Journal Club: From Etiology to Ecosystem — Reframing Cancer Biomarkers Across Mutation, Microbe, and Immune Context
Introduction
For decades, cancer biology has chased single-feature explanations. A mutated oncogene, a viral insert, a microsatellite-instability call, a tumor mutation burden above 10 mutations per megabase—each was, at some point, sold as the biomarker. Three recent papers make the same point from three very different angles: meaningful biomarkers are contextual, multilayered, and ideally tied to mechanism.
The unifying theme of this journal club is the move from descriptive, feature-by-feature biomarker discovery toward etiology-aware, ecosystem-level models of cancer. Goossens and colleagues (2026) anchor mutational signatures in experimentally validated DNA-repair knockouts. Bautista and colleagues (2025) review how microbial communities shape metastatic behavior. Liao and colleagues (2025) show that the predictive value of tumor mutation burden (TMB) for immune checkpoint inhibitor (ICI) response depends on the tumor microenvironment (TME). Read together, these papers tell a coherent story: mutations, microbes, and immune context interact, and only models that respect that interaction will deliver clinically actionable biomarkers.
In this post, we summarize each paper, highlight their methodological choices, and discuss what their convergence means for the next generation of cancer biomarker research.
Paper 1: Supervised NMF Anchors Mutational Signatures in Knockout Cell Lines
Goossens et al., NAR Genomics and Bioinformatics, 2026
Key Findings
- The authors applied supervised non-negative matrix factorization (SNMF) to mutation profiles from human induced pluripotent stem cells (hiPSCs) carrying experimentally introduced knockouts of single DNA-damage-response (DDR) genes.
- The model recovered etiology-aligned representations for homologous recombination repair (HRR), mismatch repair (MMR), and base excision repair (BER) deficiencies—pathways where canonical COSMIC signatures (e.g., SBS3, SBS6/SBS15/SBS26) are well characterized.
- Even under pathway-level supervision, SNMF captured gene-specific BER sub-mechanisms, suggesting that integrating labels gives finer granularity than unsupervised decomposition alone.
- Learned cell-line signatures showed high cosine similarity to COSMIC tumor-derived signatures and enabled high-recall classification of tumors with DDR deficiencies in pan-cancer data.
Methodology & Study Design
The core method is a supervised variant of NMF. Standard NMF decomposes a sample-by-mutation-type count matrix into a signature matrix and an exposure matrix without any biological guidance. SNMF threads pathway- or knockout-level labels into the factorization. The intuition: by forcing each latent component to align with a known etiology, the decomposition is less likely to fragment a single biological process into multiple redundant signatures, or to merge correlated processes.
Key inputs include mutation counts from isogenic hiPSC lines with single-gene knockouts spanning homologous recombination, mismatch repair and base-excision repair. The learned signatures are compared against tumour-derived COSMIC signatures and then used to flag DDR-deficient tumours. Code is available at the lab GitHub (https://github.com/joanagoncalveslab/SNMF).
Significance
This work matters for two reasons. First, it moves mutational-signature analysis away from pattern discovery in the dark toward biologically anchored learning—an important advance because most COSMIC signatures still lack experimentally confirmed etiologies. Second, by recovering BER sub-mechanisms under pathway-level labels, the paper suggests that supervised factorization can reveal within-pathway heterogeneity that pathway-level clinical assays miss. Clinically, this matters: the therapeutic implications of a BRCA1 defect (HRR) versus a MUTYH defect (BER) are very different, but both are often grouped under "DNA repair deficiency."
Paper 2: The Microbiome as Architect of Metastasis
Bautista et al., Frontiers in Medicine, 2025
Key Findings
- This review synthesizes evidence that intra-tumoral, intracellular, and gut-resident microbes modulate multiple steps of the metastatic cascade—epithelial–mesenchymal transition (EMT), invasion, angiogenesis, immune evasion, and pre-metastatic niche formation.
- Microbial signatures correlate with immunoediting and immune escape; specific taxa and microbial metabolites are linked to systemic immunosuppression that supports disseminated tumor cells.
- Spatially resolved multi-omics (spatial transcriptomics, single-cell sequencing, metagenomics) show that microbes are non-randomly distributed within tumors, often co-localizing with hypoxic, immunosuppressive, or fibrotic niches.
