Supplementary Figure 11: Positional models improve predictions.
From: Deciphering eukaryotic gene-regulatory logic with 100 million random promoters
(a,b) Position-specific pTpA+Glu model-predicted expression levels (x axes) vs. measured expression levels (y axes) for (a) high-quality test data in the pTpA promoter scaffold, grown in glucose, and (b) native yeast promoter sequences, divided into 80 bp fragments and tested in the pTpA promoter scaffold, grown in glucose. (c) Most expression variation is attributed to accessibility. Position-specific pTpA+Glu model-predicted accessibility (Ω; x axis) vs. measured expression levels (y axis) for high-quality test data in the pTpA promoter scaffold, grown in glucose. Performance is better when incorporating positional activities (a,b), and, in particular, accessibility alone cannot distinguish the highest expression levels (measured expression >12). The dominant effect of accessibility on expression likely reflects accessibility being a prerequisite for expression. (n = 9,982, 70,924, and 9,982 promoters for (a-c), respectively; Pearson’s r2 shown in bottom right; Red lines: Generalized Additive Model lines of best fit).
