Weekly reads 2/2/26

How space, context, and uncertainty reshape cancer and single-cell biology

Feb 08, 2026

This week's reads highlight a shift in our understanding of complexity in cancer and tissue biology: structured, spatially organised, and mechanistically constrained systems rather than diffuse chaos. Research has shown that disease progression is influenced by both localised niches and long-range circuits. Examples include IL-1α-driven fibroblast-neutrophil niches in head and neck cancer, functionally distinct neutrophil states shaped by genotype and space in colorectal tumours, and a surprising tumor-brain sensory-sympathetic axis that suppresses lung cancer immunity. Others demonstrate how much temporal and causal structure can be extracted from static data by carefully modelling spatial context: OSDR reconstructs tissue population dynamics from a single biopsy, and Metient reframes metastatic spread as a space of plausible histories rather than a single inferred truth. Large-scale benchmarks reveal trade-offs and generalisation gaps in spatial batch integration, while scGeneLens deconstructs single-cell foundation models to reveal distinct biological priors encoded by different architectures. These works emphasise a unified lesson: progress comes not only from more data or larger models, but also from representations and analyses that take into account space, context, uncertainty, and mechanism at multiple biological levels.

Preprints/articles that I managed to read this week

IL-1α drives a tumor-stroma-neutrophil axis through inflammatory fibroblast activation in head and neck cancer

Hanschmann, E.-M. et al. bioRxiv (2026). https://doi.org/10.64898/2026.01.20.700440

The paper in one sentence

In head and neck cancer, tumor-derived interleukin-1α (IL-1α) reprograms nearby mesenchymal stromal cells into inflammatory, neutrophil-recruiting fibroblasts, creating spatially organized “niches” in the tumor stroma that attract, activate, and sustain tumor-associated neutrophils, which are linked to poorer patient survival.

Summary

This study uncovers a key communication circuit in head and neck squamous cell carcinoma (HNSCC) that explains why high levels of tumor-associated neutrophils (TANs) are linked to worse outcomes. The researchers found that IL-1α, a cytokine released by living or dying tumor cells, acts as a master signal. It reprograms neighboring mesenchymal stromal cells (MSCs) into a type of inflammatory cancer-associated fibroblast (iCAF). These activated fibroblasts then pump out a cocktail of factors, most importantly the chemokine CXCL8 (IL-8) and the survival factor G-CSF, that act as powerful attractants and life-support signals for neutrophils. Using patient tissues, 3D models, and zebrafish, the team showed that this creates distinct inflammatory "niches": areas of the tumor where IL-1α-positive cancer cells are surrounded by a stroma rich in CXCL8 and G-CSF, which in turn is densely packed with neutrophils. These neutrophils are not just more numerous but also show signs of activation (like higher levels of elastase and myeloperoxidase). This work provides a mechanistic link between a specific tumor-derived signal (IL-1α), stromal reprogramming, and the spatial organization and pro-tumor function of neutrophils in HNSCC.

Personal highlights

IL-1α as the upstream orchestrator of stromal inflammation: tumor-derived IL-1α, whether actively secreted or released from necrotic cells, is identified as the non-redundant signal that reprograms resting mesenchymal stromal cells into inflammatory, neutrophil-recruiting cancer-associated fibroblasts (iCAFs), initiating a feed-forward loop in the tumor microenvironment.
Spatially confined inflammatory niches drive neutrophil localization: advanced spatial mapping of patient tumors reveals that IL1A-positive tumor islets are surrounded by a defined stromal zone (within 200 µm) enriched for cells co-expressing the key neutrophil factors CXCL8 and CSF3 (G-CSF), creating localized “niches” that selectively attract and accumulate neutrophils.
Stromal amplification, not direct tumor signaling, controls neutrophil recruitment: while IL-1α has little direct effect on neutrophils, its power lies in activating stromal fibroblasts, which become the dominant source of the chemokine CXCL8, the non-redundant driver of neutrophil chemotaxis in this system.
Link to clinically relevant CAF states and patient prognosis: the IL-1α-induced gene signature in MSCs strongly overlaps with published CAF subtypes in HNSCC that are associated with extracellular matrix remodeling, immunosuppression, and poor patient overall survival, connecting the molecular mechanism to clinical outcome.
Niches shape both neutrophil abundance and functional state: IL-1α-driven niches not only increase neutrophil density but also influence their phenotype, preserving granule content (MPO, NE) and promoting a transcriptionally activated state, whereas areas without IL-1α show signs of exhausted, degranulated neutrophils.

