Gut & MicrobiomeResearch PaperOpen Access

364 Metabolites Mapped for Their Power to Reshape T Cell Immune States

A high-throughput atlas reveals how gut and endogenous metabolites reprogram T cell transcription, pointing to new immunomodulatory targets for inflammatory disease.

Sunday, October 4, 2026 3 views
Published in iScience
A researcher pipetting colorful liquid samples into a multi-well plate in a modern immunology laboratory, with computer screens displaying gene expression heatmaps in the background

Summary

Gut microbes and our own cells make small molecules that tune the immune system, but few have been tested side by side. Researchers exposed activated human T cells to 364 of these metabolites and read out gene activity for each. The responses sorted into three states: resting, metabolically primed, and strongly activated through the T cell receptor. The team then trained an AI model to predict how strongly any metabolite shifts T cells, and it ranked 148 compounds as high-impact. Predictions held up in an independent experiment on 12 metabolites. In a mouse model of colitis, the prioritized candidates reduced disease severity and partly restored immune balance. This is early lab and animal work, but it offers a map for finding diet- and microbiome-derived compounds that calm inflammation.

Detailed Summary

Chronic immune dysregulation drives inflammatory bowel disease and contributes to inflammaging, and the gut is where T cells meet a dense chemical mix of microbial and host-derived metabolites. Short-chain fatty acids, tryptophan-derived indoles, and bile acids are known to steer T cells. But no one had systematically compared large numbers of metabolites on human T cells, and predictive tools for untested compounds were lacking. This study set out to fill that gap.

The team activated primary human CD3+ T cells from healthy donors and treated them with 364 endogenous or microbiota-derived metabolites or a DMSO control. They read out transcription using DRUG-seq, a low-cost 384-well method, in biological duplicate. Unsupervised clustering yielded eight transcriptional states (C0-C7). C0 was strongly activated, with enrichment of TCR signaling and Th1/Th2/Th17 differentiation pathways. C6 carried interferon and stress signatures, contained most DMSO controls, and served as the baseline.

The eight clusters collapsed into three higher-order states: baseline resting, metabolic active, and TCR-high activation. Up- and downregulated gene counts were tightly correlated (Pearson r=0.98), and perturbation grew from resting to TCR-high. Top TCR-high metabolites included 2-methylbutanoic acid, tocotrienol, resveratrol, and indole-3-pyruvic acid (IPA), which induced IL2RA, CD5, LCK, and CCR7. Regulator activity shifted stepwise, from STAT3 and PURA at rest, to KLF2 and MXD3 in the metabolic state, to ATF4, SREBF2, RELA, and JUND at high activation. CD8+ activated and effector-memory signatures rose, while CD4+ equivalents fell (p<0.0001). Th17 signatures increased progressively, and trajectory analyses showed declining transcriptional stemness along the continuum.

The authors then built DRUG-seq-PerturbFormer (DPF), a multi-task transformer trained on 768 wells and 2,541 genes. Predicted and observed values agreed closely for T cell activation score (r=0.96), CytoTRACE2 differentiation potential (r=0.98), and pseudotime (r=0.94). The model separated perturbed from resting cells with an AUC of 0.987 (95% CI 0.976-0.998). A global score combining Euclidean distance (magnitude) and cosine distance (direction) ranked all 364 metabolites and placed 148 in the high tier. It correlated with differentially expressed gene counts (Spearman r=0.76).

Generalization was tested with independent bulk RNA-seq on 12 selected metabolites. DPF assigned them higher scores than DMSO, and CD3E, LCK, and JUND were concordantly upregulated. Four candidates, IPA, folic acid, hippuric acid, and 4-methylcatechol, were examined at 1, 10, and 100 μM, with the model tracking escalating perturbation with dose. In a DSS-induced colitis model, representative candidates attenuated disease severity, were associated with suppressed inflammatory programs, and partly restored immune homeostasis.

Important caveats apply. The screen used activated T cells from healthy donors in vitro, measured transcription rather than function, and used fixed concentrations that may not match gut or blood levels. The colitis work was an acute mouse model, and the model's gene-attribution analyses were explicitly exploratory. These findings nominate candidates and targets rather than supporting clinical use of any metabolite.

Key Findings

  • Profiled primary human activated CD3+ T cells against 364 endogenous and microbiota-derived metabolites using DRUG-seq in biological duplicate, alongside DMSO controls
  • Unsupervised clustering found eight transcriptional states (C0-C7) that collapsed into three higher-order states: baseline resting, metabolic active, and TCR-high activation
  • Up- and downregulated gene counts per metabolite were strongly correlated (Pearson r=0.98, p<2.2×10^-16), with perturbation rising from resting to TCR-high states
  • The DPF model matched observed T cell activation score (r=0.96), CytoTRACE2 (r=0.98), and pseudotime (r=0.94), and distinguished perturbed from resting cells with AUC 0.987 (95% CI 0.976-0.998)
  • Ranking by the combined perturbation score placed 148 of 364 metabolites in the high tier, and the score correlated with differentially expressed gene counts (Spearman r=0.76)
  • Activated CD8+ and effector-memory CD8+ signatures rose across states while CD4+ equivalents fell (p<0.0001), and Th17 signatures increased progressively
  • Independent bulk RNA-seq on 12 metabolites confirmed higher DPF scores than DMSO, and representative candidates attenuated DSS-induced colitis in mice

Methodology

Primary human T cells (from PBMCs of healthy donors) were activated and treated with 364 metabolites or DMSO, then profiled by DRUG-seq in biological duplicate. Analysis included UMAP clustering, KEGG enrichment (hypergeometric test, BH-adjusted), DESeq2 differential expression, regulon inference, ssGSEA/GSVA, and Monocle3 and CytoTRACE2 trajectories, with Wilcoxon rank-sum and Fisher's exact tests. DPF, a multi-task transformer, was trained on 768 wells and 2,541 genes with regression and classification heads. It was validated using independent bulk RNA-seq on 12 metabolites, dose-response predictions at 1, 10, and 100 μM, and a DSS-induced mouse colitis model.

Study Limitations

The screen used in vitro activated T cells from healthy donors and measured transcription rather than protein or functional output, and the colitis validation was an acute mouse model. The authors describe model-attribution analyses (pathway trends from integrated gradients) as exploratory, and note that DPF rankings differ somewhat from simple differentially expressed gene counts. Conflict-of-interest information was not included in the supplied text.

Enjoyed this summary?

Get the latest longevity research delivered to your inbox every week.

Enter your email to subscribe: