Fig. 6: Comparison of different peak-picking methods on the 236-residue intrinsically disordered protein DDX4N1.

A An NH projection of an HNCA acquired on a 750 MHz spectrometer, superimposed on a peak list derived from UnidecNMR results from a high-resolution HSQC (Supplementary Fig. S6B)⁴⁰. As expected for a large IDP, the centre of the spectrum is heavily overlapped. B An indication of the false negatives and false positives of the different methods tested on a 2D ¹⁵N HSQC, and 3D HNCO and HNCA spectra. (i) On a high-resolution 2D ¹⁵N HSQC, similar performance was found for Sparky, WaVPeak, PICKY, NMRNet, and DEEP picker with ca. 30 false negatives and between 0 and 66 false positives. By contrast, UnidecNMR picked exactly the same resonances as an experienced user with two false positives (Supplementary Fig. S7A). For 2D analysis, the UnidecNMR tuneable parameter ‘fac’ was set to 1.4. (ii/ii) Similar performance was found in 3 dimensions, with UnidecNMR picking over 100 more resonances in the two spectra that were missed by the other methods. The tuneable parameter ‘fac’ was set to 1.6 for 3D data. The specific peaks that were missed by UnidecNMR are analysed (Supplementary Fig. 7), arising from resonances appearing in the spectrum that were not in the HSQC, arising most likely due to some sample degradation. C A slice from a heavily overlapped region in the HNCA illustrates the performance of UnidecNMR and the places where the other algorithms fail to spot resonances. Overall, the performance of UnidecNMR is almost identical to those obtained by an experienced user. As these are N-H slices through a 3D spectrum, the distance of an identified peak from the current, carbon, slice is indicated with the colour of the peak, as shown in the key (bottom right).