NanoCaller for accurate detection of SNPs and indels in difficult-to-map regions from long-read sequencing by haplotype-aware deep neural networks

Abstract

Long-read sequencing enables variant detection in genomic regions that are considered difficult-to-map by short-read sequencing. To fully exploit the benefits of longer reads, here we present a deep learning method NanoCaller, which detects SNPs using long-range haplotype information, then phases long reads with called SNPs and calls indels with local realignment. Evaluation on 8 human genomes demonstrates that NanoCaller generally achieves better performance than competing approaches. We experimentally validate 41 novel variants in a widely used benchmarking genome, which could not be reliably detected previously. In summary, NanoCaller facilitates the discovery of novel variants in complex genomic regions from long-read sequencing.

Keywords: Deep learning; Difficult-to-map regions; Long-range haplotype; Variant calling.

Fig. 1

An example on how to construct input features for a SNP candidate site. a Reference sequence and read pileups at candidate site b and at other genomic positions that share the same reads. The columns in gray are genomic positions that will not be used in input features for candidate site b as they do not satisfy the criteria for being highly likely heterozygous SNP sites. Only the columns with colored bases will be used to generate input features for site b and will constitute the set Z as described in the SNP pileups generation section of "Methods". These neighboring likely heterozygous sites can be up to thousands of bases away from candidate site b. b Reference sequence and read pileups for only the candidate site and neighboring highly likely heterozygous SNP sites. c Raw counts of bases at sites in the set Z for each read group split by the nucleotide types at site b. These raw counts are multiplied with negative signs for reference bases. d Flattened pileup image with fifth channel after reference sequence row is added. e Pileup image used as input features for NanoCaller deep convolutional neural network

Fig. 2

An example on how to construct input features for an indel candidate site. a Reference sequence and read pileup at the candidate site before and after multiple sequence alignment, and the consensus sequence. b Reference sequence and consensus sequence at the candidate site before and after pairwise alignment, and the inferred sequence. c Raw counts of each symbol at each column of multiple sequence alignment pileup. d Matrix M, showing frequency of each symbol at each column of multiple sequence alignment pileup. e First channel of input image, matrix M minus Q (one-hot encoding of realigned reference sequence). f Matrix Q, one-hot encoding of realigned reference sequence which forms the second channel of input image for NanoCaller deep convolutional neural network

Fig. 3

Performances of NanoCaller and other variant callers on ten ONT datasets. SNP performance on whole-genome high-confidence intervals: a precision, b recall, c F1 score. F1 scores of SNP performances on d “all difficult-to-map” regions and e MHC. Indel performance non-homopolymer regions: f precision, g recall, h F1 score. i: F1 score of indel performance in whole-genome high-confidence intervals

Fig. 4

Performances of NanoCaller and other variant callers on four PacBio CCS and four PacBio CLR datasets. SNP performance on whole-genome high-confidence intervals using CCS reads: a precision, b recall, c F1 score. Indel performance on whole-genome high-confidence intervals using CCS reads: d precision, e recall, f F1 score. SNP performance on whole-genome high-confidence intervals using CLR reads: g precision, h recall, i F1 score

Fig. 5

Evidence for a novel multiallelic SNP. a IGV plots of Nanopore, PacBio CCS, and Illumina reads of HG002 genome at chr3:5336450-5336480. b Sanger sequencing signal data for the same region. NanoCaller on older HG002 data correctly identified the multiallelic SNP at chr3:5336450 (A>T,C) shown in black box

Fig. 6

Evidence for novel deletions. a IGV plots of Nanopore, PacBio CCS, and Illumina reads of HG002 genome at chr9:135663780-135663850. The 40-bp-long deletion shown below in black box was identified using Sanger sequencing at chr9:135663795 or chr9:135663804 (both are correct and the difference is due to two different alignments). b Sanger sequencing signal data around the deletion