Fig. 2: Testing of peak picking algorithms on synthetic data.

Following optimisation of the UnidecNMR algorithm on 1D data (Supplementary Fig. 4) performance of a range of freely available peak pickers was assessed on simulated 2D data (7200 spectra). Two Gaussian resonances were simulated with a function of signal to noise (9 values ranging from 2 to 10) and separation (8 values, ranging from 0.85 to 2.6 in units of the full-width half maximum, FWHM of the simulated resonances). For each separation/SN combination, the seed for the random number that generates the noise was varied allowing us to derive an ensemble average over 100 repeats (see extended methods). In the UnidecNMR analysis, a Gaussian peak shape with the same width used for the simulation was used to deconvolve the data, with the one manually tuneable parameter, ‘fac’ set to 1.6. In the tests that follow, the width of the peak shape used by UnidecNMR was purposely mis-set for the purpose of testing which demonstrated that the results are reasonably agnostic of the precise value used, indicating that the peak shape chosen for analysis needs only to be ‘roughly correct’ (Supplementary Fig. 1). a Simulated data illustrate a range of signal-to-noise and separation values. The numbers 1–9 indicate where these example data were taken from the full dataset, c. b Illustrative examples of three spectra that provide a representative discrimination of the various algorithms tested that could be easily automated to iterate over the full dataset. Only NMRNet¹² and UnidecNMR were able to resolve the peak with the closest separation whereas only WaVPeak¹¹, PICKY¹⁰ and UnidecNMR were able to distinguish the peaks with the widest separation. NMRNet overpicked one of the peaks on the third spectrum, which we have indicated in blue for clarity. c Overall performance on the entire simulated 2D dataset. Only UnidecNMR, PICKY¹⁰ and WaVPeak¹¹ achieve 100% accuracy at the relatively trivial high separation/high signal-to-noise limit. The blue contour line represents the threshold above which 100% accuracy is achieved, allowing a convenient means by which to compare the different algorithms in this test. UnidecNMR outperforms the other algorithms on these synthetic datasets.