Why should we care?

This study identifies IL-1α as the "ignition switch" and stromal fibroblasts as the "signal amplifiers" as a clear, targetable axis for new anti-cancer drugs. For example, potential combination therapies such as blocking IL-1 signaling (with drugs like anakinra) alongside CXCL8 inhibition, could be used to dismantle these pro-tumor neutrophil niches, especially in aggressive, neutrophil-rich cancers.

One-Shot Tissue Dynamics Reconstruction (OSDR): Inferring Temporal Cell Population Dynamics from a Single Spatial Proteomics Snapshot

Somer, J., Mannor, S., & Alon, U. Nature (2026). https://doi.org/10.1038/s41586-025-09876-1

The paper in one sentence

OSDR is a method that uses a single spatial proteomics biopsy: measuring cell neighbourhood composition and division marker (Ki67), to reconstruct dynamical models of cell population change over time. It was validated in breast cancer to reveal fibroblast-macrophage steady states and excitable T-B cell pulses.

Summary

OSDR tackles a basic shortcoming in tissue biology: human biopsies only provide a static snapshot of cell populations, making it difficult to predict how they evolve with time. The method uses imaging mass cytometry data to model cell division likelihood based on neighbourhood cellular composition within a given radius (~80 µm). OSDR uses Ki67 as a division marker to fit logistic regression models that predict division rates from local cell-type counts. It then constructs stochastic spatial simulations or deterministic phase portraits (ODE-based) of population dynamics. The authors verified OSDR on synthetic data, replicated known fibroblast-macrophage co-culture dynamics, established an excitable pulse-generating circuit between T and B cells in the tumour microenvironment, and predicted tumour collapse in treatment responders with only early-treatment biopsies.

Personal highlights

From static snapshots to dynamical phase portraits: OSDR translates spatial neighborhood compositions into a vector field of population change, constructing phase portraits that reveal stable fixed points (e.g., hot/cold fibrosis) and excitable dynamics (e.g., T–B cell pulses) without longitudinal data.
Division-rate inference via neighborhood logistic regression: the core of OSDR is a per-cell-type logistic model that predicts Ki67⁺ probability from counts of neighboring cell types, effectively treating local cellular context as a predictor of proliferation stimulus or inhibition.
Dual output: stochastic simulations vs. deterministic ODEs: OSDR offers two complementary views, agent-based stochastic simulations that respect initial spatial configurations, and smooth ODE-based phase portraits that abstract away spatial arrangement to reveal general circuit logic.
Excitable immune circuit discovery: The method uncovered a previously undescribed excitable dynamic between T and B cells in breast cancer, where crossing a CD4⁺ T cell density threshold triggers a pulse of immune activity followed by B cell–mediated suppression, a temporal flare-and-refractory pattern akin to autoimmune dynamics.
Clinical prediction from early biopsies: In a triple-negative breast cancer trial, OSDR models fit to week‑3 biopsies predicted tumor population collapse in responders (to chemo or chemo+immunotherapy) but not in non‑responders, outperforming simple proliferation rate comparisons.

Why should we care?

OSDR links spatial context to division rates and turns a snapshot into a “tissue dynamics simulator” that can reconstruct interaction circuits (like fibroblast-macrophage co‑dependence) and discover new temporal motifs (like excitable immune pulses). It demostrated that carefull modelling of neighbourhood effects can extract temporal insights from what is considered static data.

