Extended Data Fig. 10: Interaction of antibiotic exposures, survival outcomes and response prediction modeling.

(a) Incidence curves for disease progression in patients were stratified according to high-risk antibiotics exposure in the 3-weeks pre-CAR-T cell infusion time window and according to presence of the peptidoglycan biosynthesis V - pathway in the patients’ stool metagenomes. (b) Deterministic logistic regression models were trained on clinical variables and pre-CAR-T cell infusion gut microbial species composition (mean per patient) of the German cohort and their performance was assessed on the US cohort (with all baseline samples: n = 45 patients; after excluding high-risk antibiotics samples: n = 36 patients). Area under the curve (AUC) receiver operating characteristic (ROC) curves indicating the false-positive and true-positive rates for predicting CR vs. non-CR at day 180 based on an analysis including all patient samples. (c) PCA computed for the whole dataset (that is, training and validation data including high-risk antibiotics samples) based on the features selected during the training process of response at day 180 after CAR-T cell infusion. (d) AUROC curves indicating the false-positive and true-positive rates for predicting CR vs. non-CR at day 180 after training the probabilistic model on the German dataset that excluded high-risk antibiotics exposed samples. (e) AUROC curves indicating the false-positive and true-positive rates for predicting CR vs. non-CR at day 180 based on a deterministic model analysis that excluded samples collected on or less than two weeks after exposure to high-risk antibiotics; data displayed for the US validation cohort. (f) PFS curve of patients after CAR-T cell infusion stratified by the median baseline relative abundance of Bacteroides stercoris as top feature for non-CR as revealed in the machine learning model for CR; only pre-infusion samples without high-risk antibiotic exposure were included. (g) Differences in relative abundances for the four most important species predicting non-response (no CR) grouped by response and high-risk antibiotic exposures; group comparisons carried out by pairwise Mann–Whitney tests with FDR correction of P-values (n = 38 / 41 / 4 / 12 patients [from left to right]); FDR-corrected P-values for B. stercoris are 0.089 (for all three comparisons), for B. fragilis are 0.0725 (for all four comparisons), and for E. sp. CAG38 are 0.0037 (for both comparisons). Statistics in (A) and (F) performed by log-rank tests.