Fig. 8.

ROC curves for classifiers illustrate specificity and sensitivity tuning by miRNA target set location and number. Receiver operator characteristic (ROC) curves derived from data in Figs. 6 and 7 were generated by calculating sensitivity and specificity at varying the mKate2 thresholds (previously fixed at 10²). For reference, circles along the curves show sensitivity and specificity at the fixed threshold used for Figs. 6 and 7. a Circuit diagrams and ROC curves for miRNA classifiers with varying target site position as tested in Fig. 6. Circuit shading color corresponds to ROC curve color. The classifier with miR-21-5p target sets in different UTRs (red) shows greater maximal accuracy compared to that with target sites in only the 3′ UTR (blue) as indicated by nearer approach towards the upper left and greater area under the curve (AUC = 0.9947 vs 0.9895). b Similarly, the 3-input classifier with miR-21-5p in the 5′ UTR and miR-23a-3p and miR-106b-3p in the 3′ UTR (purple) exhibits better overall classification than the variant with target sets in only the 3′ UTR (yellow), with AUCs of 0.9952 and 0.9881, respectively. c Circuit diagrams and ROC curves for miRNA classifiers with or without target site duplication across UTRs for let-7a-5p and miR-21-5p. Data for c and d correspond to that from Fig. 7. Both classifier variants show similar maximum accuracy when classifying between HEK293FT and HeLa cells, though at the example threshold of 10² the variant without duplication (red, AUC = 0.9833) prioritizes sensitivity while the variant with target set duplication (blue, AUC = 0.9750) prioritizes specificity (see inset for magnified view). d For classifiers with target sites for miR-25-3p and miR-21-5p, the circuit without duplication (purple, AUC = 0.9235) exhibits better classification compared to that with duplication (yellow, AUC = 0.7018), since repression was determined by both miR-25-3p and miR-21-5p rather than being dominated by miR-21-5p and also since potential uORFs are avoided by removing miR-25-3p targets from the 5′ UTR (both effects enhance sensitivity). Error bars indicated s.d. calculated by threshold averaging from three technical triplicates