Fig. 2: Strain-level models predicting social connections.
From: Gut microbiome strain-sharing within isolated village social networks
a,d, True-positive and false-negative network predictions for all relationships (a) and non-kin-different-household relationships only (d) for the village of Hernani (all relationships, n = 210; non-kin-different-household relationships, n = 165). As expected, the model including all relationships performs better than the non-kin, different-household model as there is increased sharing within households and amongst kin. The model also generally performs better at predicting within social clusters as there is increased strain-sharing around nodes with a high clustering coefficient. b,c,e,f, AUC predicting social or familial relationships (b), or non-kin-different-household relationships only (e), compared with un-nominated pairs living in the same village. When looking at stable ties across time (over a 2-year span), the AUC in both predictions remains stable and slightly improved in the strain-sharing-rate-only model (c and f). DeLong 95% confidence intervals are shaded surrounding each line. The diagonal dotted line indicates an AUC of 0.50, a ‘test’ no better than chance. Legends report the mean and s.d. for each classifier’s AUC. SSR, strain-sharing rate.
