Not which bacteria are present, but which metabolites they produce. Valerate helps; indole-3-carboxaldehyde and isovalerate harm. One metabolite reverses direction between the dish and the animal.
Perl et al. followed 129 patients receiving CAR-T at three German centres (Regensburg, Heidelberg, Munich) between 2020 and 2024, with a median follow-up of 360 days. Faecal metabolome was available for 111 patients, serum metabolome for 98 and metagenome for 56; samples were taken at admission, a median of four days before infusion. The design choice that drives the whole paper: metabolites are the primary variable, taxonomy is the reference.
Every reasonable taxonomic analysis was run against three endpoints (day-180 complete response, PFS, OS). The significant species barely overlap across methods and endpoints, and when previously reported species and genera were re-tested in this cohort the direction-consistency AUC came out at 0.474 — below chance. Mean between-patient Bray-Curtis distance at species level was 0.87, which explains why. Only Bifidobacterium longum reproduced (PFS HR 0.48, p = 0.034; OS HR 0.15, p = 0.016). Metabolic pathways, by contrast, showed direction: anaerobic energy-production and fatty-acid biosynthesis pathways trended toward better PFS, aerobic pathways toward worse.
Unsupervised clustering exposed a subgroup of 34 patients with globally low immunomodulatory metabolites and distinctly worse outcome (PFS HR 2.84, OS HR 6.48, both p < 0.0001). At the single-metabolite level short-chain fatty acids (acetate, butyrate, valerate) tracked with better PFS and OS, while branched-chain fatty acids (isobutyrate, 2-methylbutyrate, isovalerate) and the indoles IAA and ICA tracked with worse PFS. One detail worth keeping: indole-3-propionate (IPA) trended positive on both endpoints — same indole scaffold, opposite direction.
Serum clustering produced the same branched-chain-high subgroup (PFS HR 4.39, OS HR 5.67). But ICA was barely detectable in serum and showed no association there, although faecal ICA did. The serum version of the risk score had to drop ICA as a result, and lost discrimination (AUC 0.668 faecal vs 0.594 serum) and significance in the multivariable model. The most likely reading is anatomical: ICA acts in the gut and portal circulation rather than systemically.
Valerate and isovalerate (1,000 µM, 72 h) raised T-bet, Eomes, IFN-γ and granzyme B in CD4⁺ and CD8⁺ CAR-T cells and improved killing of SKOV3 (HER2⁺) and NALM-6 (CD19⁺). ICA and IAA (500 µM) did the opposite, dose-dependently. Single-cell transcriptomics put ICA on an AHR-dependent exhaustion route — a terminal-exhausted CD4 population (CD3D⁻, TCF7⁻, FOXP1⁺, TOX⁺), reversible by CRISPR knockout of AHR — and valerate/isovalerate on an mTOR-S6 plus HDAC2 inhibition route (H3K9 acetylation up roughly eight-fold).
In syngeneic CD19⁺ A20 lymphoma mice, oral ICA worsened burden, expansion and survival; oral valerate improved all three; oral isovalerate — despite helping in vitro — accelerated progression. The explanation is tumour-intrinsic: isovalerate upregulates MCL-1 and PD-L1 and promotes tumour growth directly, an effect that persists in immunodeficient NXG mice. Enhancing immune-cell function in a dish is not a proxy for efficacy.
Valerate below threshold, or isovalerate, IAA or ICA above threshold, each score one point (maximum four). Patients at four points had significantly shorter PFS (HR 2.9, p < 0.0001) and shorter OS (HR 2, p = 0.0654), and the score remained independent after adjusting for centre, co-stimulatory domain, disease, antibiotic exposure, CRS grade, sampling time, sex, age, LDH and ECOG. It held in Heidelberg and Regensburg and trended in Munich (n = 10).
Four caveats. The cohort is observational — human intervention data do not exist yet. ICA concentrations used in vitro and in vivo were pharmacological; gut-level concentrations were not directly measured. Metagenome covers only 56 patients (43.3%) and sampling is a single pre-treatment time point. And the isovalerate reversal is the methodological lesson: any cell-based screen of microbiota metabolites needs an intact in vivo model behind it, or the direction can come out backwards.
All data and conclusions above are taken from the published article. Paragraphs marked as editorial commentary in the Chinese version are the Eynos technical team's judgement, not the authors' conclusions. Products mentioned on this site are for research use only.
Tell us your cell type, sample source and downstream assay — we will suggest an isolation route, a culture condition and the matching catalogue numbers.