Weekly reads 3/08/26

How tumors talk to their surroundings, and how we’re learning to listen

Aug 09, 2026

This weeks reads focus on how we push the boundaries of how we model, measure, and manipulate biological systems, from spatial foundation models that capture tumor-microenvironment crosstalk to dietary paradoxes rewiring cancer cachexia. TMEformer introduces a transformer-based framework to predict tumor cell responses by integrating spatial transcriptomics, while Cross et al. reveal how a high-fat diet worsens cachexia in Lkb1-mutant lung cancer via local PGE₂ signaling to sensory neurons. On the technical front, PIP-ATAC-seq and PIP-Multiome-seq democratize single-nucleus multiomics with microfluidic-free, cost-effective workflows, and Cell-JEPA shifts representation learning toward latent-space prediction to improve robustness in noisy single-cell data. Meanwhile, PULSAR bridges molecular, cellular, and multicellular scales to classify diseases and simulate immune perturbations, and AIR-seq leverages the antiviral NHC to measure RNA dynamics without chemical conversion. Finally, scVision reframes single-cell biology as a vision problem, encoding transcriptomes as gene-expression images to unlock new interpretability and perturbation capabilities.

Preprints/articles that I managed to read this week

Mapping Tumor-Microenvironment dependencies with TMEformer: A spatial foundation framework enabling in silico perturbation

Li et al. bioRxiv (2026). 10.64898/2026.05.17.725770

The paper in one sentence

TMEformer is a tumor microenvironment-aware spatial foundation model that integrates tumor-intrinsic transcriptional programs with local microenvironmental signals to enable predictive and perturbable modeling of tumor ecosystems.

Summary

This study introduces TMEformer, a transformer-based framework that leverages high-resolution spatial transcriptomics to model tumor cells within their spatial microenvironment. By explicitly incorporating signals from neighboring cells, TMEformer captures tumor-microenvironment interactions and enables in silico perturbations to predict cellular responses. Validated across diverse prostate cancer cohorts, the model outperforms baseline approaches (e.g., Geneformer) in recapitulating key tumor transitions, such as neuroendocrine differentiation and castration resistance. It systematically identifies tumor-intrinsic transcription factors (e.g., SOX2, ASCL1, TWIST1) and TME-derived ligands (e.g., IL-23, SPP1, CD274) that drive disease progression. Additionally, TME-derived embeddings enhance spatial stratification of tumor cells, aligning more closely with pathological architecture. Notably, the framework generalizes across cancer types (prostate, breast, ovarian) and spatial platforms (Xenium, MERFISH), suggesting broad applicability.

Personal highlights

Spatial ecosystem modeling: Explicitly integrates spatial context into tumor modeling, addressing a critical gap in cell-centric approaches by capturing tumor-microenvironment interactions.
Superior predictive performance: Outperforms Geneformer-derived baselines in recapitulating tumor transitions tightly coupled to microenvironmental regulation.
Driver discovery: Recovers established regulators (e.g., SOX2, ASCL1) and identifies novel candidates (e.g., TWIST1) through large-scale in silico perturbation screening.
Microenvironmental mechanisms: Uncovers immunosuppressive myeloid states, Treg signaling, and fibroblast-derived factors as key contributors to neuroendocrine differentiation and therapy resistance.
Cross-cancer generalizability: Demonstrates applicability across multiple cancer types and spatial platforms, capturing conserved regulatory principles.

Why should we care?

This work models tumours as spatially organized ecosystems where cellular behavior emerges from structured interactions, and not as isolated cells. In this manner, it offers a systematic way to explore how tumors interact with their surroundings to drive progression and resistance, potentially accelerating the discovery of therapeutic targets. However, the model is purely computational and requires experimental validation of its predictions. In addition, it includes fixed spatial neighborhood definitions and reliance on transcriptomic data alone, which may overlook dynamic or long-range interactions and other molecular modalities (e.g., proteomics, chromatin accessibility)

A dietary switch promotes sensory neuron-dependent cancer-associated cachexia

Cross et al., Science (2026). 10.1126/science.adz4196

The paper in one sentence

In Lkb1-mutant lung cancer, a high-fat diet paradoxically worsens cachexia through tumor-derived prostaglandin E₂ (PGE₂) signaling to sensory neurons, rather than circulating factors.

