Fig. 1

EvoAug improves generalization and interpretability of Basset models. a Schematic of evolution-inspired data augmentations (left) and the two-stage training curriculum (right). b Generalization performance (area under the precision-recall curve) for Basset models pretrained with individual and combinations of augmentations, i.e., Noise+Ins+RC (Gaussian noise, insertion, reverse-complement) and all augmentations (Gaussian noise, reverse-complement, mutation, translocation, deletion, insertion), and fine-tuned on Basset dataset. Standard represents no augmentations during training. c Comparison of the average hit rate of first-layer filters to known motifs in the JASPAR database (top) and the average q-value of the filters with matches (bottom). d Comparison of the average Pearson correlation between model predictions and experimental data from CAGI5 Challenge. b–d Each box-plot represents 5 trials with random initializations