Accurate detection of m6A from direct RNA sequencing data using a Multiple Instance Learning (MIL) framework

We develop m6ANet, a computational method based on the Multiple Instance Learning (MIL) framework to detect m6A modifications from nanopore signal intensity-based features using only a single sample of direct RNA sequencing data. Unlike existing m6A detection approaches, our model considers the whole distribution of differentially labelled reads to output the probability that a given DRACH site contains m6A modification. Our model outperforms existing computational approaches and is comparable to established experimental protocols used to detect m6A modification. Additionally, our model generalizes well to different cell lines and can therefore be deployed without retraining on any biological sample of interest.