TOWARDS PRECISION MEDICINE:a cloud-based application for analysis of personal genomes

Reid J. Robison, MD MBA December 6th, 2013

$3 BILLION

$2000

THE FALLING COST of sequencing the human genome

THE SEQUENCING EXPLOSION

Boycott et al. Rare-disease genetics in the era of next-generation sequencing: discovery to translation. Nature Reviews Genetics 14, 681–691 (2013)

PACE OF DISCOVERY of novel rare-disease-causing genes

using whole-exome sequencing

35

70

105

140

2009 2010 2011 2012

40

80

120

160

2000 2005 2010 2015 2020 2025 2030

RATE OF APPROVAL of rare disease drug products

Extrapolation from January 2013 Orphanet Report Series

“We are on the tipping point of a whole new game in how we develop drugs.”

Janet Woodcock, M.D. Director, Center for Drug Evaluation and Research, FDA

GENE-FINDING

Patient w/ unknown disease

Next-gen sequencing

Data processing

Variant calling

3 million SNVs 0.5 million indels

1000 SVs

????????????????????????????????

STEPWISE REDUCTION

ANNOVAR 450 citations

>40,000 downloads

Wang K et al. ANNOVAR: Functional annotation of genetic variants from next-generation sequencing data Nucleic Acids Research, 38:e164, 2010

25523 variants

6423 variants

2935 variants

2652 variants

421 variants

17 genes

10 genes

Only non-synonymous or frameshift

Conserved variants from 44-species alignment

Remove variants in segmental duplication regions

Remove variants with MAF>1%

Apply recessive model

Remove “dispensable” genesLiterature survey identifies PKLR as candidate gene (confirmed with biochemical assay)

ADHD and anemia?

①

Genome browser shot of the PKLR gene and the location of the two causal mutations. Each of the two mutations sits within an evolutionarily conserved region, and has been reported once in patients affected with PKLR deficiency.

3910156 variants

21488 variants

10380 variants

1146 variants

935 variants

582 variants

52 genes

Keep only exonic/splicing variants

Remove synonymous & non-synonymous frameshift variants

Remove variants in 1000 genomes project

Remove variants in ESP6500 database

Remove variants in dbSNP135

Keep only genes with multiple variants

!

BOOKMAN syndrome

②

52 genes

!2 genes left:

TAF1L RBCK1!

RanBP-type and C3HC4-type zinc finger containing 1 (Mutation results in splicing error)

Remove psuedogenes & questionable calls

Remove olfactory receptor genes

Sanger sequencing validation

OGDEN SYNDROMEClinical Features Two Families

!

OGDEN syndrome

③3441 variants on X chromosome

2381 variants

136 variants

40 variants

40 variants

1 variant

NAA10

Keep only heterozygous

Keep only Stop/NonSyn/FS/Splice

Remove variants in dbSNP

Remove variants in ClinSeq

Keep only variants in shared haplotype

(Encodes the catalytic subunit of the major human N-terminal acetyltransferase)

VARIANT annotation & interpretation

• >60 annotation types (SIFT, PolyPhen, Allele Freq, HGMD…) • User-driven filtering for step-wise reduction. Run gene panels. • Robust, scalable, secure. Run case-control & family-based analyses • Machine-learning algorithms to generate TUTE score for prioritization

of variants

THE TUTE SCORE using machine-learning to prioritize disease genes

① Select a set of functional prediction scores for which coding and non-coding variants can be assigned into

② Built SVM prediction models using SVMsensus

③ Identify the optimal hyperplane for the biggest margin between training points for neutral and deleterious variants & genes

④ Test & refine prediction model using known disease variants from UniProt & synthetic data sets

COMING SOON: more accurate variant calling

Toward  more  accurate  variant  calling  for  “personal  genomes”

