Journal Club: Fibroblast Niches, Spatial Signaling, and Multi-Omics Prognosis in Liver Cancer

August 10, 2026 · 10 min read · Paper Feed

Conceptual illustration: Fibroblast Niches, Spatial Signaling, and Multi-Omics Prognosis in Liver Cancer

Introduction

Hepatocellular carcinoma (HCC) remains one of the deadliest cancers worldwide, in part because its tumor microenvironment (TME) — a tangled mix of malignant hepatocytes, immune cells, endothelial cells, and fibroblasts — actively shapes how the disease progresses, evades therapy, and recurs. For years, single-cell and bulk genomic atlases have cataloged who is in the TME. But the deeper questions are functional: where do these cells interact, what are they exchanging, and can we use those signals to predict outcome or guide treatment?

The three papers in this journal club each tackle a different layer of that problem, yet together they paint a coherent picture. Paper 1 maps the spatial-metabolic geography of HCC and shows that cancer-associated fibroblasts (CAFs) behave as metabolic hubs that drive macrophage polarization. Paper 2 digs into a specific fibroblast state — senescent CAFs — and identifies CTHRC1 as a secreted driver of cancer stemness via the SOX4–CTHRC1–Notch1 axis. Paper 3 zooms out from mechanism to modeling: SeNMo, a self-normalizing multi-omics neural network, integrates five omics layers across 32 cancers to predict survival, classify tumor type, and even estimate tertiary lymphoid structure (TLS) ratios — bridging molecular signals with clinically interpretable immune features.

The through-line: non-malignant cells, especially CAFs, act as control nodes, and learning to read their signals computationally is the next frontier in solid-tumor biology.

Paper 1: A Metabolic "Interface Zone" in HCC Revealed by Spatial Omics and MSI

Key Findings

The authors identify a discrete tumor–immune–CAF "interface" zone in HCC where lactic acid and long-chain polyunsaturated fatty acids (LC-PUFAs) are markedly elevated.
CAFs in this zone upregulate glycolysis and secrete lactate, which in coculture experiments polarizes macrophages toward an immunosuppressive M2 phenotype.
This M2-skewed, lactate-rich niche correlates with tumor invasion, pointing to a metabolic route to immune evasion.
The result reframes CAFs from structural support cells into active metabolic engineers of the immune microenvironment.

Methodology & Study Design

The team built a spatial multiomics platform that combines spatial transcriptomics/omics with single-cell mass spectrometry imaging (MSI), letting them overlay gene-expression programs onto metabolite distributions at near-cellular resolution. They complemented this with in vitro coculture systems to test causality — specifically, whether CAF-derived lactate is sufficient to push macrophages toward M2. Pairing in situ mapping with functional validation is a meaningful upgrade over descriptive spatial atlases.

Significance

Prior spatial HCC studies had mapped cellular neighborhoods but rarely tied them to specific metabolites. By putting lactate and LC-PUFAs on the map, the authors give the field actionable molecular handles — and they validate one of those handles experimentally. This is the kind of work that turns "the TME is heterogeneous" from a truism into a testable, targetable claim.

Paper 2: Senescent CAFs Secrete CTHRC1 to Drive Cancer Stemness

Key Findings

Using the GSE149614 single-cell dataset plus CytoTRACE and cell–cell communication analysis, the authors show that senescent CAFs (SCAFs) in HCC correlate strongly with cancer stemness.
A 10-gene CAF-senescence score (CSscore) — built from CTHRC1, SERPINE1, RNF11, ENG, MARCKSL1, ASAP1, FHL3, LAMB1, CD151, OLFML2B — stratifies patients in TCGA, ICGC, and NODE cohorts.
High CSscore tumors show M2 macrophage and Treg enrichment, lower CD8⁺ T cell and dendritic cell infiltration, higher tumor mutation burden, and worse responses to immunotherapy and transarterial chemoembolization (TACE).
Mechanistically, CTHRC1 is transcriptionally driven by SOX4 in SCAFs and activates Notch1 signaling in tumor cells to promote stemness and metastasis, validated both in vitro and in vivo.

