ModiCal: A Targeted Calibration Workflow for Site-Specific m ⁵ C Validation by Nanopore Direct RNA Sequencing

Accurate identification of RNA 5-methylcytidine (m5C) at the single-nucleotide resolution remains a central challenge in nanopore direct RNA sequencing (DRS). Current global scanning and modification-aware basecalling methods enable transcriptome-wide profiling but often yield high false-positive rates and lack site-specific accuracy. To address this, we repurposed ModiDeC, originally a de novo multimodification classifier, into a targeted, high-precision validation tool for RNA modification sites with prior biochemical knowledge. This was implemented through a three-step calibration workflow that alternates between biochemical and computational modules using the well-characterized m5C2278 site in 25S rRNA as a starting point. Baseline training uses short synthetic RNAs carrying either a methylated or unmodified C2278 as ground truth, followed by IVT-derived calibration and validation in methyltransferase knockout yeast. The baseline model accurately detected the bona fide m5C2278 site but initially produced off-target predictions. Iterative retraining with unmodified IVT signals progressively reduced and ultimately eliminated false positives while maintaining a strong signal at the bona fide site. The final model retained enzyme-dependent detection in wild-type versus knockout yeast and, when explicitly targeted, was also able to detect the second rRNA site, C2870, which remained invisible in the initial analysis. Application to native human prerRNA processing intermediates further resolved two distinct m5C deposition regimes on 28S rRNA, while generalization to dengue virus genomic RNA confirmed that the same calibration logic transfers across diverse RNA contexts. Together, this study establishes a reproducible and transferable framework that integrates biochemical validation with iterative neural network refinement, providing a route toward reliable site-specific m5C confirmation by nanopore direct RNA sequencing.