Whole genome trio variant calling evaluation: low complexity regions, GATK VQSR and high depth filters

Whole genome trio validation

I’ve written previously about the approaches we use to validate the bcbio-nextgen variant calling framework, specifically evaluating aligners and variant calling methods and assessing the impact of BAM post-alignment preparation methods. We’re continually looking to improve both the pipeline and validation methods and two recent papers helped advance best-practices for evaluating and filtering variant calls:

• Michael Linderman and colleagues describe approaches for validating clinical exome and whole genome sequencing results. One key result I took from the paper was the difference in assessment between exome and whole genome callsets. Coverage differences due to capture characterize discordant exome variants, while complex genome regions drive whole genome discordants. Reading this paper pushed us to evaluate whole genome population based variant calling, which is now feasible due to improvements in bcbio-nextgen scalability.
• Heng Li identified variant calling artifacts and proposed filtering approaches to remove them, as well as characterizing caller error rates. We’ll investigate two of the filters he proposed: removing variants in low complexity regions, and filtering high depth variants.

We use the NA12878/NA12891/NA12892 trio from the CEPH 1463 Pedigree as an input dataset, consisting of 50x whole genome reads from Illumina’s platinum genomes. This enables both whole genome comparisons, as well as pooled family calling that replicates best-practice for calling within populations. We aligned reads using bwa-mem and performed streaming de-duplication detection with samblaster. Combined with no recalibration or realignment based on our previous assessment, this enabled fully streamed preparation of BAM files from input fastq reads. We called variants using two realigning callers: FreeBayes (v0.9.14-7) and GATK HaplotypeCaller (3.1-1-g07a4bf8) and evaluated calls using the Genome in a Bottle reference callset for NA12878 (v0.2-NIST-RTG). The bcbio-nextgen documentation has full instructions for reproducing the analysis.

This work provides three practical improvements for variant calling and validation:

• Low complexity regions contribute 45% of the indels in whole genome evaluations, and are highly variable between callers. This replicates Heng’s results and Michael’s assessment of common errors in whole genome samples, and indicates we need to specifically identify and assess the 2% of the genome labeled as low complexity. Practically, we’ll exclude them from further evaluations to avoid non-representative bias, and suggest removing or flagging them when producing whole genome variant calls.
• We add a filter for removing false positives from FreeBayes calls in high depth, low quality regions. This removes variants in high depth regions that are likely due to copy number or other larger structural events, and replicates Heng’s filtering results.
• We improved settings for GATK variant quality recalibration (VQSR). The default VQSR settings are conservative for SNPs and need adjustment to be compatible with the sensitivity available through FreeBayes or GATK using hard filters.

Low complexity regions

Low complexity regions (LCRs) consist of locally repetitive sections of the genome. Heng’s paper identified these using mdust and provides a BED file of LCRs covering 2% of the genome. Repeats in these regions can lead to artifacts in sequencing and variant calling. Heng’s paper provides examples of areas where a full de-novo assembly correctly resolves the underlying structure, while local reassembly variant callers do not.

To assess the impact of low complexity regions on variant calling, we compared calls from FreeBayes and GATK HaplotypeCaller to the Genome in a Bottle reference callset with and without low complexity regions included. The graph below shows concordant non-reference variant calls alongside discordant calls in three categories: missing discordants are potential false negatives, extras are potential false positives, and shared are variants that overlap between the evaluation and reference callsets but differ due to zygosity (heterozygote versus homozygote) or indel representation.

• For SNPs, removing low complexity regions removes approximately ~2% of the total calls for both FreeBayes and GATK. This corresponds to the 2% of the genome subtracted by removing LCRs.
• For indels, removing LCRs removes 45% of the calls due to the over-representation of indels in repeat regions. Additionally, this results in approximately equal GATK and FreeBayes concordant indels after LCR removal. Since the Genome in a Bottle reference callset uses GATK HaplotypeCaller to resolve discrepant calls, this change in concordance is likely due to bias towards GATK’s approaches for indel resolution in complex regions.
• The default GATK VQSR calls for SNPs are not as sensitive, relative to FreeBayes calls. I’ll describe additional work to improve this below.

Practically, we’ll now exclude low complexity regions in variant comparisons to avoid potential bias and more accurately represent calls in the remaining non-LCR genome. We’ll additionally flag low complexity indels in non-evaluation callsets as likely to require additional followup. Longer term, we need to incorporate callers specifically designed for repeats like lobSTR to more accurately characterize these regions.

High depth, low quality, filter for FreeBayes

The second filter proposed in Heng’s paper was removal of high depth variants. This was a useful change in mindset for me as I’ve primarily thought about removing low quality, low depth variants. However, high depth regions can indicate potential copy number variations or hidden duplicates which result in spurious calls.

Comparing true and false positive FreeBayes calls with a pooled multi-sample call quality of less than 500 identifies a large grouping of false positive heterozygous variants at a combined depth, across the trio, of 200:

The cutoff proposed by Heng was to calculate the average depth of called variants and set the cutoff as the average depth plus a multiplier of 3 or 4 times the square root of average depth. This dataset was an average depth of 169 for the trio, corresponding to a cutoff of 208 if we use the 3 multiplier, which compares nicely with a manual cutoff you’d set looking at the above graphs. Applying a cutoff of QUAL < 500 and DP > 208 produces a reduction in false positives with little impact on sensitivity:

A nice bonus of this filter is that it makes intuitive sense: variants with high depth and low quality indicate there is something problematic, and depth manages to partially compensate for the underlying issue. Inspired by GATK’s QualByDepth annotation and default filter of QD < 2.0, we incorporated a generalized version of this into bcbio-nextgen’s FreeBayes filter: QUAL < (depth-cutoff * 2.0) and DP > depth-cutoff.

GATK variant quality score recalibration (VQSR)

The other area where we needed to improve was using GATK Variant Quality Score Recalibration. The default parameters provide a set of calls that are overly conservative relative to the FreeBayes calls. VQSR provides the ability to tune the filtering so we experimented with multiple configurations to achieve approximately equal sensitivity relative to FreeBayes for both SNPs and Indels. The comparisons use the Genome in a Bottle reference callset for evaluation, and include VQSR default settings, multiple tranche levels and GATK’s suggested hard filters:

While the sensitivity/specificity tradeoff depends on the research question, in trying to set a generally useful default we’d like to be less conservative than the GATK VQSR default. We learned these tips and tricks for tuning VQSR filtering:

• The default setting for VQSR is not a tranche level (like 99.0), but rather a LOD score of 0. In this experiment, that corresponded to a tranche of ~99.0 for SNPs and ~98.0 for indels. The best-practice example documentation uses command line parameter that specify a consistent tranche of 99.0 for both SNPs and indels, so depending on which you follow as a default you’ll get different sensitivities.
• To increase sensitivity, increase the tranche level. My expectations were that decreasing the tranche level would include more variants, but that actually applies additional filters. My suggestion for understanding tranche levels is that they specify the percentage of variants you want to capture; a tranche of 99.9% captures 99.9% of the true cases in the training set, while 99.0% captures less.
• We found tranche settings of 99.97% for SNPs and 98.0% for indels correspond to roughly the sensitivity/specificity that you achieve with FreeBayes. These are the new default settings in bcbio-nextgen.
• Using hard filtering of variants based on GATK recommendations performs well and is also a good default choice. For SNPs, the hard filter defaults are less conservative and more in line with FreeBayes results than VQSR defaults. VQSR has improved specificity at the same sensitivity and has the advantage of being configurable, but will require an extra tuning step.

Overall VQSR provides good filtering and the ability to tune sensitivity but requires validation work to select tranche cutoffs that are as sensitive as hard filter defaults, since default values tend to be overly conservative for SNP calling. In the absence of the ability or desire to tune VQSR tranche levels, the GATK hard filters provide a nice default without much of a loss in precision.

Data availability and future work

Thanks to continued community work on improving variant calling evaluations, this post demonstrates practical improvements in bcbio-nextgen variant calling. We welcome interested contributors to re-run and expand on the analysis, with full instructions in the bcbio-nextgen example pipeline documentation. Some of the output files from the analysis may also be useful:

• VCF files for FreeBayes true positive and false positive heterozygote calls, used here to improve filtering via assessment of high depth regions. Heterozygotes make up the majority of false positive calls so take the most work to correctly filter and detect.
• Shared false positives from FreeBayes and GATK HaplotypeCaller. These are potential missing variants in the Genome in a Bottle reference. Alternatively, they may represent persistent errors found in multiple callers.

