DeepMP: a deep learning tool to detect DNA base modifications on Nanopore sequencing data

Abstract

Motivation DNA Methylation plays a key role in a variety of biological processes. Recently, Nanopore long-read sequencing has enabled direct detection of these modifications. As a consequence, a range of computational methods have been developed to exploit Nanopore data for methylation detection. However, current approaches rely on a human-defined threshold to detect the methylation status of a genomic position and are not optimized to detect sites methylated at low frequency. Furthermore, most methods employ either the Nanopore signals or the basecalling errors as the model input and do not take advantage of their combination. Results Here we present DeepMP, a convolutional neural network (CNN)-based model that takes information from Nanopore signals and basecalling errors to detect whether a given motif in a read is methylated or not. Besides, DeepMP introduces a threshold-free position modification calling model sensitive to sites methylated at low frequency across cells. We comprehensively benchmarked DeepMP against state-of-the-art methods on E. coli, human and pUC19 datasets. DeepMP outperforms current approaches at read-based and position-based methylation detection across sites methylated at different frequencies in the three datasets. Availability DeepMP is implemented and freely available under MIT license at https://github.com/pepebonet/DeepMP Supplementary information Supplementary data are available at Bioinformatics online.

NEMO: Improved and accurate models for identification of 6mA using Nanopore sequencing

“…Most of the existing 6mA callers have been trained and validated on limited datasets, such as a pUC plasmid grown in presence of Escherichia coli dam (DNA adenine methylase) 19,20 . We aimed to generate more robust datasets, both to benchmark existing callers, and to train and develop better methods for 6mA identification.…”

“…The copyright holder for this preprint this version posted March 13, 2024. ; https://doi.org/10.1101/2024.03.12.584205 doi: bioRxiv preprint methylase) 19,20 . We aimed to generate more robust datasets, both to benchmark existing callers, and to train and develop better methods for 6mA identification.…”

“…Tombo, an old tool provided by Oxford Nanopore, identifies possible modified bases by testing statistically significant differences in the current levels among reads aligning to the target position 18 . DeepSignal and DeepMP are two recent Deep Neural Network based tools which provide models for 6mA identification only in the GATC context 19,20 . mCaller provides a “dinucleotide” model, which theoretically covers all possible dimers with 6mA in the 1st position 21 .…”

Detection of nucleotide modifications in bacteria and bacteriophages: Strengths and limitations of current technologies and software

“…Several strategies are available for detecting 5mC, 6mA, and 4mC methylations. Current state‐of‐the‐art strategies include Hidden Markov Models, such as in Nanopolish (Simpson et al, 2017) (5mC in CpG context) and Megalodon (https://nanoporetech.github.io/megalodon) (5mC and 6mA), machine‐learning‐based tools, such as DeepMP (Bonet et al, 2022) (5mC and 6mA), DeepSignal (Ni et al, 2019) (5mC in CpG context), DeepMod (Liu, Fang, et al, 2019) (5mC and 6mA), NanoDisco (Tourancheau et al, 2021) (5mC, 6mA, and 4mC), and mCaller (McIntyre et al, 2019) (6mA), or statistic test approaches such as tombo (Stoiber et al, 2016) (5mC and 6mA). Notably, NanoDisco does not only identify methylated positions and distinguish among them (typing), but it can also bin metagenomes based on methylation patterns.…”

Maximizing the potential of genomic and transcriptomic studies by nanopore sequencing

“…8 D). The analysis of per base modifications can be performed either by a hidden markov models ( (51)), by statistical tests ( (52)), or by neuronal networks ( ONT, ONT, (53), (54), (55), (56). Additionally, (13) combines the results from up to six other tools (Nanopolish, Tombo, DeepSignal, Guppy, Megalodon, DeepMod) using a random forest model and thus shows increased accuracy compared to using the single tools, but also increased runtime.…”