The new probe, named rsmiRFPnano, belongs to a family of photoreceptors called cyanobacteriochromes, or CBCRs, found in cyanobacteria. These proteins normally require only a single GAF domain to bind their light-sensing chromophore, making them far more compact than the bacteriophytochromes that previously dominated near-infrared photoacoustic imaging. Bacteriophytochrome-based probes, even in their truncated forms, exceed 55 kilodaltons and demand the full three-domain PAS-GAF-PHY assembly to bind biliverdin IXalpha, the heme-breakdown product that mammalian cells supply naturally. That bulk has constrained how these probes can be fused to other proteins or incorporated into reporters. The research team, led by Vladislav Verkhusha of the University of Helsinki and Albert Einstein College of Medicine together with Junjie Yao of Duke University, reasoned that a verdin-specific CBCR lineage, whose members naturally accept biliverdin-like chromophores and absorb at red-shifted wavelengths, could be coaxed into a mammalian-compatible photoswitch a third of the size.

The starting material was the third GAF domain of a CBCR called JSC1_58120 from the cyanobacterium Leptolyngbya sp. JSC1. In its native form, the domain showed promising photochemistry in bacteria, flipping reversibly between a near-infrared-absorbing ground state with an absorbance peak at 715 nanometers and a far-red-absoring photoproduct peaking at 650 nanometers. But in mammalian cells it performed poorly, a fatal flaw for any genetically encoded in vivo probe. Mutations in the biliverdin-binding pocket tend to disrupt the delicate photocycle, so the team took a different tack: iterative saturation mutagenesis aimed at structural regions outside the pocket, targeting paired residues in the N- and C-terminal helices that face each other in the folded structure. Four sequential rounds of evolution, screened first in bacteria and then validated in transfected mammalian cells, yielded rsmiRFPnano, a 163-residue protein carrying seven substitutions that collectively tripled mammalian-cell fluorescence relative to the parent while preserving, and even slightly enhancing, the light-absorption machinery.

The numbers behind the engineered protein are striking. rsmiRFPnano exhibits molar extinction coefficients of 85,000 and 88,000 per molar per centimeter in its two photoswitchable states, among the highest values reported for genetically encoded near-infrared probes. Its pKa values of roughly 4.2 in both states mean spectral properties remain stable across the physiological pH range, and native gel electrophoresis confirmed it behaves as a monomer with no detectable aggregation. In HEK293T cells, flow cytometry showed a 3.5-fold fluorescence increase over the parental protein, and the probe accumulated with kinetics comparable to enhanced green fluorescent protein, reaching half-maximal signal in about 27 hours with an intracellular half-life of roughly 18 hours. These are the characteristics of a probe built to survive and function inside living cells, not just in a cuvette.

The real magic, however, lies in how the probe exploits its photoswitching. Under alternating illumination with 635 and 725 nanometer light, rsmiRFPnano cycles between an ON state that absorbs strongly in the near-infrared and an OFF state that absorbs in the far-red. By acquiring photoacoustic images in both states and subtracting the OFF image from the ON image, the technique known as differential photoacoustic tomography isolates only the signals that actually switch, suppressing the static background from hemoglobin and water. In phantom experiments, purified rsmiRFPnano at 17 micromolar concentration embedded in whole blood produced contrast-to-noise ratios essentially identical to those measured in buffer, demonstrating that the differential contrast is virtually independent of the surrounding optical environment. The team estimated a noise-equivalent detection sensitivity of about 0.33 micromolar against a whole-blood background, a level of performance that transforms what is detectable inside a living animal.

To prove the probe could report on biology rather than just its own presence, the researchers built transcriptional reporters for two master regulators of inflammation, NF-kappaB and AP-1. Each construct placed tandem response elements upstream of a minimal promoter driving rsmiRFPnano expression, so that pathway activation directly translated into probe production. In cultured cells, the reporters showed minimal basal leakiness and strong, dose-dependent induction upon stimulation with tumor necrosis factor or phorbol ester. The team then delivered the reporters to mouse livers via hydrodynamic transfection and induced systemic acute inflammation with lipopolysaccharide. Differential photoacoustic imaging tracked the molecular response longitudinally over five days, revealing that the NF-kappaB reporter peaked rapidly at 12 hours post-treatment while the AP-1 reporter rose more gradually to a maximum at 48 hours, a temporal distinction consistent with known signaling dynamics in liver injury. Control animals receiving saline instead of the inflammatory stimulus showed peak signals fivefold to sevenfold weaker, confirming that the readouts reflected genuine inflammation rather than transfection stress.

The imaging platform went further by integrating super-resolution ultrasound localization microscopy on the same system. Microbubble tracking resolved the hepatic microvasculature in three dimensions and co-registered it with the molecular photoacoustic maps, revealing significant decreases in liver blood-flow speed and blood volume at 48 hours after lipopolysaccharide treatment. An integrated inflammation index, combining the spatial extent and intensity of reporter activation, rose in parallel with the hemodynamic decline. To validate the in vivo findings independently, the collaborators performed multiplexed spatial proteomics using CODEX imaging with a 50-antibody panel on liver sections collected at baseline and at 24 and 72 hours. Single-cell segmentation and clustering showed immune cell infiltration peaking at 24 hours, with macrophages and CD45-positive cells surging early, while inflamed hepatocytes expressing the MHC class I molecule H-2Db dominated by 72 hours. The cellular choreography mapped ex vivo matched the temporal patterns the living-imaging approach had captured noninvasively.

Perhaps the most forward-looking application came from fusing rsmiRFPnano into nanobody scaffolds. Nanobodies, the 15-kilodalton single-domain antibody fragments derived from camelid heavy chains, bind intracellular antigens with high specificity, but unbound nanobodies accumulate and generate background that obscures the signal of interest. The team’s solution, called rsNIR-Fbs, inserts rsmiRFPnano into a loop in the nanobody’s third framework region, distal from the antigen-binding surface. The design makes the fusion protein unstable unless it binds its target: unbound probes are rapidly degraded by the proteasome, while antigen-bound probes are stabilized and generate photoacoustic signal. AlphaFold3 modeling confirmed the insertion preserves the canonical VHH beta-sandwich fold and the antigen-binding loops. The researchers built rsNIR-Fbs against enhanced green fluorescent protein, HIV-1 gp41, and the SARS-CoV-2 spike receptor-binding domain, and in each case fluorescence appeared only when the cognate antigen was coexpressed.

In vivo proof of concept followed in mice whose livers expressed green fluorescent protein via hydrodynamic transfection while the anti-GFP rsNIR-Fb was delivered systemically through an adeno-associated virus. Differential photoacoustic imaging combined with ultrasound localization microscopy revealed strong, spatially confined signals in the liver, co-registered with the microvasculature, and nothing in control organs. When the antigen plasmid was omitted, virtually no differential signal appeared, demonstrating that the photoacoustic output was governed by antigen presence rather than probe delivery or nonspecific accumulation. The result establishes a new class of conditional probes in which the imaging signal itself is gated by molecular recognition inside cells, within organs whose dense blood supply would ordinarily defeat molecular specificity.

The implications extend well beyond inflammation models. Because rsmiRFPnano is threefold smaller than the most compact bacteriophytochrome-derived photoacoustic probes, it offers a flexible scaffold for reporters of virtually any transcriptional pathway, from oncogenic signaling to neuronal plasticity, and the rsNIR-Fb strategy opens photoacoustic immuno-imaging to endogenous intracellular targets including signaling intermediates, viral antigens, and cancer-associated proteins. The authors note that further improvements in biliverdin binding and switching contrast should push detection limits lower still. What the study ultimately delivers is a convergence of protein engineering, acoustic detection, and spatial omics validation: a genetically encodable molecular flashlight, small enough to fit inside a nanobody, bright enough to be heard through centimeters of blood-filled tissue, and switchable enough to be distinguished from everything the body does on its own. For a field that has long struggled to make molecular signals audible above the noise of biology, that is a genuinely resonant advance.

