Two findings sit next to each other in the literature without ever meeting. First, the gut microbiome — the community of bacteria whose metabolites enter circulation and tune host immunity — helps decide whether checkpoint immunotherapy (antibodies that release the brakes on anti-tumor T cells) works: patients whose stool is enriched for certain bacterial species and short-chain fatty acids respond better, and transplanting stool from responders into mice makes tumors shrink and T cells flood in. Second, the tumors that resist immunotherapy are often not lacking immune cells — the cells are physically walled off, sequestered behind fibroblast barriers or encapsulated interface niches that keep them away from the tumor. Nobody has asked the obvious question connecting these: do gut microbial signals control whether those walls get built, and can we detect their remodeling cheaply, from the routine stained tissue slides every patient already has?
This proposal tests that question directly. The core idea is a linking measurement: a deep-learning model that predicts, from a standard hematoxylin-and-eosin (H&E) slide, a virtual multi-marker protein map, from which we compute the same "barrier" geometry that expensive spatial platforms have shown to govern resistance. We then use this measurement twice — in mice, to test causally whether fecal transplants and microbial metabolites dismantle or erect these walls, and in patients, to test whether combining the spatial readout with stool metagenomics predicts resistance better than either alone.
Core Claim
The claim is that the diet–microbiome–treatment axis acts on immunotherapy resistance partly through spatial architectural remodeling of the tumor — a mechanism invisible to the dissociated cell counts and bulk sequencing that microbiome studies currently rely on — and that this remodeling is quantifiable from routine H&E slides via virtual protein multiplexing, making it a practical clinical and preclinical endpoint.
Three lines of work motivate this. Clinical microbiome studies show species and metabolites that track with immunotherapy response: Dou et al.'s multi-omics study of MSI-H gastrointestinal cancer found responder-enriched bacteria (B. caccae, A. putredinis) and metabolites (butyrate, L-arginine), and their stool-only XGBoost model predicted primary resistance at 0.83 accuracy on an external validation set. Mechanistic mouse work shows causal metabolite signaling: butyrate from Roseburia intestinalis binds TLR5 on CD8+ T cells and turns cold tumors hot (Kang et al.), and indole-3-carboxylic acid from Lactobacillus gallinarum competitively blocks the AHR receptor ($K_D = 39.4\ \mu$M versus $61.5\ \mu$M for its competitor kynurenine) to suppress regulatory T cells (Fong et al.). Separately, spatial-omics studies show resistance is architectural: in hepatocellular carcinoma, a graph neural network over spatial proteomics found that non-responders' lymphoid aggregates are encapsulated behind an SPP1+ interface barrier, achieving 0.93 AUROC for pre-treatment response prediction (Trevino et al.); in metastatic colorectal cancer, fibroblast walls segregate C1QC+ macrophages from CD4+ T cells in non-responders, with high algorithmic "barrier scores" tracking poor survival (Zhang et al.).
The gap is precise. The microbiome papers evaluate tumor tissue only as average cell densities, never as spatial architecture; the spatial papers model the tumor as a closed system with no microbial, dietary, or metabolomic inputs. And the tool that could connect them at scale — HistoPlexer (histopathology-based protein multiplex generation), a conditional GAN that predicts an 11-channel multiplexed protein map directly from H&E, using pyramid and contrastive losses to absorb the misalignment between serial sections — has never been adapted to detect resistance niches or barrier geometry, let alone linked to non-invasive microbial predictors. HistoPlexer improved survival prediction by 3.18% C-index and immune-subtype classification by 17.02% F1 over H&E alone in melanoma; we want the analogous gains for barrier quantification in gastrointestinal cancer, where the luminal microbiome is anatomically adjacent to the disease.
Research Program
The program has three phases, each feeding the next.
Phase 1 — a virtual barrier score from routine slides. We train an H&E-to-multiplex generator, following HistoPlexer's architecture and its three loss terms (adversarial, Gaussian pyramid, and patch-wise contrastive), on gastrointestinal cancer tissue rather than melanoma. Ground truth comes from real multiplexed imaging: the metastatic colorectal cancer spatial atlas of Zhang et al. provides 38-plex Imaging Mass Cytometry on 50 cases and 314,774 mapped cells, plus spatial proteomic and transcriptomic cohorts. From the predicted multiplexes we compute two geometric quantities: a barrier score (the degree of fibroblast encapsulation separating C1QC+ macrophage neighborhoods from CD4+ T cells, following Zhang et al.'s definition) and an interface-niche score (encapsulation versus direct exposure of lymphoid aggregates, following Trevino et al.). Phase 1's success criterion is agreement between virtual and real multiplex-derived scores on held-out slides where both exist.
