Supplementary Figure 13: Features of positional preferences.

(a) Capturing helically biased positional preferences. For each location within the promoter (x axis), this shows the learned activity bias parameters (red curve; as in Fig. 4b) for the poly-A motif, long-range trend captured by a loess fit (green), and short-range residual activity bias after subtracting loess fit (blue) with reference 10.5 bp sine waves (black) for the minus strand (top) and plus strands (bottom) for the four different models (columns). (b) Modeling positional preferences increases predictive accuracy within the same scaffold but can drastically decrease it between scaffolds. For each training data set (four sub-panels) for both model types (colors), the Pearson r² (y axis) capturing performance on each test dataset (x axis). (n = 1 set of model predictions per bar; 4 independent training sets each with 2 model types, evaluated on each of 3 test datasets total).