Extended Data Figure 8: Cross-validation accuracy for classifying total bacterial load in fecal samples from MSKCC allo-HCT recipients.

A schematic illustration of our classification model is enclosed in the dashed box. Our model uses two threshold parameters, ${\theta }_o$ and ${\theta }_t$, to convert the oral bacterial fraction and total bacterial load into binary categories, respectively. Given training data, we optimized the two cutoff parameters by minimizing the P value of Fisher’s exact test of independence, subject to the constraints ${\theta }_o\in \left[{10}^{-4},1\right]$ and ${\theta }_t\in \left[{10}^3,{10}^{10}\right]$. The optimized ${\theta }_o$. was subsequently applied to predict high or low bacterial loads in the test set by comparing the oral bacterial fractions to ${\theta }_o$. Simultaneously, we binarized the observed bacterial loads by comparing them to ${\theta }_t$. Accuracy was assessed by comparing the predicted bacterial load categories to the observed bacterial load categories. In the boxplot, each dot corresponds to a single 5-fold cross-validation split and the random train-test split was repeated 50 times. Box plots represent the median, 25^th and 75^th percentiles and whiskers represent the 95^th and 5^th percentiles.