Fig. 2.

Consensus decoding improves sequencing accuracy. Reads were run through PoreOverNet (magenta), the community basecaller DeepNano-blitz (yellow), and ONT’s Bonito basecaller (blue) to generate softmax probabilities, which were then decoded using our algorithms. Guppy accuracies (in violet) were generated entirely from running the Guppy basecaller and its 1D² basecalling mode without any additional decoding. The Guppy basecaller has the option of two neural network architectures using either smaller (fast) or larger (high accuracy, hac) recurrent layer sizes. DeepNano-blitz was run with its width64 network. The median accuracy is represented by a dashed line