Fig. 6

Combined machine learning-based prediction of aGvHD severity from the pre-HSCT microbiota composition of all three body sites (gut, mouth, nose). A Conditional inference tree (CTREE) displaying ASVs identified as significant split nodes by nonparametric regression for the prediction of aGvHD from the time points before HSCT. 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 represented samples) with aGvHD grades 0–I vs II–IV. The icons indicate at which body sites the ASVs were detected. The less predominant body sites are indicated in parenthesis. B Boxplots depict 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. The y-axis is the same for all plots and indicated once for the most upper left plot. C Comparison of the combined machine learning model with the individual body site-specific models. The results from the three body site-specific analyses (svmLinear model, Boruta feature selection, regression framework with CTREE) are indicated for the seven taxa identified in the combined machine learning model. Blue color indicates agreement, and white color indicates that the ASV was not detected as being significant in the particular analysis. While the most significant ASV_131 Parabacteroides distasonis in the combined analysis was not identified as being significant in the CTREE for the body site-specific analysis of the gut, a closely related Parabacteroides distasonis (ASV_128) was identified as being the most significant taxon in the gut by the body site-specific CTREE analysis