A Spatial and Single-Cell Atlas of Neutrophils in Colorectal Cancer Reveals Dual Roles, Functional Niches, and Tumor-Induced Reprogramming

Marteau, V. et al. Cancer Cell (2026). 10.1016/j.ccell.2025.12.003

The paper in one sentence

By integrating 4.27 million single cells from 650 colorectal cancer patients, this study uncovers the spatial organization, phenotypic plasticity, and pro- versus anti-tumor roles of neutrophils, linking specific subsets to patient survival, genotype-dependent immune remodeling, and tumor-induced granulopoiesis.

Summary

This work builds a large-scale, multi-modal atlas of colorectal cancer (CRC) by combining single-cell RNA sequencing, spatial transcriptomics, and proteomics across 1,670 samples. The atlas reveals four distinct tumor immune phenotypes (immune-deserted, T-cell, B-cell, and myeloid-dominant) and identifies consensus myeloid gene programs. Neutrophils are shown to exist in both pro-tumor and antigen-presenting anti-tumor states, with the latter associated with improved survival. Spatially, neutrophils form organized niches in the tumor core and invasive margin, interacting with cancer-associated fibroblasts via IL-1 signaling. Using patient-derived organoids and mouse models, the study demonstrates that KRAS-mutant tumors reprogram neutrophils toward a pro-tumor phenotype and that tumor signals can alter granulopoiesis in the bone marrow. These findings offer new insights into myeloid-driven immune suppression and potential therapeutic targets in CRC.

Personal highlights

Large-scale, multi-modal integration of 4.27 million cells: the atlas harmonizes scRNA-seq, spatial transcriptomics, and imaging mass cytometry across 1,670 clinical samples, providing an unprecedented resolution of CRC immune landscapes and enabling robust patient stratification into four distinct immune phenotypes.
Identification of neutrophil states with opposing functions: neutrophils are shown to exist as both pro-tumor (LOX-1+ TAN) and anti-tumor (HLA-DR+ hybrid) subsets, with the latter acting as antigen-presenting cells and correlating with better patient outcomes in early-stage disease.
Spatial mapping of neutrophil-functional niches: using 10x Xenium and deep-learning-based niche detection, the study reveals spatially organized neutrophil aggregates that interact with fibroblasts via IL-1B signaling, a mechanism linked to therapy resistance and microenvironment remodeling.
Genotype-driven neutrophil reprogramming: KRAS-mutant tumors are found to polarize neutrophils toward a pro-tumor phenotype via soluble factors, enhancing neutrophil survival and suppressing CD8+ T cell activity, providing a mechanistic link between oncogenic signaling and immune evasion.
Tumor-induced granulopoiesis along the bone marrow–tumor axis: orthotopic mouse models show that CRC reprograms neutrophil progenitors in the bone marrow, driving systemic expansion of pro-tumor Siglec-F+ neutrophils that can be selectively depleted to reduce T cell exhaustion.

Why should we care?

This work reveal how neutrophils, often overlooked as short-lived bystanders, orchestrate both tumour promotion and suppression in CRC. By linking specific neutrophil subsets to patient genotypes, spatial neighbourhoods, and clinical outcomes, it provides a roadmap for targeting myeloid cells in immunotherapy-resistant cancers. It reframes neutrophils as dynamic, targetable regulators of the tumor microenvironment, opening new avenues for myeloid-directed therapies in colorectal and other cancers.

Inferring Cancer Type-Specific Patterns of Metastatic Spread Using Metient

Koyyalagunta, D. et al. Nature Methods (2025). https://doi.org/10.1038/s41592-025-02924-8

The paper in one sentence

Metient is a gradient-based, multiobjective optimization framework that reconstructs multiple plausible migration histories of metastatic spread from tumor sequencing data, using genetic distance and organotropism priors to resolve parsimony conflicts and reveal cancer type-specific dissemination patterns.

