Poster #7: Yongji Zou

Robust Direct RNA Sequencing-based Reactivity Calling Across Nanopore Pore Generations and Probing Chemistries

Yongji Zou1,2; Umair Mian Ahsan1; Kai Wang1,3 1. Center for Cellular and Molecular Therapeutics, Children's Hospital of Philadelphia, Philadelphia, PA, USA 2. Bioengineering Graduate Program, University of Pennsylvania, Philadelphia, PA, USA 3. Department of Pathology and Laboratory Medicine, University of Pennsylvania, Philadelphia, PA, USA

Chemical probing measures RNA structural accessibility, but dominant cDNA-based methods lose native RNA modifications. Oxford Nanopore direct RNA sequencing (DRS) reads native, full-length molecules and preserves their modifications, yet current DRS reactivity callers recover less structural signal than cDNA-based mutational profiling. Methods. We compare raw nanopore currents in matched unprobed controls and SHAPE-probed samples. Single-base, single-molecule signal features form low-dimensional representations; per-locus reactivity is the probe-induced increase in representation variance. Feature mappings are estimated on reads disjoint from those scored; callers are evaluated on identical reads against orthogonal SHAPE-MaP reactivity. Results. Across six probed riboswitches, our caller achieved mean per-construct Spearman correlation 0.40 with SHAPE-MaP, versus 0.30 for a basecalling-error caller, 0.21 for segSHAPE, and 0.12 for sm-PORE-cupine, the published caller for these data, outperforming every DRS caller tested. The same procedure, with pore-specific models from two nanopore generations, produced profiles agreeing at rho=0.75; both correlated positively with SHAPE-MaP on every construct. Without retraining, correlations were 0.29-0.37 for three 2'-OH-directed reagents (NAI-N3, 1AI, 2A3). For guanine-directed kethoxal, correlation with 2'-OH-based SHAPE-MaP was near zero overall but 0.24 at guanines, where every 2'-OH reagent degraded. A scan given no nucleotide identities or variant positions localized differences between independently prepared libraries of the same molecules to the ligand-responsive region of the aptamer (p=0.007) and, for a three-nucleotide mutant, the variant site (p=0.006); scoring reads were disjoint from region-selection reads. An RNA language model fine-tuned on DRS transcriptome reactivity called by this pipeline improved RNA-binding-protein transfer by 0.043 AUROC (95% CI, 0.036-0.050; 172 proteins) over a competence-matched control when Illumina icSHAPE was supplied at inference. Conclusion. The same procedure outperforms the DRS callers tested, agrees across pore generations, distinguishes reagent specificity, localizes differences without nucleotide identities or variant positions, and supports reactivity-informed model fine-tuning.