Detecting and annotating genetic variations using the HugeSeq pipeline

To the Editor

Deciphering genome sequences is important for the mapping of genetic diseases and prediction of their risks. Advances in high-throughput DNA sequencing technologies using short read lengths have enabled rapid sequencing of entire human genomes and unlocked the potential for comprehensive identification of their underlying genetic variations. Various computational algorithms for identifying and characterizing variants have been developed; however, most of these computational methods are neither integrated nor interoperable, making it difficult for biologists to extract all the genetic information from billions of sequences generated by these sequencing technologies. Here, we present HugeSeq, an integrated computational pipeline to fully automate the process of variant detection from alignment of these genomic sequences to detection and annotation of all types of genetic variations (single nucleotide polymorphisms (SNPs), short insertions or deletions (indels) and larger structural variations (SVs)).

Compared with other popular platforms for genome data analysis that typically analyze SNPs or a limited set of variants (Supplementary Table 1), HugeSeq covers a more complete spectrum of variant types. The complete variant detection and characterization workflow of the HugeSeq pipeline (Fig. 1) is a modular framework comprising three phases: first, a mapping phase that prepares and aligns reads; second, a sorting phase that combines and sorts alignments for parallel variant detection; and third, a reduction phase that detects and annotates different variants (SNPs, indels and structural variations). It is based on a MapReduce¹ approach and runs in a parallel computational environment, making it highly efficient and scalable.

Figure 1.

A MapReduce approach for detecting genetic variants from high-throughput genome sequencing. Phase 1 is the mapping phase including sequence alignment. Phase 2 is the sorting phase including sorting alignments by mapped chromosomes. Phase 3 is the reduction phase including variant detection. chr1, chromosome one; chrM, chromosome M; SV, structural variation; SR, split-read analysis; RD, read-depth analysis; RP, read-pair mapping; JM, junction mapping; VCF, variant call format; GFF, general feature format.

HugeSeq uses sequence reads (both single end and paired end) in a FASTA or FASTQ format (optionally compressed in a GZIP format) as input for alignment. Because alignment of a single read is independent of others, HugeSeq divides the reads into smaller subsets so they can be aligned in parallel. It then distributes the reads in the computer cluster and carries out a gapped alignment against the reference genome using the Burrows-Wheeler aligner². The generated sequence alignment map (SAM)³ is then converted into its binary format, BAM, using SAMtools³ to ensure efficient storing and access of alignment information. After alignment, HugeSeq collects all the mapped reads and sorts them according to their aligned chromosomal positions with the Picard tool. The sorted reads for each chromosome are assigned to their corresponding chromosomal BAM and indexed for rapid and random access using SAMtools. Because most variant detections are intrachromosomal, the detection process can be carried out on each chromosomal BAM simultaneously. Interchromosomal translocation detection can also be enabled and run in a nonparallel mode, although it slows down the process considerably.

To enhance the quality of the alignments for more accurate variant detection, HugeSeq carries out several processing (‘cleanup’) procedures before variant calling. First, to minimize experimental artifacts, it removes potential PCR duplicates using the Picard tool. Second, it carries out a local realignment around indels and SNP clusters using the Genome Analysis ToolKit (GATK) realigner⁴. Last, based on the realignment, it recalibrates the base quality of the alignments using the GATK recalibrater⁴ so that the quality scores represent the empirical probability of mismatching to the reference genome. With the processed read alignments or any user-specified BAMs, HugeSeq detects variants of different kinds in a parallel fashion.

For SNP and small indel detection, HugeSeq uses two different well-established SNP and indel calling algorithms, the GATK UnifiedGenotyper⁴ and SAMtools³. When calling indels using GATK, it uses the Dindel⁵ model for greater sensitivity. The resulting SNPs and indels are then passed through the GATK variant filtering tool with default parameters similar to those used in the 1000 Genomes Project⁶. Structural variations and copy number variants (CNVs) are often difficult to detect, largely owing to their heterogeneous nature. A variety of different methods can be used to find them but each has distinct biases. To identify as many structural variations as possible, HugeSeq uses four major approaches: first, paired-end mapping using BreakDancer⁷; second, split-read analysis using Pindel⁸; third, read-depth analysis using CNVnator⁹ and fourth, junction mapping using BreakSeq¹⁰ (a version we modified to support BAM as input for unmapped reads). Because these structural variation and CNV callers generate variant calls in different formats, HugeSeq standardizes their outputs by converting them into the standard general feature format.

