Figure 6.

Positive predictive value–sensitivity (PPV–SEN) plots comparing our algorithm with competitive methods on the pk168 dataset (Left) and the RS-pk388 dataset (Right). For the pk168 dataset, we set ${\gamma }^{\left(1\right)}=1$, ${\gamma }^{\left(2\right)}=3$ for Neuralfold; ${\gamma }^{\left(1\right)}=2$, ${\gamma }^{\left(2\right)}=4$ for IPknot with the Dirks–Pierce (D&P) model; ${\gamma }^{\left(1\right)}=2$, ${\gamma }^{\left(2\right)}=16$ for IPknot with/without refinement; $\gamma =2$ for CentroidFold. For the RS-pk388 dataset, we set ${\gamma }^{\left(1\right)}=1$, ${\gamma }^{\left(2\right)}=3$ for Neuralfold; ${\gamma }^{\left(1\right)}=2$, ${\gamma }^{\left(2\right)}=2$ for IPknot without refinement; ${\gamma }^{\left(1\right)}=1$, ${\gamma }^{\left(2\right)}=1$ for IPknot with refinement; $\gamma =2$ for CentroidFold.