Figure 3

Performance of raw neighbour count with varying threshold for removing high-degree metabolites. The degree of each metabolite (the number of reactions it participates in as reactant or product) was calculated, and metabolites were progressively removed. After each removal, the performance of the raw neighbour count for predicting synthetic lethal interactions was assessed by the area under the receiver operating characteristic curve (AUC, panel A) and by the F-Score, defined the maximum harmonic mean of precision and recall along the precision-recall curve (F-Score, panel B). The results in the main text used a nominal threshold of the top 5% of all metabolites, corresponding to 53 metabolites with degree ≥13 of the entire set of 1061 metabolites. The results using a graph diffusion kernel (GDK) are shown to be superior to the raw neighbour regardless of the threshold used.