Skip to main content

Table 1 MAE and MEDAE results for the top 4 performing regressors (mean ± standard error)

From: Probabilistic metabolite annotation using retention time prediction and meta-learned projections

   Blender DNN DKL XGB
Features Metric desc+fgp desc fgp desc+fgp desc fgp desc+fgp desc fgp desc+fgp
All MAE (s) \(39.98\pm 1.49\) \(42.35\pm 1.13\) \(39.23\pm 1.2\) \(39.54\pm 0.97\) \(49.27\pm 2.09\) \(45.60\pm 2.37\) \(44.49\pm 1.11\) \(50.24\pm 0.99\) \(48.13\pm 0.63\) \(48.72\pm 0.90\)
MEDAE (s) \(18.63\pm 1.24\) \(20.02\pm 0.73\) \(17.22\pm 0.89\) \(18.00\pm 0.44\) \(26.73\pm 2.00\) \(23.25\pm 1.87\) \(22.51\pm 0.73\) \(27.18\pm 0.71\) \(25.64\pm 0.56\) \(25.32\pm 0.98\)
Non-retained MAE (s) \(239.02\pm 8.68\) \(240.30\pm 8.61\) \(235.46\pm 15.85\) \(228.11\pm 5.41\) \(220.07\pm 38.77\) \(228.94\pm 6.10\) \(216.61\pm 21.06\) \(243.97\pm 7.14\) \(242.19\pm 6.57\) \(244.17\pm 7.89\)
MEDAE (s) \(87.18\pm 137.45\) \(106.45\pm 180.01\) \(25.10\pm 6.89\) \(17.58\pm 3.80\) \(128.53\pm 167.60\) \(15.18\pm 5.86\) \(33.40\pm 44.41\) \(129.74\pm 117.80\) \(134.06\pm 108.68\) \(126.17\pm 97.29\)
Retained MAE (s) \(34.73\pm 1.14\) \(37.11\pm 0.67\) \(34.06\pm 0.86\) \(34.56\pm 0.67\) \(44.80\pm 2.57\) \(40.77\pm 2.39\) \(39.85\pm 1.50\) \(45.14\pm 0.93\) \(43.01\pm 0.65\) \(43.57\pm 0.94\)
MEDAE (s) \(18.42\pm 1.37\) \(19.99\pm 0.72\) \(17.21\pm 0.87\) \(18.01\pm 0.43\) \(26.79\pm 1.99\) \(23.29\pm 1.86\) \(22.58\pm 0.76\) \(26.90\pm 0.58\) \(25.35\pm 0.56\) \(25.04\pm 0.79\)
  1. In this table, desc, fgp and desc+fgp refer to the input features used by each regressor. desc means descriptors, fgp means fingerprints and desc+fgp means that both descriptors and fingerprints have been used. Note that some standard errors are larger than the mean MAE/MEDAE due to the presence of outliers. See Figs. 5 and S3 (the latter in Additional file 1) for better error estimates