Extended Data Figure 3. Hybrid reads identify RNA duplexes.

a, Analysis of hybrid reads in 3′ UTRs demonstrates significantly smaller ensemble folding energy compared to random RNA fragments as calculated by RNAhybrid. The kernel density estimate of ensemble folding energy distribution is plotted, and the ensemble folding energy for hybrid reads and random RNAs is compared by Mann-Whitney U test (n = 4492 for both hybrid reads and random RNAs). b, Similar to a, but the analysis is restricted to hybrid reads in CDS (n = 958). c, Similar to a, but the analysis is restricted to inter-molecular hybrid reads (i.e., hybrid reads whose left arm and right arm originated from different genes; n = 257). d, Similar to a, but the analysis is restricted to hybrid reads in rRNAs (n = 3502). e, Median normalised PARS scores were calculated around center of all mRNA duplexes. PARS scores were obtained from Wan et al⁴, and positions with 0 values were removed. PARS score represents a ratio between reads starts after cutting with dsRNase (positive) / ssRNase (negative). Assuming that the double-stranded RNase fully digests each duplex, it is expected that the positive values in PARS-seq will be highest at the last nucleotide of each duplex. This might explain why maximum PARS values occur at the positions closer to the 3′ end of duplexes. f, Metaprofiles of the distribution of STAU1 cross-link sites, identified by the start sites of non-hybrid hiCLIP reads (blue), or the randomly repositioned sites (black, mean value of 10 randomizations; gray, standard deviation of the 10 randomizations) around the positions of hiCLIP duplexes. g, Distribution of mRNA median probabilities to be single-stranded from −50 to 100 nucleotides around the cross-link sites.