Fig. 4

Machine learning-based prediction of aGvHD severity from the pre-HSCT gut microbiota composition. A Relative abundances of the 12 most abundant families over time in the gut in patients with aGvHD grades 0–I versus II–IV. B Importance plot of top 20 predictive gut ASVs identified by the svmLinear model with importance scores indicating the mean decrease in prediction accuracy in case the respective ASV would be excluded from the model. The final cross-validated svmLinear model predicted aGvHD (0–I versus II–IV) from the abundances of gut ASVs pre-HSCT with 86% accuracy (95% CI: 65 to 97%). The ASVs that were also confirmed by Boruta feature selection are indicated with asterisk. C Conditional inference tree (CTREE) displaying ASVs identified as significant split nodes by nonparametric regression for prediction of aGvHD. Numbers along the branches indicate split values of variance stabilized bacterial abundances. The terminal nodes show the proportion of samples originating from patients (n = number of samples) with aGvHD grade 0–I vs II–IV. D Boxplots depicting the log transformed relative abundances of the predictive ASVs at time points up to the transplantation in aGvHD grades 0–I compared with grades II–IV patients. E Trajectories of Lactobacillaceae and Tannerellaceae ASVs that were identified by tree-based sparse LDA, including ASV 3 and ASV 128 that were predictive for aGvHD (bold lines), in patients with aGvHD grades 0–I vs II–IV