Summary

Metient addresses some limitations in existing metastasis reconstruction tools, poor scalability, reliance on oversimplified parsimony models, and the return of only a single migration history, by introducing a flexible, gradient-based optimization approach that explores a Pareto front of plausible solutions. The method scores histories using three parsimony metrics (migrations, comigrations, seeding sites) and two biologically informed metastasis priors: genetic distance (favouring migrations on longer phylogenetic branches) and organotropism (favouring transitions to more common metastatic sites first). Metient can be calibrated per cancer cohort to learn cohort-specific parsimony weights, or used with a pan-cancer model. Validated on simulated data and applied to 167 patients across melanoma, ovarian, neuroblastoma and NSCLC cohorts, Metient recovered known expert-annotated histories, often identified additional plausible alternatives, and estimated higher rates of polyclonal seeding and metastasis-to-metastasis spread than previously reported. The framework scales to single-cell lineage tracing data and provides visualization tools to compare multiple hypotheses of metastatic dissemination.

Personal highlights

Pareto-front exploration of metastatic histories: instead of forcing a single “most parsimonious” history, Metient maps a Pareto front of solutions that trade off migrations, comigrations and seeding sites, explicitly capturing the ambiguity inherent in migration inference.
Biologically informed metastasis priors: Metient introduces two novel priors—genetic distance (migrations on longer mutational branches) and organotropism (common metastatic sites seeded earlier), that break ties among equally parsimonious solutions and can be calibrated per cancer type.
Cohort-specific parsimony calibration: The metient-calibrate mode learns cancer-type-specific weights for the parsimony metrics by minimizing cross‑entropy between the parsimony distribution and the prior‑based distribution across a patient cohort, moving beyond a one‑size‑fits‑all parsimony model.
Scalable gradient‑based optimization with Gumbel‑Softmax: Metient replaces mixed‑integer linear programming with a differentiable, GPU‑friendly optimization that samples labelings via Gumbel‑Softmax, enabling application to large trees (e.g., from single‑cell lineage tracing) and efficient exploration of the solution space.
Higher estimates of polyclonality and metastasis‑to‑metastasis seeding: by considering multiple plausible histories and avoiding a priori assumptions (e.g., “primary‑only seeding”), Metient infers nearly twice the previously reported rate of polyclonal seeding and highlights frequent intermediate‑site (e.g., lymph node) hubs in NSCLC and neuroblastoma.

Why should we care?

Metient reframes metastasis reconstruction from a search for a single “correct” history to an exploration of the space of plausible histories, acknowledging that current genomic data often cannot uniquely resolve the route of spread. By making this ambiguity explicit and providing tools to rank solutions with biologically grounded priors, Metient helps avoid overconfidence in models that may be biased by simplifying assumptions (e.g., that all metastases seed directly from the primary)

Comprehensive benchmarking of batch integration methods for spatial transcriptomics using a large-scale cancer atlas

Ludington et al. BioRxiv (2026). DOI: 10.64898/2026.01.12.699017

The paper in one sentence

This study benchmarks 11 spatial transcriptomics integration methods on a large, clinically diverse cancer atlas, revealing that probabilistic models excel at batch correction and biological conservation, while graph-based methods preserve spatial structure but struggle with batch effects and that generalization to unseen samples remains a major challenge.

Summary

The authors evaluated 11 integration methods, spanning linear, graph-based, and probabilistic models, on the MOSAIC Window dataset, a multi-sample, multi-cancer spatial transcriptomics atlas with strong donor-specific batch effects. Using both in-distribution and out-of-distribution (OOD) validation, they found that probabilistic methods like scVI and scVIVA lead in batch correction and biological conservation, while graph-based approaches (e.g., STAGATE) excel in spatial preservation but perform poorly on batch mixing. Notably, even top-performing methods like scVI showed significant performance drops on unseen samples, whereas simpler methods like PCA generalized more robustly. The work highlights critical trade-offs between integration quality and real-world applicability.