Subject of Research: Engineering of a compact cyanobacteriochrome-derived photoswitchable protein for photoacoustic molecular imaging of inflammation and intracellular antigens

Article Title: Near-infrared photochromic miniprotein enables photoacoustic molecular imaging of acute inflammation and intracellular antigens

Article References: Oliinyk, O. S., Xu, Y., Luo, J., Tokarchuk, K. O., Manoilov, K. Y., van de Klashorst, J., Eyring, M. P., Wang, M.-E., Chen, M., Song, P., Hickey, J. W., Yao, J., & Verkhusha, V. V. (2026). Near-infrared photochromic miniprotein enables photoacoustic molecular imaging of acute inflammation and intracellular antigens. Nature Methods. https://doi.org/10.1038/s41592-026-03255-y

Image Credits: AI Generated

DOI: 10.1038/s41592-026-03255-y

Keywords: photoacoustic tomography, rsmiRFPnano, cyanobacteriochrome, near-infrared imaging, NF-kappaB, AP-1, inflammation, nanobody, biliverdin, molecular imaging, protein engineering, spatial omics

The core idea is deceptively simple. Every tissue image produced by modern spatial proteomics, such as imaging mass cytometry, records the position of each cell and the abundance of dozens of protein markers within it. SpaCEy transforms each sample into a spatial graph using Delaunay triangulation, a geometric method that connects every cell to its nearest spatial neighbours. In this graph, each node is a cell and each node’s features are the measured levels of the protein markers at that location. Critically, the model receives no predefined cell-type labels or anatomical compartment annotations as inputs; it must learn what matters directly from raw molecular measurements and tissue geometry. Clinical annotations are used only afterwards, to help interpret what the model has found.

Once trained, the network compresses each tissue into a compact embedding, a vector that summarises both the molecular states and the spatial relationships of its cells. For survival prediction the model is trained with a censoring-aware loss adapted from the DeepSurv framework, which handles the fact that some patients are still alive at the end of a study and their true survival times are unknown. But prediction alone is where most existing graph-based methods stop. SpaCEy adds an integrated explainer, adapted from the GNNExplainer algorithm, that learns masks over the edges and features of each graph, highlighting the connected subgraphs of cells that contribute most to the prediction. Node importance is computed by aggregating edge importance values within each node’s local neighbourhood, which yields spatially contiguous, compact regions of tissue that the model deems decisive.

Applied to a lung adenocarcinoma cohort of 416 patients profiled by imaging mass cytometry, SpaCEy separated patients who later progressed from those who did not. The regions it flagged told a coherent biological story. Tumours from progressors showed important nodes enriched for FOXP3, a canonical regulatory T-cell marker, and CD163, part of an immunosuppressive macrophage axis previously linked to aggressive lung cancer. Non-progressing tumours instead showed areas rich in helper T cells, cytotoxic T cells and B cells, along with higher expression of HLA-DR and CD4, hallmarks of an active anti-tumour immune response. The model also uncovered a previously undescribed enrichment of CD163-negative macrophages in progression-associated regions, illustrating how its explanations can generate new hypotheses rather than merely confirming known biology.

In head-to-head benchmarking against two state-of-the-art spatially aware methods, SpaCEy came out ahead on the lung cancer progression task. Using fixed, stratified cross-validation folds, it achieved a mean accuracy of 0.68 compared with 0.61 for the method of Ali and colleagues and 0.55 for SPACE-GM, with corresponding advantages in F1-score and area under the ROC curve. Its performance was also the most stable across folds, with peak values reaching 0.77 accuracy and 0.78 AUC in individual runs.

The breast cancer results are where the approach demonstrates its translational ambition. On a dataset of 720 imaging mass cytometry samples from Basel and Zurich, SpaCEy learned embeddings whose clusters separated patients by overall survival with a global log-rank p-value of 0.00253, and it did so independently of the established receptor-based subtypes. One cluster was dominated by low-survival patients and enriched for triple-negative tumours, while another was skewed toward longer survivors with more hormone receptor-positive disease. Within the model-selected important regions, lower-survival samples showed locally elevated abundance of Slug, a driver of epithelial-to-mesenchymal transition, Ki-67, a proliferation marker, c-Myc, an oncogene, and CD3, which in the stromal compartment pointed to an immune-excluded phenotype. About 55 percent of the important regions fell within tumour areas and roughly 29 percent in stroma, showing that outcome-relevant signals span multiple tissue compartments.

The same framework was trained on a second independent breast cancer cohort from the METABRIC imaging study, comprising 460 samples from 405 patients. Once again, embedding clusters separated patients by survival, and the flagged markers overlapped substantially with the first cohort: fibronectin, beta-catenin and vimentin characterised lower-survival regions, while progesterone receptor and HER2-associated signals marked better outcomes. The authors were careful to distinguish generalisable signals from dataset-specific ones, noting, for example, that fibronectin enrichment was not restricted to hormone receptor-positive, HER2-negative patients and should not be read as a standalone universal biomarker. Notably, within the clinically ambiguous HR-positive, HER2-negative subtype, SpaCEy still stratified patients into groups with significantly different survival, addressing one of precision oncology’s most common blind spots.

The quantitative case for spatial information is striking. Against 24 non-spatial baseline configurations built from pseudobulk protein summaries and cell-type compositions, including Fast Survival SVM, Random Survival Forests, Gradient Boosting Survival Analysis and a Cox proportional hazards model, SpaCEy outperformed every variant. On the Zurich and Basel cohort it reached a mean concordance index of 0.720, a 16.5 percent relative improvement over the best baseline; on METABRIC it reached 0.688, a 20.7 percent improvement. A non-spatial neural network trained on the same data failed to produce significant survival separation between its embedding clusters in either cohort, underscoring that the gains come specifically from tissue geometry. In a separate CODEX colorectal cancer cohort, SpaCEy achieved a held-out AUPRC of 0.846 and retained an AUPRC of 0.801 when trained on one tissue microarray batch and tested on another, showing reasonable robustness to technical variation.

The team also stress-tested the method itself. Ablation experiments showed that removing explainer-identified important nodes degraded predictions far more than removing matched unimportant ones. A synthetic tumour-stroma benchmark with known ground truth confirmed that the explainer selectively highlights intended spatial motifs, assigning much higher importance to intratumoural CD8 cells than to identical cells scattered in stroma. Scalability tests indicated the model can process graphs up to a million cells per sample through tiling, with bounded GPU memory. The authors acknowledge limitations: the approach currently captures only protein-marker abundances, explanations inherit the ceiling of predictive performance, and importance scores identify influential features rather than proven causes.

Even with those caveats, SpaCEy signals a shift in how machine learning might enter pathology. Rather than a black box that assigns risk scores, it delivers maps: here are the cells, here are the proteins, here is the patch of tumour that tipped the prediction. That transparency matters for clinical trust, for regulatory scrutiny, and for biologists hunting new mechanisms in the tumour microenvironment. As whole-slide spatial datasets linked to outcomes accumulate, tools that connect tissue architecture to patient fate, and explain themselves while doing so, may become as familiar in the oncology lab as the microscope itself.

