Testing the Host-Genetics → Microbiome → Immune-Exclusion Causal Chain in Colon Cancer

Most colon cancers do not respond to immunotherapy, and we do not fully understand why. In mismatch-repair-deficient tumors, checkpoint inhibitors produce objective responses in 40–50% of patients, but in mismatch-repair-proficient, microsatellite-stable (pMMR/MSS) tumors — roughly 84–86% of cases — the response rate is close to 0% (Tumor-Infiltrating Lymphocytes in Colorectal Cancer). One widely suspected culprit is a metabolic dialogue: tissues and their resident microbes consume and produce small molecules that starve or reprogram immune cells, for example tumors depleting glucose via the transporter SLC2A3 to kill CD8+ T cells (SLC2A3 promotes head and neck squamous cancer developing through negatively regulating CD8+ T cell in tumor microenvironment). But every piece of this story has been observed separately, in different tissues and with different tools, and the causal chain connecting them has never been assembled or tested.

This proposal builds and tests that chain in one place. The idea, in two sentences: host genetic variants (identified as expression quantitative trait loci, or eQTLs — inherited DNA differences that dial gene expression up or down) may partly determine which bacteria dominate the gut mucosa; those bacteria, in turn, may reshape the metabolic state of specific immune and tumor cells in ways that exclude CD8+ T cells from MSS colon tumors. We will test this using germline genetic variants as causal instruments — a design called Mendelian Randomization, which uses the random assortment of inherited DNA as a natural experiment that sidesteps the confounding and reverse causation that plague observational studies — while fixing two known statistical failures in how this has been done before.

Core Claim

The specific claim is that a three-link causal chain — host cis-eQTLs in metabolic regulator genes → mucosal abundance of specific bacterial taxa → cell-type-specific metabolic-immune gene programs in the colon tumor microenvironment — is testable with existing public data, but only if two methodological defects are repaired. First, prior microbiome-to-cancer Mendelian Randomization used weak, locus-wide-significant instruments and simple two-sample regression (Identification of Host Gene-Microbiome Associations in colorectal cancer patients Using Mendelian Randomization), and standard two-step and multivariable Mendelian Randomization are known to lose confidence-interval coverage, dropping from a nominal 95% to 70–80%, under exactly the weak-instrument and selection biases that microbiome GWAS exhibit (Mediation Analysis with Mendelian Randomization and Efficient Multiple GWAS Integration). Second, prior integrations of microbial GWAS with host transcriptomes used either bulk tissue (Mendelian Randomization Integrating GWAS and eQTL Data Reveals DAAM1, a Potential Immune-Related Biomarker for Breast Cancer Prognosis) or, at best, single-cell data in a different organ system — the respiratory tract, where Wang et al. linked oral-flora GWAS to immune-cell-type-specific expression in severe influenza-like illness and found CD16+ monocytes to be the responsive subset. That analytical bridge has never been carried into the gut, where the microbes actually live. We claim that repairing both defects simultaneously — bias-corrected mediation across the three links, with single-cell-resolved outcomes — is what makes the chain falsifiable rather than merely suggestive.

Research Program

The program has three components, each reusing a technique proven in one of the source literatures and transplanting it to the intersection.

The first component is instrument hygiene. The MiBioGen consortium genotyped and 16S-sequenced fecal samples from 18,340 people across 24 cohorts, mapping 31 genome-wide significant microbiome loci ($P < 5 \times 10^{-8}$) and showing, importantly, that the LCT lactase locus drives Bifidobacterium abundance in an age-dependent way ($P = 1.28 \times 10^{-20}$) (Large-scale association analyses identify host factors influencing human gut microbiome composition). We will use only instruments passing that genome-wide bar, plus explicit modeling of the dominant technical confounder — MiBioGen reports that DNA extraction protocol alone accounted for ~29% of cross-cohort variation. A key early analysis quantifies how many of the six taxa that Xiang et al. flagged as causally linked to colorectal cancer (family Porphyromonadaceae OR = 1.26; genus Intestinibacter OR = 1.31; among others, all at locus-wide $P < 1 \times 10^{-5}$) retain usable genome-wide instruments. If few survive, that is itself a finding: it would mean the existing microbiome-to-CRC causal literature rests partly on instruments indistinguishable from extraction noise.

The second component is a cell-type-resolved outcome layer. Following the Wang et al. recipe — single-cell disease relevance scoring (scDRS) plus cell-type-specific eQTL weight models built with FUSION — but applied to gut rather than airway, we will pseudobulk the colon cancer single-cell dataset GSE178318 by cell type, build eQTL/TWAS weight models for each immune and malignant subset, and define metabolic-immune outcome scores from two curated sources: the 274 mitochondrial genes of MitoCarta 3.0 as distilled in Comprehensive analysis of mitochondrial-related gene signature for prognosis, tumor immune microenvironment evaluation, and candidate drug development in colon cancer, and the specific effectors documented in the crosstalk literature — glucose transport (SLC2A3), tryptophan catabolism (IDO1), AMPK signaling (PRKAG1, whose genetically proxied expression is causally protective against colorectal cancer, OR 0.74, 95% CI 0.63–0.87, adjusted $P = 0.001$, in Evaluating the Impact of Putative Metformin Targets on Cancer Outcomes). The question this layer answers is which cell — CD8+ T cell, Treg, macrophage, or tumor epithelium — carries the microbial-metabolic signal, something bulk deconvolution cannot determine.

