Linking Microbial Genotoxins to Tumor Mutational Signatures with Host-Microbiome Multi-Omics

Beyond Taxonomic Dysbiosis: Direct Mutagenesis vs. Secondary Inflammation

The conceptual framework linking the human microbiome to oncogenesis has historically centered on secondary, microenvironment-mediated mechanisms. In this classical paradigm, research focused predominantly on taxonomic dysbiosis: shifts in microbial diversity, loss of protective obligate anaerobes, and the expansion of pro-inflammatory pathobionts. Under this model, bacteria were viewed primarily as tumor promoters. The translocation of pathogen-associated molecular patterns (PAMPs) such as lipopolysaccharide (LPS) and flagellin across compromised mucosal barriers stimulates pattern recognition receptors like TLR4 and NOD2. Downstream activation of NF-κB, STAT3, and MAPK pathways in epithelial progenitor cells promotes hyperproliferation, suppresses physiological apoptosis, and recruits immune infiltrates that generate reactive oxygen and nitrogen species (ROS and RNS).

While sustained inflammation accelerates neoplastic progression by fostering an extrinsically mutagenic microenvironment, it does not account for the stereotypic somatic mutational scars detected across distinct solid tumors. Recent advances in chemical biology and cancer genomics have revealed a direct mechanism: microbial chemical mutagenesis. Specific mucosal microbes produce low-molecular-weight genotoxins that directly enter host epithelial nuclei to induce sequence-specific covalent adducts, crosslinks, and strand scissions. Consequently, the field is evolving from descriptive taxonomic profiling toward the quantitative dissection of functional microbial genotoxins as bona fide somatic mutagens.

Microbial Genotoxins as Etiological Drivers of Genomic Scars

Host genomic DNA continuously encounters endogenous and exogenous insults, counteracted by conserved DNA damage response (DDR) and repair pathways—including base excision repair (BER), nucleotide excision repair (NER), mismatch repair (MMR), homologous recombination (HR), and non-homologous end joining (NHEJ). When DNA lesions bypass repair via error-prone translesion synthesis (TLS) polymerases or misaligned end joining, mutations are fixed in the daughter cell genome.

Crucially, distinct genotoxic mechanisms possess stereochemical affinities for specific sequence contexts, generating characteristic distributions of single-base substitutions (SBS), doublet-base substitutions (DBS), and small insertions/deletions (ID). Cataloged into reference signatures within the Catalog of Somatic Mutations in Cancer (COSMIC), these patterns reflect the etiological history of tumorigenesis.

Among bacterial mutagens, colibactin—a hybrid polyketide-nonribosomal peptide produced by pks+ Enterobacteriaceae—is the benchmark example. Colibactin exposure in intestinal lineages generates two distinct signatures: single-base substitution signature SBS88 and indel signature ID18. Other mutagens include cytolethal distending toxin (CDT), whose CdtB subunit exhibits DNase I activity inducing double-strand breaks, and Bacteroides fragilis toxin (BFT), which stimulates host spermine oxidase (SMO) to release reactive aldehydes and peroxides that precipitate oxidative base lesions. These genotoxins act as fundamental somatic mutagens whose historical activity remains preserved in tumor genomes.

The Methodological Bottleneck: Why Bulk Cross-Sectional Studies Fail to Establish Causality

Despite mechanistic validation in cell systems, human observational studies frequently struggle to establish definitive causality between bacterial genotoxins and somatic mutations. This challenge stems from relying on static, cross-sectional cohort designs. In standard protocols, biospecimens are obtained at a single timepoint: surgical resection or biopsy of macroscopic neoplasms. Correlating microbial abundance from late-stage samples with tumor mutational signatures introduces severe confounding.

By the time a solid tumor is detected, years of tumor progression have reshaped the local ecosystem. Hypoxia, central necrosis, altered metabolite gradients, and altered luminal transit fundamentally alter microbial colonization. The microbial community identified in late-stage tumors frequently reflects secondary colonization (the “passenger” effect) rather than ancestral initiating mutagens (the “driver” effect).

Correlating late-stage metagenomic data with early mutational signatures conflates consequence with cause. Addressing this challenge requires an exposure-aware, multi-omics approach that combines deep somatic whole-genome sequencing (WGS), host transcriptomic stress profiling, and high-resolution metagenomic reconstruction. Deploying comprehensive multi-omics solutions allows researchers to systematically bridge genomic mutational scars, host cellular response states, and microbial functional genetic architecture.