- The authors argue that microbial biomarkers (16S rRNA profiles, microbial cell-free DNA, metabolite signatures) could complement existing staging systems, particularly for predicting early metastasis or ICI response.
Methodology & Study Design
As a narrative review, the paper does not introduce a single method. Instead, it surveys five converging domains: (i) microbiome–TME interactions in metastasis, (ii) immunoediting and immune escape, (iii) intratumoral/intracellular bacterial contributions to dissemination, (iv) spatial-multi-omic mapping of microbial niches, and (v) microbial biomarkers predictive of metastatic risk and therapy outcome. Cited evidence spans murine models, patient metagenomic cohorts, and single-cell studies that resolve bacterial reads in tumor scRNA-seq data.
Significance
The paper crystallizes a shift in thinking. Tumors are not sterile, and metastasis—at least in part—appears to be a microbiome-influenced process. The practical implication is that future biomarker panels may need a microbial layer, and therapeutic strategies targeting the microbiome (e.g., fecal microbiota transplant, narrow-spectrum antibiotics, engineered consortia) deserve prospective testing in the metastatic setting. The review also highlights methodological gaps: distinguishing live, replicating intracellular bacteria from contaminated or phagocytosed material remains technically challenging.
Paper 3: TMB Only Predicts ICI Response in the Right Immune Neighborhood
Liao et al., Frontiers in Immunology, 2025
Key Findings
- In a pan-cancer meta-analysis spanning TCGA and multiple anti-PD-1/PD-L1 ICI cohorts, TMB's predictive value varied dramatically across cancer types and was strongly modulated by TME composition.
- In tumors with high CD8⁺ T-cell infiltration and M1-polarized macrophages, TMB retained (or even improved) its predictive power for ICI benefit. In immunosuppressive microenvironments—those dominated by M0 macrophages, activated mast cells, and stromal signatures—TMB alone failed to stratify patients.
- Using LASSO and Cox regression on immunosuppression-related genes (ISRGs), the authors built a 10-gene risk signature that robustly prognosticated overall survival across multiple cohorts.
- RPLP0, the ribosomal protein highlighted as the most robust predictor in the model, was overexpressed in tumors; knocking it down in a bladder cancer mouse model significantly enhanced the efficacy of anti-PD-1 therapy, with elevated serum IFN-γ and TNF-α.
Methodology & Study Design
The pipeline combines bulk transcriptomic deconvolution (CIBERSORT, ESTIMATE) for TME profiling, machine-learning variable selection (LASSO) on ISRGs, and multivariate Cox regression. Validation spans multiple independent ICI-treated cohorts. The RPLP0 arm moves from clinical computational analysis to in vitro and in vivo validation, including a subcutaneous syngeneic model with intratumoral shRNA knockdown plus anti-PD-1 treatment.
Significance
The paper makes a simple but underappreciated point: TMB is necessary but not sufficient. The same TMB can mean opposite things depending on whether the tumor is "hot" (inflamed) or "cold" (excluded). Their risk signature effectively answers, given the patient's TMB, how immune-permissive is the surrounding stroma? The RPLP0 experiments are a particularly attractive translational angle—suggesting that even a ribosomal protein can have non-canonical roles in shaping anti-tumor immunity and may be targetable to convert cold tumors into hot ones.
Synthesis & Discussion
Three Papers, One Story
These three papers converge on a single thesis: single-feature biomarkers are inadequate, and the field needs to move toward mechanism-aware, context-dependent integration.
| Paper | Core Question | Contribution |
|---|---|---|
| Goossens et al. | Is the signature biologically meaningful? | Anchors signatures in experimental knockouts via SNMF |
| Bautista et al. | Does the microbiome drive metastasis? | Synthesizes mechanistic and correlative evidence for microbial oncology |
| Liao et al. | Why does TMB fail in some patients? | Demonstrates TME-dependent TMB utility and proposes an ISRG-based risk score |
Complementary Insights
A natural integrating hypothesis emerges: DNA-repair deficiency (Paper 1) shapes mutational burden (Paper 3), which in turn conditions immune visibility, while microbial ecology (Paper 2) modulates how that visibility translates into effector response. HRR-deficient tumors are known to accumulate more SBS3-like mutations, and microbial metabolites (e.g., short-chain fatty acids) can either promote or suppress anti-tumor immunity. Whether these axes are causally linked remains untested—an open question worth pursuing.