Jason  O’Rawe1,2, Tao Jiang3, Guangqing Sun3, Yiyang Wu1,2, Wei Wang4, Jingchu Hu3, Paul Bodily5, Lifeng Tian6, Hakon Hakonarson6, W. Evan Johnson7, Reid J. Robison9, Zhi Wei4, Kai Wang8,9, Gholson J. Lyon1,2,9

Background To facilitate the clinical implementation of genomic medicine by next-generation sequencing, it will be critically important to obtain accurate and consistent variant calls on personal genomes. Multiple software tools for variant calling are available, but it is unclear how comparable these tools are or what their relative merits in real-world scenarios might be. Under conditions where “perfect” pipeline parameterization is un-attainable, researchers and clinicians stand to benefit from a greater understanding of the variability introduced into human genetic variation discovery when utilizing many different bioinformatics pipelines or different sequencing platforms.

1) Stanley Institute for Cognitive Genomics, Cold Spring Harbor Laboratory, NY, USA; 2) Stony Brook University, Stony Brook, NY, USA; 3) BGI-Shenzhen, Shenzhen, China; 4) Department of Computer Science, New Jersey Institute of Technology, Newark, NJ, USA; 5) Department of Computer Science, Brigham Young University, Provo, UT, USA; 6) Center for Applied Genomics, Children’s Hospital of Philadelphia, Philadelphia, PA, USA; 7) Department of Medicine, Boston University School of Medicine, Boston MA, USA; 8) Zilkha Neurogenetic Institute, Department of Psychiatry and Preventive Medicine, University of Southern California, Los Angeles, CA, USA; 9) Utah Foundation for Biomedical Research, Salt Lake City, UT, USA.

References 1. Li, H. & Durbin, R. Fast and accurate short read alignment with Burrows-Wheeler transform. Bioinformatics 25, 1754-1760, doi:btp324 [pii]10.1093/bioinformatics/btp324 (2009). 2. DePristo, M. A. et al. A framework for variation discovery and genotyping using next-generation DNA sequencing data. Nature genetics 43, 491-498, doi:10.1038/ng.806 (2011). 3. Li, H. et al. The Sequence Alignment/Map format and SAMtools. Bioinformatics 25, 2078-2079, doi:btp352 [pii]10.1093/bioinformatics/btp352 (2009). 4. Li, R. et al. SOAP2: an improved ultrafast tool for short read alignment. Bioinformatics 25, 1966-1967 (2009). 5. Li, R. et al. SNP detection for massively parallel whole-genome resequencing. Genome Res 19, 1124-1132, (2009). 6. Li, R. et al. De novo assembly of human genomes with massively parallel short read sequencing. Genome Res 20, 265-272, doi:gr 097261.109 [pii] 7. Clement, N. L. et al. The GNUMAP algorithm: unbiased probabilistic mapping of oligonucleotides from next-generation sequencing. Bioinformatics 26, 38-45, doi:btp614 [pii]10.1093/bioinformatics/btp614 (2010). 8. Wei, Z., Wang, W., Hu, P., Lyon, G. J. & Hakonarson, H. SNVer: a statistical tool for variant calling in analysis of pooled or individual next-generation sequencing data. Nucleic acids research 39, e132, doi:10.1093/nar/gkr599 (2011).

A) SNV concordance was measured between all SNV calls made by the five illumina data pipelines. Overall concordance is low: 57.4%.

B) SNV concordance is higher for already described variation (present in

dbSNP135). C) SNV concordance is lower for novel, un-described, human genetic

variation (absent in dbSNP135).

• All Illumina exomes have at least 20 reads or more per base pair in >80% or more of the 44 MB target region.

• Concordance rates with common SNPs genotyped on Illumina 610K genotyping chips were calculated.

• All pipelines are very good with

identifying already known, common SNPs.