Phase 2 — clinical fusion. In a prospective cohort modeled on Dou et al.'s 77-patient MSI-H design (paired stool metagenomics, plasma metabolomics, pre-treatment biopsy H&E, and anti-PD-1/PD-L1 outcomes), we test two things: whether responder-associated species and metabolites correlate with low barrier scores at baseline, and whether a model fusing stool features with virtual spatial features predicts resistance better than stool features alone. This phase also addresses the pharmacokinetic blind spot that Fong et al. themselves flag — that metabolite concentrations needed in vitro exceed measured plasma levels — by measuring metabolites directly in resected tumor tissue with targeted mass spectrometry, asking whether locally accumulated metabolites align with barrier status.
Phase 3 — causal test in mice. Using the syngeneic MC38 (microsatellite-instability-high) and CT26 (stable) colorectal models that all three mechanistic microbiome papers employed, we run fecal microbiota transplantation from human responders versus non-responders (as Dou et al. did), oral butyrate (following Kang et al.), and — as a predicted negative perturbation — a high-salt diet, which Deng et al. showed deranges bacterial tryptophan metabolism and impairs chemotherapy. All arms receive anti-PD-1. The primary endpoint is not tumor volume alone but the change in virtual barrier score and lymphoid-aggregate exposure across treatment, with terminal multiplexed immunofluorescence confirming the virtual readout. The decisive statistical question is mediation: does the microbiome's effect on response travel through barrier remodeling, or around it? We test this by asking whether the effect of microbiome arm on response shrinks toward zero once change in barrier score is included as a mediator.
If the causal arm succeeds, the translational payoff is the one the spatial-resistance literature explicitly lacks: a non-invasive route (diet, probiotics, metabolites) to dismantling the physical walls that current spatial papers can only describe. The nearest demonstration that multi-omics profiling can change treatment decisions is the melanoma TuPro cohort, where multiomics-guided therapy achieved a 38% objective response rate and a median progression-free survival of 8.34 months versus 2.0 months in matched controls (Miglino et al.) — notably, with no microbiome profiling at all.
Baselines
The naive reference point is the current clinical standard: microsatellite-instability status plus PD-L1 Combined Positive Score. This is the baseline every cited study agrees is inadequate — Dou et al. document substantial primary resistance in MSI-H patients, and Jiang et al.'s biology-guided deep learning (BgDL) outperformed PD-L1 CPS for predicting objective response with an AUC of 0.753, though no reported CPS number was found on Lacuna. Any fused model must clear 0.753-equivalent performance to be clinically interesting, and MSI status alone defines the floor.
The strongest competitor on the exact clinical question we pose — predicting primary resistance in MSI-H gastrointestinal cancer from non-invasive samples — is Dou et al.'s XGBoost over metagenomic species and functional pathways, which reported an accuracy of 0.83 on an external validation set of 29 stool samples. We will run this exact model on our cohort's stool data with the same feature space, the same nested cross-validation, and the same tuning budget, so the only difference between it and our model is the added virtual-spatial features. Beating it means higher AUROC for resistance prediction on a held-out split with confidence intervals excluding chance-level overlap, plus a pre-registered check that the gain survives when controlling for tumor mutational burden and stage.
The strongest spatial baseline, and the realistic ceiling, is the MIF-GNN of Trevino et al., which achieved 0.93 AUROC for response prediction from pre-treatment biopsies — but it requires Akoya Phenocycler spatial proteomics, not routine slides, so it cannot run on the same inputs our clinical cohort provides. We will nevertheless train and evaluate it on the Phase 1 subset where real multiplex data exists, on the same slides and the same response labels, to establish what the expensive platform buys; our H&E-only fused model is compared against it honestly as a cheaper approximation, not an equal. For spatial niche classification we also run Zhang et al.'s ResNet18-on-spatial-features approach on virtual multiplexes.
The key ablation isolates our methodological ingredient: the same fusion model trained with raw H&E features (no virtual multiplex intermediate) versus virtual-multiplex-derived barrier features. This directly tests whether the multiplex intermediate — with its explicit fibroblast, macrophage, and T-cell channels — is what carries the barrier signal, echoing HistoPlexer's demonstration in melanoma that virtual multiplexes beat H&E alone by 3.18% C-index and 17.02% F1. A second ablation replaces the virtual barrier score with simple immune-cell density counts, testing whether geometry matters or just cell quantity — the exact confound the spatial papers warn about.