Personal highlights

Probabilistic methods dominate batch-aware integration: variational inference models (scVI, scVIVA, sysVI) consistently outperformed linear and graph-based methods in removing batch effects while preserving biological signals, reinforcing a shift toward probabilistic frameworks in spatial omics.
scVIVA balances batch correction, biology, and spatial structure: by extending scVI with spatial-aware components, scVIVA achieved top-tier performance across batch correction and biological conservation while maintaining competitive spatial preservation, a uniquely balanced profile among tested methods.
Graph-based models preserve spatial layout but fail at batch mixing:mMethods like STAGATE and SEDR excelled in spatial conservation metrics but ranked worst in batch correction, suggesting that prioritizing local neighborhoods may amplify sample-specific noise rather than invariant biological signals.
PCA is surprisingly robust to unseen samples: despite poor batch correction in all-at-once evaluation, PCA generalized robustly in OOD tests, with near-zero performance drops, highlighting that simpler models can be more reliable when deploying to new patient samples.
Generalization remains an underappreciated challenge: The novel OOD evaluation revealed that even state-of-the-art methods like scVI suffer significant performance degradation on unseen donors, emphasizing that benchmark rankings should account for real-world deployment scenarios.

Why should we care?

This benchmark tests integration methods not on toy datasets but on a real and heterogeneous cancer one. The findings will be useful for everyone who uses a multi-sample spatial dataset: if your goal is removing batch effects across patients, probabilistic models are currently best; if spatial clustering is the focus, graph-based tools may suffice. Crucially, the study sounds a warning that even the best-performing models can falter on new samples, meaning method choice must align with whether you’re exploring a fixed dataset or building tools for prospective use.

Tumour‑brain crosstalk restrains cancer immunity via a sensory‑sympathetic axis

Wei, H.K. et al. Nature (2026). https://doi.org/10.1038/s41586‑025‑10028‑8

The paper in one sentence

This study reveals that lung tumours hijack a specific vagal sensory‑to‑sympathetic neural circuit, signalling through the brainstem to boost local sympathetic activity, which then suppresses anti‑tumour immunity via β₂‑adrenergic signalling in alveolar macrophages.

Summary

Using genetically engineered mouse models of lung adenocarcinoma combined with neural tracing, single‑cell transcriptomics, and chemogenetic manipulations, the authors uncovered a bidirectional tumour–brain communication axis. Tumour‑derived factors promote the innervation and activation of vagal NPY2R/TRPV1 sensory neurons; these afferents relay signals to rostral ventrolateral medulla (RVLM) neurons in the brainstem, which in turn drive elevated sympathetic outflow to the tumour microenvironment. The resulting noradrenaline release acts via β₂‑adrenergic receptors (ADRB2) on alveolar macrophages, polarising them toward an immunosuppressive state and inhibiting tumour‑reactive T‑cell responses. Genetic or pharmacological disruption of this sensory‑sympathetic pathway or of ADRB2 signalling, restores anti‑tumour immunity and strongly suppresses tumour growth. Clinical data from lung‑cancer patients show that high expression of vagal‑sensory and sympathetic‑nerve gene signatures correlates with poor survival and reduced CD8⁺ T‑cell infiltration, underscoring the translational relevance of the mechanism.

Personal highlights

Lung tumours selectively engage vagal NPY2R/TRPV1 sensory neurons: using anterograde tracing and single‑cell RNA‑seq, the authors show that lung adenocarcinoma recruits and transcriptionally reprograms a specific subset of vagal sensory neurons that express NPY2R and TRPV1, while sparing neighbouring P2RY1⁺ neurons.
A functional vagal‑sensory‑to‑sympathetic brainstem circuit drives tumour progression: chemogenetic activation and ablation experiments demonstrate that vagal NPY2R/TRPV1 afferents signal to glutamatergic RVLM pre‑sympathetic neurons, which then elevate sympathetic tone in the tumour microenvironment, a pathway that is essential for tumour growth.
β₂‑adrenergic signalling in alveolar macrophages is the key immunosuppressive effector: sympathetic‑derived noradrenaline acts through ADRB2 receptors on tumour‑associated alveolar macrophages to upregulate ARG1 and downregulate MHC‑II, promoting an immunosuppressive phenotype that blunts T‑cell responses.
Genetic and pharmacological interruption of the axis unleashes anti‑tumour immunity: ablation of vagal NPY2R/TRPV1 neurons, chemogenetic silencing of RVLM neurons, or ADRB2 blockade all restore CD8⁺ and CD4⁺ T‑cell activity, reduce tumour burden, and extend survival, effects that are abolished by alveolar‑macrophage depletion or T‑cell depletion.
Clinical relevance: a combined vagal‑sympathetic gene signature predicts poor outcomes in human lung cancer: analysis of TCGA data reveals that patients with high expression of both vagal‑sensory and sympathetic‑nerve marker genes exhibit shorter survival and lower CD8⁺ T‑cell signatures, mirroring the mouse mechanistic findings.

