DeepNano: Deep Recurrent Neural Networks for Base Calling in MinION Nanopore Reads

Vladimír Boža, Broňa Brejová, Tomáš Vinař. DeepNano: Deep recurrent neural networks for base calling in MinION nanopore reads. PLoS One, 12(6):e0178751. 2017.

DeepNano is an alternative basecaller for Oxford Nanopore MinION reads based on deep recurrent neural networks. Currently it works with SQK-MAP-006 and SQK-MAP-005 chemistry and as a postprocessor for Metrichor. A newer version for R9 and R9.4 chemistry can also be downloaded.

Here are our benchmarks for SQK-MAP-006 chemistry, which compare mapping accuracy of reads produced by DeepNano and Metrichor from Oxford Nanopore. We trained on reads which align to half of the E. coli reference and tested on the other half of E. coli and on Klebsiela pneumoniae:

	Ecoli Metrichor	Ecoli DeepNano	Klebsiella Metrichor	Klebsiella DeepNano
Template reads	71.3%	77.9%	68.1%	76.3%
Complement reads	71.4%	76.4%	69.5%	75.7%
2D reads	86.8%	88.5%	84.8%	86.7%

Download DeepNano at bitbucket repository

• DeepNano bitbucket repository

Datasets used for evaluation

• http://www.ebi.ac.uk/ena/data/view/ERR1147230
• https://www.ebi.ac.uk/ena/data/view/SAMEA3713789