Mechanistic Foundations: Microbial Genotoxins and Their Mutational Footprints

Figure 1. Molecular mechanism of colibactin genotoxicity and mutational imprint landscape, detailing DNA crosslinking and COSMIC SBS88/ID18 signature induction.

Colibactin and the Genesis of SBS88 and ID18 Signatures

Synthesized by the 54-kilobase pks pathogenicity island, colibactin biosynthesis involves 19 enzymes (clbA through clbS), including nonribosomal peptide synthetases (NRPS), polyketide synthases (PKS), and maturation peptidases. Due to chemical instability, it is produced as an inactive prodrug (precolibactin) carrying an N-terminal acyl-D-asparagine protecting group, cleaved periplasmically by the peptidase ClbP before direct, contact-dependent delivery into host epithelial cells.

Colibactin contains dual electrophilic cyclopropane warheads that alkylate host DNA at the N3 position of adenine bases on opposite strands, forming covalent interstrand crosslinks (ICLs) preferentially within adenine-rich motifs (such as 5'-AAATTT-3' and poly-A tracts). During S-phase in proliferating epithelial stem cells, these crosslinks stall replication forks, triggering fork collapse and generating double-strand breaks (DSBs).

When translesion synthesis (TLS) polymerases bypass unrepaired colibactin monoadducts, nucleotide misincorporation predominantly causes T>C and T>A substitutions (represented on the pyrimidine partner strand) within a 5'-W[T>C]W-3' or 5'-W[T>A]W-3' sequence context (W = A or T). In COSMIC nomenclature, this pattern is designated SBS88. Simultaneously, replication slippage and misaligned end joining at adenine tracts induce single thymidine deletions within homopolymeric T-tracts (e.g., (T)n where n ≥ 4), classified as indel signature ID18. Driver genes harboring adenine-rich homopolymeric runs—such as APC in colorectal models—frequently display recurrent SBS88 and ID18 mutations, demonstrating that colibactin alkylation directly drives initiating oncogenic events.

Beyond Colibactin: Diverse Bacterial Genotoxins and DNA Damage Mechanisms

Beyond pks+ Escherichia coli, several bacterial genotoxins introduce distinct mutagenic damage:

• Cytolethal Distending Toxins (CDTs): Produced by Gram-negative pathogens including Campylobacter jejuni, Helicobacter hepaticus, and Salmonella enterica, CDTs are tripartite holotoxins (CdtA, CdtB, CdtC). The catalytic CdtB subunit possesses type I DNase activity, directly generating single- and double-strand breaks without host metabolic activation. Repair via non-homologous end joining and homologous recombination can produce small deletions, microhomology-mediated rearrangements, and localized chromosomal instability.
• Bacteroides fragilis Toxin (BFT): Secreted by enterotoxigenic B. fragilis (ETBF), BFT is a zinc-dependent metalloprotease that cleaves the extracellular domain of E-cadherin (CDH1), releasing β-catenin and triggering Wnt pathway activation. Concurrently, BFT induces colonocyte spermine oxidase (SMO) upregulation. Catalytic spermine oxidation generates hydrogen peroxide and reactive aldehydes, producing 8-oxo-dG lesions that induce G>T transversions (mirroring SBS18/SBS36 profiles) and base excision repair stress.
• Tilimycin and Tilivalline: Produced by Klebsiella oxytoca, these metabolites exhibit dual cytotoxicity: tilimycin covalently alkylates the DNA minor groove, while tilivalline acts as a tubulin-binding spindle disruptor, combining localized base damage with chromosomal missegregation.

Mutational Signature Deconvolution: WGS vs. WES Resolution

A fundamental requirement for attributing mutational signatures to bacterial genotoxins is sufficient statistical power during computational deconvolution. Whole-exome sequencing (WES) has historically been favored due to lower sequencing costs, but for microbial signature deconvolution, WES presents significant statistical limitations.

Exons comprise approximately 1.5% to 2% of the human genome (~30 Mb). In solid tumors with average mutational burdens (2 to 5 mutations/Mb), an exome captures only 60 to 150 somatic mutations per sample. Deconvolving a 96-dimensional trinucleotide substitution matrix from fewer than 200 mutations violates the power requirements of non-negative matrix factorization (NMF). Low-prevalence signatures like SBS88—often representing 2% to 15% of total mutations in exposed tumors—frequently drop below algorithmic detection thresholds.

