Journal Club: Reading Cancer's Signatures from Stress, Spectra, and cfDNA

· 10 min read · Paper Feed

Journal Club: Reading Cancer's Signatures from Stress, Spectra, and cfDNA

Conceptual illustration: Reading Cancer's Signatures from Stress, Spectra, and cfDNA

Introduction

Cancer leaves traces. You find them in stressed mitochondria that rewire their metabolism under pressure, in the genomes of tumor cells as patterns of point mutations that betray specific mutagenic processes, and in the bloodstream as vanishingly rare fragments of tumor-derived cell-free DNA. Each of these "signatures" carries information about what a cancer has experienced, where it came from, and how it might be caught early. But signatures are only as useful as our ability to read them correctly.

This journal club looks at three recently published studies that, at first glance, sit in different subfields of cancer biology. Kelley et al. (2025) build integrated transcriptomic and metabolomic maps of mitochondrial stress and introduce a deconvolution tool called SQUID. Kazachkova et al. (2026) re-examine a previously hidden mutational signature in colorectal cancer and argue that what was called SBS18 is actually two processes in disguise. Chen et al. (2026) simulate hundreds of millions of cfDNA whole-genome sequencing samples to map the tradeoff between sequencing coverage and per-base error rate. What ties them together is a common question: How do we systematically characterize complex biological signals so they can be detected, deconvoluted, and acted upon?

Each paper tackles a different layer of this problem: cellular stress physiology, the mutational history written in tumor genomes, and the engineering of liquid biopsy assays. Their solutions share a methodological ethos: when signatures are subtle or entangled, you need richer data, sharper models, or better quantification of your detection limit.

Paper 1: Decoding Mitochondrial Stress with SQUID

Key Findings

  • Using primary human fibroblasts and a panel of mitochondrial stressors, Kelley and colleagues generated integrated transcriptional and metabolomic profiles that distinguish between metabolic, oxidative, and proteotoxic mitochondrial insults.
  • They developed SQUID (Stress Quantification Using Integrated Datasets), a computational tool that deconvolutes these multi-omic signatures and can be applied to external datasets.
  • Applied to the TCGA PanCancer Atlas, SQUID uncovered a signature consistent with pyruvate import deficiency in IDH1-mutant glioma, pointing to a metabolic vulnerability tied to mutant isocitrate dehydrogenase activity.
  • Different stressors evoke overlapping but distinguishable adaptive programs, with shared ATF4/ATF5-driven branches and stressor-specific metabolic rewiring.

Methodology & Study Design

The authors treated primary human fibroblasts with a curated panel of mitochondrial inhibitors (targeting the electron transport chain, mtDNA replication, and protein import) and performed bulk RNA-seq alongside steady-state LC-MS metabolomics. They then used a multi-omic integration framework to identify shared and unique responses, encoding the result into a model that SQUID can score against any new transcriptomic dataset. The TCGA application shows cross-platform transferability: SQUID inferences remain interpretable on bulk tumor RNA-seq data despite being trained on fibroblasts.

Why this matters

This study fills a stubborn gap. "Mitochondrial stress" has often been treated as a single state. By tightly coupling metabolite measurements to gene-expression programs, Kelley et al. produce a mechanistically interpretable taxonomy of mitochondrial stress rather than a generic stress score. SQUID is a step toward quantitative, repurposable stress phenotyping in large patient cohorts, relevant to aging, neurodegeneration, and tumor metabolism.

Paper 2: A Missed Mutational Signature in Colorectal Cancer

Key Findings

  • In a re-analysis of 2,616 microsatellite-stable (MSS) colorectal cancers from three independent whole-genome cohorts, the authors confirm that a previously decomposed de novo signature ("SBS_D") is reproducibly distinct from SBS18, the canonical ROS-associated signature.
  • SBS_D was provisionally assigned COSMIC identifier SBS111 based on its consistent prevalence and pattern across cohorts.
  • The SBS_D mutational pattern aligns more closely with signatures tied to DNA polymerase δ (POLD1) proofreading defects and mismatch-repair-related processes, even though tumors carrying SBS_D largely lack pathogenic POLD1 exonuclease or canonical MMR mutations.
  • The authors therefore propose that SBS_D reflects noncanonical DNA repair infidelity emerging late in tumor evolution rather than inherited or early driver repair defects.

