Fig 1.
A. Anatomy of the hsa-mir-17 hairpin, showing the duplex (grey)and its four ends (k55, k53, k33, k35). B. Vector representation of the true miRNA duplex of the hsa-mir-17 with 10 nucleotides flanking region. B.1. Flanking region of 5′ strand 5′ end (top), and 3′ strand 3′ end (bottom), no zero padding needed. B.2. Duplex sequence, 5′ strand (top), 3′ strand (bottom). Zero padding occurs in the middle to reach the maximum (27nts top, 26nts bottom) length observed in the training set. B.3 Flanking region of 5′ strand 3′ end (top), and 3′ strand 5′ end. Zero padding is performed at the end as flanking region extends beyond the loop tip. Adopted from [15] with modifications.
Fig 2.
Flowcharts of the training and testing procedures.
Fig 3.
Prediction accuracy of MiRduplexSVM and six other methods on duplex identification.
Panels A-F show the prediction accuracy (y-axis) of MiRduplexSVM (in black) and a second compared tool (in colour) as a function of the All Corners Sum Absolute Error (ACSAE, x axis) for errors of 0-8nts. The performance of the Simple Geometric Locator (S.G.L.), MatureBayes, miRPara, MaturePred-Plants, MaturePred-Mammals and MiRdup is shown in A—blue bars, B—pink bars, C—cyan bars, D—green bars, E-red bars and F-purple bars, respectively. Panel G shows the prediction accuracy of MiRduplexSVM (y axis) against the prediction accuracy of each compared tool (x axis). The colour code is the same as in A-F. Symbols (upward triangle, diamond, downward triangle, circle, right pointed triangle, square, left pointed triangle, pentagram star and hexagram star) correspond to errors less than or equal to 0, 1, 2, 3, 4, 5, 6, 7, 8 nucleotides, respectively. All points above the diagonal in G are statistically significant at level 0.05.
Fig 4.
Prediction accuracy of MiRduplexSVM versus six other methods on corner identification.
Performance accuracies are estimated using the EAE for errors of 0-5nts. In each panel, the y axis shows the prediction accuracy of MiRduplexSVM (in %) and the x axis shows the prediction accuracy of other methods (in %). The colour and symbol scheme is the same as in Fig 3G. Statistically significant results are indicated with filled symbols.
Table 1.
End Absolute Error (EAE) in nts.
Table 2.
End Absolute Error (EAE) in nts.
Table 3.
End Absolute Error (EAE) in nts.
Fig 5.
Identification of high confidence miRNAs.
As shown in the figure MiRduplexSVM assigned a higher score to 554 high confidence miRNAs (blue bars, median = 0.53 and mean = 0.44) than to 554 randomly selected miRNAS (red bars, median = -0.24 and mean = -0.14) with the observed differences being statistically significant (ranksum: p = 8.3084e-47 and t-test: p = 3.9577e-50). The x axis shows MiRduplexSVM’s scores and the y axis shows the percentage of hairpins assigned with the respective scores.