Methodology & Study Design

The study chains together single-cell RNA-seq senescence scoring, machine-learning feature selection (four algorithms), and bulk-cohort validation across three independent databases before moving to the bench: ChIP confirmation of SOX4 binding to the CTHRC1 promoter, plus functional assays for stemness and metastasis. The pipeline is a textbook example of the modern compute-to-bench workflow.

Significance

CAFs are famously heterogeneous, and lumping them together obscures the biology. By isolating the senescent subset and tying it to a single secreted factor with a defined transcription factor upstream and a Notch effector downstream, the authors give the field one of the cleanest CAF→cancer stemness axes described in HCC. It also offers an explanation for why some HCCs grow more aggressive over time — senescent stromal cells accumulate and secrete stemness-promoting ligands.

Paper 3: SeNMo — A Self-Normalizing Neural Network for Pan-Cancer Multi-Omics

Key Findings

SeNMo is a self-normalizing deep neural network that ingests five omics layers — gene expression, DNA methylation, miRNA abundance, somatic mutations, and protein expression — plus clinical variables, and tolerates missing modalities.
Trained on >10,000 TCGA patients across 32 tumor types, it reaches a concordance index of 0.758 for overall survival on a held-out TCGA test set.
External validation: CPTAC lung squamous cell carcinoma cohort (c-index 0.73) and 108-patient Moffitt cohort (c-index 0.66).
After fine-tuning on Moffitt data, SeNMo predicted TLS ratios that aligned with expert annotations (p < 0.05) and separated high- vs low-TLS survival groups.
A single linear head classified primary cancer type with 99.8% accuracy across 33 classes — no backbone changes needed.

Methodology & Study Design

SeNMo uses self-normalizing neural networks (SNNs), which employ scaled exponential linear units (SELU) and a specific initialization scheme to keep activations normalized across layers — useful when integrating modalities of wildly different scales and missingness. The architecture supports modality dropout so the model can be trained and deployed even when one omics layer is unavailable. The downstream heads are deliberately simple (a single linear layer for classification), so the heavy lifting happens in the shared representation.

Significance

Most multi-omics prognostic models are tumor-specific. SeNMo is pan-cancer, modality-robust, and multi-task — a foundation-model-style approach to oncology. The TLS prediction result stands out because TLS ratios are usually estimated by pathologists; a model that recovers them from omics alone opens the door to immune-feature quantification when tissue is scarce or unavailable. For a field juggling sparse, heterogeneous clinical-genomic data, this is a useful kind of architecture.

Synthesis & Discussion

How the papers complement each other

These three studies form a near-complete pipeline from mechanism to biomarker to model:

Layer	Paper	Contribution
Spatial mechanism	Paper 1	Where metabolism happens; CAFs → lactate → M2 macrophages
Stromal signaling	Paper 2	Who is signaling; senescent CAFs → CTHRC1 → Notch1 → stemness
Predictive modeling	Paper 3	How to compute from it all; multi-omics → survival, type, TLS

Read together, they argue that CAFs are central nodes in HCC — controlling metabolism in one paper and stemness in another. That convergence is itself a finding: it suggests fibroblast state (senescent vs. glycolytic vs. quiescent) may be a key variable that explains much of HCC's heterogeneity.

Points of tension and nuance

The papers don't directly contradict each other, but they emphasize different fibroblast functions. Paper 1's CAFs are metabolic hubs secreting lactate; Paper 2's CAFs are senescent signal factories secreting CTHRC1. Are these the same cells in different states, different subsets, or sequential phenotypes? The field doesn't yet know. It's also worth flagging that Paper 3 is pan-cancer and HCC-agnostic — its strength is generalizability, not tissue specificity. Tying its representations back to the CAF mechanisms in Papers 1 and 2 is a future direction.