We plan to continue to explore variant calling improvements in bcbio-nextgen. Our next steps are to use the trio population framework to compared pooled population calling versus the incremental joint discovery approach introduced in GATK 3. We’d also like to compare with single sample calling followed by subsequent squaring off/backfilling to assess the value of concurrent population calling. We welcome suggestions and thoughts on this work and future directions.

Improving reproducibility and installation of genomic analysis pipelines with Docker

Motivation

bcbio-nextgen is a community developed, best-practice pipeline for genomic data processing, performing automated variant calling and RNA-seq analyses from high throughput sequencing data. It has an automated installation script that sets up the code and third party tools used during analysis, and we’ve been working on improving the process to make getting started with bcbio-nextgen easier. The current approach of installing tools in a separate semi-isolated directory is non-optimal for a couple of reasons:

• A separate directory does not give full isolation from system programs and libraries. It’s possible to disrupt processing by unintentionally including other command line programs into your PATH. Additionally it is not easy to recreate a snapshot of the current environment for reproducibility without manual re-installation of specific versions of software.
• The automated installation script needs to deal with the peculiarities of heterogeneous cluster environments. Different system characteristics can be tricky to anticipate and automate, and lead to more tickets devoted to install problems than we’d like. The goal is to do more science and spend less time dealing with installation woes.

Docker lightweight linux containers help solve both of these issues. By isolating tools and software involved in processing, installation is as easy as downloading a pre-built image containing the software. By containerizing the running process, software does not interfere with other installed programs. Docker containers provide the isolation and deployment advantages of Virtual Machines without the associated overhead. Additionally they allow easy export of the full software environment used to run an analysis, improving our ability to reproduce results.

This post describes bcbio-nextgen-vm, a wrapper around bcbio-nextgen that runs analyses using pre-created Docker containers. The implementation is feature compatible with bcbio-nextgen but provides improved installation, isolation and reproducibility. I’ll also discuss future work to further improve provenance and traceability of analysis runs with the Arvados platform, and a fun chance to work on reproducibility and provenance at an Arvados hackathon on Tuesday March 11th.

Implementation

We reused the existing bcbio-nextgen installation scripts to create easily distributed Docker images with pipeline code and external tools. In fact, the bcbio-nextgen Dockerfile replicates current best practice recommendations for setting up the pipeline on a local system. CloudBioLinux drives installation of the software, using packaging work from existing communities such as Bio-Linux, DebianMed and homebrew-science. The advantage over the previous installation approach is that this Docker installation takes place in a defined environment and we distribute the pre-built images, avoiding the need to configure and build software on individual systems.

The pre-built Docker image contains a full manifest of installed software, from the system libraries to custom scientific packages. Coupled with the ability to export and save Docker images, this creates a reproducible run environment. Special thanks for the manifest implementation are due to the DebianMed community and Tony Travis. I had time to finish the manifest implementation while at the DebianMed Hackathon in Aberdeen. This critical component helped enable external version queries for Docker isolated software.

Tying all these parts together, the bcbio-nextgen-vm wrapper drives processing of individual run components using isolated Docker containers. The Python wrapper script uses the existing work in bcbio-nextgen for defining workflows, and it runs on distributed cluster systems using the IPython parallel framework. Using Conda and Binstar to handle installation of Python dependencies results in a streamlined installation procedure for all the wrapper software.

The diagram below shows the parts of bcbio-nextgen handled within each of the components of the system. bcbio-nextgen-vm drives the workflow and parallel runs, interacting with a cluster scheduler, and lives outside of Docker on a central server. The wrapper code manages the work of starting Docker containers and mounting external filesystems to local mounts within the Docker container. On each processing node, execution happens within isolated Docker containers with external biological software and bcbio-nextgen processing-specific code.

Availability

The initial v0.1.0 release of bcbio-nextgen-vm contains full support for all bcbio-nextgen functionality using isolated Docker containers. It runs on clusters using IPython parallel and on single machines using multiple cores, and has minimal external requirements beyond Docker. See the full installation instructions, and bcbio-nextgen-vm run instructions to get started with processing your samples. It uses the same infrastructure and input files as bcbio-nextgen, so the bcbio-nextgen documentation contains much more detail on defining the biological pipelines to run.

With the new isolated framework, you can install bcbio-nextgen on a system with only Docker installed. Conda handles installation of the Python dependencies, ideally inside of an isolated minimal Anaconda Python environment, and is the only non-Docker-contained infrastructure required. The install script will also download and prepare biological data required for processing, including genomes, index files and annotations.

We’re hoping to migrate bcbio-nextgen to this Docker enabled framework over time and welcome feedback on installation or usage challenges that still exist. Reporting problems on the GitHub issue tracker would be a major help as we continue to develop and improve the wrapper framework.

One area of particular interest is installation and security on cluster systems. While patiently waiting for the ability to run Docker as a non-root user, we recommend installing bcbio-nextgen-vm to run with the docker group id on execution. The internal scripts within the bcbio-nextgen Docker container run all commands as the calling user to mitigate security issues.

Provenance and further work

Adding Docker isolated containers provides the pipeline with improved reproducibility. Maintaining the full state of all the tools and software only requires exporting and gzipping the Docker image and storing it alongside the final processed result. The 1Gb stored image can be later reconstituted and rerun to reproduce earlier results, or shared with collaborators to ensure identical processing pipelines across multiple locations. Saving the initial input data plus the Docker image provides the ability to re-run an analysis at any point in the future.

With this framework in place, the next step for improving reproducibility is enabling full provenance to trace processing steps. bcbio-nextgen currently has extensive log files of command lines and program output, but in parallel environments it requires work to deconvolute these to establish the full set of steps leading up to production of files of interest.

Arvados is an promising open source framework designed to help handle provenance and run tracking. Curoverse provides commercial support and development for the Arvados platform and recently closed a round of financing as they continue to expand and develop the framework.

If you’re interested in reproducibility and provenance, and live in the Boston area, Curoverse is hosting an Arvados hackathon next Tuesday evening, March 11th at their offices. I’ll be there learning about ways to integrate bcbio-nextgen with the work they’re doing and would be happy to talk with anyone about the Docker work or reproducible pipelines in general.

Updated comparison of variant detection methods: Ensemble, FreeBayes and minimal BAM preparation pipelines

Variant evaluation overview

I previously discussed our approach for evaluating variant detection methods using a highly confident set of reference calls provided by NIST’s Genome in a Bottle consortium for the NA12878 human HapMap genome, In this post, I’ll update those conclusions based on recent improvements in GATK and FreeBayes.

The comparisons use bcbio-nextgen, an automated open-source pipeline for variant calling and evaluation that identifies concordant and discordant variants with the XPrize validation protocol. By having an automated validation workflow attached to a regularly updated, community developed, variant calling pipeline, we can actively track progress of variant callers and provide updates as algorithms improve.

Since the initial post, There have been two new GATK releases of UnifiedGenotyper and HaplotypeCaller, as well as multiple improvements to FreeBayes. Additionally we’ve enchanced our ensemble calling method, which combines inputs from multiple callers into a single final set of calls, to better handle comparisons with inputs from three callers.

The goal of this post is to re-evaluate these variant detection approaches and provide an updated set of recommendations:

• FreeBayes detects more concordant SNPs and indels compared to GATK approaches, including GATK’s HaplotypeCaller method.
• Post-alignment BAM processing steps like base quality recalibration and realignment have little impact on the quality of variant calls with variant callers that perform local realignment, including FreeBayes and GATK HaplotypeCaller.
• The Ensemble calling method provides the best variant detection by combining inputs from GATK UnifiedGenotyper, HaplotypeCaller and FreeBayes.

Avoiding the post-alignment BAM recalibration and realignment steps allows us to save significant time and pipeline complexity. Combined with the improvements in FreeBayes, this enables a variant calling pipeline that can be freely used for academic, clinical and commercial work with equal quality variant calls compared to current GATK best-practice approaches.

Calling and evaluation methods

We called variants on a NA12878 exome dataset from EdgeBio’s clinical pipeline and assessed them against the NIST’s Genome in a Bottle reference material. Full instructions for replicating the analysis and installing the pipeline are available from the bcbio-nextgen documentation site. Following alignment with bwa-mem (0.7.5a), we post-processed the BAM files with two methods:

• GATK’s best practices (2.7-2): This involves de-duplication with Picard MarkDuplicates, GATK base quality score recalibration and GATK realignment around indels.
• Minimal post-processing, with de-duplication using samtools rmdup and no realignment or recalibration.

