Fig. 3: Site-level evaluation.

a A table presenting site-level evaluation for six different datasets. Only positions with coverage higher than 5x for each ONT-based tool and WGBS (whole genome bisulfite sequencing) were included. For the ground truth, we use only fully unmethylated or fully methylated positions concerning WGBS. R9.4.1 Rockfish models outperform Megalodon and Nanopolish on most metrics and for most datasets. Furthermore, R10.4.1 Rockfish model outperforms Remora on both evaluation datasets. The metrics are standardized with FPR denoting false positive rate, and the best scores bolded in the table. b positives and c negatives for different methods based on nanopore signal and the ground truth (Illumina) for the R9.4.1 NA12878 dataset. Rockfish is represented with the small model and Megalodon is represented with the Rerio model. Sample space is defined as the set of all fully unmethylated or methylated sites called by Illumina with at least 5x. Rockfish calls the highest number of true positives and true negatives and achieves high precision and recall. d, e show the same analysis for the R10.4.1 NA12878 dataset. Rockfish is represented with the base model and evaluated in all fully unmethylated or methylated sites called by Illumina with at least 5x. Rockfish calls the more true positives and negatives than Remora while achieving high precision and recall.