Subject of Research: Explainable graph neural networks linking spatial proteomics tissue patterns to clinical outcomes in cancer

Article Title: SpaCEy links spatial tissue patterns to clinical outcomes using explainable graph neural networks

Article References: Rifaioglu, A. S., Ervin, E. H., Sarigun, A., Germen, D., Bodenmiller, B., Tanevski, J., & Saez-Rodriguez, J. (2026). SpaCEy links spatial tissue patterns to clinical outcomes using explainable graph neural networks. Nature Communications, 17(1), Article 10232. https://doi.org/10.1038/s41467-026-77924-z

Image Credits: AI Generated

DOI: 10.1038/s41467-026-77924-z

Keywords: spatial omics, graph neural networks, explainable AI, cancer, tumour microenvironment, survival prediction, spatial proteomics, precision oncology, imaging mass cytometry, patient stratification, machine learning, Nature Communications

The biological heart of the discovery lies in a long-recognized but difficult-to-study distinction between two tumor cell states. Neuroblastoma cells can exist in an adrenergic state, which reflects the tumor’s origin in sympathetic nervous tissue and is generally associated with lower-risk disease, or in a mesenchymal state, which resembles connective-tissue-like cells and has been linked to high-risk disease, resistance to therapy and relapse. The problem, and it was a formidable one, is that a child’s body contains abundant healthy mesenchymal cells. Standard laboratory techniques that profile the cells within a tumor cannot easily distinguish a malignant mesenchymal cancer cell from a normal mesenchymal stromal cell sitting alongside it in the same tissue. That ambiguity has hampered efforts to quantify these dangerous cells, to study their biology and to determine how much they actually contribute to a patient’s prognosis.

To break through that barrier, the team analyzed 54 tumors from 50 patients using an arsenal of complementary technologies, including single-cell RNA sequencing, which reads the gene activity of individual cells, and spatial omics methods, which map where specific cells and molecules sit within intact tissue. But the decisive methodological move involved patient-derived xenografts, tumor cells taken from children and grown in mice, generated as part of the St. Jude Childhood Solid Tumor Network. Because healthy mesenchymal cells from the patient’s tumor do not survive and grow in xenografts, any mesenchymal-like cells that persisted in the mouse models were, by definition, malignant. By comparing the original patient tumors with the xenografts, the researchers could isolate a clean gene expression profile belonging specifically to the cancer-related mesenchymal population, free of the contaminating signal from normal stromal cells.

That purified profile became the basis for a new gene signature, a defined set of marker genes whose combined activity identifies malignant mesenchymal cells in a tumor sample. Co-corresponding author Michael Dyer, PhD, chair of the St. Jude Department of Developmental Neurobiology, emphasized the clinical significance of the advance, noting that with a reliable gene expression signature for mesenchymal cells in neuroblastoma, researchers can begin to dig into why outcomes vary so dramatically between different children with the same cancer. The signature effectively converts a previously invisible cell population into a measurable variable, one that can be scored in existing patient datasets and correlated with how each child fared.

The validation process was deliberately rigorous. The researchers first tested the signature against an independent database of RNA sequencing data from patient tumors, a dataset the team had not used to build the signature. The result was striking: the new mesenchymal signal significantly improved predictions of patient outcomes, whereas a pre-existing signature derived from decades-old laboratory cell lines did not. That comparison matters because much of what the field believed about neuroblastoma cell states had been built on long-established cell lines that, as the new work makes clear, failed to capture the malignant mesenchymal population as it actually exists in children. The old tools, in other words, were looking in the wrong place.

To confirm the signature across every platform available, the scientists created organoids, three-dimensional laboratory-grown assemblies of cancer cells, along with new cell lines, in addition to their xenografts. They then applied spatial transcriptomics, spatial proteomics, electron microscopy and chromatin profiling, techniques that probe, respectively, gene activity in place, protein abundance in place, ultrastructural anatomy and the regulatory landscape of the genome. In every system and with every method, the signature cleanly separated the two neuroblastoma cell states. The separation went deeper than gene expression: the adrenergic and mesenchymal cells displayed distinct cellular shapes, different internal spatial organization and different behaviors, indicating that the two states represent genuinely different cell types within the tumor rather than subtle shifts in a single program.

First author Anand Patel, MD, PhD, of the St. Jude Department of Oncology, described the robustness of the marker as endlessly fascinating, observing that regardless of which test the team applied, the signature held up, and that these markers will be powerful tools for understanding neuroblastoma better. That kind of cross-platform consistency is rare in cancer genomics, where signatures identified by one technology often dissolve when measured another way. The convergence of evidence from sequencing, imaging, proteomics and electron microscopy gives the field unusually strong confidence that the malignant mesenchymal population is real, reproducible and clinically meaningful.

With the marker validated, the group went a step further and built the research infrastructure the field has lacked. They generated additional in vitro models and new cell lines from xenografts derived from patient samples containing varying proportions of mesenchymal to adrenergic cells, creating a spectrum of models that mirror the diversity seen in actual patients. All of these resources are available to other researchers upon request through the Childhood Solid Tumor Network, a St. Jude initiative that distributes pediatric cancer models freely to the scientific community. Co-corresponding author Elizabeth Stewart, MD, of the St. Jude Department of Oncology, called the new models a launching point for neuroblastoma research, saying that with the dataset as a foundation, the field is poised to explore the biology of this pediatric cancer more deeply and, ultimately, to move toward improving outcomes for patients.

The implications extend well beyond a single tumor type. The study demonstrates a general strategy for a problem that plagues pediatric oncology broadly: when a tumor’s dangerous cells resemble normal cells of the host tissue, single-cell analysis alone can mislead. By exploiting the selective growth behavior of xenografts to separate malignant from benign populations, then locking down the finding with spatially resolved, multi-omic validation, the St. Jude-led team has provided a template other researchers can adapt. For neuroblastoma specifically, the immediate consequences are concrete. Clinicians and researchers can now measure the malignant mesenchymal fraction of a tumor accurately, incorporate it into risk assessment, investigate why those cells resist therapy and screen candidate drugs against models that actually contain them. The study was supported by the National Cancer Institute, the National Institutes of Health, the Howard Hughes Medical Institute, the Tully Family Foundation, the Peterson Foundation, Hyundai Hope on Wheels, the Damon Runyon Cancer Foundation, Alex’s Lemonade Stand Foundation and the American Lebanese Syrian Associated Charities. What began as an effort to clean up a noisy dataset has ended by giving the neuroblastoma field something it has needed for decades: a clear view of the enemy within.

Subject of Research: Identification of a malignant mesenchymal cell state and gene signature associated with poor prognosis in pediatric neuroblastoma

Article Title: Scientists identify the neuroblastoma cells that signal poor outcomes

Article References: Scientists identify the neuroblastoma cells that signal poor outcomes. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: neuroblastoma, pediatric cancer, gene signature, single-cell RNA sequencing, spatial omics, mesenchymal cell state, adrenergic cell state, patient-derived xenografts, treatment resistance, cancer prognosis, organoids, St. Jude Children's Research Hospital

The starting point of the review is the physical assault that radiotherapy delivers. Ionizing radiation deposits energy that breaks DNA directly and, indirectly, floods cells with reactive oxygen species that damage proteins, lipids, and genomes alike. The dose, the fractionation schedule, and the type of radiation used all determine the character of this pressure. Cells that survive the initial insult do so by activating their DNA damage response machinery, a coordinated network of sensors, transducers, and effectors that pauses the cell cycle, repairs double-strand breaks through pathways such as non-homologous end joining and homologous recombination, and, when damage is overwhelming, triggers senescence or apoptosis. But the review emphasizes that DNA repair is only the first layer of a much deeper program of tolerance.

