Detection of SNPs, indels and SVs using whole-haplotype genome alignment

To call heterozygous sites between the two haploid sequences, independent of the GenomeScope calculation, we first performed a Mummer (v.3.23) alignment with the parameters of ‘nucmer -maxmatch -l 100 -c 500’. Because our assemblies span most repetitive sequences, repeat-masking treatment was not necessary before conducting the Mummer alignment. A series of custom scripts (https://github.com/comery/marmoset) identified and sorted our SNPs and indels in the alignments. We used svmu (v.0.4-alpha)⁷¹, Assemblytics (v.1.2)⁷², and SyRi (v.1.0)⁷³, to detect SVs from Mummer alignment. After several test rounds, we found that svmu reported more accurate large indels, and Assemblytics detected CNVs, particularly tandem repeats, whereas SyRi detected other SVs well. We used these three methods and combined the results as confident SVs. We used default parameters for svmu, Assemblytics, and recommended nucmer alignment for SyRi (https://schneebergerlab.github.io/syri/).

To generate a high-quality SV dataset, we manually checked all inversions and translocations with the following steps: (1) clip 300 bp of upstream/downstream flanking sequence of each break point between the two haplotypes, blast against local PacBio reads with threshold identity >96% and aligned length >550 bp, and require the SV region where the maternal and paternal sequences aligned to have high similarity (>90%); (2) if (1) failed, then check the 10X linked-read count between a 5-kb flanking region; (3) if any break point is not supported by 10X linked-reads, check the Hi-C heat map of this region; if it shows an inversion or translocation pattern on heat map or an ambiguous situation, then remove it.

To evaluate the accuracy of SV detection, we searched the binned PacBio reads around the break points of both maternal and paternal assemblies for all indels in chromosome 1. We looked for one of the following three features to determine the indel as accurate: (1) at least one single PacBio long read from each haplotype that spans the entire indel region with the variation found in each haplotype; (2) overlapping PacBio reads that span the two break points; or (3) manually validated PacBio read alignment by the Integrative Genomics Viewer (IGV)⁷⁴. Finally, we found that 95.7% of indels are correct when considering the breakage location; however, 74.2% are accurate when considering both boundary and location.