Accurate prediction of multiple RNA modifications from nanopore direct RNA sequencing data with RNANO

Abstract

Various types of RNA modifications exert vital regulatory functions across a broad array of biological processes. Nanopore direct RNA sequencing (DRS) has emerged as a promising technique holding the unique potential to capture signals from multiple types of RNA modifications in a single sequencing run. However, high-performance computational methods are demanded to recognize multiple modification types from complicated native DRS data from human cells. To this end, we developed RNANO, a deep learning-based method for predicting potential RNA modification sites and their modification states. This approach leverages an attention-enhanced multi-instance learning framework with the dynamic time warping algorithm to fine-tune the alignment between electrical signal events and reference sequences. RNANO efficiently detects seven common RNA modification types (m6A, m5C, {Psi}, m1A, Nm, ac4C, and m7G) in DRS data, exhibiting superior performance on both held-out testing samples and cross-cell-line independent benchmarks. A case study focusing on a cancer cell line also indicates that RNANO can pinpoint RNA modifications within key oncogenic pathways. In summary, RNANO enables accurate and efficient identification of multiple RNA modifications in human cell lines, offering useful clues for disease researches and therapeutic development related to RNA modifications.