That deeper program is cellular plasticity. Rather than dying or fully recovering, a subset of surviving cells appears to enter altered, quasi-stable states that echo phenomena described in drug-treated tumors. The authors catalog several of these: persister-like states in which cells reversibly suspend proliferation; cancer stem cell-like states marked by molecules such as CD44 and aldehyde dehydrogenase, with enhanced self-renewal and repair capacity; senescence-like states, detectable in part by senescence-associated beta-galactosidase, in which cells remain metabolically alive but stop dividing and can secrete inflammatory signals; and hybrid epithelial-mesenchymal transition states that combine adhesive and migratory traits. Crucially, these states are not permanent identities. They are plastic, meaning the same residual cell may drift between them, and that fluidity makes them extraordinarily difficult to eliminate with any single targeted agent.

Alongside chromatin remodeling and transcriptional stress memory, which allow cells to carry a kind of biochemical recollection of the radiation they endured, these plastic states form the tumor-cell-intrinsic half of the radioresistance story. The review is careful to note that this intrinsic component complements rather than replaces classical radiosensitivity, which depends on factors such as oxygenation, cell-cycle phase, and repair capacity. What is new is the recognition that intrinsic tolerance alone cannot explain why recurrences so often emerge in predictable locations and time windows. For that, the authors argue, one must look outward, to the microenvironment that surrounds and stabilizes the surviving cells.

The concept of the tolerant niche is the conceptual heart of the paper. The authors describe several recurring microenvironmental configurations that act as sanctuaries for residual disease. Hypoxic niches are perhaps the best known: low oxygen reduces the radiation-induced free radical chemistry that makes oxygen so potent a radiosensitizer, and hypoxia simultaneously drives stem-like and invasive transcriptional programs. Perivascular niches, the zones around surviving blood vessels, shelter quiescent tumor cells and are rich in protective signaling. Niches dominated by cancer-associated fibroblasts and remodeled extracellular matrix provide both mechanical and biochemical support, secreting growth factors and stiffening tissue architecture in ways that dampen radiation lethality. Metabolic niches, shaped by nutrient scarcity and altered metabolite exchange, and myeloid niches, populated by tumor-associated macrophages, myeloid-derived suppressor cells, and regulatory T cells that suppress antitumor immunity, complete the picture. In each case, the niche does not merely coexist with tolerant cells; it actively maintains their state.

One of the review’s most useful contributions is its insistence that these mechanisms vary substantially among tumor types. The relative weight of hypoxia, stromal shielding, immune suppression, or stem-like plasticity differs between, for example, glioblastoma, pancreatic ductal adenocarcinoma, and lung cancer, and the authors caution against assuming that a strategy validated in one context will transfer to another. This heterogeneity has practical consequences for how the field measures residual disease. The authors assess the tools available and find them strikingly unequal in maturity. Functional imaging, including hypoxia-targeted PET tracers and advanced MRI sequences, and liquid biopsy based on circulating tumor DNA, offer selected clinical readouts that can already be deployed in trials to track residual disease and recurrence risk over time.

Spatial omics technologies, by contrast, which map gene expression and other molecular features across intact tissue sections, currently serve mainly as discovery and validation tools rather than clinical diagnostics. They are indispensable for identifying which cells occupy which niches and how those relationships change after irradiation, but sampling limitations, cost, and analytical complexity keep them out of routine practice. The review is candid about this gap: the field can describe tolerant niches in exquisite detail in resected specimens, yet it largely lacks the longitudinal, minimally invasive measurements needed to watch them form and dissolve in living patients. Bridging that gap, the authors suggest, will require combining serial ctDNA monitoring with imaging and, where feasible, repeat tissue sampling.

On the therapeutic side, the review outlines both evidence-based and investigational strategies for what the authors term niche-informed radiotherapy. Established approaches include hypoxia modification, dose escalation and altered fractionation, and combinations with DNA damage response inhibitors that disable the repair programs tolerant cells rely on. More speculative candidates target the niche itself: agents that reprogram cancer-associated fibroblasts, normalize tumor vasculature, modulate myeloid suppressor cells, or relieve immune suppression so that checkpoint blockade can act on residual disease. Senolytic strategies aimed at clearing senescence-like cells, and agents designed to push plastic persister states out of their protective equilibrium, also feature among the investigational options. The unifying principle is that candidate combinations should be validated against the specific niche dependencies of a given tumor before clinical adaptation, rather than applied indiscriminately.

The authors are equally clear about the obstacles standing between this framework and the clinic. Conceptually, tolerant niches are dynamic and heterogeneous, so static snapshots may mislead. Sampling remains a fundamental problem, since the most informative regions of a treated tumor are often inaccessible. Implementation faces the usual barriers of combination-therapy trials, toxicity management, and the absence of validated biomarkers that identify which patients carry which niche dependencies. None of these limitations, in the authors’ view, undermines the framework; they simply define the work that remains. The review, which was supported by numerous Chinese national and institutional research programs and published open access on 9 October 2026, is intended as a roadmap rather than a final verdict.

If the tolerant niche concept holds up under experimental and clinical scrutiny, its implications for oncology could be considerable. Recurrence after radiotherapy would no longer be viewed as a mysterious resurgence but as the predictable product of a documented sequence: radiation pressure, followed by the emergence of plastic residual states, followed by their stabilization within hypoxic, stromal, metabolic, and immune sanctuaries. Each step in that sequence offers an interception point, a moment when a rationally chosen combination therapy might extinguish the embers before they reignite. Turning that vision into practice will demand the longitudinal monitoring tools and niche-specific combination trials the review calls for, but it reframes one of radiotherapy’s oldest frustrations as a solvable, spatially defined problem in tumor ecology.

Subject of Research: Radioresistance and tumor recurrence through cellular plasticity and radiation-induced tolerant niches in the tumor microenvironment

Article Title: Radioresistance and tumor recurrence: cellular plasticity, tolerant niches, and therapeutic opportunities

Article References: Zhang, Y., Li, H., Lv, S., Peng, L., Meng, W., Guo, J., Lu, Y., Chen, W., Yang, W., & Peng, X. (2026). Radioresistance and tumor recurrence: cellular plasticity, tolerant niches, and therapeutic opportunities. Molecular Cancer. https://doi.org/10.1186/s12943-026-02806-3

Image Credits: AI Generated

DOI: 10.1186/s12943-026-02806-3

Keywords: radioresistance, radiotherapy, tumor recurrence, tolerant niches, cellular plasticity, DNA damage response, tumor microenvironment, cancer stem cells, hypoxia, residual disease, spatial omics, ctDNA

The research, led by Xinyu Song and colleagues working across institutions in China and the United States, tackles a problem that has long hampered precision oncology: most studies of drug resistance rely on samples taken from different patients at different stages, making it impossible to distinguish true temporal evolution from background variation between individuals. What makes this study remarkable is its longitudinal design. The team assembled a cohort of six patients with EGFR-mutant NSCLC who underwent serial re-biopsies at three critical time points: at baseline before any TKI treatment, after developing resistance to a first-generation EGFR-TKI, and after subsequently failing a third-generation EGFR-TKI. These consecutive samples from the same patients are extraordinarily rare, because repeated lung biopsies are invasive and many patients decline them, yet they are the only way to watch the same tumor evolve in real time within the same body.