The third component is the mediation estimator. We will adapt the MAGIC estimating-equation framework of Lyu et al., which maintained 95% confidence-interval coverage in simulations where standard two-step MR and multivariable MR fell to 70–80%, and which in a cardiovascular case study detected mediators (diastolic blood pressure, HDL) that two-step MR missed entirely. The causal estimand is the indirect effect of host genotype on the cell-type-specific outcome score that flows through genus abundance:

$$\beta_{\text{indirect}} = \theta_{\text{taxon}\to\text{outcome}} \cdot \gamma_{\text{eQTL}\to\text{taxon}}$$

where $\gamma$ is the effect of a host cis-eQTL on genus abundance (instrumented via mbQTL summary statistics) and $\theta$ is the effect of genus abundance on the outcome score (instrumented via the same mbQTLs). Both stages use genome-wide-significant instruments only, with estimator-level corrections for winner's-curse and loser's-curse selection bias. We will calibrate a simulation to the observed instrument strength and the ~29% technical-variance figure before running anything on real data, so the coverage claim is verified in our own setting rather than inherited.

Proposed Experiments

The first experiment is the positive-control test. The LCT → Bifidobacterium → Bifidobacterium reduces ulcerative colitis risk (OR = 0.51) chain from MiBioGen is the one host-genetics-to-microbe-to-outcome path with population-scale support; our pipeline must recover it before anyone should trust it on novel chains. Failure here kills the project early and cheaply.

The second experiment runs the full mediation scan in MSS colorectal cancer: host cis-eQTLs for metabolic genes (GTEx v10 colon tissue, F-statistic > 10, following the instrument discipline of the metformin-target study) as exposures, MiBioGen genus abundances as mediators, and CD8-exhaustion / Treg / M2-macrophage metabolic scores in specific cell types as outcomes. The primary output is a ranked list of significant eQTL–taxon–cell-type triples with bias-corrected indirect effects. The scientific prize is any triple in which a metabolic-regulator eQTL (for example in AMPK subunits) acts on CD8+ T-cell metabolic state through a taxon rather than directly — that would be evidence that part of the "cold tumor" phenotype is routed through the gut lumen, which no study in the literature has tested.

The third experiment is validation and robustness. Predicted triples are checked against TCGA-COAD: tumors from individuals whose genotype predicts higher mediator-taxon abundance should show the predicted cell-type-specific metabolic signature in deconvolved RNA-seq, stratified by microsatellite status. Because preprocessing choices are a documented failure mode — top-ranked genes from linkage analysis overlap only 10–43% across normalization methods (RMA vs. GCRMA vs. PLIER), with heritability estimates flipping from high to near-zero (Impact of normalization and filtering on linkage analysis of gene expression data) — every mediation result must survive recomputation under alternative normalization and filtering pipelines, a sensitivity check that existing tumor-transcriptomic studies in this space (Comprehensive analysis of mitochondrial-related gene signature, DAAM1) have never performed. As a benchmark anchor, the 9-gene mitochondrial risk signature validated on TCGA-COAD reaches an AUC of 0.723 for 3-year overall survival, which sets the predictive bar any competing prognostic use of our triples must clear. We will also probe the "Treg paradox" — Bai et al. note that high Treg infiltration sometimes predicts better survival in CRC, in tension with the uniformly immunosuppressive reading in the mitochondrial-signature paper — by estimating cell-type-specific effects rather than bulk infiltrate fractions, since the paradox may be an artifact of averaging.

Baselines

The naive reference point is plain observational association: Spearman correlation between genus abundance and deconvolved immune fractions in TCGA-COAD, with no genetic instruments. No reported result found on Lacuna for this specific correlation in MSS colon cancer, which is itself part of the motivation: the field has not even established the correlational version at scale. Beating it means showing that genetically instrumented effects disagree with naive correlations in both magnitude and direction for at least some triples — the expected outcome given that baseline inter-individual microbiome variation dominates biological effects and many taxonomic shifts fail false-discovery-rate control even in a controlled drug trial (Novel Gut-Based Pharmacology of Metformin in Patients with Type 2 Diabetes Mellitus).