Methodology & Study Design

The team applied non-negative matrix factorization (NMF) and related deconvolution approaches to whole-genome sequencing data, comparing naïve (unconstrained) decompositions with prior fits that forced SBS_D into SBS18. They validated the signature across three MSS CRC cohorts to assess reproducibility, then compared SBS_D's trinucleotide context to known COSMIC reference signatures (SBS1, SBS5, SBS15, SBS18, SBS36, and others). By examining the absence of canonical driver mutations in POLD1 and MMR genes in SBS_D-high tumors, they infer a mechanistically novel, late-arising source of mutagenesis.

Why this matters

Mutational signatures underpin precision oncology: they help identify hereditary cancer syndromes (MUTYH, Lynch), reveal mutagen exposures (SBS88 from pks+ E. coli), and classify tumors. If SBS_D is real and common, then a sizeable chunk of MSS CRCs may have been assigned the wrong dominant mutational process. Clinically, that means misattribution of DNA repair deficiency, with downstream consequences for therapy selection and surveillance. The work is also a striking example of how statistical decomposition can quietly merge two distinct processes into one.

Paper 3: How Low Can You Go? Coverage, Error Rates, and ctDNA Detection

Key Findings

  • The authors built a large-scale simulation framework generating roughly 480 million in silico cfDNA whole-genome sequencing samples across 10 cancer types, each with realistic mutational burdens derived from tumor WGS.
  • For high-mutation-burden cancers (e.g., melanoma, lung), tumor fractions below 0.1% can be reliably detected with only ~3× genome coverage when low-error methods are used.
  • For low-mutation-burden cancers, achieving comparable sensitivity requires roughly 6× higher coverage (≥18×).
  • Improving base quality from Q30 to Q55 at 30× coverage extends detection down to tumor fractions of ~1 × 10⁻⁵, the boundary regime relevant for minimal residual disease (MRD) monitoring.
  • The lowest detectable tumor fraction varies substantially by cancer type even at identical sequencing parameters, driven almost entirely by mutational load.

Methodology & Study Design

Chen and colleagues simulate cfDNA by overlaying somatic SNVs from real tumor WGS profiles onto in silico background genomes, then modeling the asymmetric noise structure of cfDNA WGS (mono-/di-nucleotide errors, strand bias, sequencing-platform-specific error profiles). They scan coverage and quality-score parameter grids to map detection limits as a function of tumor fraction, mutation burden, and platform error rate. The code is publicly available at https://github.com/UMCUGenetics/cfdetect/tree/main.

Why this matters

Liquid biopsy for MRD and recurrence monitoring is now bottlenecked less by biology than by signal-to-noise engineering. Chen et al. make this quantitative. Their framework gives assay designers a defensible answer to: Is it worth paying for an ultra-low-error platform, or should I just sequence more? The answer depends on cancer type, and by extension on which mutational processes dominate that tumor's genome, a thread that loops back to Papers 1 and 2.

Synthesis & Discussion

How these papers fit together

Each paper can be read as a layer in a stack. Paper 1 gives a high-resolution view of how cells respond to stress, including the metabolic and transcriptional rewiring that can themselves be mutagenic (e.g., ROS bursts, nucleotide imbalances). Paper 2 asks whether one mutational readout usually attributed to oxidative damage, SBS18, is being read correctly in tumor genomes, and concludes that the answer is partly no. Paper 3 asks how detectably those genomic scars survive the journey into the bloodstream, and shows that for low-mutation cancers, detecting subtle late-arising processes like SBS_D may require both higher coverage and tighter error control.

Together, they trace a loop: a biological process → a mutational signature in tumors → a limit-of-detection question in circulating DNA. None of the three papers connects mitochondrial stress to SBS_D, and Kazachkova et al. point to noncanonical repair infidelity instead; whether SQUID-style stress scores track SBS_D at all is our speculation and nothing more.

Tensions and disagreements

There is a quiet tension in how the three papers treat "shared vs specific" signatures. Paper 1 explicitly distinguishes overlapping yet distinguishable mitochondrial stress programs, treating spectral similarity as a modeling challenge to be resolved. Paper 2 shows that statistical decomposition can merge two processes, and uses that as the paper's central cautionary tale. Paper 3 sidesteps the biology but reveals how mutational burden alone creates order-of-magnitude detectability gaps, implying that biology is doing as much work as assay engineering in cfDNA-based MRD.