Sample Software Compared Sites Concordance Sites

Concordance rate

Mother-1 SOAPsnp 6088 6074 99.77% GATK 6249 6224 99.60% SNVer 5723 5708 99.74% GNUMAP 5458 5434 99.56% SAMTools 5885 5848 99.37%

Son-1 SOAPsnp 6366 6353 99.80% GATK 6341 6323 99.72% SNVer 6255 6239 99.74% GNUMAP 5850 5828 99.62% SAMTools 6383 6362 99.67%

Son-2 SOAPsnp 6412 6401 99.83% GATK 6426 6413 99.80% SNVer 6336 6325 99.83% GNUMAP 5906 5889 99.71% SAMTools 6477 6450 99.58%

Father-1 SOAPsnp 6247 6238 99.86% GATK 6304 6288 99.75% SNVer 6205 6192 99.79% GNUMAP 5805 5786 99.67% SAMTools 6344 6327 99.73%

• Sensitivities and specificities were calculated for each pipeline using the Illumina 610k genotyping chips as a golden standard.

• All pipelines show relatively high sensitivity and specificity

when detecting known and common SNPS. • Specificity generally increases for sets of variants detected by

more than a single pipeline.

Specificity Sensitivity Known SNPs Novel SNPs

Mean* SD Mean* SD #Total #cSNP Ti/Tv #Total #cSNP Ti/Tv

SOAPsnp 99.82 0.039 94.53 2.287 30,022 17,409 2.77 875 419 1.94

GATK 99.72 0.085 95.33 1.161 29,620 17,306 2.8 365 206 2.34

SNVer 99.78 0.044 92.32 4.339 28,242 17,111 2.85 490 253 2.52

GNUMAP 99.64 0.065 86.67 3.286 24,893 15,144 3.03 1,091 659 1.28

SAMTools 99.59 0.158 94.45 4.221 29,577 17,449 2.78 949 539 1.33

ANY pipeline 99.62 0.113 97.72 1.215 33,947 19,638 2.68 2,163 1,182 1.23

>=2 pipelines 99.69 0.074 96.68 2.298 31,099 18,108 2.77 639 323 2.17

>=3 pipelines 99.73 0.045 95.65 3.143 29,363 17,257 2.84 416 230 2.56

>=4 pipelines 99.82 0.041 92.63 3.412 26,772 16,097 2.91 318 193 2.67

5 pipelines 99.87 0.015 80.61 5.266 21,174 13,320 3.12 234 149 2.83

Methods We sequenced 15 exomes from four families using the Illumina HiSeq 2000 platform and Agilent SureSelect v.2 capture kit, with ~120X coverage on average. We analyzed the raw data using near-default parameters with 5 different alignment and variant calling pipelines (SOAP, BWA-GATK, BWA-SNVer, GNUMAP, and BWA-SAMTools). We additionally sequenced a single whole genome using the Complete Genomics (CG) sequencing and analysis pipeline (v2.0), with 95% of the exome region being covered by 20 or more reads per base. Finally, we attempted to validate 919 SNVs and 841 indels, including similar fractions of GATK-only, SOAP-only, and shared calls, on the MiSeq platform by amplicon sequencing with ~5000X average coverage.

Results SNV concordance between five Illumina pipelines across all 15 exomes is 57.4%, while 0.5-5.1% variants were called as unique to each pipeline. Indel concordance is only 26.8% between three indel calling pipelines, even after left-normalizing and intervalizing genomic coordinates by 20 base pairs. 2085 CG v2.0 variants that fall within targeted regions in exome sequencing were not called by any of the Illumina-based exome analysis pipelines, likely due to poor capture efficiency in those regions. Based on targeted amplicon sequencing on the MiSeq platform, 97.1%, 60.2% and 99.1% of the GATK(v.15)-only, SOAPsnp(v1.03)-only and shared SNVs can be validated, yet 54.0%, 44.6% and 78.1% of the GATK-only, SOAP-only and shared indels can be validated.

• SNP concordance between the illumina data calls and the Complete Genomics v2.0 data calls was calculated for a single sample, “k8101-49685”.

• There are 2085 SNVs that Complete Genomics v2.0 detected but are not detected by any of the

five Illumina data pipelines, despite high mappability among these variants.

