Figure 4: Antibiotic properties and cross-resistance.
(a) Weak association between chemical structural similarity between antibiotic pairs and cross-resistance frequency (Spearman’s ρ=0.40, P<10−3, N=66), which disappears when aminoglycosides are excluded (ρ=0.21, P=0.18, N=45). Structural similarity between antibiotics was estimated by the Tanimoto similarity of their molecular fingerprints. (b) Correlation between chemogenomic profile similarity and overlap in the set of accumulated mutations during laboratory evolution (Spearman’s ρ=0.67, P<10−5, N=36). (c) Antibiotic pairs that frequently display cross-resistance interactions show relatively high overlap in their chemogenomic profiles (Spearman’s ρ=0.77, P<10−7, N=36). Dashed red curves on scatterplots A–C indicate smooth curves fitted by Loess regression56. (d) Predicting antibiotic resistance phenotypes from genome sequences. Prediction performance for each antibiotic based on the set of accumulated mutations was measured by the area under the receiver operating characteristic (ROC) curve (AUC). This gives an overall measure of accuracy by taking into account both true positive and false positive rates across all possible cutoffs of the prediction score. Random prediction gives an AUC of 0.5. Variation in resistance among evolved strains can be predicted with 55–88% (76% average) accuracy, depending on the antibiotic studied. Special care was taken to avoid circularity in the predictions.