Why should we care?

This hwork uncovers a hijacked interoceptive circuit that allows tumours to “talk” to the brain and, in turn, receive immunosuppressive instructions back. Beyond adding a new layer to cancer neuroscience, the study points to actionable therapeutic targets: disrupting the vagal‑sensory input (e.g., via TRPV1 inhibitors), dampening sympathetic output, or blocking β₂‑adrenergic signalling on macrophages could each potentially restore anti‑tumour immunity. For clinicians, it reinforces the potential of beta‑blockers as adjuncts to immunotherapy, while for neuro‑immunology researchers, it provides a mechanistic blueprint for how sensory‑afferent circuits can shape the tissue immune landscape in chronic disease

How different AI models understand cells differently

Zhao, Y. et al. bioRxiv (2026). https://doi.org/10.64898/2026.01.29.702682

The paper in one sentence

Researchers introduce scGeneLens, an interpretability framework that reveals how different single-cell foundation models encode distinct biological priors, with scFoundation focusing on cell-type identity and scGPT on cross-condition pathway activity.

Summary

The paper presents scGeneLens, a model-agnostic framework designed to interpret how Transformer-based single-cell foundation models (scFMs) process gene expression data. By replacing dense attention with a block-sparse attention mechanism, the method makes attention patterns interpretable and traceable across layers. The authors apply scGeneLens to two leading models, scFoundation and scGPT, and uncover fundamentally different inductive biases: scFoundation organizes representations around cell-type marker genes, enabling strong cell-type separation, while scGPT emphasizes shared pathway activity, leading to representations that generalize across biological conditions. The framework also uses integrated gradients to disentangle the contributions of gene identity versus expression level and traces gene influence propagation across layers, offering a unified lens for model comparison and future design.

Personal highlights

Block-sparse attention for interpretable gene–gene dependencies: scGeneLens restructures dense self-attention into a compact set of dominant gene–gene relations using block-level sparsification, making attention patterns analyzable without compromising representational capacity.
Comparative model profiling reveals distinct biological priors: applied to scFoundation and scGPT, the framework exposes systematic differences, scFoundation anchors on cell-type marker genes for strong identity separation, while scGPT prioritizes pathway-level activity for cross-condition generalization.
Integrated gradients disentangle gene identity vs. expression signals: by attributing model outputs to static gene embeddings and expression value embeddings, the method quantifies whether representations are driven more by which genes are present or by how highly they are expressed.
Gene influence propagation traces attention flow across layers: using a cross-layer propagation scheme, scGeneLens visualizes how attention to specific genes accumulates with depth, revealing model-specific convergence onto small subsets of biologically meaningful genes.
Model-agnostic, architecture-preserving interpretability: the framework can be applied to various scFMs with minimal modification, offering a standardized way to probe internal representations and compare inductive biases across architectures.

Why should we care?

scGeneLens opens a window into how these single-cell foundation models “think” about biology. By revealing that models with similar architectures can learn fundamentally different representations, some tuned for cell identity, others for pathway activity, this work helps researchers select and design models based on the biological questions they want to answer.

Other papers that peeked my interest and were added to the purgatory of my “to read” pile

Thanks for reading.

Cheers,

Seb.

Sebcentrism

Discussion about this post

Ready for more?