To interrogate these precious samples, the researchers employed GeoMx Digital Spatial Profiling, a technology that bridges the gap between bulk sequencing, which averages signals across an entire tissue piece, and single-cell methods, which lose the architectural context of where each cell sits. With GeoMx, the team segmented each biopsy into tumor-enriched and stroma-enriched areas of interest, effectively separating the malignant cells themselves from the surrounding supportive tissue of fibroblasts, immune cells, and blood vessels. For each compartment, they quantified paired RNA and protein profiles, allowing them to ask not only which genes were being transcribed but also which proteins were actually being produced. This dual-layer approach matters because gene expression and protein abundance do not always correspond, and drugs typically target proteins rather than transcripts.

The first major finding was a confirmation that the spatial segmentation was biologically meaningful: tumor-enriched and stroma-enriched compartments proved consistently distinct from one another, both at the transcriptional and the protein level. This might sound like an expected result, but it validates the analytical framework and establishes that the technology reliably captures genuine biological differences rather than technical noise. More intriguing was what happened over time. The researchers found that spatial heterogeneity, the degree to which different regions of the tumor differ from one another, remained relatively stable as patients moved from baseline to first-generation TKI resistance. But between first-generation resistance and third-generation resistance, heterogeneity increased substantially, and the stroma-enriched compartments accounted for most of that expansion. In other words, the ecosystem surrounding the tumor, rather than the tumor cells alone, appears to become increasingly diverse and disorganized as the disease advances through sequential therapies.

This stromal expansion of heterogeneity carries important implications. The tumor microenvironment is not a passive backdrop; it actively shapes drug penetration, immune surveillance, and the selective pressures that drive resistance. If the stroma becomes progressively more heterogeneous after each line of therapy, it may create an increasingly complex landscape in which resistant clones can find refuge and in which immune cells become progressively less capable of recognizing and attacking malignant cells. The finding also suggests that therapeutic strategies targeting the microenvironment, such as stromal modulation or combinations with immunotherapy, might need to be timed differently depending on how many lines of TKI treatment a patient has already received.

The immune story that emerged from the longitudinal analysis was equally striking and followed what the authors describe as a biphasic trajectory. Early remodeling, occurring as tumors transitioned from baseline to first-generation TKI resistance, featured a loss of T cell activation programs, reduced neutrophil signatures, increased myeloid remodeling, and impairment of antigen presentation. Each of these changes points in the same troubling direction: toward an immune system that is progressively less able to mount an effective response against the tumor. T cell activation is the engine of anti-cancer immunity, antigen presentation is the mechanism by which tumor cells display their abnormal proteins for immune inspection, and myeloid cells, when remodeled toward immunosuppressive states, can actively suppress the immune response. The early phase of resistance thus appears to involve a coordinated dismantling of the immune recognition machinery.

The late phase of the trajectory, from first-generation to third-generation TKI resistance, was characterized instead by stromal checkpoint reprogramming, a shift in the inhibitory signaling molecules expressed within the supportive tissue surrounding the tumor. This temporal separation is significant because it suggests that resistance is not a single event but a staged process, with different biological mechanisms dominating at different points along the treatment course. A therapy designed to counteract early myeloid remodeling would need to be deployed at a different time than one aimed at late stromal checkpoint changes. The concept of stage-dependent immune remodeling, demonstrated here at spatial resolution within individual patients, provides a template for thinking about resistance as a choreographed sequence rather than a monolithic switch.

Perhaps the most tantalizing observation came from comparing tumors that later acquired the T790M mutation with those that did not. T790M is the classic gatekeeper mutation that renders first-generation EGFR-TKIs ineffective and is the very reason third-generation inhibitors such as osimertinib were developed. The researchers found that early spatial remodeling patterns showed potential differences between samples that subsequently acquired T790M and those that resisted through other mechanisms. While the authors are careful to frame this as preliminary and hypothesis-generating rather than definitive, the implication is profound: the microenvironmental changes visible early in treatment might eventually serve as predictive markers, telling clinicians which resistance pathway a tumor is likely to take before it does so. Such foresight could allow therapeutic interventions to be staged preemptively rather than reactively.

It is important to appreciate the limitations that the authors themselves acknowledge. With six patients and fifteen samples, this is an exploratory study, and its explicit purpose is to generate testable, time-aware hypotheses rather than to establish clinical guidelines. Rare consecutive re-biopsy cohorts are inherently small, and the findings will need validation in larger, independent datasets before they can influence practice. Nevertheless, the study provides something genuinely new: a preliminary spatially resolved, longitudinally matched atlas of how the tumor microenvironment changes across sequential EGFR-TKI therapies, capturing both RNA and protein layers within anatomically defined compartments. As spatial omics technologies mature and re-biopsy practices expand, atlases of this kind could become the foundation for a new generation of treatment strategies that anticipate, rather than merely react to, the evolutionary maneuvers of resistant tumors.

Subject of Research: Tumor microenvironment remodeling during sequential EGFR-TKI resistance in EGFR-mutant non-small-cell lung cancer

Article Title: Longitudinal spatial multi-omics delineates tumor microenvironment remodeling across sequential EGFR-TKIs in EGFR-mutant NSCLC

Article References: Song, X., Pan, C., Li, Z., Zhou, X., Xiong, A., Zhao, Z., Luo, H., & Zhou, C. (2026). Longitudinal spatial multi-omics delineates tumor microenvironment remodeling across sequential EGFR-TKIs in EGFR-mutant NSCLC. Journal of Translational Medicine, 24(1), Article 1212. https://doi.org/10.1186/s12967-026-08897-2

Image Credits: AI Generated

DOI: 10.1186/s12967-026-08897-2

Keywords: NSCLC, EGFR-TKIs, drug resistance, tumor microenvironment, spatial omics, digital spatial profiling, longitudinal re-biopsy, spatial heterogeneity, T790M, immune remodeling, tumor stroma, precision oncology

Mass spectrometry imaging has become one of the most powerful tools in modern biology because it can map hundreds of molecules directly within a tissue slice, revealing where lipids, metabolites, proteins and other biomolecules reside without labels or stains. Yet the technique has long been constrained by its spatial resolution. Conventional instruments raster a laser across a tissue section with a defined step size, and features smaller than that step blur together. Boosting resolution usually demands costly instrumentation such as specialized lasers, transmission-mode optics or next-generation ion optics, placing subcellular molecular imaging out of reach for many laboratories.

TEMI takes a different path. Instead of refining the instrument, the team enlarges the tissue itself. Drawing on principles from expansion microscopy, a technique introduced more than a decade ago, the researchers embed tissue in a water-rich hydrogel polymer network. When the gel swells, it carries the tissue along with it, spreading the molecules apart and effectively magnifying the sample before any mass spectrometry takes place. A laser raster that would have been too coarse for the original tissue becomes fine enough to resolve individual cells in the expanded version, yielding more than a 3.5-fold improvement in effective imaging resolution with standard equipment.

The critical challenge was chemistry. Traditional expansion protocols often rely on harsh denaturation conditions, including high heat and strong detergents, to homogenize tissue and allow gels to stretch uniformly. Those conditions would destroy the very molecules mass spectrometry aims to detect. The TEMI workflow solves this by using a re-embedding strategy that expands tissue under mild, harsh-condition-free conditions, preserving the chemical integrity of lipids, metabolites, N-glycans, peptides and proteins. The result is a sample that is both physically enlarged and chemically faithful to its original molecular composition.

The published protocol walks readers through every stage of the process. It begins with constructing a gelation chamber, then describes hydrogel-based tissue gelation and expansion, followed by cryosectioning of the expanded tissue-hydrogel composite. Because expanded samples are soft and hydrated, cutting them into thin sections requires careful handling, and the protocol provides the specific parameters that make reproducible sectioning possible. Detailed guidance covers matrix application, data acquisition and visualization pipelines, along with troubleshooting tips accumulated through the team’s extensive experimentation.

