a, Effect sizes of predictors on alpha- and beta-diversity. Maximum pairwise effect size (difference between means divided by standard deviation) between categories of each predictor plotted for observed tag sequences (alpha-diversity) and unweighted and weighted UniFrac distance (beta-diversity). Response variables (alpha- and beta-diversity) were derived from the QC-filtered subset of the 90-bp Deblur table containing n = 23,828 biologically independent samples. Numeric predictor variables were converted to quartiles (categorical predictors). Categories within each predictor had a minimum of 75 samples per category. b, Cumulative variation explained by the optimal model of stepwise redundancy analysis (RDA) of predictors: study ID, EMPO level 3, ENVO biome level 3, latitude, and longitude (predictors with values for less than half of samples, including host scientific name, were excluded). Environment (EMPO level 3) and biome (ENVO biome level 3) explained as much variation as study ID when study ID was excluded from the RDA. c, Confusion matrix for random forest classifier of samples to environment (EMPO level 3). The classifier was trained on the 2,000-sample subset, which was then tested on the remaining samples (QC-filtered samples minus 2,000-sample subset). Squares are coloured relative to 100 classification attempts for each true label. Overall success rate was 84%, with the most commonly misclassified sample environments being Surface (non-saline), Animal secretion, Soil (non-saline), and Aerosol (non-saline). d, Receiver operating characteristic (ROC) curve for classification of samples to environment (EMPO level 3). The AUC (area under curve) indicates the probability that the classifier will rank a randomly chosen sample of the given class higher than a randomly chosen sample of other classes. e, Classification success, using a random forest classifier, to EMPO levels 1–3, ENVO material, ENVO feature, and ENVO biome levels 1–3. f, Microbial source tracking: mean predicted proportion of tag sequences from each source environment (EMPO level 3) that occurs in each sink environment. The model was trained on a subset of samples (~20% of each environment), and tested to predict tag sequence source composition in all remaining samples. Aerosol (non-saline), Surface (saline), and Hypersaline samples were not included in this analysis because there were insufficient sample numbers. g, Microbial source tracking: which other environments a sample type most resembles. The model was trained on all source environments except one using a leave-one-out cross-validated model, and then used to classify each sample included in that group. Hence, the predicted classification proportion of environment X to environment X is zero. h, Correlation of microbial richness with latitude. Richness of 16S rRNA tag sequences per sample across EMPO level 2 environmental categories as a function of absolute latitude. Samples from studies that span at least 10° latitude are highlighted in colour, with linear fits displayed per-study as matching coloured lines. Samples from studies with narrower latitudinal origens are shown in grey. The global fit for all samples per category is indicated by a dashed black line.