A second tension: Papers 1 and 2 both find that late, noncanonical processes matter: noncanonical DNA repair infidelity in tumors, and stress programs that don't fit the textbook UPRmt. Paper 3 suggests such late, subtle processes may be the hardest to detect in cfDNA, especially in low-mutation cancers where the signal is already sparse.

Emerging patterns

  1. Integration beats single-omics. All three studies depend on combining multiple data types or dimensions to see structure that a single view would blur.
  2. Statistical deconvolution is a sharp tool, easily dulled. Both SQUID's multi-omic integration and the SBS18/SBS_D re-analysis turn on whether one allows the model enough flexibility to separate truly distinct processes.
  3. Biology determines the engineering limits. Mutational burden, not just sequencing depth, sets the ceiling on ctDNA sensitivity.
  4. Late, subtle biology is increasingly the frontier. From noncanonical DNA repair to nuanced mitochondrial stress states, what was previously lumped together is now being carefully partitioned.

For Your Lab Meeting

  1. Methodology choice: Paper 2's success hinges on comparing constrained (forced SBS18) vs naïve decompositions. In your own work, where might forcing decompositions into a known reference set be hiding a real signal?
  2. Generalizability: SQUID was trained in primary fibroblasts but applied to TCGA tumors. What assumptions about cell-type specificity of stress programs does this break, and how would you test them?
  3. Cross-paper integration: Could a SQUID-style stress score, computed from tumor RNA-seq, predict SBS_D prevalence in matched WGS data? What confounders would you expect?
  4. Engineering vs biology: If you are designing an MRD assay for low-mutation-burden colorectal cancers, would you prioritize duplex sequencing, more depth, or a multi-analyte approach (methylation, fragmentomics)?
  5. Future direction: What would a "SQUID for mutational signatures" look like: a tool that deconvolutes overlapping mutational processes in individual tumors the way SQUID deconvolutes overlapping stress programs?

Key Terms

  • SQUID (Stress Quantification Using Integrated Datasets): A computational method that scores how strongly an external transcriptomic sample resembles reference mitochondrial stress states derived from integrated RNA-seq + metabolomics data.
  • Mutational signatures (SBS): Characteristic patterns of single-base substitutions inferred from tumor genomes (e.g., SBS18 = ROS-related, SBS111 = newly proposed); modeled using non-negative matrix factorization over trinucleotide contexts.
  • MSS colorectal cancer (microsatellite-stable): Colorectal tumors that retain functional mismatch repair, accounting for ~85% of CRC cases and generally lacking the high mutation burden of MSI tumors.
  • cfDNA / ctDNA: Cell-free DNA circulating in blood plasma; the tumor-derived fraction (ctDNA) is detected via tumor-informed or tumor-naïve approaches and is increasingly used for MRD and recurrence monitoring.
  • Phred quality score (Q-score): A logarithmic measure of per-base sequencing accuracy (Q30 ≈ 99.9% accuracy; Q55 ≈ 99.9997%); a central parameter in the coverage-vs-error tradeoff explored by Chen et al.

References

  1. Liam P. Kelley, Song-Hua Hu, Sarah A. Boswell, Peter K. Sorger, A. Ringel, & Marcia C. Haigis (2025). Integrated analysis of transcriptional and metabolic responses to mitochondrial stress. Cell Reports Methods. https://doi.org/10.1016/j.crmeth.2025.101027

  2. Kazachkova M, Otlu B, Díaz-Gay M, Abbasi A, Moody S, Jiang Z, et al. (2026). Identification and validation of a previously missed mutational signature in colorectal cancer. Nature communications. https://doi.org/10.1038/s41467-026-76472-w

  3. Chen LT, de Ridder J, & Jager M (2026). Large-scale simulation of coverage and error rate tradeoffs for cancer detection in cell-free DNA whole-genome sequencing. Bioinformatics (Oxford, England). https://doi.org/10.1093/bioinformatics/btag460

Cancer Genomics Oncology Tumor