Furthermore, exons are subject to strong purifying selection and transcription-coupled nucleotide excision repair (TC-NER), which selectively cleans lesions on transcribed strands and skews mutation spectra. In contrast, deep whole-genome sequencing (WGS) surveys ~3.2 Gb of genomic real estate, typically capturing 5,000 to 50,000 somatic variants per tumor. This provides the statistical leverage required to resolve overlapping signatures, calculate strand bias, detect sparse indels (ID18), and map structural variant breakpoints. For detailed analytical guidance on platform capabilities, review WGS vs. WES vs. targeted sequencing panels.

The “Hit-and-Run” Dilemma: Temporal Decoupling of Bacterial Exposure and Tumor Genesis

Figure 2. The “Hit-and-Run” temporal decoupling model of microbial carcinogenesis, illustrating the decades-long divergence between ancestral mutagen exposure and contemporary tumor detection.

The Divergence Between Current Microbiome and Historical Mutational Scars

A primary conceptual obstacle in microbial oncology is temporal decoupling: the latency between initial genotoxin exposure and tumor resection. In adult solid tumors, the latency between an initiating gatekeeper mutation (e.g., in APC or KRAS) and symptomatic malignancy often spans 10 to 30 years.

Bacterial genotoxins primarily function as initiating mutagens within self-renewing epithelial stem cells (e.g., Lgr5+ crypt base stem cells). A transient colonisation event by pks+ E. coli in early life can alkylate critical driver loci. Following repair misincorporation, the mutated stem cell expands into a clonal patch.

Decades later, when the tumor is sampled, the local microenvironment has altered completely. Neoplastic tissue produces elevated antimicrobial peptides, alters luminal pH, and experiences ischemia. Concurrently, lifetime dietary shifts, immune changes, and antibiotic exposures alter the microbiome. In many instances, the ancestral pks+ strain has been eliminated from the host flora. Under this “hit-and-run” model, the microbial driver damages the host genome early, departs, and leaves behind only its mutational scars (SBS88/ID18). A cross-sectional sequencing study evaluating only late-stage fecal metagenomics would observe zero pks reads, erroneously dismissing microbial etiology.

Mucosa-Adherent vs. Luminal/Fecal Communities: Spatial Topography Considerations

A related methodological pitfall involves spatial topography. Most microbiome studies rely exclusively on fecal samples, which capture the luminal ecosystem—shed bacterial biomass and unabsorbed dietary components reflecting pooled distal contents.

However, genotoxin delivery is contact-dependent. Because colibactin degrades rapidly in physiological buffers, pks+ bacteria must penetrate the mucus barrier and adhere directly to the host epithelial apical membrane. Consequently, the mucosa-adherent microbiota (tissue-associated biofilm) differs in composition, density, and virulence gene expression from luminal stool populations. Studies examining only stool may miss low-abundance, mucosa-adherent pathogens sequestered within epithelial crypts. Robust study designs must incorporate spatially matched mucosal biopsies alongside fecal samples.

Defining the Exposure-Aware Analytical Mindset

To address temporal and spatial decoupling, researchers must adopt an exposure-aware analytical mindset. In this framework, multi-omics modalities act as temporally complementary biological recorders:

• Tumor Whole-Genome Sequencing serves as an archaeological recorder: it permanently preserves lifetime cumulative DNA damage, differentiating early clonal initiating mutations (trunk events) from recent subclonal mutations.
• Host Tissue Transcriptomics serves as a functional response recorder: it captures intermediate-to-long-term biological states, revealing persistent DNA damage response (DDR) signaling, senescence-associated secretory phenotypes (SASP), or chronic inflammatory pathways (e.g., NF-κB) that outlast active bacterial colonization.
• Metagenomic Shotgun Sequencing serves as a contemporary exposure recorder: it evaluates whether active genotoxic strains and biosynthetic gene clusters are presently colonizing the host, distinguishing active exposure from past hit-and-run dynamics.

Integrating these modalities enables cohort classification into meaningful exposure categories (e.g., Historical Hit-and-Run: elevated SBS88, elevated DDR, absent current pks; Active Ongoing Exposure: elevated SBS88, elevated DDR, high pks; Non-Exposed: baseline profiles across all metrics).

The Multi-Omics Study Design Blueprint

Figure 3. Comprehensive architecture of an exposure-aware multi-omics study design, combining tumor deep WGS, host transcriptomics, and fecal/mucosal metagenomics.

