The microbes in our gut release small molecules — short-chain fatty acids, indoles, secondary bile acids — that travel into host tissue and behave like drugs: they bind host receptors, reshape chromatin, and steer immune cells toward or away from inflammation. A recent screen mined a large inflammatory-bowel-disease cohort and predicted 983 interactions between 135 such metabolites and 80 human proteins Expanding the drug discovery space with predicted metabolite–target interactions, which is exactly the kind of list a drug-discovery pipeline wants. But the screen's own authors flag the flaw that blocks everything downstream: knowing a metabolite binds a target tells you nothing about whether it activates or blocks it, and their follow-up assays were bulk measurements that average over all the cell types in a well. Meanwhile, the machine-learning side of biology has learned to read single-cell transcriptomes as fingerprints of perturbation — models like CPA predict how cells respond to unseen drugs and gene edits with $R^2$ of 0.92–0.94 on held-out conditions Learning interpretable cellular responses to complex perturbations in high-throughput screens — yet no single-cell perturbation compendium anywhere contains a single microbial metabolite scPerturb: Harmonized Single-Cell Perturbation Data.
This proposal builds the missing measurement: a single-cell perturbation atlas in which a panel of microbial metabolites is applied, one at a time and at graded doses, to co-cultured human intestinal epithelium and matched donor immune cells, with the transcriptome of each cell read out individually. The atlas is then used to answer the question screening cannot: for each predicted metabolite–target pair, does the metabolite act as an agonist or an antagonist, in which cell types, and is the target necessary for the response? That answer is what converts a binding prediction into a validated therapeutic hypothesis, and it comes with a built-in guard against the field's most common artifact — generic stress signatures that masquerade as specific biology.
Core Claim
The central claim is that a cell-type-resolved, dose-resolved differential-expression signature, corrected against an empirical background of "what differential expression looks like everywhere," is sufficient to infer the functional direction of a metabolite–target interaction — agonism versus antagonism — and to test the target's necessity, without first running a dedicated biochemical assay for each pair. Three ingredients make this possible where prior work stopped short. First, the perturbation panel includes anchor metabolites whose targets and directions are already established by mechanism (butyrate inhibiting histone deacetylases to drive Foxp3 and regulatory T-cell expansion Direction Summary; indoles activating the aryl hydrocarbon receptor Direction Summary; secondary bile acids such as isoalloLCA inhibiting RORγt, the master switch of inflammatory Th17 cells Gut microbial metabolites in inflammatory bowel disease: immunological mechanisms regulating Treg/Th17 balance and therapeutic potential; and 4-hydroxybenzeneacetic acid activating JAK2/STAT3 to induce CXCL3 in colorectal cancer cells Gut microbial metabolite 4-hydroxybenzeneacetic acid drives colorectal cancer progression via accumulation of immunosuppressive PMN-MDSCs). These anchors calibrate the mapping from signature to direction. Second, generic stress is subtracted explicitly: DEET's reanalysis of 3,162 human differential-expression comparisons showed that a recurring set of inflammatory and stress genes — the "DE Prior" — appears differentially expressed in almost every experiment regardless of condition Differential Expression Enrichment Tool (DEET): An interactive atlas of human differential gene expression, and uncorrected enrichment reads this noise as specific mechanism. Third, necessity is tested genetically: if knocking down the predicted receptor collapses the metabolite's residual signature, the target is on the causal path; if the signature persists, the interaction was a correlation.
The proposal deliberately does not attempt in vivo efficacy models for every pair. Its scope is the step the field is missing — directional, cell-type-specific target engagement — with in vivo work reserved for the two or three pairs that survive.