Summary

Using genetically engineered mouse models (GEMMs) of Kras-driven lung adenocarcinoma, this study investigates how tumor genotype and diet interact to drive cancer-associated cachexia (CAC). The authors demonstrate that Lkb1 mutations promote cachexia, characterized by weight loss, anorexia, and reduced activity, while P53 or Cdkn2a/2b mutations do not. Surprisingly, switching to a high-fat diet (HFD) exacerbates cachexia in Lkb1-mutant models, accelerating weight loss and reducing survival despite the diet’s caloric density. Mechanistically, the study identifies local tumor-derived PGE₂ as the key driver of cachexia, acting on sensory neurons innervating the lung to induce sickness behaviors. Genetic ablation of PGE₂ synthesis (Ptges knockout) or inhibition of sensory neuron signaling (via vagotomy or chemogenetic tools) ameliorates cachexia, while clinical data show elevated PGE₂ in the bronchoalveolar lavage fluid (BALF) of cachectic NSCLC patients. The work challenges the prevailing dogma that CAC is driven primarily by circulating factors (e.g., IL-6, GDF15) and instead highlights a critical role for local tumor-nerve signaling.

Personal highlights

Genotype-specific cachexia: Lkb1 mutations in lung cancer drive cachexia and anorexia, unlike P53 or Cdkn2a/2b mutations, revealing a genetic basis for susceptibility to metabolic dysfunction.
Dietary paradox: High-fat or ketogenic diets, typically used to combat weight loss, worsen cachexia in Lkb1-mutant models, reducing survival without increasing tumor burden.
Local PGE₂ as the culprit: Tumor-derived PGE₂, not circulating cytokines, mediates cachexia by signaling through sensory neurons, as shown by rescue experiments with Ptges knockout or COX-2 inhibition.
Sensory neuron dependency: Ablation of vagal sensory neurons or chemogenetic inhibition of lung-innervating neurons restores food/water intake and attenuates weight loss, establishing a neural mechanism for cachexia.
Clinical correlation: Cachectic NSCLC patients exhibit elevated PGE₂ in BALF, and LKB1-mutant human tumors show increased PTGES expression, supporting the translational relevance of the findings.

Why should we care?

This works demomstates that local tumor-nerve interactions, rather than systemic circulating factors, can drive cachexia, or at least in LKB1-mutant lung cancer. The counterintuitive finding that high-fat diets exacerbate cachexia in this context underscores the need for precision nutrition in oncology, as blanket dietary recommendations may harm subsets of patients. Therapeutically, the work identifies actionable targets: PGE₂ synthesis (COX-2/Ptges pathway) and sensory neuron signaling, both of which are druggable. For example, NSAIDs (e.g., aspirin) or omega-3 fatty acid supplementation (fish oil) partially rescued cachexia in mice, suggesting repurposable strategies. However, the study’s scope is limited to Lkb1-mutant models, and while human data are supportive, they remain correlative.

A Simple, Cost-Effective, High-Throughput Method for Measuring Chromatin Accessibility and Gene Expression in Single Nuclei

Luo and Greenleaf, bioRxiv (2026). 10.64898/2026.06.29.735326

The paper in one sentence

The study introduces PIP-ATAC-seq and PIP-Multiome-seq, microfluidic-free, droplet-based methods for high-throughput, cost-effective single-nucleus chromatin accessibility and multiomic profiling, offering a scalable alternative to existing technologies.

Summary

The authors present Particle-templated Instant Partitioning (PIP) as a foundation for two new assays: PIP-ATAC-seq (for chromatin accessibility) and PIP-Multiome-seq (for simultaneous chromatin accessibility and gene expression profiling). Unlike current methods: plate-based (low-throughput, costly), microfluidics-based (expensive, specialized equipment), or split-pool (technically complex): PIP-seq uses uniformly sized beads and simple vortexing to generate monodispersed droplets, eliminating the need for microfluidic hardware. The workflows are straightforward, affordable, and produce high-quality data comparable to gold-standard platforms like 10x Genomics, but at a fraction of the cost and with higher throughput (2,000–100,000 cells per sample). While PIP-Multiome-seq shows slightly lower unique fragment counts than 10x Multiome-seq, it compensates with higher TSS enrichment, indicating reduced background noise. The approach’s simplicity, scalability, and cost-efficiency (up to 10x cheaper) make it a compelling option for labs without specialized single-cell infrastructure.

