Figure 6: Models trained on RAMPAGE data can learn PRO-cap signal similar to ProCapNet.

(A) Relationship between total PRO-cap signal and total RAMPAGE signal within PRO-cap peaks. r, Pearson correlation. (B) Counts task performance of RAMPAGE-Net on RAMPAGE peaks from held-out test chromosomes (C) Counts predictions from RAMPAGE-Net vs. predictions from ProCapNet within PRO-cap peaks. (D) Same as C, but with measured PRO-cap signal on the y-axis. (E) PRO-cap and RAMPAGE signal, ProCapNet and RAMPAGE-Net model predictions, contribution scores from both models, and 100-way PhyloP sequence conservation scores¹⁰⁵ for an example promoter. (F) Position-probability matrix (PPM) and position-weight matrix (PWM) summarizing all sequences where TSSs (5’ read ends) were identified by PRO-cap (left) or RAMPAGE (right), weighted by the number of reads at each TSS. (G) Pearson correlations between contribution scores from ProCapNet and RAMPAGE-Net for the profile and counts tasks across 1kb windows over all PRO-cap peaks.