We then called variants with three general purpose callers:

• FreeBayes (v0.9.9.2-18): A haplotype-based Bayesian caller from the Marth Lab. We filter calls with a hard filter based on depth, quality and strand bias.
• GATK UnifiedGenotyper (2.7-2): GATK’s widely used Bayesian caller. Since this is single sample exome data, we filter calls using GATK’s recommended hard filters, instead of Variant Quality Score Recalibration (VQSR).
• GATK HaplotypeCaller (2.7-2): GATK’s more recently developed haplotype caller which provides local assembly around variant regions. We also filtered these calls using recommended hard filters.

Finally, we evaluated the calls from each combination of variant caller and BAM post-alignment preparation method using the bcbio.variation framework. This provides a summary identifying concordant and discordant variants, separating SNPs and indels since they have different error profiles. Additionally it classifies discordant variants. where the reference material and evaluation variants differ, into three categories:

• Extra variants, called in the evaluation data but not in the reference. These are potential false positives or missing calls from the reference materials.
• Missing variants, found in the NA12878 reference but not in the evaluation data set. These are potential false negatives.
• Shared variants, called in both the evaluation and reference but differently represented. This results from allele differences, such as heterozygote versus homozygote calls, or variant identification differences, such as indel start and end coordinates.

Variant caller comparison

Using this framework, we compared the 3 variant callers and combined ensemble method:

• FreeBayes outperforms the GATK callers on both SNP and indel calling. The most recent versions of FreeBayes have improved sensitivity and specificity which puts them on par with GATK HaplotypeCaller. One area where FreeBayes performs better is in correctly resolving heterozygote/homozygote calls, reflected in the lower number of discordant shared variants.
• GATK HaplotypeCaller is all around better than the UnifiedGenotyper. In the previous comparison, we found UnifiedGenotyper performed better on SNPs and HaplotypeCaller better on indels, but the recent improvements in GATK 2.7 have resolved the difference in SNP calling. If using a GATK pipeline, UnifiedGenotyper lags behind the realigning callers in resolving indels, and I’d recommend using HaplotypeCaller. This mirrors the GATK team’s current recommendations.
• The ensemble calling approach provides the best overall resolution of both SNPs and indels. The one area where it lags slightly behind is in identification of homozygote/heterozygote calls, especially in indels. This is due to positions where HaplotypeCaller and FreeBayes both call variants but differ on whether it is a heterozygote or homozygote, reflected as higher discordant shared counts.

In addition to calling sensitivity and specificity, an additional factor to consider is the required processing time. Rough benchmarks on family-based calling of whole genome sequencing data indicate that HaplotypeCaller is roughly 7x slower than UnifiedGenotyper and FreeBayes is 2x slower. On multiple 30x whole genome samples, our experience is that calling can range from 10 hours for GATK UnifiedGenotyper to 70 hours for HaplotypeCallers. Ensemble calling requires running all three callers plus combining into a final call set, and for family-based whole genome samples can add another 100 hours of processing time. These estimates fluctuate greatly depending on the compute infrastructure and presence of longer difficult genomic regions with deeper coverage, but give some estimates of timing considerations.

Post-alignment BAM preparation comparison

Given the improved accuracy of local realignment haplotype-based callers like FreeBayes and HaplotypeCaller, we explored the accuracy cost of removing the post-alignment BAM processing steps. The recommended GATK best-practice is to follow up alignment with identification of duplicate reads, followed by base quality score recalibration and realignment around indels. Based on whole genome benchmarking work, these steps can take as long as the initial alignment and scale poorly due to the high IO costs of manipulating large BAM files. For multiple 30x whole genome samples running on 16 cores per sample, this can account for 12 to 16 hours of processing time.

To compare the quality impact of avoiding recalibration and realignment, we performed the identical alignment and variant calling steps as above, but did minimal post-alignment BAM preparation. Following alignment, the only step performed was deduplication using samtools rmdup. Unlike Picard MarkDuplicates, samtools rmdup handles piped streaming input to avoid IO penalties. This is at the cost of not handling some edge cases. Longer term, we’d like to explore biobambam’s markduplicates2, which implements a more efficient streaming version of the Picard MarkDuplicates algorithm.

Suprisingly, skipping base recalibration and indel realignment had almost no impact on the quality of resulting variant calls:

While GATK UnifiedGenotyper suffers during indel calling without recalibration and realignment, both HaplotypeCaller and FreeBayes perform as good or better without these steps. This allows us to save on processing time and complexity without sacrificing call quality when using a haplotype aware realigning caller.

Caveats and conclusions

Taken together, the improvements in FreeBayes and ability to avoid post-alignment BAM processing allow use of a commercially unrestricted GATK-free pipeline with equal quality to current GATK best practices. Adding in GATK’s two callers plus our ensemble combining method provides the most accurate overall calls, at the cost of additional processing time.

It’s also important to consider potential drawbacks of this analysis as we continue to design future evaluations. The comparison is in exome regions for single sample variant calling. In future work it would be helpful to have population or family based inputs. We’d also like to prepare test datasets that focus specifically on evaluating the quality of calls in more difficult repetitive regions within the whole genome. Using populations or whole genomes would also allow use of GATK’s Variant Quality Score Recalibration as part of the pipeline, which could provide improved filtering compared to the hard-filtering approach used here.

Another consideration is that the reference callset prepared by the Genome in a Bottle consortium makes extensive use of GATK tools during preparation. Evaluation of the reference materials with FreeBayes and other callers can help reduce potential GATK-specific biases when continuing to develop reliable reference materials.

All of these pipelines are freely available, open-source, community developed projects and we welcome feedback and contributors. By integrating validation into a scalable analysis pipeline, we hope to build a community interested in widely accessible calling pipelines coupled with well-evaluated reference datasets and methods.

Scaling variant detection pipelines for whole genome sequencing analysis

Scaling for whole genome sequencing

Moving from exome to whole genome sequencing introduces a myriad of scaling and informatics challenges. In addition to the biological component of correctly identifying biological variation, it’s equally important to be able to handle the informatics complexities that come with scaling up to whole genomes.

At Harvard School of Public Health, we are processing an increasing number of whole genome samples and the goal of this post is to share experiences scaling the bcbio-nextgen pipeline to handle the associated increase in file sizes and computational requirements. We’ll provide an overview of the pipeline architecture in bcbio-nextgen and detail the four areas we found most useful to overcome processing bottlenecks:

• Support heterogeneous cluster creation to maximize resource usage.
• Increase parallelism by developing flexible methods to split and process by genomic regions.
• Avoid file IO and prefer streaming piped processing pipelines.
• Explore distributed file systems to better handle file IO.

This overview isn’t meant as a prescription, but rather as a description of experiences so far. The work is a collaboration between the HSPH Bioinformatics Core, the research computing team at Harvard FAS and Dell Research. We welcome suggestions and thoughts from others working on these problems.

Pipeline architecture

The bcbio-nextgen pipeline runs in parallel on single multicore machines or distributed on job scheduler managed clusters like LSF, SGE, and TORQUE. The IPython parallel framework manages the set up of parallel engines and handling communication between them. These abstractions allow the same pipeline to scale from a single processor to hundreds of node on a cluster.

The high level diagram of the analysis pipeline shows the major steps in the process. For whole genome samples we start with large 100Gb+ files of reads in FASTQ or BAM format and perform alignment, post-alignment processing, variant calling and variant post processing. These steps involve numerous externally developed software tools with different processing and memory requirements.

Heterogeneous clusters

A major change in the pipeline was supporting creation of heterogeneous processing environments targeted for specific programs. This moves away from our previous architecture, which attempted to flatten processing and utilize single cores throughout. Due to algorithm restrictions, some software requires the entire set of reads for analysis. For instance, GATK’s base quality recalibrator uses the entire set of aligned reads to accurately calculate inputs for read recalibration. Other software operates more efficiently on entire files: the alignment step scales better by running using multiple cores on a single machine, since the IO penalty for splitting the input file is so severe.

To support this, bcbio-nextgen creates an appropriate type of cluster environment for each step:

• Multicore: Allocates groups of same machine processors, allowing analysis of individual samples with multiple cores. For example, this enables running bwa alignment with 16 cores on multiprocessor machines.
• Full usage of single cores: Maximize usage of single cores for processes that scale beyond the number of samples. For example, we run variant calling in parallel across subsets of the genome.
• Per sample single core usage: Some steps do not currently parallelize beyond the number of samples, so require a single core per sample.

