Fig. 3

Predicting microbial composition in real metagenomic data. The prediction error of the kNN versus k (blue lines), compared with the prediction error of the null model(orange lines). Each point represents the average prediction error over 100 ‘leave-one-out’ realizations, where a single microbial sample is randomly selected as a test sample. Different panels show the results of 13 microbial communities from different body sites, obtained from the HMP dataset: anterior nares (a), attached keratinized nares (b), buccal mucosa (c), hard palate (d), left retroauricular crease (e), right retroauricular crease (f), palatine tonsils (g), saliva (h), subgingival plaque (i), supragingival plaque (j), throat (k), tongue dorsal (l), and stool (m). In each case, we define \(\Delta\) as the difference between the prediction errors of the kNN for \(k=k_{min}\) and of the null model, as demonstrated in a. n Purple dots represent the relation between \(\Delta\) and \(k_{min}\) calculated for the different body sites. The solid curve represents a polynomial fit