Detection of m6A from direct RNA sequencing using a Multiple Instance Learning framework

Abstract

RNA modifications such as m6A methylation form an additional layer of complexity in the transcriptome. Nanopore direct RNA sequencing captures this information in the raw current signal for each RNA molecule, enabling the detection of RNA modifications using supervised machine learning. However, experimental approaches provide only site-level training data, whereas the modification status for each single RNA molecule is missing. Here we present m6Anet, a neural network-based method that leverages the Multiple Instance Learning framework to specifically handle missing read-level modification labels in site-level training data. m6Anet outperforms existing computational methods, shows similar accuracy as experimental approaches, and generalises to different cell lines with almost identical accuracy. We demonstrate that m6Anet captures the underlying read-level stoichiometry that can be used to approximate differences in modification rates. m6Anet achieves this without retraining model parameters, enabling the transcriptome-wide identification and quantification of m6A from a single run of direct RNA sequencing.

Direct RNA Sequencing Reveals SARS-CoV-2 m6A Sites and Possible Differential DRACH Motif Methylation among Variants

“…Although some positions are consistent between two experiments (e.g., 27,764), others vary (e.g., 28,616). The analysis of the output of m6Anet program shows that the positions putatively methylated have a significantly higher probability (>25%) than the putatively non-methylated (<5%), which is consistent with the model training and benchmarking [33]. Therefore, positions that, in at least one experiment, have a ≥50% probability are indicative that a substantial proportion of modified bases, over unmodified, are present in the sample.…”

“…Therefore, positions that, in at least one experiment, have a ≥50% probability are indicative that a substantial proportion of modified bases, over unmodified, are present in the sample. The m6A detection method employed here uses a model that takes into account the mixture of modified and unmodified RNAs and outputs the m6A-modification probability at any given site for all the DRACH 5-mers represented in the neural network training data and this might, at least in part, explain the variance among experiments as observed with different cell lines [33]. As a future perspective, the m6Anet model can be trained for the SARS-CoV-2 genome, which might adjust the probability values for m6A prediction in these specific RNAs.…”

“…The probability of methylation in DRACH motifs was calculated with m6anet (v-0.1.1-pre) [33], as recommended, (I) by preprocessing the segmented raw signal file with "m6anet-dataprep", and (II) running m6anet over data using "m6anet-run_inference".…”

Beyond sequencing: machine learning algorithms extract biology hidden in Nanopore signal data

“…These m 6A -labeled data sets can be obtained through wet lab assays such as m6ACE-seq and CLIP-seq [42,61] or artificially methylating adenosines with methyltransferases [43]. Such supervised methods include MINES [42], EpiNano [43], Nanom6A [44], and m6anet [45]. As m 6A is preferentially found in DRACH/RRACH motifs [12], MINES uses this aspect by modeling m 6A sites using a random forest classifier for each of the DRACH motifs using CLIP-seq-identified m 6A sites as positive samples [42].…”

“…EpiNano trains a support vector machine that predicts candidate m 6A sites from basecalling errors that are presumably caused by the presence of m 6A [43] while Nanom6A trains an eXtreme Gradient Boosting (XGBoost) model with the raw signal features [44]. Predicting the probability of a site being modified with m 6A , m6anet employs a multiple instance learning framework and takes the entire differentially labeled reads into account [45].…”