One of the most striking capabilities of TEMI is multiomics mapping on a single tissue section. The protocol describes a sequential workflow in which N-glycans are released in situ by treatment with the enzyme PNGase F, after which proteins are digested with trypsin and imaged, all following lipid and metabolite analysis on the same expanded section. This means a single slice of brain tissue can yield spatially resolved maps of multiple molecular classes, each anchored to the same anatomical context, an efficiency that conventional workflows struggle to match.

Quality control receives particular attention. The protocol includes a dedicated workflow for measuring deformation maps and expansion factors, which quantify how uniformly the gel has expanded. Uneven expansion would distort molecular maps, so the researchers provide computational tools, released through a public code repository, that quantify tissue expansion non-uniformity. They also demonstrate that analyte delocalization during sample preparation is minimal: lipid and small-metabolite signals were detected exclusively in tissue regions, with no corresponding signals in adjacent blank hydrogel areas, confirming that molecules stay where they belong during the swelling process.

The demonstration experiments showcase the method’s power on the mouse cerebellum, a tissue with exquisitely organized layers of cells. Comparing an unexpanded control cerebellum imaged with a 50-micrometer laser raster against a double-embedded, expanded sample at the same step size reveals a dramatic difference in molecular detail. Pushing further, the team performed TEMI after three cycles of gel embedding and expansion with a 10-micrometer raster step, resolving biomolecular heterogeneity that remains invisible in unexpanded tissue. The protocol also demonstrates mapping of small metabolites, N-glycans and proteins across cerebellar structures, illustrating the breadth of the multiomics capability.

The work builds on the team’s primary research paper published in Nature Methods, which introduced TEMI and established its foundational performance, and on a rich history of expansion microscopy methods developed by collaborators at the Howard Hughes Medical Institute’s Janelia Research Campus, including protein-retention expansion microscopy and ten-fold robust expansion. Raw datasets from the protocol’s figures are publicly available in the MassIVE repository, and the deformation-measurement code is hosted on GitHub, lowering barriers for laboratories that want to adopt or adapt the approach.

The implications extend across basic biology and medicine. Single-cell spatial resolution for untargeted molecular imaging could illuminate how metabolic gradients shape tissue function, how lipid compositions vary between neighboring cells in the nervous system, and how disease states such as cancer or neurodegeneration alter molecular architecture at scales previously accessible only to antibody-based imaging. Because TEMI requires no instrument modifications, it promises broad accessibility: laboratories with standard mass spectrometry imaging systems can now reach a resolution regime once reserved for a handful of specialized facilities. The protocol’s authors, spanning the University of Wisconsin-Madison and HHMI Janelia, and supported in part by the National Institutes of Health, position the method as a practical bridge between the spatial-omics revolution and the everyday mass spectrometry lab, inviting a wide research community to stretch its view of tissue, quite literally, into sharper focus.

Subject of Research: Tissue expansion combined with mass spectrometry imaging for high-spatial-resolution multiomics molecular mapping

Article Title: Tissue expansion mass spectrometry imaging (TEMI) for high-spatial-resolution multiomics molecular mapping

Article References: Zhang, H., Ding, L., Duong, T., Selby, K. G., Lee, W.-J., Hu, A., Lu, H., Tillberg, P. W., Wang, M., & Li, L. (2026). Tissue expansion mass spectrometry imaging (TEMI) for high-spatial-resolution multiomics molecular mapping. Nature Protocols. https://doi.org/10.1038/s41596-026-01427-w

Image Credits: AI Generated

DOI: 10.1038/s41596-026-01427-w

Keywords: mass spectrometry imaging, tissue expansion, TEMI, spatial omics, single-cell imaging, lipidomics, metabolomics, N-glycans, proteomics, hydrogel expansion, Nature Protocols, multiomics

The central insight driving the review is that the tumor microenvironment is not a passive backdrop but an active determinant of both treatment response and clinical outcomes. In colorectal cancer, the composition of immune cells infiltrating a tumor, the ratio of tumor cells to stroma, the presence of tertiary lymphoid structures and the burden of tumor-infiltrating lymphocytes all carry prognostic and predictive weight. Yet these features are traditionally measured on tissue samples obtained through biopsy or surgery, which capture only a fragment of the tumor and cannot be repeated freely over time. Radiomics offers an alternative. By extracting hundreds of quantitative features from routine medical images, such as computed tomography and magnetic resonance imaging, radiomics aims to characterize the whole tumor, including its internal heterogeneity, without touching the patient.

Technically, radiomics pipelines convert standard acquisitions such as T2-weighted imaging, contrast-enhanced T1-weighted imaging and diffusion-weighted imaging into high-dimensional feature sets. First-pass texture statistics capture local gray-level variation, while higher-order features extracted through filtered images and deep learning networks probe patterns that the human eye cannot resolve. Apparent diffusion coefficient maps derived from diffusion-weighted imaging, for example, reflect tissue cellularity and can serve as indirect surrogates of tumor density and stromal content. The review organizes the colorectal cancer radiomics literature into three thematic clusters: features that correlate directly with microenvironment components, features associated with vascular-invasion-related phenotypes, and features tied to tumor-intrinsic properties that are themselves shaped by the microenvironment. Each cluster, the authors argue, contributes a different piece of the puzzle linking what radiologists see on a screen to what pathologists see under a microscope and what molecular biologists sequence in the lab.

The vascular-invasion theme is particularly consequential clinically. Extramural venous invasion and microvascular invasion are established markers of poor prognosis in colorectal cancer, signaling that tumor cells have entered the circulatory system and raised the risk of metastasis. Radiomic models trained on CT and MRI can flag these phenotypes before surgery, potentially informing decisions about neoadjuvant therapy and surgical planning. Meanwhile, radiomic signatures predicting microsatellite instability and deficient mismatch repair status offer a non-invasive proxy for the single most important biomarker in modern colorectal cancer immunotherapy, since patients with dMMR or MSI-high tumors are the ones most likely to benefit from immune checkpoint inhibitors. In locally advanced rectal cancer, radiomic models have also been used to predict pathological complete response after chemoradiotherapy, a finding that could eventually help identify patients for organ-preserving strategies.

Pathomics extends the same quantitative logic to the microscopic scale. Whole-slide imaging digitizes histopathology slides, and computational methods then profile tumor architecture and the spatial distribution of immune cells at high throughput. Where a pathologist might estimate tumor-infiltrating lymphocyte density visually, pathomic pipelines can quantify it precisely, map the spatial relationships between tumor nests and stromal compartments, and compute the tumor-stroma ratio automatically. Deep learning models trained on whole-slide images can even predict molecular alterations, such as microsatellite instability, directly from hematoxylin and eosin-stained tissue. Because pathomics operates on resected or biopsied tissue, it provides the microscopic ground truth that radiomics lacks, and the two approaches are natural partners: radiomics sees the whole tumor in vivo, while pathomics resolves the cellular detail of the sampled regions.

The most ambitious portion of the review describes cross-scale integration, in which radiomic and pathomic features are fused with genomics and transcriptomics to trace a continuous chain from macroscopic phenotype to molecular mechanism. Radiomic features that predict microsatellite instability, for instance, can be connected to the immune-inflamed transcriptional programs that accompany deficient mismatch repair, including upregulated checkpoint molecules and enriched cytotoxic T-cell signatures. Consensus molecular subtypes of colorectal cancer, which stratify tumors by their gene-expression patterns, also leave imaging fingerprints, and studies reviewed by the authors show that radiomic models can distinguish between these molecular classes with useful accuracy. At the single-cell and spatial-omics frontier, technologies such as single-cell RNA sequencing and spatially resolved transcriptomics reveal the precise cellular neighborhoods within the microenvironment, including interactions between tumor-associated macrophages, myeloid-derived suppressor cells, cancer-associated fibroblasts and lymphocytes, offering mechanistic explanations for the imaging features that models detect.