Why not just…?
Why not just extend the stool-only XGBoost predictor? This is the most serious alternative, and we treat it as the primary baseline rather than arguing around it. A stool-only model already predicts at 0.83 accuracy, needs no tissue at all, and could plausibly absorb most of the predictive signal. What it cannot do is answer why specific patients resist, or point to an intervention — a score from stool features gives no handle on whether the failure is an architectural wall, and the MSI-H resistance problem is precisely that biomarkers predict without explaining. Our design settles this empirically rather than rhetorically: if the fused model does not beat the stool-only baseline, the predictive claim fails and we will say so; the preclinical causal phase, which no stool-only model can address, would stand or fall independently.
Why not just use real multiplexed platforms instead of a virtual one? Imaging Mass Cytometry and Phenocycler give the ground truth directly, and if tissue, budget, and time were unlimited, we would simply use them everywhere. The file's own literature explains why that does not scale: standard clinical immunohistochemistry handles only one to three markers per section, while high-plex platforms carry high reagent costs, low throughput, destructive sampling, and limited availability outside specialized centers (HistoPlexer; TuPro). Our mouse experiments need barrier measurements across many animals and time points, and the clinical phase needs to work on archival routine slides. Still, this alternative is a real threat to validity: if virtual barrier scores disagree with real ones, the whole chain fails, which is why Phase 1 validates against real multiplex scoring on a paired subset before anything downstream runs, and why every mouse experiment terminates with confirmatory multiplexed immunofluorescence rather than relying on the virtual readout alone.
Why not just repeat the existing mouse fecal-transplant experiments with flow cytometry and cell-density counts? All three mechanistic microbiome papers already did this — Dou et al. measured intratumoral T-cell infiltration by mIHC after fecal transplant, Kang et al. and Fong et al. used flow cytometry and conventional IHC scoring. The problem is what dissociation destroys: position. The spatial studies show two tumors can hold identical average densities of T cells and macrophages while one keeps them sequestered behind fibroblast walls and the other lets them contact tumor — and the sequestered one resists (Zhang et al.; Trevino et al.). Our central hypothesis is that the microbiome acts on exactly this positional variable, so a density-only endpoint could return a null result even when the mechanism is real. We include density endpoints as an ablation arm so the experiment itself shows whether geometry adds anything over counts.
Decisive Checks
Four results would make or break the proposal. First, Phase 1's agreement between virtual and real multiplex-derived barrier scores on held-out paired slides; without it, nothing downstream is interpretable. Second, the Phase 2 comparison against Dou et al.'s 0.83-accuracy stool-only baseline under matched conditions — improvement would show spatial information adds predictive value; failure would confine the contribution to mechanism. Third, the Phase 3 mediation analysis: if the microbiome arm's effect on tumor response largely disappears once barrier-score change is included as a mediator, the causal chain runs through spatial remodeling, which would be the first direct evidence that gut microbes govern the physical architecture of immunotherapy resistance. Fourth, the high-salt-diet negative control, where deranged tryptophan metabolism (Deng et al.) should, on our account, raise barrier scores and worsen response — a directional prediction that a density-only story cannot make.
References
- Multi-omics of the gut microbial ecosystem in the immunotherapy resistance in microsatellite instability-high gastrointestinal cancer patients
- Roseburia intestinalis generated butyrate boosts anti-PD-1 efficacy in colorectal cancer by activating cytotoxic CD8 + T cells
- Lactobacillus gallinarum-derived metabolites boost anti-PD1 efficacy in colorectal cancer by inhibiting regulatory T cells through modulating IDO1/Kyn/AHR axis
- 478 Spatial omics and deep learning define an immunotherapy resistance niche in hepatocellular carcinoma
- Spatially resolved C1QC+ macrophage-CD4+ T cell niche in colorectal cancer microenvironment: implications for immunotherapy response
- Histopathology-based protein multiplex generation using deep learning
- High-salt diet decreases FOLFOX efficacy via gut bacterial tryptophan metabolism in colorectal cancer
- Feasibility of multiomics tumor profiling for guiding treatment of melanoma
- Biology-guided deep learning predicts prognosis and cancer immunotherapy response