Cohort Stratification, Sample Triads, and Metadata Capture

An exposure-aware study design linking microbial genotoxins to mutational signatures requires prospective or curated retrospective cohorts built on a harmonized sample triad:

• Malignant Tumor Tissue: High-quality fresh-frozen (FF) resection tissue is preferred over formalin-fixed paraffin-embedded (FFPE) material to avoid artificial deamination artifacts.
• Matched Germline Normal Control: Peripheral blood mononuclear cells (PBMCs) or histologically verified normal mucosa (≥10 cm from tumor margins) to enable clean subtraction of germline variants.
• Paired Microbial Specimens: A dual-niche collection protocol including (a) pre-treatment, pre-bowel-prep fresh-frozen stool, and (b) surgically harvested tumor-adjacent mucosal biopsies.

Standardized metadata tracking is essential. Covariates known to modulate the microbiome or mutation frequencies must be captured: age, tumor anatomical subsite (e.g., proximal vs. distal colon), microsatellite instability (MSI) / mismatch repair (MMR) status, prior antibiotic/chemotherapy treatments, chronic NSAID use, and dietary habits.

Tumor Deep Whole Genome Sequencing (WGS) Protocol

Accurate signature deconvolution requires high-depth, high-fidelity whole-genome sequencing. Deploying dedicated whole-genome sequencing services ensures uniform coverage and low base-calling error rates across complex genomic regions.

Recommended sequencing depth parameters:

• Tumor Specimen: ≥60x–80x mean deduplicated coverage, ensuring subclonal mutations present at low variant allele frequencies (VAF 0.10–0.20) are captured with statistical confidence.
• Matched Germline Normal: ≥30x–40x mean deduplicated coverage for accurate germline subtraction.

Bioinformatics pipelines follow standardized quality gates. Paired-end reads (2×150 bp) are aligned to GRCh38/hg38 using BWA-MEM2, followed by Picard deduplication and GATK Base Quality Score Recalibration (BQSR). Somatic variant calling employs an ensemble consensus approach combining Mutect2, Strelka2, and Vardict for SNVs and indels, with Manta and SvABA for structural variants (SVs).

Mutational signature extraction maps mutations into the 96 trinucleotide substitution matrix based on the mutated pyrimidine base and flanking context. Deconvolution is performed using non-negative matrix factorization (NMF) via SigProfilerExtractor or MutationalPatterns. Reference signature refitting against COSMIC v3.4 focuses on SBS88, SBS18, SBS36, and ID18 using strict cosine similarity thresholds (cosine > 0.85) and bootstrap resampling.

Furthermore, integrating purity- and ploidy-corrected VAFs (via Absolute or Battenberg) yields the cancer cell fraction (CCF) for each mutation. Mutations with CCF values near 1.0 are designated clonal (trunk) events, while lower CCF values indicate subclonal evolution. Confirming SBS88 as an exclusively clonal event demonstrates that colibactin exposure acted as an ancestral initiating hit. Researchers can structure these workflows using standardized bioinformatics workflow for whole-genome sequencing protocols.

Fecal & Tissue Metagenomic Shotgun Sequencing Architecture

Because genotoxin synthesis varies at the strain level rather than the genus level, shotgun metagenomics is essential. Protocols outlined in the gut microbiome shotgun metagenomics guide provide standardized extraction and balanced library construction.

Key metagenomic parameters:

• Sequencing Yield: ≥10 to 20 Gb clean paired-end data per sample (30–60 million paired-end reads).
• Host DNA Depletion: In mucosal biopsies, human host DNA can exceed 80–90% of extracted nucleic acids. Applying differential enzymatic lysis (e.g., saponin treatment) or computational filtering (Bowtie2 alignment to GRCh38 with unmapped read extraction) preserves microbial read depth.

Taxonomic profiling via MetaPhlAn4 or Kraken2/Bracken establishes species-level abundance. Concurrently, reads are aligned against curated virulence databases to identify genotoxin operons, implementing specialized sequencing techniques in identification of bacterial virulence genes.

For colibactin, reads are mapped against the 54-kb pks pathogenicity island, calculating mean depth and breadth of coverage across all 19 genes (clbA through clbS). Samples are designated pks+ only if ≥80% of the island exhibits uniform coverage, avoiding false positives from homologous housekeeping genes. De novo assembly (via MEGAHIT) and biosynthetic gene cluster (BGC) annotation with antiSMASH enable detection of novel genotoxins in uncharacterized taxa.

Host Tissue Transcriptomics (RNA-Seq) for Biological Grounding

Matched host transcriptome sequencing (RNA-seq) on fresh-frozen tumor tissue and adjacent normal mucosa provides functional context to genomic scars.