Personal highlights

Microfluidic-free partitioning: Uses bead-templated vortexing to create droplets in minutes, removing the need for expensive microfluidic devices while maintaining high throughput and data quality.
Cost and scalability: 10x cheaper than 10x Genomics, with the ability to process 2,000–100,000 cells per sample, democratizing access to single-cell multiomics.
High data quality: Comparable to 10x Genomics in terms of unique fragments, gene counts, and TSS enrichment, with lower background noise in multiomic assays.
Cross-contamination control: Minimal cross-species contamination in species-mixing experiments (e.g., <0.5% in RNA libraries), confirming high single-nucleus purity.
Versatility: Successfully applied to frozen tissues (e.g., mouse brain), identifying 16 distinct cell types with concordant RNA and ATAC clustering, and adaptable to other single-cell assays (e.g., CUT&TAG, DNA methylation).

Why should we care?

The authors tackle here the high cost and techncial complexity of scATAC and multiome-seq from 10x Genomics. While PIP-seq matches 10x in many metrics, its slightly lower unique fragment counts in multiome mode and reliance on kit-based protocols (which may still require optimization) mean it is not a universal replacement for all use cases. Additionally, the throughput per run (while high) may not yet match the absolute scale of some industrial platforms. Nonetheless, for labs seeking a balance of affordability, accessibility, and performance, PIP-seq represents a practical step forward

Cell-JEPA: Latent Representation Learning for Single-Cell Transcriptomics

ElSheikh et al., arxiv (2026). https://arxiv.org/abs/2602.02093

The paper in one sentence

Cell-JEPA introduces a joint-embedding predictive architecture that shifts single-cell representation learning from reconstructing sparse, noisy gene expression counts to predicting robust cell-level embeddings in latent space, improving zero-shot transferability and absolute-state prediction.

Summary

Single-cell transcriptomics enables high-resolution studies of cellular heterogeneity but suffers from extreme sparsity and technical noise, with dropout rates often exceeding 90%. Existing foundation models (e.g., scGPT, Geneformer) address this by reconstructing masked gene expression, but this can inadvertently encode measurement artifacts as biological signal. Cell-JEPA proposes a complementary approach: a joint-embedding predictive architecture (JEPA) that learns by predicting cell-level embeddings from partial observations, forcing the model to capture stable cellular programs rather than noisy inputs. The framework builds on scGPT with a student-teacher transformer architecture, where the student encoder processes masked inputs and predicts the teacher’s latent representations (updated via exponential moving average). Pre-trained on 800,000 human kidney cells, Cell-JEPA demonstrates superior performance in zero-shot cell-type clustering (36% relative AvgBIO improvement over scGPT: 0.72 vs. 0.53) and perturbation-response prediction (e.g., +24% Pearson correlation on the Norman dataset). However, it does not consistently improve delta-based metrics (effect-size estimation), suggesting a trade-off between recognizing cellular states and modeling their changes.

Personal highlights

Latent-space prediction for robustness: By predicting cell embeddings rather than raw counts, Cell-JEPA reduces sensitivity to dropout artifacts, learning features that generalize better to unseen data.
Zero-shot transferability: Achieves a 36% relative improvement in AvgBIO for cell-type clustering without task-specific fine-tuning, indicating stronger representation of stable biological structure.
Hybrid objective: Combines gene-level reconstruction (to preserve fine-grained expression information) with JEPA-style latent prediction (to enforce semantic consistency), balancing fidelity and robustness.
Perturbation insights: Improves absolute post-perturbation state prediction but struggles with delta-based metrics, revealing that state recognition and change modeling are complementary challenges.

Why should we care?

This work adresses how to learn how to learn meaningful representations from data dominated by technical noise. By demonstrating that latent-space prediction can outperform reconstruction-only approaches in zero-shot settings, it suggests that focusing on abstract cellular states rather than raw measurements may be key to building more generalizable models. However, the study also exposes an important trade-off: while Cell-JEPA excels at recognizing cell states (clustering, absolute-state prediction), it does not improve the prediction of changes in cell state following perturbations. This implies that representation learning and perturbation modeling may require distinct architectures or objectives, a nuance often overlooked in the rush to build “foundation models” for biology.

PULSAR: a Foundation Model for Multi-scale and Multicellular Biology

Pang et al. bioRxiv (2025). 10.1101/2025.11.24.685470

The paper in one sentence

PULSAR is a multi-scale foundation model that integrates molecular, cellular, and multicellular biology to enable disease classification, clinical prediction, and in silico simulation of immune perturbations from single-cell transcriptomes.

