy. The study has limitations inherent to its retrospective design, including the modest external cohort of 51 patients and the 20-month progression window, and the authors note that the published version is an accepted manuscript shared early for faster access, subject to further editorial edits. Still, the work represents a meaningful step toward personalized, imaging-guided management of brain metastases, turning the pretreatment MRI from a static snapshot of tumor burden into a quantitative map of the microenvironments that will decide whether the drug wins or the tumor does.

Subject of Research: MRI-based intratumoral and peritumoral heterogeneity for predicting intracranial progression in EGFR-mutant lung adenocarcinoma patients with treatment-naive brain metastases receiving first-line third-generation EGFR-TKIs

Article Title: MRI-based intratumoral and peritumoral heterogeneity for predicting intracranial progression in EGFR-mutant lung adenocarcinoma patients with treatment-naive brain metastases receiving first-line third-generation EGFR-TKIs

Article References: Deng, G., Li, H., Dai, Z., Fan, J., Song, X., Pang, G., Li, Z., & Zhao, F. (2026). MRI-based intratumoral and peritumoral heterogeneity for predicting intracranial progression in EGFR-mutant lung adenocarcinoma patients with treatment-naive brain metastases receiving first-line third-generation EGFR-TKIs. Journal of Translational Medicine. https://doi.org/10.1186/s12967-026-09109-7

Image Credits: AI Generated

DOI: 10.1186/s12967-026-09109-7

Keywords: MRI-based, intratumoral, peritumoral, heterogeneity, predicting, intracranial, progression, EGFR-mutant, lung, adenocarcinoma, patients, treatment-naive

The researchers set out to test what happens when so-called bystander T cells, which recognize viral rather than tumor targets, are reactivated within the tumor microenvironment. Using a mouse model of melanoma, they transferred memory CD8+ T cells specific to an irrelevant viral antigen and then delivered the matching viral peptide directly into the tumor, alongside anti-PD-L1 checkpoint blockade. The result was striking: tumors were cleared even though the activated T cells could not, in any conventional sense, see the cancer. In experiments where mice lacked any tumor-specific TCRαβ+ T cells whatsoever, tumor elimination still proceeded, demonstrating that classical recognition-dependent killing was not required.

The mechanism, the authors show, is paracrine. Activated T cells flood the tumor microenvironment with effector cytokines, chiefly interferon-γ and tumor necrosis factor, which act on surrounding cells rather than on the tumor directly through T cell receptors. These signals recruit waves of innate immune cells, including Ly6c-high monocytes and neutrophils, and induce the enzyme iNOS in myeloid cells, driving local production of nitric oxide. The combination of interferon-γ, TNF and nitric oxide proved lethal to tumor cells, triggering caspase-dependent death pathways that recapitulated melanoma clearance observed in living animals.

Technical detail from the single-cell work reinforces the picture. Using CITE-seq, the team profiled tens of thousands of cells from the tumor microenvironment before and after treatment, mapping how activated virus-specific T cells reshape the entire cellular ecosystem. The adhesion molecule VCAM-1 emerged as essential, apparently by anchoring and coordinating the influx of myeloid cells, and depletion experiments confirmed that innate leukocytes, not just the cytokines themselves, are indispensable to the killing program. Notably, natural killer cells were not required, pointing instead to recruited monocytes and neutrophils as the critical innate effectors.

The tumor cell death observed was not a quiet, orderly apoptosis alone. The researchers found evidence of panoptotic pathways, the interconnected family of inflammatory death programs that includes pyroptosis, necroptosis and apoptosis, converging on caspase-dependent execution. This matters because inflammatory cell death can further amplify immune recruitment, potentially converting a localized activation event into a self-reinforcing tumoricidal cascade. The synergy of interferon-γ and TNF in driving this form of death echoes findings from other recent studies linking cytokine cooperation to inflammatory tumor cell killing.

Perhaps the most clinically provocative result came from translational analysis. The gene expression signatures associated with this bystander-activation response in mice were predictive of survival among human patients with melanoma, suggesting that the same biology operates, or at least leaves traces, in human disease. In vitro, the cytokine-and-nitric-oxide cocktail killed human melanoma cell lines, including A375 and SK-MEL-2 cells, through the same caspase-dependent mechanism, bolstering the case that the mouse findings are not an artifact of the model system.

The study builds on a growing body of work showing that tumors are infiltrated by large numbers of T cells that have nothing to do with the cancer. Earlier research established that virus-specific memory T cells populate tumors and can be repurposed for immunotherapy, and that bystander CD8+ T cells are abundant and phenotypically distinct in human tumor infiltrates. Strategies have already been proposed to exploit this, from oncolytic viruses carrying tumor-irrelevant epitopes to lipid nanoparticle RNA approaches that leverage SARS-CoV-2-specific immunity for cancer treatment. The new work provides the mechanistic foundation for why such approaches might succeed: productive activation, not antigen specificity, is the trigger.

The implications for immunotherapy design are considerable. Current approaches such as personalized neoantigen vaccines, adoptive T cell transfer and checkpoint blockade all aim, in different ways, to generate or rescue tumor-specific T cell responses, an endeavor that is expensive, slow and often thwarted by tumor immune evasion. If intratumoral T cell activation alone can suffice, then simpler strategies become conceivable: delivering activation signals directly into tumors to wake up whatever unexhausted bystander T cells happen to be present, and letting the paracrine storm of cytokines, nitric oxide and recruited innate cells do the killing. Intratumoral CpG oligonucleotides and STING agonists, which already show clinical promise, may partly work through exactly this kind of bystander mechanism.

Cautions remain. The experiments were performed largely in mouse melanoma models, and the requirement for VCAM-1, myeloid cells and specific cytokine combinations may vary across tumor types and tissue contexts. The balance between tumoricidal inflammation and harmful tissue damage will also need careful calibration, particularly given the known role of interferon-γ and TNF synergy in cytokine shock syndromes. Still, the conceptual shift is profound: the tumor microenvironment may be less a fortress requiring a precisely targeted key and more a tinderbox awaiting a spark, provided enough activated T cells are standing by inside it.

For a field that has spent decades chasing tumor antigens, the message of this study is liberating and unsettling in equal measure. Immunotherapy, the authors conclude, may not need to induce or rescue cancer-specific responses at all. Triggering productive T cell activation within tumors can be sufficient, and the immune system’s own inflammatory machinery will handle the rest.

Subject of Research: Paracrine tumor killing by activated bystander T cells independent of tumor antigen recognition

Article Title: Intratumoral T cell activation kills tumors regardless of T cell specificity

Article References: Ghirardelli Smith, O. C., Dao, T. T., Gavil, N. V., O’Flanagan, S. D., Rubin, A. J., Nguyen, S., Watowich, M. B., Liu, N., Weyu, E., Quarnstrom, C. F., Soerens, A. G., Joag, V., Rosato, P. C., Krummel, M. F., Geller, M. A., Miller, J. S., Giubellino, A., Vezys, V., Shalek, A. K., & Masopust, D. (2026). Intratumoral T cell activation kills tumors regardless of T cell specificity. Nature Immunology. https://doi.org/10.1038/s41590-026-02642-z

Image Credits: AI Generated

DOI: 10.1038/s41590-026-02642-z

Keywords: T cell activation, bystander T cells, tumor immunology, interferon-gamma, nitric oxide, melanoma, checkpoint blockade, innate immune cells, panoptosis, immunotherapy, Intratumoral, cell