De novo basecalling of RNA modifications at single molecule and nucleotide resolution

RNA modifications influence RNA function and fate, but detecting them in individual molecules remains challenging for most modifications. Here we present a novel methodology to generate training sets and build modification-aware basecalling models. Using this approach, we develop the m⁶ABasecaller, a basecalling model that predicts m⁶A modifications from raw nanopore signals. We validate its accuracy in vitro and in vivo, revealing stable m⁶A modification stoichiometry across isoforms, m⁶A co-occurrence within RNA molecules, and m⁶A-dependent effects on poly(A) tails. Finally, we demonstrate that our method generalizes to other RNA and DNA modifications, paving the path towards future efforts detecting other modifications.