IPython parallel provides the distributed framework for creating these processing setups, working on top of existing schedulers like LSF, SGE and TORQUE. It creates processing engines on distributed cores within the cluster, using ZeroMQ to communicate job information between machines.

Cluster schedulers allow this type of control over core usage, but an additional future step is to include memory and disk IO requirements as part of heterogeneous environment creation. Amazon Web Services allows selection of exact memory, disk and compute resources to match the computational step. Eucalyptus and OpenStack bring this control to local hardware and virtual machines.

Parallelism by genomic regions

While the initial alignment and preparation steps require analysis of a full set of reads due to IO and algorithm restrictions, subsequent steps can run with increased parallelism by splitting across genomic regions. Variant detection algorithms do require processing continuous blocks of reads together, allowing local realignment algorithms to correctly characterize closely spaced SNPs and indels. Previously, we’d split analyses by chromosome but this has the downside of tying analysis times to chromosome 1, the largest chromosome.

The pipeline now identifies chromosome blocks without callable reads. These blocks group by either genomic features like repetitive hard to align sequence or analysis requirements like defined target regions. Using the globally shared callable regions across samples, we fraction the genome into more uniform sections for processing. As a result we can work on smaller chunks of reads during time critical parts of the process: applying base recalibration, de-duplication, realignment and variant calling.

Streaming pipelines

A key bottleneck throughout the pipeline is disk usage. Steps requiring reading and writing large BAM or FASTQ files slow down dramatically once they overburden disk IO, distributed filesystem capabilities or ethernet connectivity between storage nodes. A practical solution to this problem is to avoid intermediate files and use unix pipes to stream results between processes.

We reworked our alignment step specifically to eliminate these issues. The previous attempt took a disk centric approach that allowed scaling out to multiple single cores in a cluster. We split an input FASTQ or BAM file into individual chunks of reads, and then aligned each of these chunks independently. Finally, we merged all the individual BAMs together to produce a final BAM file to pass on to the next step in the process. While nicely generalized, it did not scale when running multiple concurrent whole genomes.

The updated pipeline uses multicore support in samtools and aligners like bwa-mem and novoalign to pipe all steps as a stream: preparation of input reads, alignment, conversion to BAM and coordinate sorting of aligned reads. This results in improved scaling at the cost of only being able to increase single sample throughput to the maximum processors on a machine.

More generally, the entire process creates numerous temporary file intermediates that are a cause of scaling issues. Commonly used best-practice toolkits like Picard and GATK primarily require intermediate files. In contrast, tools in the Marth lab’s gkno pipeline handle streaming input and output making it possible to create alignment post-processing pipelines which minimize temporary file creation. As a general rule, supporting streaming algorithms amenable to piping can ameliorate file load issues associated with scaling up variant calling pipelines. This echos the focus on streaming algorithms Titus Brown advocates for dealing with large metagenomic datasets.

Distributed file systems

While all three of CPU, memory and disk speed limit individual steps during processing, the hardest variable to tweak is disk throughput. CPU and memory limitations have understandable solutions: buy faster CPUs and more memory. Improving disk access is not as easily solved, even with monetary resources, as it’s not clear what combination of disk and distributed filesystem will produce the best results for this type of pipeline.

We’ve experimented with NFS, GlusterFS and Lustre for handling disk access associated with high throughput variant calling. Each requires extensive tweaking and none has been unanimously better for all parts of the process. Much credit is due to John Morrissey and the research computing team at Harvard FAS for help performing incredible GlusterFS and network improvements as we worked through scaling issues, and Glen Otero, Will Cottay and Neil Klosterman at Dell for configuring an environment for NFS and Lustre testing. We can summarize what we’ve learned so far in two points:

• A key variable is the network connectivity between storage nodes. We’ve worked with the pipeline on networks ranging from 1 GigE to InfiniBand connectivity, and increased throughput delays scaling slowdowns.
• Different part of the processes stress different distributed file systems in complex ways. NFS provides the best speed compared to single machine processing until you hit scaling issues, then it slows down dramatically. Lustre and GlusterFS result in a reasonable performance hit for less disk intensive processing, but delay the dramatic slowdowns seen with NFS. However, when these systems reach their limits they hit a slowdown wall as bad or worse than NFS. One especially slow process identified on Gluster is SQLite indexing, although we need to do more investigation to identify specific underlying causes of the slowdown.

Other approaches we’re considering include utilizing high speed local temporary disk, reducing writes to long term distributed storage file systems. This introduces another set of challenges: avoiding stressing or filling up local disk when running multiple processes. We’ve also had good reports about using MooseFS but haven’t yet explored setting up and configuring another distributed file system. I’d love to hear experiences and suggestions from anyone with good or bad experiences using distributed file systems for this type of disk intensive high throughput sequencing analysis.

A final challenge associated with improving disk throughput is designing a pipeline that is not overly engineered to a specific system. We’d like to be able to take advantage of systems with large SSD attached temporary disk or wonderfully configured distributed file systems, while maintaining the ability scale on other systems. This is critical for building a community framework that multiple groups can use and contribute to.

Timing results

Providing detailed timing estimates for large, heterogeneous pipelines is difficult since they will be highly depending on the architecture and input files. Here we’ll present some concrete numbers that provide more insight into the conclusions presented above. These are more useful as a side by side comparison between approaches, rather than hard numbers to predict scaling on your own systems.

In partnership with Dell Solutions Center, we’ve been performing benchmarking of the pipeline on dedicated cluster hardware. The Dell system has 32 16-core machines connected with high speed InfiniBand to distributed NFS and Lustre file systems. We’re incredibly appreciative of Dell’s generosity in configuring, benchmarking and scaling out this system.

As a benchmark, we use 10x coverage whole genome human sequencing data from the Illumina platinum genomes project. Detailed instructions on setting up and running the analysis are available as part of the bcbio-nextgen example pipeline documentation.

Below are wall-clock timing results, in total hours, for scaling from 1 to 30 samples on both Lustre and NFS fileystems:

	primary	1 sample	1 sample	1 sample	30 samples	30 samples
	bottle	16 cores	96 cores	96 cores	480 cores	480 cores
	neck	Lustre	Lustre	NFS	Lustre	NFS
alignment	cpu/mem	4.3h	4.3h	3.9h	4.5h	6.1h
align post-process	io	3.7h	1.0h	0.9h	7.0h	20.7h
variant calling	cpu/mem	2.9h	0.5h	0.5h	3.0h	1.8h
variant post-process	io	1.0h	1.0h	0.6h	4.0h	1.5h
total		11.9h	6.8h	5.9h	18.5h	30.1h

Some interesting conclusions:

• Scaling single samples to additional cores (16 to 96) provides a 40% reduction in processing time due to increased parallelism during post-processing and variant calling.
• Lustre provides the best scale out from 1 to 30 samples, with 30 sample concurrent processing taking only 1.5x as along as a single sample.
• NFS provides slightly better performance than Lustre for single sample scaling.
• In contrast, NFS runs into scaling issues at 30 samples, proceeding 5.5 times slower during the IO intensive alignment post-processing step.

This is preliminary work as we continue to optimize code parallelism and work on cluster and distributed file system setup. We welcome feedback and thoughts to improve pipeline throughput and scaling recommendations.

Framework for evaluating variant detection methods: comparison of aligners and callers

Variant detection and grading overview

Developing pipelines for detecting variants from high throughput sequencing data is challenging due to rapidly changing algorithms and relatively low concordance between methods. This post will discuss automated methods providing evaluation of variant calls, enabling detailed diagnosis of discordant differences between multiple calling approaches. This allows us to:

• Characterize strengths and weaknesses of alignment, post-alignment preparation and calling methods.
• Automatically verify pipeline updates and installations to ensure variant calls recover expected variations. This extends the XPrize validation protocol to provide full summary metrics on concordance and discordance of variants.
• Make recommendations on best-practice approaches to use in sequencing studies requiring either exome or whole genome variant calling.
• Identify characteristics of genomic regions more likely to have discordant variants which require additional care when making biological conclusions based on calls, or lack of calls, in these regions.

This evaluation work is part of a larger community effort to better characterize variant calling methods. A key component of these evaluations is a well characterized set of reference variations for the NA12878 human HapMap genome, provided by NIST’s Genome in a Bottle consortium. The diagnostic component of this work supplements emerging tools like GCAT (Genome Comparison and Analytic Testing), which provides a community platform for comparing and discussing calling approaches.

