RNNHC: A hybrid error correction algorithm for long reads based on Recurrent Neural Network

Abstract

Abstract Long reads generated by third-generation sequencing technologies show irreplaceable advantages in downstream analysis by virtue of their ultra-long read lengths. However, their high error rate also affects the accuracy of downstream analysis. Meanwhile, deep learning has shown its talents in the processing of long gene sequence problems. In this work, a hybrid error correction algorithm is proposed, which uses the idea of processing multi-class tasks with a recurrent neural network to capture the long-term dependencies in the long reads to solve the problem of long-read error correction. RNNHC first aligns the long reads to the high-precision short reads to generate the corresponding feature vectors and labels, then adopts a recurrent neural network to capture the dependencies between bases, and finally trains the model so that it can be used later for prediction and error correction. We demonstrate that the performance of RNNHC is better than that of state-of-the-art hybrid error correction methods on real-world PacBio and ONT data sets including E. coli, S. cerevisiae, and Drosophila melanogaster. As shown in our experimental results that RNNHC can improve the alignment identity while maintaining read length and continuity of the sequence, and spending less user time than other hybrid error correction algorithms. Furthermore, RNNHC is applicable to data from two mainstream sequencing platforms.