Multimodal fusion is the methodological glue holding this framework together. Rather than treating imaging, pathology and molecular data as parallel silos, fusion approaches combine them within a single predictive model, allowing each modality to compensate for the blind spots of the others. Deep learning architectures can ingest radiomic features from CT or MRI, pathomic features from whole-slide images, and genomic or transcriptomic profiles from the same patient, learning joint representations that outperform any single data type. The review highlights early studies demonstrating that such combined models improve prediction of prognosis, immunotherapy response and treatment-related outcomes in colorectal cancer compared with unimodal baselines, suggesting that the cross-scale framework is more than the sum of its parts.

The authors are explicit, however, about the caveats. Much of the evidence reviewed is retrospective, derived from single-center cohorts with limited sample sizes, and few radiomic models have been validated prospectively or across diverse populations. Standardization of image acquisition, feature definitions and model reporting remains inconsistent across the field, raising concerns about reproducibility. The relationship between imaging features and microenvironment biology is often correlational rather than mechanistically established, and the review repeatedly emphasizes that clinical utility, and any role in therapeutic decision-making, remain to be established through prospective validation. These are not trivial hurdles; they are the same obstacles that have slowed the translation of radiomics in other cancer types.

Even so, the trajectory described in the review is striking. What began as an effort to squeeze extra information out of images that radiologists already acquire routinely has matured into a multi-scale program that connects the radiology suite to the pathology lab and the sequencing core. If prospective studies bear out the promise of imaging-driven multi-omics integration, clinicians could one day profile a patient’s tumor microenvironment repeatedly, cheaply and non-invasively, tracking its evolution under therapy and selecting patients for immunotherapy with far greater precision than today’s single-timepoint biomarkers allow. For a disease that kills hundreds of thousands of people each year and frustrates clinicians with its heterogeneity, that would represent a genuinely transformative shift, one that this review maps out with unusual technical clarity and admirable restraint about what has, and has not, yet been proven.

Subject of Research: Decoding the tumor microenvironment in colorectal cancer through radiomics and multi-omics integration

Article Title: From imaging to multi-omics: decoding the tumor microenvironment in colorectal cancer

Article References: Gong, L., Wu, X., Zhang, W., Lai, B., Yuan, J., Gu, Y., Shen, H., Liu, X., Xiong, Y., Zheng, J., Wang, L., Han, X., Zhang, B., & Zhang, S. (2026). From imaging to multi-omics: decoding the tumor microenvironment in colorectal cancer. Journal of Translational Medicine. https://doi.org/10.1186/s12967-026-08958-6

Image Credits: AI Generated

DOI: 10.1186/s12967-026-08958-6

Keywords: colorectal cancer, tumor microenvironment, radiomics, pathomics, multi-omics, immunotherapy, microsatellite instability, deep learning, spatial omics, tumor heterogeneity, precision oncology, biomarkers

The framework, developed by Seowon Chang, Alexander Fleischmann and Ying Ma, addresses a problem that has grown acute as spatial assays have diversified. Transcriptome-plus-protein platforms such as SPOTS and spatial CITE-seq, epigenome-transcriptome methods that jointly capture chromatin accessibility or histone marks alongside RNA, metabolomic imaging and multiplexed approaches such as spatial-Mux-seq each carry modality-specific signals that generic integration tools tend to wash out. Earlier software could typically merge only two modalities, struggled with the massive spot counts of Visium HD, and returned a single consensus answer with no indication of where the model was guessing. SCIGMA’s designers set out to build a system that was scalable, generalizable across platforms and honest about its own uncertainty.

Architecturally, SCIGMA combines two ideas that have proven powerful in machine learning but had not been married in quite this way for spatial biology. The first is a multiview graph neural network built on graph attention layers, which represents each spatial location as a node connected to its tissue neighbors, letting the model learn from both molecular measurements and the physical arrangement of the tissue. Each modality is encoded through its own branch, so transcriptomic, proteomic, epigenomic and metabolic information is transformed into a shared latent space without being forced into a single flattened feature matrix. The second idea is an uncertainty-aware contrastive learning objective: by treating temperature as a learnable, uncertainty-linked quantity, the model learns to pull together representations of the same location seen through different molecular lenses while pushing apart mismatched views, and it modulates how strongly it aligns views depending on how reliable the data at that location appear to be.

This uncertainty machinery is more than a statistical nicety. After training, SCIGMA produces spatially resolved uncertainty maps that highlight regions of biological or technical heterogeneity — tumor margins, regions with mixed cell populations, or spots where assay quality degrades. The authors show that uncertainty estimates flag locations where feature reconstruction error is highest, meaning researchers can see precisely where the integrated representation is least trustworthy rather than accepting a smoothed-over consensus. Interpretability was a design goal throughout: the framework preserves modality-specific signals within its joint embedding, so users can trace which molecular layer drives a given spatial domain and recover regulatory programs that are visible only in, say, the chromatin channel and not the transcriptome.

The evaluation was unusually broad. The team benchmarked SCIGMA across 19 datasets spanning eight modalities, ten tissue types and nine technological platforms, ranging from spatial epigenome-transcriptome profiling of the postnatal mouse brain to protein-plus-RNA measurements of the mouse spleen, single-cell-resolution Xenium Prime datasets from human ovarian and cervical cancers, and enormous Visium HD sections of mouse intestine and human colorectal cancer with more than one million spots each. Where ground truth allowed comparison, SCIGMA outperformed existing methods on spatial domain detection, preservation of modality-specific information, feature reconstruction and reproducibility across repeated runs — a metric that matters given growing concern about the fragility of machine learning models in biomedical data science.

The biological case studies illustrate what multimodal integration buys that single-modality analysis cannot. In the mouse brain, combining chromatin accessibility or histone modification data with transcriptomics let SCIGMA refine cortical layers and white matter boundaries beyond what RNA alone revealed, and to link transcriptional identities with the regulatory elements that govern them. In the spleen, joint transcriptome-proteome analysis sharpened the demarcation of follicles, marginal zones and T cell zones, tying protein-level markers such as CD19 and CD3 components to their local transcriptional contexts. Analyses of Xenium Prime tumor sections demonstrated that the framework resolves intratumoral heterogeneity, distinguishing tumor epithelial niches, stromal compartments and immune infiltrates while its uncertainty layer highlights the unstable boundary regions where tumor biology is most contested.

Scalability was addressed head-on. Naive graph neural networks choke when every spot in a Visium HD section becomes a node in a graph with millions of vertices. SCIGMA incorporates efficient sampling strategies and nearest-neighbor retrieval, borrowing ideas from the Faiss library, to keep training tractable, and the authors demonstrate full-length analysis of Visium HD datasets exceeding one million spatial locations. Equally important is extensibility: the framework is modular, so when a future technology profiles six, seven or more molecular layers simultaneously, additional modality encoders can be attached without redesigning the core model. The team demonstrated this flexibility on spatial-Mux-seq mouse brain data, simultaneously integrating five modalities into a coherent joint representation.