Summary

This study introduces PULSAR (Patient Understanding Leveraging Single-cell universAl Representation), a hierarchical foundation model that connects three biological scales: molecular (using ESM2 protein embeddings), cellular (using Universal Cell Embeddings), and multicellular (using a Multicellular Transformer) to generate unified donor-level representations from single-cell RNA sequencing (scRNA-seq) data. Trained on 36.2 million cells from 6,807 donors, PULSAR demonstrates state-of-the-art performance in disease classification (0.954 mean accuracy across 41 studies and six disease categories), zero-shot prediction of plasma proteomics, and forecasting clinical events such as rheumatoid arthritis onset and flu vaccine responsiveness. The model also enables generative simulation of cytokine perturbations across biological scales and provides interpretable insights into cell-type-specific drivers of disease, such as plasma cell activity in severe COVID-19. A reference database (DONOR×EMBED) with 2,804 donors supports rapid disease classification via k-nearest neighbors search.

Personal highlights

Multi-scale biological integration: PULSAR’s hierarchical architecture explicitly links molecular, cellular, and multicellular scales, preserving the natural organization of biological systems.
Population-scale training: The model leverages 36.2 million cells from 6,807 donors, capturing human immune diversity and enabling robust, generalizable representations.
High-accuracy disease classification: Achieves 0.954 mean accuracy across 41 independent studies and six disease categories, with strong external validation (0.862 accuracy).
Generative and predictive capabilities: Simulates cytokine perturbations across physical scales and predicts future clinical events (e.g., rheumatoid arthritis conversion) from baseline immune profiles.
Interpretable disease mechanisms: Attention analysis reveals cell-type-specific signatures, such as plasma cell and cytotoxic T cell involvement in COVID-19 severity, linking model predictions to biological mechanisms.

Why should we care?

PULSAR demonstrates that foundation models can bridge the gap between molecular biology and clinical phenotypes, offering a scalable framework for precision medicine. By connecting single-cell transcriptomics to donor-level outcomes, it enables early disease detection (e.g., identifying at-risk individuals for rheumatoid arthritis) and personalized treatment strategies (e.g., predicting cytokine responses). The model’s ability to generalize across diverse datasets and diseases suggests it captures fundamental biological principles rather than dataset-specific artifacts. However, its reliance on dissociated single-cell data ignores spatial context, which is critical for solid tissues, and the focus on transcriptomics omits other important layers, such as genetic variation, immune receptor repertoires, or clonal expansion. While the clinical translation of such models is promising, it will require addressing these limitations, validating predictive power in prospective studies, and integrating additional data modalities. The work underscores both the potential and the challenges of using AI to unify multi-scale biological data for actionable medical insights.

Analog intrinsic recoding measures RNA dynamics without chemical conversion

Hansen et al. bioRxiv (2026). 10.64898/2026.07.27.741041

The paper in one sentence

AIR-seq uses the antiviral compound NHC (N4-hydroxycytidine) to mark newly synthesized RNA with intrinsic mismatches, enabling RNA dynamics measurements in standard bulk and single-cell sequencing workflows without chemical conversion or enrichment.

Summary

This study introduces Analog Intrinsic Recoding Sequencing (AIR-seq), a method that repurposes the base-pairing ambiguity of N4-hydroxycytidine (NHC), the active metabolite of the COVID-19 antiviral molnupiravir, to directly mark newly synthesized RNA. Unlike existing metabolic labeling methods (e.g., TimeLapse-seq), AIR-seq eliminates the need for post-labeling chemical conversion or enrichment, which can damage RNA and complicate workflows. The authors demonstrate that NHC incorporation into RNA by RNA polymerase II generates C-to-T and T-to-C mismatches during reverse transcription, allowing the distinction of new and old RNA in standard RNA-seq libraries.

Personal highlights

Chemical-free RNA dynamics: AIR-seq leverages the intrinsic base-pairing ambiguity of NHC to introduce C-to-T and T-to-C mismatches in newly synthesized RNA, eliminating the need for chemical conversion or enrichment, a major bottleneck in current metabolic labeling methods.
Compatibility with standard workflows: The method integrates seamlessly with standard bulk and single-cell RNA-seq protocols, including droplet-based platforms, making it broadly accessible for routine transcriptomic studies.
Preserved RNA integrity: Unlike periodate-based conversion methods (e.g., TimeLapse-seq), NHC treatment maintains high RNA integrity (RIN 9.2–9.6), minimizing artifacts and improving data reliability.
Single-cell resolution of RNA kinetics: AIR-seq resolves cell-cycle-dependent RNA dynamics at hourly resolution, revealing temporal offsets between new, total, and old RNA that improve trajectory inference (e.g., RNA velocity) and uncover hidden regulatory programs (e.g., OXPHOS gene coordination).
Clinical potential: Since NHC is the active form of the FDA-approved antiviral molnupiravir, AIR-seq raises the possibility of in vivo RNA dynamics measurements in clinical or preclinical settings, pending further validation.