I’ll show a 12 way comparison between 2 different aligners (novoalign and bwa mem), 2 different post-alignment preparation methods (GATK best practices and the Marth lab’s gkno pipeline), and 3 different variant callers (GATK UnifiedGenotyper, GATK HaplotypeCaller, and FreeBayes). This allows comparison of openly available methods (bwa mem, gkno preparation, and FreeBayes) with those that require licensing (novoalign, GATK’s variant callers). I’ll also describe bcbio-nextgen, the fully automated open-source pipeline used for variant calling and evaluation, which allows others to easily bring this methodology into their own work and extend this analysis.

Aligner, post-alignment preparation and variant calling comparison

To compare methods, we called variants on a NA12878 exome dataset from EdgeBio’s clinical pipeline and assessed them against the NIST Genome in a Bottle reference material. Discordant positions where the reference and evaluation variants differ fall into three different categories:

• Extra variants, called in the evaluation data but not in the reference. These are potential false positives.
• Missing variants, found in the NA12878 reference but not in the evaluation data set. These are potential false negatives. The use of high quality reference materials from NIST enables identification of genomic regions we fail to call in.
• Shared variants, called in both the evaluation and reference but differently represented. This could result from allele differences, such as heterozygote versus homozygote calls, or variant identification differences, such as indel start and end coordinates.

To further identify causes of discordance, we subdivide the missing and extra variants using annotations from the GEMINI variation framework:

• Low coverage: positions with limited read coverage (4 to 9 reads).
• Repetitive: regions identified by RepeatMasker.
• Error prone: variants falling in motifs found to induce sequencing errors.

We subdivide and restrict our comparisons to help identify sources of differences between methods indistinguishable when looking at total discordant counts. A critical subdivison is comparing SNPs and indels separately. With lower total counts of indels but higher error rates, each variant type needs independent visualization. Secondly, it’s crucial to distinguish between discordance caused by a lack of coverage, and discordance caused by an actual difference in variant assessment. We evaluate only in callable regions with 4 or more reads. This low minimum cutoff provides a valuable evaluation of low coverage regions, which differ the most between alignment and calling methods.

I’ll use this data to provide recommendations for alignment, post-alignment preparation and variant calling. In addition to these high level summaries, the full dataset and summary plots available below providing a starting place for digging further into the data.

Aligners

We compared two recently released aligners designed to work with longer reads coming from new sequencing technologies: novoalign (3.00.02) and bwa mem (0.7.3a). bwa mem identified 1389 additional concordant SNPs and 145 indels not seen with novoalign. 1024 of these missing variants are in regions where novoalign does not provide sufficient coverage for calling. Of those, 92% (941) have low coverage with less than 10 reads in the bwa alignments. Algorithmic changes impact low coverage regions more due to the decreased evidence and susceptibility to crossing calling coverage thresholds, so we need extra care and consideration of calls in these regions.

Our standard workflow uses novoalign based on its stringency in resolving large insertions and deletions. These results suggest equally good results using bwa mem, along with improved processing times. One caveat to these results is that some of the available Illumina call data that feeds into NIST’s reference genomes comes from a bwa alignment, so some differences may reflect a bias towards bwa alignment heuristics. Using non-simulated reference data sets has the advantage of capturing real biological and process errors, but requires iterative improvement of the reference materials to avoid this type of potential algorithmic bias.

Post-alignment preparation and quality score recalibration

We compared two methods of quality recalibration:

• GATK’s best practices (2.4-9): This involves de-duplication with Picard MarkDuplicates, GATK base quality score recalibration and GATK realignment around indels.
• The Marth Lab’s gkno realignment pipeline: This performs de-duplication with samtools rmdup and realignment around indels using ogap. All commands in this pipeline work on streaming input, avoiding disk IO penalties by using unix pipes. This piped approach improves scaling on large numbers of whole genome samples. Notably, our implementation of the pipeline does not use a base quality score recalibration step.

GATK best practice pipelines offer an advantage over the gkno-only pipeline primarily because of improvements in SNP calling from base quality recalibration. Specifically we lose ~1% (824 / 77158) of called variations. These fall into the discordant missing “other” category, so we cannot explain them by metrics such as coverage or genome difficulty. The simplest explanation is that initial poor quality calculations in those regions result in callers missing those variants. Base quality recalibration helps recover them. These results match Brendan O’Fallon’s recent analysis of base quality score recalibration.

This places a practical number on the lost variants when avoiding recalibration either due to scaling or GATK licensing concerns. Some other options for recalibration include Novoalign’s Quality Recalibration and University of Michigan’s BamUtil recab, although we’ve not yet tested either in depth as potential supplements to improve calling in non-GATK pipelines.

Variant callers

For this comparison, we used three general purpose callers that handle SNPs and small indels, all of which have updated versions since our last comparison:

• FreeBayes (0.9.9 296a0fa): A haplotype-based Bayesian caller from the Marth Lab, with filtering on quality score and read depth.
• GATK UnifiedGenotyper (2.4-9): GATK’s widely used Bayesian caller, using filtering recommendations for exome experiments from GATK’s best practices.
• GATK HaplotypeCaller (2.4-9): GATK’s more recently developed haplotype caller which provides local assembly around variant regions, using filtering recommendations for exomes from GATK’s best practices.

Adjusting variant calling methods has the biggest impact on the final set of calls. Called SNPs differ by 4577 between the three compared approaches, in comparison with aligner and post-alignment preparation changes which resulted in a maximum difference of 1389 calls. This suggests that experimenting with variant calling approaches currently provides the most leverage to improve calls.

A majority of the SNP concordance differences between the three calling methods are in low coverage regions with between 4 and 9 reads. GATK UnifiedGenotyper performs the best in detecting SNPs in these low coverage regions. FreeBayes and GATK HaplotypeCaller both call more conservatively in these regions, generating more potential false negatives. FreeBayes had the fewest heterozygote/homozygote discrimination differences of the three callers.

For indels, FreeBayes and HaplotypeCaller both provide improved sensitivity compared to UnifiedGenotyper, with HaplotypeCaller identifying the most, especially in low coverage regions. In contrast to the SNP calling results, FreeBayes has more calls that match the expected indel but differ in whether a call is a heterozygote or homozygote.

No one caller outperformed the others on all subsets of the data. GATK UnifiedGenotyper performs best on SNPs but is less sensitive in resolving indels. GATK HaplotypeCaller identifies the most indels, but is more conservative than the other callers on SNPs. FreeBayes provides intermediate sensitivity and specificity between the two for both SNPs and indels. A combined UnifiedGenotyper and HaplotypeCaller pipeline for SNPs and indels, respectively, would provide the best overall calling metrics based on this set of comparisons.

Low coverage regions are the key area of difference between callers. Coupled with the alignment results and investigation of variant changes resulting from quality score binning, this suggests we should be more critical in assessing both calls and coverage in these regions. Assessing coverage and potential false negatives is especially critical since we lack good tools to summarize and prioritize genomic regions that are potentially missed during sequencing. This also emphasizes the role of population-based calling to help resolve low coverage regions, since callers can use evidence from multiple samples to better estimate the likelihoods of low coverage calls.

Automated calling and grading pipeline

Method comparisons become dated quickly due to the continuous improvement in aligners and variant callers. While these recommendations are useful now, in 6 months there will be new releases with improved approaches. This rapid development cycle creates challenges for biologists hoping to derive meaning from variant results: do you stay locked on software versions whose trade offs you understand, or do you attempt to stay current and handle re-verifying results with every new release?

Our goal is to provide a community developed pipeline and comparison framework that ameliorates this continuous struggle to re-verify. The analysis done here is fully automated as part of the bcbio-nextgen analysis framework. This framework code provides full exposure and revision tracking of all parameters used in analyses. For example, the ngsalign module contains the command lines used for bwa mem and novoalign, as well as all other tools.

To install the pipeline, third-party software and required data files:

wget https://raw.github.com/chapmanb/bcbio-nextgen/master/scripts/bcbio_nextgen_install.py
python bcbio_nextgen_install.py /usr/local /usr/local/share/bcbio-nextgen

The installer bootstraps all installation on a bare machine using the CloudBioLinux framework. More details and options are available in the installation documentation.

