The gut microbiome appears to press two opposing epigenetic buttons in the same tissue. Certain pathogens turn the mute button on: Xia et al. showed that Fusobacterium nucleatum (22.34-fold enriched) and Hungatella hathewayi (9.67-fold) directly upregulate host DNA methyltransferases in colonic epithelium, hypermethylating promoters of tumor suppressor genes like CDX2 and MLH1. Commensals press the opposite button: butyrate-producing Faecalibacterium prausnitzii inhibits histone deacetylases and keeps promoters like SOCS1 in an open, acetylated state (Shi et al.), while butyrate broadly sustains regulatory T cell differentiation through the same HDAC-inhibition route (Yuanyuan Liu et al.). Nobody has measured these two pressures in the same tissue, asked whether they act on different gene programs, or tested whether the "open chromatin" dial can rescue genes already locked shut by dense promoter methylation — the very lock that, per Ng et al., recruits HDAC-containing co-repressor complexes and compacts chromatin.
The obstacle is that the relevant measurements rarely coexist in one cohort. Genome-wide methylation assays destroy spatial context; histology slides carry morphology but no molecular readout; metabolomics demands fresh-frozen tissue and small samples (Taira et al. managed only n=24 per group); and microbiome sequencing has been paired with methylation only once, in disaggregated biopsies without histomorphometry (Xia et al.). This proposal builds a computational route around that obstacle and then settles the mechanism in mice.
Core Claim
I claim three things. First, that the two microbial pressures are separable in archival clinical tissue: an imputation model trained on RNA-seq can estimate a sample's butyrate-exposure and pathogen-load signature (extending UnitedMet, which imputed 48% to 67% of metabolites from transcriptomes), and a spatial graph model over H&E slides (extending SlideGraphmethyl) can predict which gene groups carry promoter hypermethylation. Second, that combining the two — feeding microbial-metabolite exposure estimates into the slide model as global covariates — improves methylation prediction enough to matter, which no existing model tests. Third, that functional enrichment applied to the predicted methylated gene groups, stratified by the balance of the two pressures, will show they silence different pathway classes, and that a gnotobiotic mouse experiment will show HDAC inhibition cannot unlock already-hypermethylated promoters even though it prevents new silencing.
The simplest formal statement of the central hypothesis uses one contrast. For sample $i$, let $\hat{P}_i$ be the imputed abundance of DNMT-inducing taxa and $\hat{B}_i$ the imputed butyrate exposure; define
$$d_i = z(\hat{P}_i) - z(\hat{B}_i),$$
where $z(\cdot)$ standardizes each quantity across the cohort. The claim is that $d_i$ predicts the fraction of predicted-hypermethylated gene groups better than either pressure alone, and that the gene groups silenced at high $d_i$ are pathway-enriched differently from those that merely lose acetylation at low $\hat{B}_i$.
Research Program
The first component is a metabolite-exposure imputer. UnitedMet established that RNA–metabolite covariation carries enough signal to impute unmeasured metabolites in renal cancer cohorts — 59 metabolites were reproducibly well predicted, including kynurenine and fumarate, and imputed succinate and malate predicted sunitinib response with $P < 10^{-4}$ in IMmotion151 (n=823). But UnitedMet was built on bulk RNA-seq with no microbial inputs, and the file records no result on whether short-chain fatty acids are among the well-predicted metabolites — no reported result found on Lacuna for SCFA recovery by UnitedMet. That is itself the first research question. I will train an imputer of the same Bayesian rank-covariation family on cohorts with paired LC-MS/MS and RNA-seq (the Taira et al. BRAF-mutant CRC cohort provides paired metabolomics and transcriptomics, albeit at n=24 per group), augmented with host-side butyrate-response gene signatures as anchors. Where 16S data exists — above all the 33-biopsy cohort of Xia et al. — the imputed butyrate exposure will be validated against taxonomic butyrate-production capacity, the strongest available proxy since no SCFA quantification exists in that cohort.
The second component is a spatial methylation predictor built on SlideGraphmethyl, which divides H&E whole-slide images into patches, embeds them with a pretrained encoder, connects patches by Delaunay triangulation into a graph, and uses graph convolutions to predict differential methylation of predefined gene groups — reporting over 20% AUROC improvement over methods like CLAM on TCGA-GBMLGG (no absolute AUROC value found on Lacuna; I will reproduce it under my own splits before comparing). My extension is specific: the imputed $\hat{P}_i$ and $\hat{B}_i$ enter as global node features, so the model can learn, for instance, that crypt hyper-proliferation morphology predicts CDX2-group methylation only under pathogen pressure. Training uses TCGA-COAD and TCGA-READ, which pair H&E slides with Infinium 450K arrays and RNA-seq. A stated weakness: TCGA has no metagenomics, so the microbial features there are imputed, which is exactly why the Xia cohort serves as the external check of the imputer rather than of the slide model.