Why should we care?

The key takeaway of this work is that RNA abundance alone is a poor proxy for gene activity. Cells regulate gene expression not just by controlling how much RNA is made, but also by tuning how quickly it is degraded. AIR-seq provides a simpler, more robust tool to study these processes, which could accelerate discoveries in fields ranging from cancer biology to developmental dynamics. However, the reliance on NHC, whose long-term effects on cellular physiology are not fully understood, warrants caution. While the method is elegant, its broad adoption will depend on further validation, particularly in complex tissues and in vivo systems where NHC’s off-target effects (e.g., mutagenicity) could complicate interpretations. Nonetheless, AIR-seq represents a significant step toward making RNA dynamics a standard dimension of transcriptomic analysis.

A vision foundation model for single-cell biology via spatial gene cartography

Yesiloglu et al., arxiv (2026). https://arxiv.org/abs/2607.14163

The paper in one sentence

scVision reframes single-cell representation learning as a vision problem by encoding each cell as a continuous gene-expression image, preserving both quantitative expression levels and biological gene relationships.

Summary

This work introduce scVision, a vision foundation model that transforms single-cell transcriptomics data into images using optimal transport to spatially arrange genes based on co-expression patterns. Unlike token-based models that discretize expression values and treat genes as unordered, scVision preserves the continuous nature of gene expression and the relational structure among genes. The model, based on a vision transformer (ViT-base, ~86M parameters), is pretrained via masked image modeling on 72 million human cells from public atlases. In zero-shot evaluations across six independent held-out studies: kidney cortex, ovary, focal cortical dysplasia cortex, Crohn’s disease ileum, human retina, and a multi-organ reference, scVision outperforms existing foundation models (scGPT, scFoundation, Geneformer) and classical baselines in cell-type annotation, gene-program discovery, and multi-study integration. The spatial organization also improves interpretability, with attention maps revealing biologically coherent gene programs (e.g., proximal tubule injury in kidney, unfolded-protein response in ovarian pericytes), and enables unique operations like spatial masking of gene neighborhoods for in silico perturbation studies.

Personal highlights

Novel spatial representation: Genes are assigned to fixed positions on a 104×104 lattice using Gromov-Wasserstein optimal transport, so co-expressed genes become spatial neighbors, turning each cell’s transcriptome into an image where co-regulation appears as local texture and cellular identity as global pattern.
Large-scale pretraining: A ViT-base encoder is trained via masked image modeling on 72M human cells, creating one of the largest pretrained models for single-cell analysis, with end-to-end encoding throughput 37–300× faster than token-based models.
Zero-shot superiority: Across all six held-out atlases, scVision achieves the highest balanced accuracy and present-class macro-F1 for cell-type annotation without task-specific fine-tuning, including in challenging settings like rare cell types and multi-organ references.
Interpretable and transferable biology: Attention maps recover cell-type-specific gene programs (e.g., p53 signaling in microglia, TREM2-DAP12 inflamed-stroma in Crohn’s disease) without pathway supervision, and these programs recur across tissues (e.g., a four-gene myeloid program shared across five organs).
Label efficiency: One labeled cell per type with scVision often outperforms other foundation models given 50 labeled cells, reducing annotation effort by up to 50-fold in some settings.

Why should we care?

The key takeaway is that how we encode biological data can be as transformative as the models we train on it. The spatial gene layout unlocks new analytical possibilities, such as perturbing entire gene programs in silico by masking neighborhoods of co-regulated genes, an operation with no direct counterpart in token-based models. However, the approach is not without limitations: the pan-tissue gene layout may not fully capture tissue-specific relationships, and performance degrades with shallow sequencing depth, highlighting the need for further refinement. Ultimately, scVision offers a compelling alternative to tokenization, showing that connecting single-cell biology to modern computer vision can retain more biological signal while opening the door to novel analyses. The work underscores that foundation models in biology must be evaluated not just on scale, but on how faithfully they preserve the underlying relationships that make the data meaningful.

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?