Software accessibility rounded out the release. The SCIGMA package, along with scripts reproducing every published analysis and a suite of tutorials, is openly available through GitHub and archived on Zenodo, supporting the reproducibility standards the paper explicitly engages with. Funding came from the National Science Foundation and the National Institutes of Health, and the work was carried out at Brown’s Center for Computational Molecular Biology with computational support from the university’s Center for Computation and Visualization. The authors report no competing interests, and the article underwent peer review by researchers including Mengjie Chen, Zhaoheng Li and Wei Sun.

For the field, SCIGMA arrives at a moment when the volume and variety of spatial data are outpacing analysis tools. High-resolution platforms such as Visium HD and Xenium Prime are making million-spot, subcellular-resolution datasets routine, and multiplexed assays are stacking molecular layers that earlier software simply could not combine. A framework that handles five modalities, scales to the largest current datasets, runs across platforms without platform-specific tuning, and quantifies its own confidence offers a kind of common analytical ground that spatial biology has lacked. If it holds up in broad community use, the practical effect will be that researchers spend less time wrestling incompatible data formats into fragile pipelines and more time asking biological questions — about how tumors orchestrate their microenvironments, how cortical architecture is patterned, and how genomic regulation plays out across the geography of living tissue.

Subject of Research: Development of SCIGMA, a scalable uncertainty-aware deep learning framework for integrating spatial multiomics data across modalities and platforms

Article Title: Scalable, generalizable and uncertainty-aware integration of spatial multiomics across diverse modalities and platforms with SCIGMA

Article References: Chang, S., Fleischmann, A., & Ma, Y. (2026). Scalable, generalizable and uncertainty-aware integration of spatial multiomics across diverse modalities and platforms with SCIGMA. Nature Genetics, 58(9), 2284-2302. https://doi.org/10.1038/s41588-026-02706-8

Image Credits: AI Generated

DOI: 10.1038/s41588-026-02706-8

Keywords: spatial omics, multiomics integration, deep learning, graph neural networks, contrastive learning, uncertainty estimation, spatial transcriptomics, tumor heterogeneity, spatial domain detection, Visium HD, Xenium Prime, computational biology

A team of Austrian researchers believes they have a practical answer. In a study published in BMC Bioinformatics, Markus Steiner, Stephan Drothler, Jan P. Höpner, Roland Geisberger and Nadja Zaborsky, based at Paracelsus Medical University, the Salzburg Cancer Research Institute and Paris-Lodron University Salzburg, introduce Spatialgater, an open, web-based tool built in R Shiny that lets scientists select individual cells directly on a zoomable image of their tissue. Rather than forcing analysts to write custom scripts every time they want to isolate, say, a cluster of immune cells sitting at the edge of a tumor nest, Spatialgater allows them to draw polygon gates around cells of interest with a few clicks of a mouse — in situ, exactly where the cells actually live.

The problem the tool addresses is rooted in how spatial omics data are currently processed. The dominant data structure in the R ecosystem for this kind of work is the SpatialExperiment class, a container that stores molecular expression measurements alongside the spatial coordinates of each cell. Standard analysis workflows built on this framework typically proceed by clustering cells according to their expression of biomolecules — proteins, transcripts or other markers — and then subsetting or annotating cells based on those molecular groupings. What these workflows largely ignore is geography. A tumor cell and its lookalike elsewhere on the slide may share nearly identical expression profiles, yet occupy radically different microenvironments, one nestled against a blood vessel, the other buried in dense stroma. Clustering alone cannot tell them apart by location.

Computational biologists have developed patch-detection and neighborhood-analysis methods that identify recurring spatial patterns, such as regions where particular cell types consistently co-occur. These approaches are powerful for systematic surveys of tissue architecture, but they are fundamentally automated and population-level. None of them gives a researcher the simple, immediate ability to say: show me these cells, right here, and let me select exactly the ones I mean. That gap matters in practice. Pathologists and immunologists often spot something visually striking — an unusual accumulation of T cells at an invasive margin, a suspicious ring of macrophages around a necrotic core — and want to interrogate those specific cells without writing bespoke code or approximating their region of interest with crude coordinate filters.

Spatialgater fills that gap with a deliberately simple interface. The tool renders cells as dots overlaid on a zoomable image of the tissue, so users can navigate the sample much as they would navigate a digital slide under a microscope. Drawing a polygon directly on the image selects every cell whose coordinates fall inside the drawn boundary. The selection is fully interactive: researchers can zoom in to single-cell resolution, refine a gate, add additional polygons and inspect the molecular profiles of the cells they have captured. Because the tool operates on SpatialExperiment objects, it slots naturally into existing R-based spatial omics pipelines rather than requiring researchers to export their data into an unfamiliar format.

Two features elevate the tool beyond a simple lasso. The first is an integrated k-nearest-neighbor function that can automatically extend a manually drawn gate across spatially similar microenvironments elsewhere in the tissue. If a researcher delineates a distinctive cellular neighborhood in one region, the algorithm can propagate that selection to comparable regions, dramatically reducing the manual labor of annotating large tissue sections while keeping the human judgment that defined the original gate. The second is traceability. Every polygon a user draws is recorded in a log file, creating an auditable record of exactly how each selection was made — a small but significant safeguard for reproducibility in a field where manual choices have traditionally gone undocumented.

Export and integration are equally considered. Selected cell identifiers can be written out as a standard CSV file for use in any downstream software, or saved directly back into the original SpatialExperiment object as a new logical column, meaning manual selections become first-class citizens in subsequent statistical analyses. This design choice reflects the authors’ core motivation: boosting interactivity while reducing the programming burden of image analysis. A researcher no longer needs to be a fluent R programmer to translate a visual observation into a computable cell set; the web interface handles the translation.

To demonstrate the tool in action, the team applied Spatialgater to a publicly available imaging mass cytometry dataset of breast cancer tissue, using it to characterize and compare T cells according to their spatial location within the tumor microenvironment. The demonstration speaks directly to one of the hottest questions in cancer immunology: how the precise positioning of immune cells — whether T cells are excluded from tumor nests, patrolling the invasive margin or dispersed through stroma — shapes antitumor immunity and predicts response to immunotherapy. A tool that makes spatially targeted cell selection fast and intuitive could accelerate exactly this kind of location-dependent immune analysis. The team also validated the package’s gating functionality using test data derived from a wild-type mouse, with the animal work approved by the Austrian Federal Ministry of Education, Science, and Research under approval number BMBWF 2023-0.644.528.

The significance of the work may lie less in any single algorithm than in what it signals about the maturing of spatial omics. The field’s first generation of tools focused on generating and processing data; the next generation is about making those data genuinely explorable by working biologists. By lowering the barrier between seeing something interesting in a tissue image and acting on it computationally, Spatialgater embodies a broader shift toward interactive, human-in-the-loop bioinformatics. The software is open access, published under a Creative Commons Attribution license, and funded by the Austrian Science Fund, WISS 2025 through the Cancer Cluster Salzburg, and the Province of Salzburg. For laboratories drowning in multiplexed images and struggling to connect visual insight with computational rigor, a free browser-based gate-drawing tool may prove to be one of those deceptively simple additions that changes daily practice.

Subject of Research: An interactive R Shiny web tool for spatially selecting and gating individual cells in spatial omics datasets.

Article Title: Spatialgater: an R Shiny webtool for in situ gating of cells in spatial omics experiments

Article References: Steiner, M., Drothler, S., Höpner, J. P., Geisberger, R., & Zaborsky, N. (2026). Spatialgater: an R Shiny webtool for in situ gating of cells in spatial omics experiments. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06620-y

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06620-y

Keywords: Spatialgater, R Shiny, spatial omics, single cell gating, imaging mass cytometry, SpatialExperiment, multiplexed imaging, breast cancer, T cells, tumor microenvironment, bioinformatics software, Shiny