Libraries (sequenced to ≥50–80 million reads per sample) are aligned using STAR and quantified with RSEM or Salmon. Differential expression (DESeq2/edgeR) and gene set enrichment analysis (GSEA) evaluate specific biological axes:

• DNA Damage Response (DDR): Quantifying ATM/ATR activation, CHEK1/CHEK2 signaling, homologous recombination (BRCA1/2, RAD51), and base excision repair components.
• Alkylation and Crosslink Repair Pathways: Expression profiling of enzymes addressing alkylation and interstrand crosslinks, including MGMT, ALKBH2/3, and Fanconi Anemia core complex members (FANCA/C/D2).
• Cellular Senescence and SASP: Colibactin can induce senescence rather than immediate apoptosis. Profiling SASP markers (IL6, IL8, CXCL1, MMP3, CDKN1A/p21, CDKN2A/p16) identifies persistent cellular stress.
• Epithelial Barrier and Immune Repertoire: Assessing mucins (MUC2, MUC5AC), tight junctions (CLDN1, TJP1), and infiltrating immune subsets via deconvolution (CIBERSORTx) characterizes local mucosal defense.

Exposure-Aware Computational Integration & Causal Modeling

Figure 4. Computational mutational signature deconvolution and clonality timing analysis, resolving clonal trunk vs subclonal branch mutations.

Multi-Omics Factor Analysis (MOFA+) and Joint Matrix Decomposition

Integrating disparate multi-omics layers—continuous somatic mutation loads, discrete signature burdens, non-negative microbial gene counts, and high-dimensional transcriptomic matrices—presents significant statistical challenges. Pairwise correlation testing introduces multi-testing penalties and ignores complex multi-variable interactions.

To resolve this, researchers deploy multi-view factorization frameworks such as Multi-Omics Factor Analysis (MOFA+). MOFA+ models the multi-omics data matrices as a product of shared latent factors and view-specific weight matrices:

Y_m = Z · W_m^T + ε_m

where Y_m represents the omics data matrix for view m (WGS mutational signatures, host RNA-seq, or metagenomics), Z denotes the shared latent factor matrix across all samples, W_m is the view-specific weight matrix, and ε_m represents residual variance.

MOFA+ identifies latent factors that jointly capture variance across the colibactin signature (SBS88/ID18), host DDR/SASP expression, and clb cluster abundance. It natively handles missing data views across patient subsets, preserving cohort completeness.

Causal Mediation Analysis: Decoupling Bacterial Abundance from Persistent DDR

To test whether host transcriptomic stress pathways bridge past microbial exposure to current mutational burdens, researchers apply statistical mediation analysis:

Equation 1 (Mediator Model): M = α0 + α1·X + γ1·C + e1

Equation 2 (Outcome Model): Y = β0 + β1·X + β2·M + γ2·C + e2

where X represents bacterial genotoxin gene abundance, Y denotes tumor SBS88/ID18 mutational burden, M represents host DDR pathway activation or SASP score, and C represents baseline covariates (age, sex, anatomical site, MSI status).

The indirect effect (α1 × β2) evaluates whether microbial exposure influences somatic mutation accumulation via persistent host pathway activation. In historical hit-and-run scenarios—where contemporary fecal pks abundance (X) is zero—mediation analysis reveals that historical mutational scars (Y) correlate with persistent host transcriptomic and epigenetic mediators (M), corroborating an ancestral exposure event. Methodological considerations for these models are detailed in discussions of integrated analysis of microbiome and transcriptome data.

Experimental Ground-Truth Validation: The Human Organoid Exposure Assay

Figure 5. Experimental triangulation via human intestinal organoid micro-injection and clonal WGS, establishing causal mutational ground truth.

Because human observational cohorts inherently demonstrate correlation rather than definitive causation, rigorous study designs incorporate an experimental validation arm: the human intestinal organoid exposure assay.

Primary human intestinal organoids derived from non-neoplastic crypts maintain physiological architecture and diploid karyotypes. In this workflow:

• Organoids are microinjected intraluminally with living isogenic bacterial pairs—wild-type genotoxic strains (e.g., pks+ E. coli) versus isogenic knockout mutants lacking essential biosynthetic enzymes (e.g., ΔclbP or ΔclbA), alongside sham controls.
• Controlled exposures are maintained over defined timeframes (4 to 12 hours).
• Exposed organoids are dissociated to single cells and clonally expanded into daughter lines over 8 to 12 weeks to amplify genomic material.
• Clonal lines undergo deep whole-genome sequencing (≥40x).