Research Program
The panel comprises roughly forty metabolites: the anchor set above; structurally diverse short-chain fatty acids, indoles, and secondary bile acids; and the specific novel pairs nominated by the screening study — oleanolic acid with GABRG2 and alpha-CEHC with THRB, plus its known positives nicotinic acid–GPR109A and linoleoyl ethanolamide–GPR119 Expanding the drug discovery space with predicted metabolite–target interactions. Each metabolite is applied at three doses to a co-culture of intestinal organoid-derived epithelium and autologous peripheral immune cells from each donor. Readout is single-cell RNA sequencing, sized per the practical threshold documented in scPerturb of at least 300 cells per perturbation scPerturb: Harmonized Single-Cell Perturbation Data. Differential expression per cell type is fit with negative-binomial models of the DESeq2 type rather than fold-change cutoffs, because count variance grows with the mean and naive thresholds treat a noisy two-fold change as equal to a tightly regulated one Bioinformatics Analysis of the Effects of Tobacco Smoke on Gene Expression; donor-level confounding is handled with counterfactual adjustment in the style of CoCoA, which separated 215 true disease genes from donor effects in a 70,634-nucleus Alzheimer's cohort and beat pseudo-bulk on precision-recall Counterfactual inference for single-cell gene expression analysis.
The analytic core is a two-step decomposition of every perturbation signature. For metabolite $m$ in cell type $c$, let $\Delta_{m,c,g}$ be the log-fold-change of gene $g$, and let $p_g$ be the gene's DE-Prior frequency — the fraction of DEET's 3,162 comparisons in which gene $g$ is differentially expressed Differential Expression Enrichment Tool (DEET): An interactive atlas of human differential gene expression. Fit
$$r_{m,c,g} \;=\; \Delta_{m,c,g} - \beta_c\, p_g,$$
and keep the residuals $r_{m,c}$ as the metabolite-specific signature. Functional enrichment (GSEA over Gene Ontology and KEGG, and a protein-interaction-augmented variant following MOMENTA, which expanded metabolic gene sets five-fold and increased significant pathway calls 25-fold under glucose starvation Multiomic Metabolic Enrichment Network Analysis Reveals Metabolite–Protein Physical Interaction Subnetworks Altered in Cancer) is then run on $r_{m,c}$, not on the raw $\Delta$. Directional inference compares each residual signature to a library of anchor signatures with known direction: a novel metabolite whose residual signature is anti-correlated with the isoalloLCA anchor, concentrated in Th17-like cells, and enriched for RORγt-target genes is predicted to antagonize Th17 differentiation. Prediction quality is scored on held-out anchors by leave-one-metabolite-out cross-validation, and the strongest novel predictions go to luciferase reporter and flow-cytometry assays for the paired receptor or lineage marker, then to CRISPR-interference knockdown of the predicted target in the relevant cell type to test whether the signature collapses.
Two design choices respond directly to tensions in the literature. The cell-type-resolved readout exists because the same metabolite can be protective or pathogenic depending on the metabolic state of the cell: Treg plasticity work shows Foxp3-positive cells drifting into pathogenic ex-Tregs under succinate and hypoxia while butyrate stabilizes them Gut microbial metabolites in inflammatory bowel disease: immunological mechanisms regulating Treg/Th17 balance and therapeutic potential, so a target nominated in one context can be wrong in another. And the necessity test exists because the "prognostic trap" is well documented: proteins upregulated as downstream consequences of stress look exactly like causal drivers in expression data, and blocking them does nothing Differential Expression of Lipid Metabolism-Related Proteins in Different Breast Cancer Subtypes.
Baselines
The naive reference point is the field's default: raw differential-expression lists selected by fold-change threshold, fed to standard over-representation analysis with no background correction. This is the practice DEET explicitly showed to be distorted, because the generic DE Prior dominates such lists Differential Expression Enrichment Tool (DEET): An interactive atlas of human differential gene expression. It will be run on the identical atlas data with the same enrichment databases; beating it means recovering anchor mechanisms at a strictly higher true-positive rate while flagging fewer stress pathways as metabolite-specific.