The resulting SNP and indel call sets, which are in a standard variant call format (VCF), are combined and merged using VCFtools¹¹. HugeSeq also uses VCFtools to concatenate variants from different BAMs for each algorithm and to merge calls from different algorithms into a single VCF. SNPs called by both GATK and SAMtools are of particularly high confidence. For the structural variation and CNV call sets, BEDtools¹² is used to intersect and merge the calls. Calls detected by two or more algorithms (50% reciprocal overlap) are regarded as high confidence¹³. Finally, HugeSeq carries out a functional annotation on the variants using Annovar¹⁴. Because both the input and output formats of Annovar are nonstandard variant formats, HugeSeq converts variants into the Annovar input format and converts the Annovar output into a tab-delimited format that can be easily merged and interpreted with the original call sets. The annotation includes gene intersection, exonic variations, repeat elements and mutation information (for example, SIFT score¹⁵). It can be expanded easily to include different databases, such as PolyPhen¹⁶.

We initially applied HugeSeq to a single human genome sequenced (~48×) with Illumina HiSeq. Overall, HugeSeq reported >8 million raw variant calls, or ~4.5 million after merging (Table 1). For SNPs, GATK UnifiedGenotyper reported 3,570,658 calls (after filtering), and SAMtools reported 3,632,090 calls. There were 3,399,561 concordant SNP calls based on positions between GATK and SAMtools. For small indels, GATK (with the Dindel model) reported 523,445 calls and SAMtools reported 553,360 calls. There were 422,305 concordant indel calls based on matching the positions. For structural variation or CNV (≥50 bp), there were 11,043 paired-end calls, 11,911 read-depth calls, 1,741 split-read calls and 1,003 junction calls. There were 21,381 structural variation union calls with 1,639 calls reported by two or more algorithms (50% reciprocal overlap). Because of the ascertainment bias toward deletions, the majority (>90%) of the structural variation calls were deletions; however, duplications, insertions and inversions were also detected (Supplementary Table 2). We observed enrichments of ~300-bp and ~6,000-bp structural variation deletions, possibly owing to Alu and long interdispersed element 1 elements. The majority of indel calls were 1 bp, whereas there was enrichment of 4-bp and 8-bp indel calls (Supplementary Fig. 1).

Table 1.

Summary of detected variant calls

Variant	Tool (algorithm)	Original	Merged (union)	Concordant (intersection)
SNP	GATK	3,570,658	3,803,187	3,399,561
	SAMtools	3,632,090
Indel	GATK SAMtools	523,445 553,360	654,500	422,305
Structural variation or CNV	BreakDancer (paired-end mapping)	11,043	21,381	1,639
	CNVnator (read-depth analysis)	11,911
	Pindel (split-read analysis)	1,741
	BreakSeq (junction mapping)	1,003
Total		8,305,251	4,479,068	3,823,505

We carried out benchmarking on HugeSeq to assess its performance. We measured its run time and memory usage at various sequencing coverage ranging from 6× to 96× (Supplementary Figs. 2 and 3). For run time, we found that HugeSeq spent relatively less clock time on variant detection, as it benefited from parallel processing (Supplementary Fig. 2a,b). We also found that its run time (~25 hours) was ~10 times faster than the time of processing individual steps in a nonparallel mode on a single computer (~250 hours) at 30× coverage, whereas the run time increased almost at a linear rate at different coverage (Supplementary Fig. 2c). For memory usage, most processes took at most 6 GB of physical memory. The duplicate removal, recalibration and GATK UnifiedGenotyper were bound by the maximum heap size of 6 GB preallocated to the Java Virtual Machine and the realignment and sorting were allocated a maximum of 12 GB heap size for a more efficient caching of high-coverage data (Supplementary Fig. 2d). When focusing on variant detection, we observed similar performance as in the overall process (Supplementary Fig. 3). We found that BreakDancer (read-pair mapping) and BreakSeq (junction mapping) were fastest and used the least memory, whereas GATK UnifiedGenotyper was slowest and used the most memory.