Because parental genomes are pre-sequenced, mutations accumulating in clonal lines can be directly attributed to the bacterial exposure window. The resulting mutational spectra establish the experimental ground-truth signature for that genotoxin, providing benchmark validation for clinical tumor deconvolution.

Quality Control, Confounding Factors, and Technical Pitfalls

Low-Biomass Contamination and Batch Effect Mitigation

Low microbial biomass poses a serious technical hurdle when profiling tissue-associated microbiomes. While stool contains ~10¹¹ bacterial cells/gram, mucosal biopsies and resection margins yield minimal microbial biomass. In low-biomass shotgun sequencing, background contaminants in DNA extraction columns and reagents (the “kitome”) can overwhelm biological signal. Common contaminants such as Ralstonia, Burkholderia, and Pseudomonas risk misidentification as tumor-associated pathobionts.

Standard contamination controls include running no-template extraction and library negatives, executing in silico decontamination via frequency and prevalence modeling (decontam), and randomizing sample batches across flowcells.

FFPE Artifacts vs. Genuine Signature Mutational Noise

Tumor biobanks often store samples as formalin-fixed paraffin-embedded (FFPE) blocks. Formalin fixation induces chemical damage, particularly hydrolytic deamination of cytosine to uracil. During library amplification, deaminated cytosines are read as thymines, introducing artificial C>T transitions that distort signature analysis and compromise detection of low-frequency signatures like SBS88.

To protect data integrity, fresh-frozen tissue is strongly recommended. If FFPE samples must be utilized, enzymatic Uracil-DNA Glycosylase (UDG) treatment prior to library prep excises artificial uracils, complemented by bioinformatic orientation-bias filtering.

Statistical Power and Sample Size Requirements for Signature Attribution

Underpowered cohorts frequently result in false-negative signature attribution. Resolving mutational signatures via NMF depends on total mutation counts per sample (≥1,000 somatic SNVs recommended for low-frequency signature deconvolution), signature prevalence across the cohort (requiring 200–500 cases to capture a 5% prevalent signature across 20+ positive samples), and bootstrap validation (1,000 iterations) to control false discovery rates (FDR < 0.05).

Synthesis: Translating Mechanistic Multi-Omics into Preclinical Discovery

Investigating microbial genotoxins and their permanent mutational signatures represents a major advance in molecular oncology. By addressing the “hit-and-run” dilemma and moving beyond descriptive dysbiosis, researchers can clarify the direct mutagenic contributions of the microbiome to neoplastic initiation.

Executing this vision requires a study design that treats the host-microbe interface as an integrated biological system. Deep tumor whole-genome sequencing provides the permanent somatic record; host transcriptomics captures active cellular stress and repair states; and fecal and mucosal metagenomics profiles contemporary microbial genetic potential. When combined with mediation modeling and validated in human organoid systems, this exposure-aware framework provides robust causal insights.

As preclinical research advances toward early biomarker discovery and cancer prevention models, understanding historical microbial exposures will be essential. Executing these multi-omics studies requires robust sequencing depth, standardized library preparation, and specialized bioinformatic integration across every layer of the analytical workflow.

Frequently Asked Questions (FAQs)

References:

1. Pleguezuelos-Manzano C, Puschhof J, Rosendahl Huber A, et al. “Mutational signature in colorectal cancer caused by genotoxic pks+ E. coli.” Nature. 2020 580(7802): 269-273 https://doi.org/10.1038/s41586-020-2080-8
2. Cuevas-Ramos G, Petit CR, Marcq I, et al. “Escherichia coli induces DNA damage in vivo and triggers genomic instability.” Science. 2010 328(5977): 515-519 https://doi.org/10.1126/science.1188188
3. Alexandrov LB, Kim J, Haradhvala NJ, et al. “The repertoire of mutational signatures in human cancer.” Nature. 2020 578(7793): 94-101 https://doi.org/10.1038/s41586-020-1943-3
4. Arthur JC, Perez-Chanona E, Mühlbauer M, et al. “Intestinal inflammation targets cancer-inducing activity of the microbiota.” Science. 2012 338(6103): 120-123 https://doi.org/10.1126/science.1224820
5. Iftekhar A, Berger H, Bouznad N, et al. “Genomic aberrations after short-term exposure to colibactin-producing Escherichia coli transform primary colon epithelial cells.” Nat Commun. 2021 12(1): 1003 https://doi.org/10.1038/s41467-021-21162-y