Fig. 5

MAE results in seconds. The different estimators have been grouped by families, which are also highlighted with different colors, and approximately sorted by performance. The different panels indicate which molecules where considered for evaluating the performance: all includes all molecules, non-retained includes only those molecules considered as non-retained (Retention Time (RT) smaller than 5 minutes), and retained includes only retained molecules. The different shapes indicate the features used for feeding the regressors (in the legend, fgp and desc indicate fingerprints and descriptors, respectively). In the case of the blender, the features represented are descriptors+fingerprints since it is using all predictions from the base-regressors (i.e, it is using predictions from regressors using fingerprints, regressors using descriptors, and regressors using both). Error bars correspond to the 99% confidence interval of the MAE. In the figure, XGBoost, lightGBM and CatBoost have been shortened to XGB, LGB and CB, respectively. The numbers near the CB regressors correspond to the total contribution of non-retained molecules compared to retained molecules during training. For example, assigning a weight of 80 to each non-retained molecule (see "Gradient boosting" Section) results in these molecules influencing the loss function 80 times more than the retained ones. Since the relative abundance is 1/40, this leads to non-retained molecules having twice the influence of retained molecules during learning, which is shown as CB 2/1