The strongest existing competitor is the Xiang et al. pipeline: two-sample Mendelian Randomization (inverse-variance weighted, MR-Egger, weighted median, MR-PRESSO) between microbiome GWAS and colorectal cancer GWAS. It is the strongest because it is the only published method connecting microbiome, host genetics, and colorectal cancer outcomes in one causal analysis, and it reports concrete causal estimates (Porphyromonadaceae OR = 1.26, Intestinibacter OR = 1.31, Anaerotruncus OR = 1.17). We will run it under identical data access, instrument pools, and compute as our pipeline, then extend the comparison in two directions our method must win: it must maintain simulation-calibrated confidence-interval coverage where the IVW pipeline's weak-instrument behavior degrades (following the coverage-drop demonstration in Lyu et al.), and it must localize effects to cell types, which the Xiang design cannot do by construction since its outcome is a binary GWAS trait. A secondary competitor is the Wang et al. flora-TWAS pipeline, which reported 19 significant flora–cell-type gene pairs in influenza-like illness; we run it verbatim on our gut data as the cell-type-resolution baseline, and our method must beat it by adding the mediation structure and instrument hygiene it lacks.

The key ablation swaps the bias-corrected mediation estimator for standard two-step Mendelian Randomization on identical instruments, data, and tuning. Lyu et al. report coverage collapse to 70–80% and missed mediators in exactly this configuration; if, in our microbiome-calibrated simulations, two-step MR performs equivalently to MAGIC, then the estimator is not the active ingredient and the proposal reduces to an application paper — we would report that honestly and the novelty claim would shift entirely to the cell-type outcome layer and the tripartite chain.

Why not just…?

Why not just correlate 16S abundances, metabolites, and immune markers directly? Because the causal direction is precisely what is disputed: the literature explicitly flags that 16S profiling cannot distinguish dysbiosis that drives pathology from dysbiosis that is a secondary adaptation to the inflamed, metabolically altered tumor environment (Xiang et al.). The trial evidence reinforces this — in the metformin study, individual baseline variation in principal-coordinate space dwarfed drug-induced shifts and many taxa changes failed FDR correction (Napolitano et al.), so an observational genus–immune correlation in colon cancer would likely be dominated by person-to-person noise. Our first experiment doubles as the test: if naive correlations and instrumented estimates agree everywhere, the genetic machinery added nothing and the simpler design was right.

Why not just use standard two-step or multivariable Mendelian Randomization? This is the most serious objection, because the software is mature and the reviewers' familiarity is real. The evidence against it is direct rather than rhetorical: Lyu et al.'s simulations show coverage falling from 95% to 70–80% under weak-instrument and selection bias (MAGIC paper), and microbiome GWAS is a weak-instrument regime by construction — MiBioGen found only 31 genome-wide significant loci across 410 genera, and only 9 genera are present in over 95% of samples. Still, we treat this as a live threat, not a settled one, which is why the abation above is decisive rather than rhetorical: if two-step MR matches MAGIC in our calibrated simulations, we will report that the estimator upgrade is unnecessary and the contribution stands or falls on the chain itself.

Why not just use bulk immune deconvolution (CIBERSORT-style) on TCGA-COAD instead of single-cell eQTL models? Because bulk averages destroy the one signal we are hunting: whether a metabolic change sits in the tumor cell, the CD8+ T cell, or the Treg. Wang et al. showed that flora selectively modulate expression in discrete immune subsets (CD16+ monocytes), a phenomenon invisible in bulk (Wang et al.), and Beyene et al. showed that even coarser preprocessing choices than cell-type averaging swing linkage discoveries across a 10–43% overlap range (Beyene et al.). Concretely, our design includes bulk deconvolution as a co-analysis on the same samples: any triple significant in bulk but absent at single-cell resolution, or vice versa, gets reported, and if the two agree everywhere the cheaper method wins and we say so.

Why not just prove the mechanism in mice? Mouse knockouts give clean causality for one link at a time — silencing Ccrk halted lipid accumulation and myeloid-derived suppressor cell recruitment in obese mice (Sun et al.) — but the translational record in this exact disease is a warning: checkpoint therapy had strong preclinical rationale yet yields ~0% responses in pMMR-MSS patients (Bai et al.), and mouse gut microbiomes do not recapitulate human community structure. Our proposal is deliberately human-data-first; mouse follow-up belongs in a future proposal for whichever single triple survives all the checks below.

Decisive Checks

The project succeeds if, in order: the pipeline recovers the known LCT → Bifidobacterium → colitis chain as a positive control; the instrument-hygiene audit shows which published microbiome-CRC causal claims survive genome-wide instruments; the mediation scan on MSS colorectal cancer yields at least one replicated, normalization-robust, cell-type-specific eQTL–taxon–metabolic-program triple with confidence intervals whose coverage was verified in microbiome-calibrated simulation; and the ablation demonstrates that the bias-corrected estimator, not just the single-cell outcome layer, is load-bearing. It fails cleanly — and informatively — if genome-wide-significant microbiome instruments are too few to power any three-link chain, a result that would itself argue that the field's causal claims about the gut microbiome and colorectal cancer currently rest on statistical footing too weak to support them.