To re-run this analysis, retrieve the input data files and configuration as described in the bcbio-nextgen example documentation with:

$ mkdir config && cd config
$ wget https://raw.github.com/chapmanb/bcbio-nextgen/master/config/\
   examples/NA12878-exome-methodcmp.yaml
$ cd .. && mkdir input && cd input
$ wget https://dm.genomespace.org/datamanager/file/Home/EdgeBio/\
   CLIA_Examples/NA12878-NGv3-LAB1360-A/NA12878-NGv3-LAB1360-A_1.fastq.gz
$ wget https://dm.genomespace.org/datamanager/file/Home/EdgeBio/\
   CLIA_Examples/NA12878-NGv3-LAB1360-A/NA12878-NGv3-LAB1360-A_2.fastq.gz
$ wget https://s3.amazonaws.com/bcbio_nextgen/NA12878-nist-v2_13-NGv3-pass.vcf.gz
$ wget https://s3.amazonaws.com/bcbio_nextgen/NA12878-nist-v2_13-NGv3-regions.bed.gz
$ gunzip NA12878-nist-*.gz
$ wget https://s3.amazonaws.com/bcbio_nextgen/NGv3.bed.gz
$ gunzip NGv3.bed.gz

Then run the analysis, distributed on 8 local cores, with:

$ mkdir work && cd work
$ bcbio_nextgen.py bcbio_system.yaml ../input ../config/NA12878-exome-methodcmp.yaml -n 8

The bcbio-nextgen documentation describes how to parallelize processing over multiple machines using cluster schedulers (LSF, SGE, Torque).

The pipeline and comparison framework are open-source and configurable for multiple aligners, preparation methods and callers. We invite anyone interested in this work to provide feedback and contributions.

Full data sets

We extracted the conclusions for alignment, post-alignment preparation and variant calling from analysis of the full dataset. The visualizations for the full data are not as pretty but we make them available for anyone interested in digging deeper:

• Summary CSV of comparisons split by methods and concordance/discordance types, easily importable into R or pandas for further analysis.
• Code for preparing and plotting results
• Full comparisons of all 12 methods, stratified by concordance and discordance: SNPs and indels
• Boxplots of differences between alignment methods: SNPs and indels
• Boxplots of differences between post-alignment preparation methods: SNPs and indels
• Boxplots of differences between variant calling methods: SNPs and indels

The comparison variant calls are also useful for pinpointing algorithmic differences between methods. Some useful subsets of variants:

• Concordant variants called by bwa and not novoalign, where novoalign did not have sufficient coverage in the region. These are calls where either novoalign fails to map some reads, or bwa maps too aggressively: VCF of bwa calls with low or no coverage in novoalign.
• Discordant variants called consistently by multiple calling methods. These are potential errors in the reference material, or consistently problematic calling regions for multiple algorithms. Of the 9004 shared discordants, the majority are potential false negatives not seen in the evaluation calls (7152; 79%). Another large portion is heterozygote/homozygote differences, which make up 1627 calls (18%). 6652 (74%) of the differences have low coverage in the exome evaluation, again reflecting the difficulties in calling in these regions. The VCF of discordants found in 2 or more callers contains these calls, with a ‘GradeCat’ INFO tag specifying the discordance category.

We encourage reanalysis and welcome suggestions for improving the presentation and conclusions in this post.

The influence of reduced resolution quality scores on alignment and variant calling

BAM file size reduction and quality score binning

We have a large upcoming whole genome sequencing project with Illumina, and they approached us about delivering BAM files with reduced resolution base quality scores. They have a white paper describing the approach, which involves binning scores to reduce resolution. This reduces the number of scores describing the quality of a base from 40 down to 8.

The advantage of this approach is a significant reduction in file size. BAM files use BGZF compression, and the underlying gzip DEFLATE algorithm compresses based on shared text regions. Reducing the number of quality values increases shared blocks and improves compression. This reduces BAM file sizes by 25-35%: an exome BAM file reduced from 5.7Gb to 3.7Gb after quality binning.

The potential downside is that the reduction in quality resolution may impact alignment and variant calling approaches that rely on base quality scores. To assess this, I implemented quality score binning as part of the bcbio-nextgen analysis pipeline using the CRAM toolkit and ran alignment, recalibration, realignment and variant calling on:

• The original unbinned 40-resolution base quality BAM from an NA12878 exome.
• The BAM binned into 8-resolution base qualities before alignment.
• The BAM binned into 8-resolution base qualities before alignment and binned again following base quality score recalibration.

A comparison of alignment and variant calls from the three approaches indicates that binning has nearly no impact on alignment and a small impact on variant calls, primarily in low depth regions.

Alignment differences

We aligned 100bp paired end reads with Novoalign, a quality aware aligner. Comparison of mapped reads showed nearly no impact on total mapped reads. The plot below shows a generic delta of changes in mapped reads across the 22 autosomes alongside the increase in unmapped pairs. Out of 73 million total reads, the changes account for ~0.003% of the total reads. There also did not appear to be any worrisome patterns of loss for specific chromosomes. Overall, there is a minimal impact of quality score binning on the ability to align the reads.

Variant call differences

We called variants using the GATK Unified Genotyper following the best practice recommendations for exomes and then compared calls from original and binned quality scores. Both approaches for binning — pre-binning, and pre-binning plus post-quality recalibration binning — showed similar levels of concordance to non-binned quality scores: 99.81 and 99.78, respectively. Since the additional binning after recalibration provides a smaller prepared BAM file for storage and has a similar impact to pre-binning only, we used this for additional analysis of discordant variants.

The table below shows the discordant differences between the 40 quality score resolution and binned, 8 quality score resolution BAMs. 40 quality discordant variants are those called with full quality score resolution but not called, or called differently, after binning to 8 quality score resolution. Conversely, the 8-quality discordants are those called uniquely after quality binning:

Overall genotype concordance	99.78
concordant: total	117887
concordant: SNPs	109144
concordant: indels	8743
40-quality discordant: total	821
40-quality discordant: SNPs	759
40-quality discordant: indels	62
8-quality discordant: total	1289
8-quality discordant: SNPs	1240
8-quality discordant: indels	49
het/hom discordant	259

We investigated the discordant variants further since 1.5% of the total variant calls change as a result of binning, Of the 1851 unique discordant variants, approximately half (928) fall into reproducible variants identified by looking at ensemble combinations of replicates. Of these potentially problematic discordant variants more than half are in low coverage regions with less than 10 reads:

The major influence of quality score binning is resolution of variants in low coverage regions. This manifests as differences in heterozygote and homozygote calling, indel representation and filtering differences related to quality and mappability. To assess the potential impact, we looked at the loss in callable bases on a 30x whole genome sequence when moving from a minimum of 5 reads to a minimum of 10, using GATK’s CallableLoci tool. Regions with read coverage of 5 to 9 make up 4.7 million genome positions, 0.17% of the total callable bases.

	5 read minimum	10 read minimum
Callable bases	2,775,871,235	2,771,109,000
Percent callable	96.90%	96.73%
Low coverage	17,641,980	22,404,215
No coverage/ poor mapping	71,272,008	71,272,008

In conclusion, quality score binning provides a useful reduction in input file sizes with minimal impact on alignment. For variant calling, use additional caution in low coverage regions with less than 10 supporting reads. Given the rapid increases in read throughput that are driving the need for file size reduction, quality score binning is a worthwhile tradeoff for high-coverage recalling work.

An automated ensemble method for combining and evaluating genomic variants from multiple callers

Overview

A key goal of the Archon Genomics X Prize infrastructure is development of a set of highly accurate reference genome variants. I’ve described our work preparing these reference genomes, and specifically defined the challenges behind merging genomic variant calls from multiple technologies and calling methods. Comparing calls from two different calling methods, for example GATK and samtools mpileup, produces a large number of differing variants which need reconciliation. Taking the overlapping subset from multiple callers is too conservative and will miss real variations, while including all calls is too liberal and introduces false positives.

Here I’ll describe a fully automated approach for preparing an accurate set of combined variant calls. Ensemble machine learning methods are a powerful way to incorporate inputs from multiple models. We use a heuristic and support vector machine (SVM) algorithm to consolidate variants, producing a final set of calls with better sensitivity and specificity than current best practice methods. The approach is open source, fully automated and generalizable to both human diploid sequencing as well as X Prize haploid reference fosmids.

