WaveNano: a signal-level nanopore base-caller via simultaneous prediction of nucleotide labels and move labels through bi-directional WaveNets

Sheng Wang, Zhen Li, Yizhou Yu, Xin Gao
2018 Quantitative Biology  
The Oxford MinION nanopore sequencer is the recently appealing third-generation genome sequencing device that is portable and no larger than a cellphone. Despite the benefits of MinION to sequence ultra-long reads in real-time, the high error rate of the existing base-calling methods, especially indels (insertions and deletions), prevents its use in a variety of applications. Methods: In this paper, we show that such indel errors are largely due to the segmentation process on the input
more » ... l current signal from MinION. All existing methods conduct segmentation and nucleotide label prediction in a sequential manner, in which the errors accumulated in the first step will irreversibly influence the final base-calling. We further show that the indel issue can be significantly reduced via accurate labeling of nucleotide and move labels directly from the raw signal, which can then be efficiently learned by a bi-directional WaveNet model simultaneously through feature sharing. Our bi-directional WaveNet model with residual blocks and skip connections is able to capture the extremely long dependency in the raw signal. Taking the predicted move as the segmentation guidance, we employ the Viterbi decoding to obtain the final base-calling results from the smoothed nucleotide probability matrix. Results: Our proposed base-caller, WaveNano, achieves good performance on real MinION sequencing data from Lambda phage. Conclusions: The signal-level nanopore base-caller WaveNano can obtain higher base-calling accuracy, and generate fewer insertions/deletions in the base-called sequences. Author summary: Oxford nanopore sequencing is a rapidly developed sequencing technology in recent years. Despite the benefits of this technique to sequence ultra-long reads in real-time, the high error rate of the existing base-calling methods, especially indels (insertions and deletions), prevents its use in many applications. Here we show that such indel errors are largely due to the segmentation process on the input electrical current signals, and propose a new deep learning model bi-directional WaveNet to perform the base-calling directly on the signal level. The experimental result suggests that our method achieves good performance on real nanopore sequencing data from Lambda phage. † These authors contributed equally to this work.
doi:10.1007/s40484-018-0155-4 fatcat:wvtefbkfenczhlueglmeomlafe