Supplementary Figure 10: Downsampling analysis determines the extent of data needed to train models of varying complexity.

Model performance as reflected by prediction accuracy on the held-out validation data (Pearson’s r², y axis, top) and by the relative performance on validation-vs-training data (y axis, bottom, ratio of validation-vs-training Pearson r²; when <1.00, model is overfit) for models trained using sub-samples of the available training data (x axis) and learning different parameters (colors; parameters are cumulative from top to bottom of legend). Dotted lines: number of training examples needed to eliminate overfitting ((validation r² / training r²) > 0.999). Solid points: number of examples at which the maximal validation performance is achieved.