We use a pair of replicates from EdgeBio’s clinical exome sequencing pipeline to prepare ensemble variant calls in the widely studied HapMap NA12878 genome. Compared to variants from a single calling method, the ensemble method produced more concordant variants when comparing the replicates, with fewer discordants. The finalized ensemble calls also provide a useful method to compare strengths and weaknesses of individual calling methods. The implementation is freely available and I’ll discuss how to get it running on your data so you can use, critique and extend the methods. This work is a collaboration between Harvard School of Public Health, EdgeBio and NIST.

Comparison materials and algorithm

A difficult aspect of evaluating variant calling methods is establishing a reference set of calls. For X Prize we use three established methods, each of which comes with tradeoffs. Metrics like transition/transversion ratios or dbSNP overlap provide a global picture of calling but are not fine grained enough to distinguish improvements over best practices. Sanger validation restricts you to a manageable subset of calls. Comparisons against public resources like 1000 genomes bias results towards technologies and callers used in preparing those variant callsets.

Here we employ a fourth method by comparing replicates from EdgeBio’s clinical exome sequencing pipeline. These are NA12878 samples independently prepared using Nimblegen’s version 3.0 kit and sequenced on an Illumina HiSeq. By comparing the replicates in regions with 4 or more reads in both samples, we identify the ability of variant calling algorithms to call identical variations with differing coverage and error profiles.

We aligned reads with novoalign and performed deduplication, base recalibration and realignment using GATK best practices. With these prepared reads, we called variants with five approaches:

• GATK UnifiedGenotyper – Bayesian approach to call SNPs and indels, treating each position independently.
• GATK HaplotypeCaller – Performs local de-novo assembly to call SNPs and indels on individual haplotypes.
• FreeBayes – Bayesian calling approach that handles simultaneous SNPs and indel calling via assessment of regional haplotypes.
• samtools mpileup – Uses an approach similar to GATK’s UnifiedGenotyper for SNP and indel calling.
• VarScan – Calls variants using a heuristic/statistic approach eliminating common sources of bias.

We took a combined heuristic and machine learning approach to consolidate these five sets of variant calls into a final ensemble callset. The first step is to prepare the union of all variant calls from the input callers, identifying calling methods that support each variant. Secondly, we annotate each variant with metrics including strand bias, allele balance, regional sequence entropy, position of calls within reads, regional base quality and overall genotype likelihoods. We then filter this prepared set of all possible variants to produce a final set of trusted calls.

The first filtering step is to heuristically identify trusted variants based on the number of callers supporting them. This configurable parameter allow you to make an initial conservative cutoff for including variants in the final calls: I trust variants supported by N or more callers.

For the remaining calls that fall below the trusted support cutoff, we distinguish true and false positives using a support vector machine (SVM). The annotated metrics described above are the input parameters and we prepare true and false positives for the classifier using a multi-step process:

• Use variants found in all callers as the true positive set, and variants found in a single caller as false positives. Use these training variants to identify an initial set of below-cutoff variants to include and exclude.
• With this initial set of below-cutoff true/false variants, re-train multiple classifiers stratified based on variant characteristics: variant type (indels vs SNPs), zygosity (heterozygous vs homozygous) and regional sequence complexity.
• Use these final classifiers to identify included and excluded variants falling below the trusted calling support cutoff.

The final set of calls includes the trusted variants and those that pass the SVM filtering. An input configuration file for variant preparation and assessment allows adjustment of the trusted threshold as well as defining which metrics to use for building the SVM classifiers.

Ensemble calling improvements

We assess calling sensitivity and specificity by comparing concordant and discordant variant calls between the replicates. To provide a consistent method to measure both SNP and indel correctness, we use the positive predictive value as the percentage of concordant calls between duplicates (concordant variants / (concordant variants + discordant variants)). This is different than the overall concordance rate, which also includes non-variant regions where both replicates do not call a variation. As a result these percentages will be lower if you expect the 99% values that result when including reference calls. The advantage of this metric is that it’s easily interpreted as the percentage of concordant called variants. It also allows individual comparisons of SNPs and indels, since the overall number of indels are low compared to the total bases considered. GATK’s VariantEval documentation has a nice discussion of alternative metrics to genotype concordance.

As a baseline we used calls from GATK’s UnifiedGenotyper to represent a current best practice approach. GATK calls 117079 SNPs, 86.6% of which are concordant. It also calls 14966 indels, with 64.6% concordant. Here are the full concordant and discordant numbers, broken down by variant type and replicate:

concordant: total	111159
concordant: SNPs	101495
concordant: indels	9664
rep1 discordant: total	9857
rep1 discordant: SNPs	7514
rep1 discordant: indels	2343
rep2 discordant: total	11029
rep2 discordant: SNPs	8070
rep2 discordant: indels	2959
het/hom discordant	4181

Our ensemble method produces improvements in both total concordant variants detected and the ratio of concordant to discordants. For SNPs, the ensemble calls add 5345 additional variants to a total of 122424, with an increase in concordance to 87.4%. For indels the major improvement is in removal of discordants: We identify 14184 indels with 67.2% concordant. Here is the equivalent table for the ensemble method:

concordant: total	116608
concordant: SNPs	107063
concordant: indels	9545
rep1 discordant: total	9555
rep1 discordant: SNPs	7581
rep1 discordant: indels	1974
rep2 discordant: total	10445
rep2 discordant: SNPs	7780
rep2 discordant: indels	2665
het/hom discordant	3975

For scientists who have worked to increase sensitivity and specificity of individual variant callers, it’s exciting to be able to improve both simultaneously. As mentioned above, you can also tune the method to increase specificity or sensitivity by adjusting the support needed for including trusted variants.

The final ensemble callsets from both replicates are available as VCF files from GenomeSpace in the xprize/NA12878-exome-v_03 folder:

• NA12878 exome ensemble calls, replicate 1 – Variants (VCF), Callable regions (BED)
• NA12878 exome ensemble calls, replicate 2 – Variants (VCF), Callable regions (BED)

Comparison of calling methods

Calling the same samples with multiple callers allows direct comparisons between calling methods. The advantage of producing an accurate final set of ensemble calls is that this provides a baseline to evaluate the strengths and weaknesses of different calling methods. The figure below compares concordant, missing variants and additional variants called by each of the 5 methods in comparison with the consolidated ensemble calls:

• GATK UnifiedGenotyper provides the best SNP calling, followed closely by samtools mpileup.
• For indel calling, the GATK HaplotypeCaller produces the most concordant calls followed by UnifiedGenotyper and FreeBayes. UnifiedGenotyper does good as well, but is conservative and has the fewest additional indels. FreeBayes and GATK HaplotypeCaller both provide resolution of individual haplotypes which helps in regions with heterozygous indels or closely spaced SNPs and indels.
• If you want to use a single variant caller, GATK UnifiedGenotyper does the best overall job.
• If you wanted to choose free open-source tools for calling, I would recommend samtools for SNP calling and FreeBayes for indel calling.

Variant calling methods with recommendations for both calling and filtering provide the best out of the box performance. An advantage of GATK and samtools is they provide calling, variant quality metrics, and filtering. On the other side, FreeBayes is a good example of a powerful tool that takes some time to learn to filter optimally. One potential source of bias in producing the individual calls is that I personally have more experience with GATK tools so may have room to improve with the other callers.

Availability and usage

Combining multiple calling approaches improves both sensitivity and specificity of the final set of variants. The downside is the need to run and coordinate calls from all of the different callers. To mitigate this, we developed an automated pipeline that ties together multiple open-source tools using two custom components:

• bcbio-nextgen – A Python framework to run a full sequencing analysis pipeline from input fastq files to consolidated ensemble variant calls. It supports multiple aligners and variant callers, and enables distributed work over multiple cores on a large machine or multiple machines in a cluster environment.
• bcbio.variation – A Clojure toolkit build on top of GATK’s variant API that provides ensemble call preparation as well as more general functionality for normalizing and comparing variants produced by multiple callers.

bcbio-nextgen has a script, built on functionality in the CloudBioLinux project, that automates installation of associated variant callers and data dependencies:

wget https://raw.github.com/chapmanb/bcbio-nextgen/master/scripts/bcbio_nextgen_install.py
python bcbio_nextgen_install.py install_directory data_directory

With the dependencies installed, you describe the input files and analysis with a YAML formatted input file. The NA12878 ensemble configuration file used for this analysis provides a useful starting point. Run the analysis, distributed on multiple cores, with:

bcbio_nextgen.py bcbio_system.yaml ensemble_sample.yaml -n 8

The bcbio-nextgen documentation provides additional details about configuration inputs and distributed processing. The framework generally handles the automation and processing involved with high throughput sequencing analysis.