The third component is the enrichment analysis that turns predictions into mechanism. Following SlideGraphmethyl's demonstration that grouping coherently methylated genes and running GSEA exposes cancer-hallmark pathways invisible to single-gene assays, I will rank gene groups by predicted methylation, stratify samples by $d_i$, and test whether the pathway classes differ. The motivating hypothesis comes from the candidate-gene literature: the DNMT pressure should concentrate on tumor-suppressor and checkpoint programs (SOCS-1/CDX2/MLH1-class, per To et al. and Xia et al.), while SCFA depletion should hit immune-tolerance programs whose chromatin openness depends on HDAC inhibition (Liu et al.; Siyu Li et al.). If the two gene-set collections overlap heavily, the "two dials" story collapses into one undifferentiated dysbiosis signal — and I would report that honestly.
The fourth component settles the mechanistic tension the computational work can only gesture at. Gnotobiotic mice, four arms: vehicle; H. hathewayi gavage; H. hathewayi plus butyrate supplementation; and H. hathewayi followed by delayed butyrate (rescue arm). Endpoints reproduce and extend Xia et al.: bisulfite sequencing of Cdx2 and Mlh1 promoters, DNMT1/3A/3B expression, H3K27ac at the same promoters (the mark Shi et al. showed butyrate can restore), Ki-67-positive crypt-cell counts, and quantitative histomorphometry — crypt depth, goblet-cell density, and cell-neighborhood topology, the measurements Xia et al. omitted and that histomorphometry studies routinely quantify (Ullah et al.; Zhang et al.). The decisive prediction: co-administered butyrate prevents hypermethylation and restores H3K27ac, but the rescue arm reopens chromatin (H3K27ac returns) without demethylating dense CpG islands, leaving expression only partially restored — because methylation's steric block persists independent of HDAC activity. This distinguishes my proposal sharply from Shi et al., who showed butyrate raising acetylation at a non-hypermethylated promoter; nobody has tested the methylated-promoter case.
Baselines
The naive reference point is the cohort base rate: predict every gene group's methylation status from its frequency in the training set, ignoring the slide entirely. This matters because single-gene methylation is often rare — To et al. found SOCS-1 promoter hypermethylation in only 27.5% (11 of 40) of primary gastric tumors, 10% of adjacent tissue, and 0% of normal mucosa — so a model that ignores the slide and predicts "mostly unmethylated" is deceptively hard to beat, and AUROC gains must be reported per gene group against this prior.
The strongest existing competitor is SlideGraphmethyl itself (SlideGraphmethyl), the graph-convolution architecture that is currently the best published method for predicting methylation status from H&E morphology; it reported over 20% AUROC improvement over SlideGraph∞, CLAM, and classical machine learning on TCGA-GBMLGG, though no absolute AUROC number is reported on Lacuna, and it performed worse on TCGA-KIRC. I will re-implement and tune it on TCGA-COAD/READ under my own splits, same patch encoder, same patch size, same compute and hyperparameter-search budget as my model. Beating it means a paired per-gene-group AUROC improvement on held-out slides whose bootstrap confidence interval excludes zero. A second serious competitor is MethylProphet, which predicts genome-wide CpG beta values from expression and sequence alone and achieved a median across-sample Pearson correlation of 0.72 on ENCODE; I will run it on the same RNA-seq, aggregate its CpG predictions to gene groups, and compare directly, since if expression alone suffices, the histology layer is redundant.
The key ablation removes exactly the new ingredient: the identical graph model trained without the imputed $\hat{P}_i$ and $\hat{B}_i$ covariates. That ablation is SlideGraphmethyl-equivalent by construction, so any gain my model shows is attributable to the microbial-metabolite features and nothing else. Two further ablations locate the gain: removing the spatial graph leaves a CLAM-style multiple-instance learner, and replacing imputed with raw RNA-seq features tests whether the imputation step adds signal beyond what a generic expression covariate provides.
Why not just…?
Why not just run EPIC arrays or WGBS on the tissue directly? These are ground truth, and the mouse arm of this proposal uses them. But as MethylProphet documents, arrays cover only 1% to 3% of human CpG sites and WGBS is cost-prohibitive for large cohorts; more importantly, bulk assays grind away the spatial architecture, and my claim is partly about where in the tissue silencing happens. The computational model exists precisely to make methylation readable in archival slides where arrays were never run.