Emerging patterns

Spatial niches matter. Uniform tumor-wide effects are giving way to interface zones, invasive margins, and fibroblast-rich hotspots as the functionally relevant units of disease.
Fibroblasts are control nodes. Across metabolic, immune, and stemness dimensions, CAFs (and especially their senescent subset) keep emerging as upstream regulators.
The atlas era is ending. The shift from "what cells are present?" to "what are they doing, and can we predict it?" is well underway.

Open questions

Are CAF metabolic and senescent programs coexpressed in the same cells, or do they define distinct subpopulations?
Can spatial metabolic markers (Paper 1) or the CSscore (Paper 2) be fed into architectures like SeNMo (Paper 3) to improve HCC-specific prognostication?
How transferable are these CAF mechanisms across HCC etiologies (HBV, HCV, MASH) and across other cancers?

For Your Lab Meeting

Spatial-mechanistic integration: Paper 1 pairs spatial omics with MSI but stops at lactate. If you had unlimited budget, would you add spatial proteomics or single-cell metabolic flux to disambiguate cause from correlation in the interface zone?
State vs. subset: Senescent CAFs (Paper 2) and glycolytic CAFs (Paper 1) may overlap. What experiment would definitively test whether they are the same cells in different activation states or distinct lineages?
Generalizability of CTHRC1: CTHRC1 is mechanistically clean in HCC. Would you bet on it playing the same role in cholangiocarcinoma, pancreatic cancer, or colorectal cancer? What's your prior?
Model design choices in SeNMo: The authors use SELU-based self-normalizing layers. How would you compare this to transformer-based or graph-neural-network approaches for multi-omics integration? When would each win?
Clinical translation: If you had to deploy one of these findings in a clinical trial tomorrow, which would have the shortest path — a CSscore assay, a lactate-targeted therapy, or a SeNMo-style prognostic report?

Key Terms

Cancer-associated fibroblast (CAF): A stromal cell in the TME that actively shapes tumor metabolism, immune response, and stemness rather than acting as passive support. Highly heterogeneous; subsets include senescent, glycolytic, and myofibroblast-like CAFs.
Senescence-associated secretory phenotype (SASP): The cocktail of cytokines, growth factors, and matrix-remodeling enzymes secreted by senescent cells. Can be tumor-suppressive early but pro-tumorigenic later in the stromal compartment.
Self-normalizing neural network (SNN): A deep architecture using SELU activation and a specific initializer (LeCun normal) that preserves mean and variance across layers, making it well-suited to heterogeneous, possibly missing omics inputs.
Tertiary lymphoid structure (TLS): Ectopic lymphoid aggregates that form in or near tumors; their density correlates with immunotherapy response and survival in many cancers.
CytoTRACE: A computational method that infers a cell's developmental "stemness" or differentiation potency from single-cell RNA-seq data by counting gene counts per cell as a proxy for transcriptional diversity.

References

Panpan Chen, Haoyuan Geng, Bangzhen Ma, Yaqi Zhang, Zihan Zhu, Min Li, et al. (2025). Integrating spatial omics and single-cell mass spectrometry imaging reveals tumor–host metabolic interplay in hepatocellular carcinoma. Proceedings of the National Academy of Sciences of the United States of America. https://doi.org/10.1073/pnas.2505789122
Hai Huang, Wang Peng, Qiaodan Zhou, Yuchong Zhao, Luyao Liu, H. Cui, et al. (2025). Senescent fibroblasts secrete CTHRC1 to promote cancer stemness in hepatocellular carcinoma. Cell Communication and Signaling. https://doi.org/10.1186/s12964-025-02369-8
A. Waqas, A. Tripathi, Sabeen Ahmed, Ashwin Mukund, Hamza Farooq, Joseph O. Johnson, et al. (2025). Self-Normalizing Multi-Omics Neural Network for Pan-Cancer Prognostication. International Journal of Molecular Sciences. https://doi.org/10.3390/ijms26157358

Cancer Genomics Oncology Tumor