Supplementary Information: Improving Interpretability of Deep Learning Models

The plot shows PC1 (horizontal axis) versus PC2 (vertical axis) of original feature space (1, 357 features) trained on splicing data. The scattered gray points are subset of input data in the PC space. Black, red, blue paths show linear, latent-linear, and neighbors paths between the same source and destination points (matching conceptual/illustrative figure 1a, top panel). Source point is picked randomly and the destination point is picked to maximize distance from source point in PC space (10 components). Neighbors path is approximation, i.e., computing neighbors distances on first 10 principal components due to computational overhead. Random seeds were manually selected to highlight differences between linear and latent linear paths.