Fig. 1: DeepSignal-plant for 5mC detection using Nanopore sequencing.
From: Genome-wide detection of cytosine methylations in plant from Nanopore data using deep learning
a Architecture of DeepSignal-plant. BiLSTM: a sequence processing network that uses long short-term memory layer to take the input from forward and backward direction to learn order dependence; Full Connection: a fully connected layer that connects all the inputs from the former layer to every activation unit of the next layer; Softmax: an activation function which normalizes a vector of real numbers into a vector of probabilities that sum to 1. b Schema of denoise training samples in DeepSignal-plant. c Signal comparison of different kinds of samples of a k-mer after denoising training. positive_kept: positive samples kept by the denoising step; positive_removed: positive samples removed by the denoising step; negative: negative samples; n = number of signal values for each base; Boxplots indicate 50th percentile (middle line), 25th and 75th percentile (box), the smallest value within 1.5 times interquatile range below 25th percentile and largest value within 1.5 times interquatile range above 75th percentile (whiskers), and outliers (dots). d Effectiveness of training samples selection on 5mC detection. The training samples were extracted from ~500× Nanopore reads of A. thaliana. Pearson correlations were calculated using the results from ~20× Nanopore reads and three bisulfite replicates of A. thaliana.