C A B

C

A

B

• Indel concordance between the three indel calling Illumina data pipelines (A) is low, 26.8%.

• Concordance is much better for known indels (B), and conversely much lower for novel, unknown, indels (C) (as defined by presence or absence in dbSNP135).

• MiSeq validation was performed on a combination of SNPs and indels chosen (1756 in total) from sequencing data from the sample “k8101-49685”.

• SNVs that were uniquely called by the SOAP-SNP v.1.03/Soap indel v2.01 and GATK v1.5 pipeline validated relatively well, with the SNVs called by both pipelines being better validated.

• Indels validated poorly for both unique to GATK(v.1.5) and SOAPindel (v2.01) calls. Overlapping indel calls validated better, though still relatively poorly.

Conclusions

We have shown that there remains significant discrepancy in SNV and indel calling between many of the currently available variant calling pipelines when applied to the same set of Illumina sequence data under near-default software parameterizations, thus demonstrating fundamental, methodological, variation between these commonly used bioinformatics pipelines. In spite of this inter-methodological variation, there exists a set of robust calls that are shared between all pipelines even under lax parameterization. However, the false negative rate is relatively high, and we agree that sequencing and analyzing samples with multiple platforms and methodologies is needed to attain a high accuracy “personal genome”.

The similarity between SNV and indel calls made between two versions of GATK, v1.5 and v2.3-9, was measured. SNV and indel calls were made using both the UnifiedGenotyper and HaplotypeCaller modules on the same k8101-49685 participant sample.

• 15 exomes from 4 families • 1 whole genome from Complete Genomics • Illumina HiSeq platform and Agilent SureSelect capture kit • 120X mean coverage • Five NGS alignment+variant calling pipelines are tested (SOAP,

BWA-GATK, BWA-SNVer, GNUMAP, and BWA-SAMtools) • Illumina 610k SNP array used as gold standard • ~60% SNVs are called by all five pipelines • 0.5 to 5.1% of variants were called as unique to each pipeline. !

#variantcallingproblems

Specificity Sensitivity Known SNPs Novel SNPs

Mean* SD Mean* SD #Total #cSNP Ti/Tv #Total #cSNP Ti/Tv

SOAPsnp 99.82 0.039 94.53 2.287 30,022 17,409 2.77 875 419 1.94

GATK1.5 99.72 0.085 95.33 1.161 29,620 17,306 2.8 365 206 2.34

SNVer 99.78 0.044 92.32 4.339 28,242 17,111 2.85 490 253 2.52

GNUMAP 99.64 0.065 86.67 3.286 24,893 15,144 3.03 1,091 659 1.28

SAMTools 99.59 0.158 94.45 4.221 29,577 17,449 2.78 949 539 1.33

ANY pipeline 99.62 0.113 97.72 1.215 33,947 19,638 2.68 2,163 1,182 1.23

>=2 pipelines 99.69 0.074 96.68 2.298 31,099 18,108 2.77 639 323 2.17

>=3 pipelines 99.73 0.045 95.65 3.143 29,363 17,257 2.84 416 230 2.56

>=4 pipelines 99.82 0.041 92.63 3.412 26,772 16,097 2.91 318 193 2.67

5 pipelines 99.87 0.015 80.61 5.266 21,174 13,320 3.12 234 149 2.83

• All pipelines are ‘good’ with known, common SNPs • Specificity increases for variants detected by more than one

pipeline • Indel concordance between all 3 platforms was very low

(26.8%) • Complete Genomics picked up >2000 variants that weren’t

seen on Illumina, despite high mappability in these regions

VARIANT CALLING CONCLUSIONS ① Significant discrepancy among all pipelines when applied to

the same Illumina datasets ② There exists a set of robust calls that are shared among all

pipelines even under lax parameters (although false negative rate is high)

To get an accurate genome, you need to run multiple algorithms.

A results portal that lets doctors, labs & researchers give their patients access to important genetic findings

Reid J. Robison, MD MBA reid@tutegenomics.com