Abstract
Rapid and accurate identification of transfer RNA (tRNA) modifications is crucial for understanding their role in protein translation and disease. However, their detection on tRNAs is challenging due to high modification density. With the release of the nanopore direct RNA sequencing kit SQK-RNA004, de novo modification calling models became available for pseudouridine (Ψ), m6A, inosine, and m5C, as part of the Dorado basecaller. By applying the Ψ model to tRNAs, we mapped both known and previously uncharacterized Ψ sites in Schizosaccharomyces pombe, and identified the corresponding pseudouridine synthases. This led to the discovery of two novel modification sites, Pus7-dependent Ψ8 and Pus1-dependent Ψ22. Furthermore, we have developed MoDorado, an algorithm to detect modifications beyond those used in model training ("off-label use"). It does so by assessing differences in modification predictions between modified and nonmodified samples using pre-trained, modification-specific models. By repurposing the Ψ/m6A/inosine/m5C models, MoDorado detected seven additional modifications (ncm5U, mcm5U, mcm5s2U, m7G, queuosine, m1A, and i6A), thus generating an expanded map of these tRNA modifications in S. pombe. This work provides a generalized framework for leveraging pre-trained models in determining the intricate landscape of tRNA modifications.