Contradictions and Cautions
- Goossens et al. rely on isogenic cell-line knockouts; tumor signatures arise in heterogeneous backgrounds with multiple co-occurring defects. Transferring SNMF to patient data still requires external validation.
- Bautista et al. flag the difficulty of ruling out contamination in low-biomass microbial studies—a known reproducibility concern.
- Liao et al. show correlations across cohorts but acknowledge that prospective validation in larger, more diverse populations is needed before clinical adoption.
What Patterns Emerge?
- Causality vs. association trade-offs: Paper 1 is causally anchored (knockouts); Papers 2 and 3 are largely associative. The strongest clinical models will probably combine causal anchors with associative layers.
- Mechanism-anchored features outperform random features: SNMF labels > unsupervised NMF; ISRGs > arbitrary genes; microbial pathways > taxonomic reads alone.
- Multi-omic integration is the obvious next move, but it raises real statistical challenges (overfitting, batch effects, multiple-hypothesis burden).
For Your Lab Meeting
- Could SNMF be re-applied to microbial or transcriptomic data? What changes if we treat SNMF as a framework rather than a tool for SNVs only, using CRISPR screen phenotypes or metabolite perturbations as labels?
- Is TMB obsolete as a biomarker? Liao et al. suggest no—TMB needs a TME qualifier. Should clinical reporting standards evolve to always include immune-context annotations?
- What experiments would causally link the microbiome to mutational signatures? Germ-free or gnotobiotic models crossed with repair-deficient backgrounds could directly test whether microbiota modulate mutation accumulation or repair efficiency.
- How do we handle the multiple-hypothesis burden of multi-omic integration? Are there rigorous correction frameworks (or Bayesian hierarchical priors) appropriate for combining microbe, signature, and immune features in one model?
- What are the ethical and equity dimensions of these biomarkers? Many TCGA and ICI cohorts over-represent European-ancestry populations. As microbial signatures in particular vary by geography and diet, how do we ensure generalizability?
Key Terms
- Mutational signature: A characteristic pattern of mutations (e.g., C>A in TpCpN context) attributable to a specific mutagenic or repair-deficient process.
- DDR (DNA Damage Response): The cellular network of pathways—including HRR, MMR, BER, NHEJ—that detect, signal, and repair DNA lesions. Deficiencies in DDR drive both tumorigenesis and therapeutic vulnerability.
- TMB (Tumor Mutation Burden): Total number of nonsynonymous somatic mutations per megabase of coding genome; used as a surrogate for neoantigen load and ICI responsiveness, typically at a threshold of 10 mut/Mb.
- TME (Tumor Microenvironment): The non-cancerous cellular and molecular milieu of a tumor, including immune cells, fibroblasts, vasculature, extracellular matrix, and soluble factors.
- SNMF (Supervised Non-negative Matrix Factorization): A variant of NMF in which auxiliary labels (e.g., pathway annotations) constrain the decomposition, anchoring latent components to known biology.
References
Goossens S, Tepeli YI, Seale C, & Gonçalves JP (2026). Etiology-guided mutational signature learning from DNA repair knockouts in cell lines using supervised NMF. NAR genomics and bioinformatics. https://doi.org/10.1093/nargab/lqag078
Jhommara Bautista, María Paula Fuentes-Yépez, Joseth Adatty-Molina, & A. López-Cortés (2025). Microbial signatures in metastatic cancer. Frontiers in Medicine. https://doi.org/10.3389/fmed.2025.1654792
Wuyuan Liao, Xin-Wei Zhou, Han-cui Lin, Zi-Hao Feng, Xinyan Chen, Yu-Hang Chen, et al. (2025). Interplay between tumor mutation burden and the tumor microenvironment predicts the prognosis of pan-cancer anti-PD-1/PD-L1 therapy. Frontiers in Immunology. https://doi.org/10.3389/fimmu.2025.1557461