The strongest existing competitor is the full pipeline of Expanding the drug discovery space with predicted metabolite–target interactions: ensemble machine-learning prioritization of cohort metabolomics, ligand-based virtual screening against ChEMBL, and bulk primary-cell assays. That study is the closest prior work — it already connects metabolite signaling, differential expression, enrichment, and target validation — and it reported 983 candidate metabolite–target interactions across 135 metabolites and 80 targets, with 11 metabolites carried into in vitro bulk assays; no reported result found on Lacuna for its accuracy at predicting direction of target engagement, because direction was out of its scope by design. To compare fairly, we run its screening stage on our panel to obtain its ranked target list, and its bulk assay as a pseudo-bulk readout of the same co-cultures, with the same donor cells, doses, and sequencing depth. Beating it means two things on held-out anchors and novel pairs: correct direction calls where screening is silent, and cell-type resolution where bulk assays cannot speak — for instance detecting an immune-cell-specific effect that the bulk average dilutes to nothing.
The third baseline is transfer from existing deep perturbation models: CPA, which reports out-of-distribution $R^2$ of 0.92 for Nutlin and 0.94 for BMS on sci-Plex 2 against linear baselines of 0.85 and 0.89 Learning interpretable cellular responses to complex perturbations in high-throughput screens, and CRADLE-VAE, which reaches synthetic-cell QC pass rates of 93–98% versus 70–80% for models without artifact disentanglement Cradle-VAE: Enhancing Single-Cell Gene Perturbation Modeling with Counterfactual Reasoning-based Artifact Disentanglement. Both will be pretrained on their published benchmarks, then zero-shot and fine-tuned on our atlas with the same compute and tuning budget as our model, and evaluated on held-out metabolites with the same $R^2$ metric on predicted expression profiles.
Two ablations isolate the key ingredients. Removing the DE-Prior subtraction (using raw $\Delta_{m,c}$ throughout) isolates the background-correction contribution; pooling all cell types into pseudo-bulk before analysis isolates the cell-type-resolution contribution, and is additionally motivated by CoCoA's finding that cell-type-stratified causal inference beat pseudo-bulk on precision-recall Counterfactual inference for single-cell gene expression analysis.
Why not just…?
Why not just run more virtual screening, as the closest prior work did? Screening scales cheaply, but its own authors state that structural similarity to known ligands cannot distinguish agonism from antagonism, and that its cell readouts were bulk assays Expanding the drug discovery space with predicted metabolite–target interactions. The atlas answers the question screening cannot even pose, and includes the anchor metabolites that make the answer auditable. Screening remains part of this proposal — as a baseline for ranking — not as an alternative.
Why not just fine-tune CPA on existing perturbation data? This is a real threat, and the experiments will settle it directly. The scPerturb compendium harmonizes 44 datasets spanning CRISPR perturbations, small molecules, and cytokines, and contains no microbial metabolites scPerturb: Harmonized Single-Cell Perturbation Data; CPA's authors also document degraded fidelity on perturbations far from the training distribution Learning interpretable cellular responses to complex perturbations in high-throughput screens. Microbial metabolites are precisely such a distribution shift. But if fine-tuning on a small slice of the atlas predicts held-out metabolites as well as our approach, that is a useful finding, and the atlas then stands on its own value as the first benchmark for this perturbation class.
Why not just do bulk RNA-seq on treated primary cells, which is far cheaper? Bulk was the readout of the prior screening study's validation step Expanding the drug discovery space with predicted metabolite–target interactions, and for purely epithelial, purely cell-autonomous effects it may suffice. But the mechanistically important metabolite effects here are lineage-specific — Treg versus Th17 skewing, MDSC recruitment — and pseudo-bulk averaging demonstrably loses such signals relative to cell-type-stratified analysis Counterfactual inference for single-cell gene expression analysis. The pseudo-bulk ablation on our own data will quantify exactly how much is lost, and if the loss is negligible for a metabolite class, future screens of that class can justifiably drop to bulk.
