Integrating mapping-, assembly- and haplotype-based approaches for calling variants in clinical sequencing applications

Abstract

High-throughput DNA sequencing technology has transformed genetic research and is starting to make an impact on clinical practice. However, analyzing high-throughput sequencing data remains challenging, particularly in clinical settings where accuracy and turnaround times are critical. We present a new approach to this problem, implemented in a software package called Platypus. Platypus achieves high sensitivity and specificity for SNPs, indels and complex polymorphisms by using local de novo assembly to generate candidate variants, followed by local realignment and probabilistic haplotype estimation. It is an order of magnitude faster than existing tools and generates calls from raw aligned read data without preprocessing. We demonstrate the performance of Platypus in clinically relevant experimental designs by comparing with SAMtools and GATK on whole-genome and exome-capture data, by identifying de novo variation in 15 parent-offspring trios with high sensitivity and specificity, and by estimating human leukocyte antigen genotypes directly from variant calls.

Figure 1

Simplified flow diagram of the integrated calling algorithm. The three stages of the algorithm are pipelined without using intermediate files or separate processes. Mapped and sorted BAM files are used as input; merging, sample demultiplexing and read deduplication are all performed by Platypus. The resulting variant calls require no post-processing, except for a Bayesian filtering stage for de novo mutations. (a) Candidate variants are obtained from read alignments, local assembly and external sources (not shown), and candidate haplotypes are formed. (b) The support of each read for any candidate haplotype is computed by alignment, and population haplotype frequencies are fitted to a diploid segregation model. (c) Variants are called by first calling haplotypes, followed by marginalization over secondary variation. Filtering on the variant and sample levels results in a final call set. See the supplementary note for full details of the algorithm.

Figure 2

Size distribution of indel calls in the NA12878 trio. (a) Histogram of small indel calls (up to 50 bp) by size (negative size, deletion with respect to the reference sequence) for three calling algorithms. UG, UnifiedGenotyper; HC, HaplotypeCaller. (b) Smoothed histograms (10-bp bins) showing larger indels and peaks around ~300 bp corresponding to insertions and deletions of Alu transposable elements. Local assembly allows Platypus to detect insertions up to a few hundred basepairs in size and deletions of over 1 kb in size.

Figure 3

Genotypes of the HLA-A, HLA-B and HLA-C loci at two- and four-digit resolution. Combined genotype alignment scores relative to the maximum-scoring genotypes are shown as a heat map; gray boxes indicate correct genotypes. Correct and unique genotypes at four-digit resolution were estimated from Platypus reference and variant calls for HLA-B (HLA-B*56:01/HLA-B*08:01) and HLA-C (HLA-C*07:01/HLA-C*01:02) (heat maps for genotypes at two-digit resolution shown; see supplementary Fig. 2 for heat maps at four-digit resolution). For HLA-A, the correct genotype HLA-A*11:01/HLA-A*01:01 was among the highest scoring genotypes, but it could be resolved uniquely at two-digit resolution only (lower middle panel); ambiguities remained at four-digit resolution for both haplotypes (right; genotypes with identical scores are indicated).