To assess the sensitivity of variant detection in HugeSeq, we detected SNPs and structural variations or CNVs using the Illumina (San Diego, CA) Human Omni1Quad genotyping array with ~1 million markers. For SNPs, we assessed the true positives by taking all 260,112 heterozygous calls from the array. We intersected these calls with the 3,399,561 SNP calls reported by both GATK and SAMtools (high confidence) and all 3,803,187 merged SNP calls (Fig. 2a). There were 254,700 and 258,654 concordant calls, corresponding to a sensitivity of 97.9% and 99.4% for the high-confidence and total sets, respectively. For the individual call sets, GATK had 257,243 and SAMtools had 256,111 concordant calls, corresponding to a sensitivity of 98.9% and 98.5%, respectively. We also observed that the GATK- and SAMtools-specific calls had a relatively lower transition-to-transversion rate (<1.6) compared with the high-confidence calls (~2.1). For structural variation or CNVs, we generated a list of true positives by taking the 482 deletion calls reported by Illumina GenomeStudio and CNVision¹⁷ based on the Omni1Quad array. We intersected these calls with the 1,594 deletion calls reported by two or more structural variation/CNV algorithms (high confidence) and all 19,809 merged deletion calls. There were 383 and 471 concordant calls (≥1 bp overlapping), corresponding to a sensitivity of 79.5% and of 97.7% for the high-confidence and total sets, respectively. By requiring an overlap of 50% on the array calls, we found that there were 358 and 444 concordant calls, giving a sensitivity of 74.3% and 92.1% for the high-confidence and total sets, respectively.

Figure 2.

Accuracy and sensitivity of variant detection. (a) Concordant and specific calls in SNP detection by GATK and SAMtools. Original calls, total number of calls from methods; specific calls, total number of method-specific calls. Validated calls, number of calls validated by the array. Ti/Tv, transition-to-transversion rate. (b) Overall sensitivity of SNP and structural variation or CNV detection over different sequencing coverage. X is the average number of reads representing a given nucleotide in a haploid human genome. 50% indicates that detected structural variation calls overlap ≥50% the array calls.

We extended our sensitivity test for SNP or structural variation detection to various sequencing coverage as in the benchmarking. The test was done using the total call sets. We found that SNP calls reached saturation at a sequencing coverage of ~40× with a sensitivity of ~98.5%, whereas structural variation calls reached saturation at ~30× with a sensitivity of ~97% (Fig. 2b). We also intersected the 19,809 deletion calls detected in HugeSeq with 22,025 deletion calls reported by the 1000 Genomes Project in the pilot phase. With 50% reciprocal overlap, we found that structural variations detected by three or more algorithms had up to 93.8% (98.2% by 1 bp) concordance with the 1000 Genomes calls, whereas those detected by two or more algorithms had 88.7% (92.3% by 1 bp) and those detected by any algorithm had only 17.7% (27.6% by 1 bp; Table 2). Thus, we think that the majority of structural variations detected by two or more algorithms are probably correct.

Table 2.

Concordant deletion calls between structural variations detected and from 1000 Genomes Project

Number of algorithms	Deletion calls	1-bp overlapping	50% reciprocal overlapping
≥3 algorithms	451	443 (98.2%)	423 (93.8%)
≥2 algorithms	1,594	1,472 (92.3%)	1,414 (88.7%)
Any algorithm	19,809	5,468 (27.6%)	3,516 (17.7%)

Given the interest in genome sequencing, the availability of an integrated platform to analyze genomes is expected to be useful. The implementation of HugeSeq using widely accessible code with output in a common format could facilitate genome analysis and clinical interpretation¹⁸. In addition to whole-genome sequencing, HugeSeq can be used for other types of sequencing such as exome-seq (currently not applicable for CNV detection). It can also be used for Illumina’s long mate-pair library. We envision that comprehensive variant detection pipelines, such as the one we present here, will facilitate the analysis of biological information extracted from the large amounts of sequencing data currently being generated from thousands of genomes.

HugeSeq is open source and available for download at http://hugeseq.snyderlab.org/. A description of the methods used in our assessment of HugeSeq is in Supplementary methods.