BasecRAWller: Streaming Nanopore Basecalling Directly from Raw Signal

Abstract

All current nanopore basecalling applications begin with the segmentation of raw signal into discrete events, which are ultimately processed into called bases. We propose the basecRAWller algorithm, a pair of unidirectional recurrent neural networks that enables the calling of DNA bases in real time directly from the rawest form of nanopore output. This shift in nanopore basecalling provides a number of advantages over current processing pipelines including: 1) streaming basecalling, 2) tunable ratio of insertions to deletions, and 3) potential for streaming detection of modified bases. Key to the streaming basecalling capability is sequence prediction at a delay of less than 1/100th of a second, allowing future signal to continuously modulate sequence prediction. BasecRAWller is computationally efficient enabling basecalling at speeds faster than current nanopore instrument measurement speeds on a single core. Further, basecalling can be paused and resumed without any change in the resulting predicted sequence, transforming the potential applications for dynamic read rejection capabilities. The basecRAWller algorithm provides an alternative approach to nanopore basecalling at comparable accuracy and provides the community with the capacity to train their own basecRAWller neural networks with minimal effort.