EdgeBio kindly made the NA12878 datasets used in this analysis publicly available:

• NA12878-NGv3-LAB1360-A (replicate 1)
• NA12878-NGv3-LAB1363-A (replicate 2)

I welcome feedback on the approach, data or tools and am actively working to extend this and make it easier to use. As re-sequencing becomes increasingly important for human health applications it is critical that we develop open, shared best-practice workflows to handle the data processing. This allows us to focus back on the fun and difficult work of understanding the biology.

Genomics X Prize public phase update: variant classification and de novo calling

Background

Last month I described our work at HSPH and EdgeBio preparing reference genomes for the Archon Genomics X Prize public phase, detailing methods used in the first version of our NA19239 variant calls. We’ve been steadily improving the calling approaches, and released version 0.2 on the X Prize validation website and GenomeSpace. Here I’ll describe the improvements we’ve made over the last month, focusing on two specific areas:

• De novo calling: Zam Iqbal suggested using his cortex_var de novo variant caller in addition to the current GATK, FreeBayes and samtools callers. With his help, we’ve included these calls in this release, and provide comparisons between de novo and alignment based methods.
• Improved variant classification: Consolidating variant calls from multiple callers involves making tough choices about when to include or exclude variants. I’ll describe the details of selecting metrics for use in SVM classification and filtering of variants.

Our goal is to iteratively improve our calling and variant preparation to create the best possible set of reference calls. I’d be happy to talk more with anyone working on similar problems or with insight into useful ways to improve our current callsets. We have a Get Satisfaction site for discussion and feedback, and have offered a $2500 prize for helpful comments.

As a reminder, all of the code and data used here is freely available:

• The variant analysis infrastructure, built on top of GATK, automates genome preparation, normalization and comparison. It provides a full pipeline, driven by simple configuration files, for consolidating multiple variant calls.
• The combined variant calls, including training data and potential true and false positives, are available from GenomeSpace: Public/chapmanb/xprize/NA19239-v0_2.
• The individual variant calls for each technology and calling method are also available from GenomeSpace: Public/EdgeBio/PublicData/Release1.

de novo variant calling with cortex_var

de novo variant calling performs reference-free assembly of either local or global genome regions, then subsequently uses these assemblies to call variants relative to a known reference. The advantage is that assemblies can avoid errors associated with mapping to the reference, helping resolve large variations as well as small variations near problem alignments or low complexity regions.

Hybrid approaches that use localized de novo assembly in variant regions help mitigate the extensive computational requirements associated with whole-genome assembly. Complete Genomics variant calling and GATK 2.0’s Haplotype Caller both provide pipelines for hybrid de novo assembly in variant detection. The fermi and SGA assemblers are also used in variant calling, although the paths from assembly to variants are not as automated.

Thanks to Zam’s generous assistance, we used cortex_var for localized de novo assembly and variant calling within individual fosmid boundaries. As a result, CloudBioLinux now contains automated build instructions for cortex_var , handling binary builds for multiple k-mer and color combinations. An automated cortex_var pipeline, part of the bcbio-nextgen toolkit, runs the processing workflow:

• Start with reads aligned to fosmid regions using novoalign.
• For each fosmid region, extract all mapped reads along with a local reference genome for variant calling.
• de novo assemble all reads in the fosmid region and call variants against the local reference genome using cortex_var’s Bubble Caller.
• Remap regional variant coordinates back to the full genome.
• Combine all regional calls into the final set of cortex_var calls.

Directly comparing GATK and cortex_var calls shows similar levels of concordance and discordance as the GATK/samtools comparison from the last post:

concordant: total	153787
concordant: SNPs	130913
concordant: indels	22874
GATK discordant: total	20495
GATK discordant: SNPs	6522
GATK discordant: indels	13973
cortex_var discordant: total	26790
cortex_var discordant: SNPs	21342
cortex_var discordant: indels	5448

11% of the GATK calls and 14% of the cortex_var calls are discordant. The one area where cortex_var does especially well is on indels: 19% of the cortex_var indels do not agree with GATK, in comparison with 37% of the GATK calls and 25% of the samtools calls. The current downside to this is SNP calling, where cortex_var has 3 times more discordant calls than GATK.

Selection of classification metrics

Overlapping variant calls from different calling methods (GATK, FreeBayes, samtools and cortex_var) and sequencing technologies (Illumina, SOLiD and IonTorrent) produces 170,286 potential calls in the fosmid regions. 58% (99,227) of these are present in all callers and technologies, so we need to do better than the intersection in creating a consolidated callset.

As detailed in the previous post, we filter the full set in two ways. The first is to keep trusted variants based on their presence in a defined number of technologies or callers. We currently have an inclusive set of criteria: keep variants present in either 4 out of the 7 callsets, 2 distinct technologies, or 3 distinct callers. This creates a trusted set containing 95% (162,202) of the variants. Longer term the goal is to reduce the trusted count and rely on automated filtering approaches based on input metrics.

This second automated filtering step uses a support vector machine (SVM) to evaluate the remaining variants. We train the SVM on potential true positives from variants that overlap in all callers and technologies, and potential false positives found uniquely in one single caller. The challenge is to find useful metrics associated with these training variants that will help provide discrimination.

In version 0.1 we filtered with a vanilla set of metrics: depth and variant quality score. To identify additional metrics, we used a great variant visualization tool developed by Keming Labs alongside HSPH and EdgeBio. I’ll write up more details about the tool once we have a demonstration website but the code is already available on GitHub.

To remove variants preferentially associated with poorly mapping or misaligned reads, we identified two useful metrics. ReadPosEndDist, written as a GATK annotation by Justin Zook at NIST, identifies variants primarily supported by calls at the ends of reads. Based on visual examination, these associate with difficult to map regions, as identified by Genome Mappability Scores:

Secondly, we identified problematic allele balances that differ from the expected ratios. For haploid fosmid calls, we expect 100% of reads to support variants and 0% to support reference calls (in diploid calls, you also need to handle heterozygotes with 50% expected allele balance). In practice, the distribution of reads can differ due to sequencer and alignment errors. We use a metric that measures deviation from the expected allele balance and associates closely with variant likelihoods:

Improved consolidated calls

To assess the influence of adding de novo calls and additional filtering metrics on the resulting call set, we compare against whole genome Illumina and Complete Genomics calls for NA19239. Previously we’d noticed two major issues during this comparison: a high percentage of discordant indel calls and a technology bias signaled by better concordance with Illumina than Complete.

The comparison between fosmid and Illumina data shows a substantial improvement in the indel bias. Previously 46% of the total indel calls were discordant, indicative of a potential false positive problem. With de novo calls and improved filtering, we’ve lowered this to only 10% of the total calls.

concordant: total	147684
concordant: SNPs	133861
concordant: indels	13823
fosmid discordant: total	7519
fosmid discordant: SNPs	5856
fosmid discordant: indels	1663
Illumina discordant: total	5640
Illumina discordant: SNPs	1843
Illumina discordant: indels	3797

This improvement comes with a decrease in the total number of concordant indel calls, since we moved from 17,816 calls to 13,823. However a large number of these seemed to be Illumina specific since 60% of the previous calls were discordant when compared with Complete Genomics. The new callset reduces this discrepancy and only 22% of the indels are now discordant:

concordant: total	139155
concordant: SNPs	127243
concordant: indels	11912
fosmid discordant: total	15484
fosmid discordant: SNPs	12028
fosmid discordant: indels	3456
Complete Genomics discordant: total	7311
Complete Genomics discordant: SNPs	4972
Complete Genomics discordant: indels	2273

These comparisons provide some nice confirmation that we’re moving in the right direction on filtering. As before, we extract potential false positives and false negatives to continue to refine the calls: potential false positives are those called in the fosmid dataset and in neither of the Illumina or Complete Genomics sets. Potential false negatives are calls that both Illumina and Complete agree on that the fosmid calls lack.

In the new callsets, there are 5,499 (3.5%) potential false positives and 1,422 (0.9%) potential false negatives. We’ve reduced potential false positives in the previous set from 10% with a slight increase in false negatives. These subsets are available along with the full callset on GenomeSpace. We’re also working hard on an NA12878 callset with equivalent approaches and will make that available soon for community feedback.

I hope this discussion, open source code, and dataset release is useful to everyone working on problems of improving variant calling accuracy and filtering. I welcome feedback on calling, consolidation methods, interesting metrics to explore, machine learning or any of the other topics discussed here.