Why not skip the atlas and validate the top targets in mice now? In vivo work is expensive, and its interpretation depends on the direction question being settled first — inhibiting versus mimicking a metabolite's effect are opposite experiments. The pleiotropy literature is a warning about jumping to targets from association: Mendelian randomization nominated TNFSF12 as protective against intracranial aneurysm (odds ratio 1.30) yet flagged that inhibiting it raises risk of cardioembolic stroke and atrial fibrillation Systematic Mendelian Randomization Exploring Druggable Genes for Hemorrhagic Strokes. This proposal reserves animal models for the pairs whose direction and necessity the atlas establishes.
Proposed Experiments and Decisive Checks
Experiment one builds the atlas: roughly forty metabolites, three doses each, co-cultures from at least four donors, single-cell readout, yielding per-cell-type, per-dose differential signatures. The decisive check is anchor recovery under leave-one-anchor-out cross-validation: a held-out anchor's predicted target and direction must match its established mechanism — butyrate to HDAC inhibition and Treg skewing Direction Summary, 4-HPA to JAK2/STAT3 activation and CXCL3 induction Gut microbial metabolite 4-hydroxybenzeneacetic acid drives colorectal cancer progression via accumulation of immunosuppressive PMN-MDSCs, bile-acid antagonism of RORγt Gut microbial metabolites in inflammatory bowel disease: immunological mechanisms regulating Treg/Th17 balance and therapeutic potential — at a rate clearly above the uncorrected and pseudo-bulk ablations.
Experiment two tests held-out metabolite prediction: train the signature model on a subset of the panel, predict expression profiles for held-out metabolites, and score $R^2$ against the observed profiles, using the linear baseline (0.85/0.89 on sci-Plex 2 Learning interpretable cellular responses to complex perturbations in high-throughput screens) as the reference scale rather than a target, since metabolites are a different perturbation class. Experiment three tests the novel pairs: direction predictions for oleanolic acid–GABRG2 and alpha-CEHC–THRB Expanding the drug discovery space with predicted metabolite–target interactions go to receptor reporter assays, and CRISPR-interference knockdown of each predicted target tests whether the metabolite's residual signature collapses — the necessity criterion that separates a causal target from the prognostic trap. The final check transfers the surviving signatures to one in vivo model: metabolite supplementation in DSS colitis with tissue single-cell sequencing, asking whether the atlas signature reappears in the matched murine cell types. If the anchors recover, the novel pairs confirm, and the necessity test holds, the field gains what it currently lacks: a transcriptome-readable, directionally resolved map of how microbial chemistry talks to human targets, and a reusable template for any perturbation class too poorly annotated for screening alone.
References
- Expanding the drug discovery space with predicted metabolite–target interactions
- Learning interpretable cellular responses to complex perturbations in high-throughput screens
- scPerturb: Harmonized Single-Cell Perturbation Data
- Gut microbiota mediated T cells regulation and autoimmune diseases
- Gut microbiota metabolites in the immunoregulation of enteritis: research progress
- Gut microbial metabolites in inflammatory bowel disease: immunological mechanisms regulating Treg/Th17 balance and therapeutic potential
- Gut microbial metabolite 4-hydroxybenzeneacetic acid drives colorectal cancer progression via accumulation of immunosuppressive PMN-MDSCs
- Differential Expression Enrichment Tool (DEET): an interactive atlas of human differential gene expression
- Bioinformatics Analysis of the Effects of Tobacco Smoke on Gene Expression
- Counterfactual inference for single-cell gene expression analysis
- Multiomic Metabolic Enrichment Network Analysis Reveals Metabolite–Protein Physical Interaction Subnetworks Altered in Cancer
- Differential Expression of Lipid Metabolism-Related Proteins in Different Breast Cancer Subtypes
- Cradle-VAE: Enhancing Single-Cell Gene Perturbation Modeling with Counterfactual Reasoning-based Artifact Disentanglement
- Systematic Mendelian Randomization Exploring Druggable Genes for Hemorrhagic Strokes