Abstract
In plants, cytosine DNA methylations (5mCs) can happen in three sequence contexts as CpG, CHG, and CHH (where H = A, C, or T), which play different roles in the regulation of biological processes. Although long Nanopore reads are advantageous in the detection of 5mCs comparing to short-read bisulfite sequencing, existing methods can only detect 5mCs in the CpG context, which limits their application in plants. Here, we develop DeepSignal-plant, a deep learning tool to detect genome-wide 5mCs of all three contexts in plants from Nanopore reads. We sequence Arabidopsis thaliana and Oryza sativa using both Nanopore and bisulfite sequencing. We develop a denoising process for training models, which enables DeepSignal-plant to achieve high correlations with bisulfite sequencing for 5mC detection in all three contexts. Furthermore, DeepSignal-plant can profile more 5mC sites, which will help to provide a more complete understanding of epigenetic mechanisms of different biological processes.
Introduction
Cytosine methylation (5mC) is a critical epigenetic modification that regulates gene expression and maintains genome stability. In plants, 5mC occurs in three distinct sequence contexts: CpG, CHG, and CHH. Each context is maintained by specific enzymatic pathways and serves unique biological functions, ranging from developmental regulation to stress responses.
Traditionally, whole-genome bisulfite sequencing (WGBS) has been the gold standard for mapping 5mC. However, WGBS relies on short-read sequencing, which struggles to map repetitive genomic regions like transposons, centromeres, and telomeres. Additionally, the bisulfite treatment itself can damage DNA. Nanopore sequencing offers a compelling alternative by providing long reads of native DNA. While tools exist for detecting 5mC from Nanopore signals, they have primarily focused on the CpG context prevalent in mammals, leaving the complex plant methylome largely unaddressed.
Methods: The DeepSignal-plant Architecture
To overcome these challenges, we developed DeepSignal-plant, a deep learning framework designed to detect 5mC across all three plant contexts. The core of the model is a Bidirectional Recurrent Neural Network (BRNN) with Long Short-Term Memory (LSTM) units.
The model processes two types of input features:
- Sequence Features: Nucleotide bases surrounding the target cytosine.
- Signal Features: Raw electrical current metrics, including mean, standard deviation, and signal length, derived from the Nanopore sequencer.
Iterative Denoising for Training
A significant hurdle in training models for non-CpG contexts, particularly CHH, is the lack of "pure" ground-truth data. In nature, CHH sites are rarely 100% methylated, leading to noisy training sets. We implemented an iterative denoising process. The algorithm initially trains on a balanced dataset, then uses the model to score the samples, removes those with low confidence (potential false positives), and retrains. This cycle refines the training data, significantly boosting the model's accuracy in the challenging CHH context.
Results and Discussion
We validated DeepSignal-plant using Arabidopsis thaliana and Oryza sativa (rice) datasets, comparing its performance against WGBS and existing tools like Megalodon.
Performance Across Contexts
DeepSignal-plant achieved high Pearson correlations with WGBS across all three contexts. Notably, the denoising process was instrumental for the CHH context, where correlation improved from 0.3696 (random sampling) to 0.7840 (balance + denoise).
Illuminating the "Dark Matter" of the Genome
By leveraging long Nanopore reads, DeepSignal-plant successfully profiled methylation in genomic regions that are typically "unmappable" with short-read sequencing. In rice, the tool identified 5.3% more 5mC sites than WGBS, providing a more comprehensive view of the epigenetic landscape in centromeric and telomeric regions.
Conclusion
DeepSignal-plant represents a significant advancement in plant epigenomics. By accurately detecting 5mC in CpG, CHG, and CHH contexts from long-read Nanopore data, it enables researchers to study methylation in complex, repetitive regions of the plant genome. Future work will focus on adapting this framework for other modifications, such as 6mA, and optimizing it for newer Nanopore chemistries.
Summary updated