Why not just measure metabolites by LC-MS/MS instead of imputing them? Where paired tissue exists, we should, and the Taira training cohort does. But UnitedMet exists because LC-MS/MS requires fresh-frozen tissue, suffers inter-laboratory batch effects, and caps cohorts at dozens of samples — Taira et al. reached only n=24 per group. The imputer is what lets the microbial-metabolite hypothesis be tested on cohorts of hundreds. The risk is real, though: if the imputer cannot recover SCFAs (no reported result found on Lacuna for UnitedMet recovering short-chain fatty acids), this component fails, and the first planned experiment is designed to detect exactly that before anything else is built on it.
Why not just use MethylProphet and skip histology entirely? MethylProphet at 0.72 correlation on ENCODE is a genuine threat to the slide model, and I treat it as a baseline rather than arguing it away. What it cannot produce is spatial output or microbial sensitivity: it operates on bulk homogenates, and HEST-1k showed that morphological features carry biological signal (nuclear shape correlating with markers like GATA3) that bulk RNA-seq misses. If MethylProphet matches my model on gene-group AUROC, the honest conclusion is that the slide adds nothing here, and the experiments are built to surface that outcome rather than bury it.
Why not just do the mouse experiment and candidate-gene MSP, with no computational model at all? This is the sharpest objection, because the mouse arm alone answers the mechanistic question. My answer is that the mouse experiment tells you the mechanism in one defined setting, while the model answers a different question — whether microbial epigenetic pressure is legible in archival human tissue at population scale, which a 33-biopsy study (Xia et al.) cannot. Candidate-gene approaches also have a documented failure mode: To et al. found one tumor (Ca36) with SOCS-1 methylation but no expression change, attributed to heterogeneity — precisely the noise that gene-group-level prediction plus enrichment is designed to absorb, as SlideGraphmethyl demonstrated when its GSEA recovered hallmarks like EMT and hypoxia from gene groups.
Proposed Experiments and Decisive Checks
The first experiment gates everything else: train the SCFA imputer on paired metabolomics–RNA-seq data and report per-metabolite held-out Pearson correlation, with success defined as butyrate or at least two other short-chain fatty acids ranking among reproducibly well-imputed metabolites at a coverage rate comparable to UnitedMet's 48% to 67%. If it fails, the proposal pivots to host gene-expression signatures of butyrate exposure as a weaker proxy, and says so.
The second experiment trains the extended graph model on TCGA-COAD/READ (roughly 600 slides with paired 450K arrays) against the baselines above, reporting per-gene-group AUROC with paired bootstrap intervals. The third applies GSEA to predicted methylated gene groups stratified by the contrast $d_i$, testing whether DNMT-pressure and SCFA-depletion gene sets are statistically disjoint in pathway composition. The fourth is the four-arm gnotobiotic mouse study, with bisulfite sequencing of Cdx2 and Mlh1 promoters, H3K27ac ChIP, DNMT expression, Ki-67 crypt counts, and full histomorphometry as endpoints, powered to detect the difference between the prevention and rescue arms.
The decisive checks, in order: the imputer must recover SCFAs; the model must beat the SlideGraphmethyl-equivalent ablation with the microbial covariates as the only difference; the two pathway collections must not fully overlap; and the rescue arm must show restored H3K27ac without promoter demethylation. Each check can fail independently and informatively, and the design reports any of those failures as a finding rather than a footnote.
References
- Bacteria pathogens drive host colonic epithelial cell promoter hypermethylation of tumor suppressor genes in colorectal cancer
- Butyrate-producing Faecalibacterium prausnitzii suppresses natural killer/T-cell lymphoma by dampening the JAK-STAT pathway
- The crosstalk between probiotics and T cell immunity
- MicroRNA-143 targets DNA methyltransferases 3A in colorectal cancer
- Comprehensive metabolomic and epigenomic characterization of microsatellite stable BRAF-mutated colorectal cancer
- UnitedMet harnesses RNA–metabolite covariation to impute metabolite levels in clinical samples
- A Novel Approach to Linking Histology Images with DNA Methylation
- Constitutional activation of IL-6-mediated JAK/STAT pathway through hypermethylation of SOCS-1 in human gastric cancer cell line
- The role of complex interactions between the intestinal flora and host in regulating intestinal homeostasis and inflammatory bowel disease
- Effect of Dietary Supplementation of Glycerol Monolaurate on Growth Performance, Digestive Enzymes, Serum Immune and Antioxidant Parameters, and Intestinal Morphology in Black Sea Bream
- Oregano Oil Combined with Macleaya Cordata Oral Solution Improves the Growth Performance and Immune Response of Broilers
- A New Paradigm for Genome-wide DNA Methylation Prediction Without Methylation Input
- HEST-1k: A Dataset For Spatial Transcriptomics and Histology Image Analysis