Fig. 1: Predicting kmer characteristic ionic currents from chemical structures.

A Graphic overview of the proposed deep learning framework for DNA analysis. B Goodness of fit of DNA canonical random downsample, base-dropout, position-dropout, and model combination analyses. Specifically, “downsample” denotes the random dropout experiment, where we create random train-test splits. “Base” denotes base-dropout experiment, where we drop DNA 6mers that contain a specific base in any given position during training. “Position” denotes positional base-dropout experiment, where we drop DNA 6mers that contain a specific base in a given position during training. As for “combine,” we drop DNA 6mers that contain both of the specified bases during training. C Goodness of fit of 5mC-containing DNA 6mer imputation analysis. D Goodness of fit of de novo 5mC-containing DNA 6mer prediction. C and 5mC refer to the goodness of fit of canonical DNA 6mers and 5mC-containing DNA 6mers, respectively. In B–D, Train (red) and Test (blue) refer to the goodness of fit of the training and test DNA 6mers, respectively. E Predictive accuracy of C/5mC status quantified by balanced accuracy. Nanopolish, predictive analysis with the nanopolish model as baseline control. De Novo, predictive analysis with 5mC-containing DNA 6mer models described in (D), which were predicted from canonical training. 0.01–0.9, predictive analysis with different imputation 5mC-containing DNA 6mer models as described in (C). FAB39088 (cyan) and FAF01164 (purple) refer to two independent NA12878 cell line native genomic DNA nanopore sequencing datasets. Throughout (B–E), the median, minimum/maximum (excluding outliers), and first/third quartile values were shown by the boxplots.