From squiggle to basepair: computational approaches for improving nanopore sequencing read accuracy

Abstract

Nanopore sequencing is a rapidly maturing technology delivering long reads in real time on a portable instrument at low cost. Not surprisingly, the community has rapidly taken up this new way of sequencing and has used it successfully for a variety of research applications. A major limitation of nanopore sequencing is its high error rate, which despite recent improvements to the nanopore chemistry and computational tools still ranges between 5% and 15%. Here, we review computational approaches determining the nanopore sequencing error rate. Furthermore, we outline strategies for translation of raw sequencing data into base calls for detection of base modifications and for obtaining consensus sequences.

Fig. 1

Timeline of reported MinION read accuracies and Oxford Nanopore Technologies (ONT) technological developments. Nanopore chemistry updates and advances in base-caller software are represented as colored bars. The plotted accuracies are ordered on the basis of the chemistry and base-calling software used, not according to publication date. Based on data from 1 [9]; 2 [10]; 3 [50]; 4 [51]; 5 [33]; 6 [28]; 7 [52]; 8 [53]; 9 [54]; 10 [29]; 11 [31]; 12 [48]; 13 [46]; 14 [55]; 15 [11]; 16 [5]; 17 [13]; 18 [3]. HMM Hidden Markov Model, RNN Recurrent Neural Network

Fig. 2

Overview of MinION nanopore sequencing. The left panel shows sources of errors during MinION sequencing and base calling. The right panel shows computational strategies that have been used to improve accuracy. HMM Hidden Markov Model, RNN Recurrent Neural Network

Fig. 3

Schematic overview of the algorithms underlying nanopore base callers. a Nanocall uses a Hidden Markov Model (HMM) for base calling. b DeepNano was the first base caller to use Recurrent Neural Networks (RNN). h1–h3 represent three hidden layers in the RNN. c BasecRAWller uses two RNNs, one to segment the raw measurements and one to infer k-mer probabilities. d Chiron makes use of a Convolutional Neural Network (CNN) to detect patterns in the data, followed by an RNN to predict k-mer probabilities, which are evaluated by a Connectionist Temporal Classification (CTC) decoder. LSTM long-short-term memory