GATK

 Best  Prac)ces  for  Variant  Discovery  

Introduc)on  to  Variant  Discovery  

All  you  need  to  know  to  get  started  

http://software.broadinstitute.org/gatk/
 Discover  “variants”  rela)ve  to  a  reference  genome  

• Gene)c  changes  in  individuals  rela%ve  to  a  reference  genome  

• Germline  (inherited)  
• Soma)c  (cancer)  

• Reference  genome  =  a  standardized  genomic  sequence  

• Human  genome  reference  sequence  

• Current  standard:  hg19  /  b37  
• New  standard  (on  the  rise):  hg38  

• Other  organisms    
• Many  have  a  fully  assembled  reference  available  
• Many  s)ll  do  not  -­‐>  SOL  
 Different  types  of  variants  

Germline  

Soma%c  

SNP/SNV   Indel   CNV/CNA   SV  

 Different  types  of  experimental  design  

Intergenic Exon I Intron I Exon II Intergenic

Whole    
Genome    
Variant site

Exome  
(+  gene  panels)  
Part  1  

BEST  PRACTICES  
 Data  genera)on  1:  Library  prep  
Varies  depending  on  experimental  design  
DNA  
[…]  
Fragments  
RT-­‐PCR  

RNA  

1) Extract  nucleic  acids  from  blood,  )ssue,  saliva  

2) RNA:  Make  cDNA  
3) Shear  dsDNA  into  fragments  
4) A`ach  fragments  to  adapters  (-­‐>  library)  

Library  prep  
 Data  genera)on  2:  Sequence  the  library  

Library  preps  

Bitesizebio.com  
Flowcell  
Enormous  pile  of  
short  reads  

Lanes  

Illumina  

HTS  machine  processes  a  flowcell  containing  lanes;    

each  lane  cons)tutes  a  read  group  (RG)    
(unless  mul)plexed)  
 Raw  sequence:  typically  in  FASTQ  format  

• Sequence  Name  (read  name,  group,  etc.)  

Accuracy  
• Sequence    
• +  (op)onal:  Sequence  name  again)  
20  and  higher  is    
• Associated  quality  score       generally  a  good  score  
 

Example  record  
 
• @EAS54_6_R1_2_1_413_324
• CCCTTCTTGTCTTCAGCGTTTCTCC Error  
• +
• ;;3;;;;;;;;;;;;7;;;;;;;88
Phred  value  =  −10  *  log10(ε)  
-­‐>  ASCII  code  translates  to  Phred-­‐scale  Q  scores  
 

90%  confidence  (10%  error  rate)  =  Q10  

99%  confidence  (1%  error  rate)  =  Q20  
99.9%  confidence  (.1%  error  rate)  =  Q30  
Format  specifica)on  h`p://maq.sourceforge.net/fastq.shtml  
Part  2  

ANALYSIS  WORKFLOWS  
 Data  pre-­‐processing  

Data   Variant   Callset    

Pre-­‐Processing   Discovery   Refinement  

FASTQ  -­‐>  BAM   BAM  -­‐>  VCF  

Step  1:  Map  the  reads  produced  by  the  sequencer  to  the  reference    

Mapping  and  
alignment  
algorithms  

• BWA  for  DNA  

• STAR  for  RNAseq  

Enormous  
pile  of  short  
reads  from  
HTS   Reference  genome  
Reads  
mapped  to  
reference  
 Output  format:  Sequence/Binary  Alignment  Map  (SAM/BAM)  

HEADER  containing  metadata  (sequence  dic)onary,  read  group  defini)ons  etc)  

RECORDS  containing  structured  read  informa)on  (1  line  per  read  record)  

read name position CIGAR read sequence metadata

SLX1:1:127:63:4 99 1 10052169 60 23M6N10M = 14 10 GAAGATACTGGTT 768832'48:::: SM:Z:JPTGBMN01 …

flags MAPQ mate quality scores

information

• Added  mapping  info  summarizes  posi%on,  quality,  and  structure  for  each  read  
 

h`p://samtools.github.io/hts-­‐specs/SAMv1.pdf  
 CIGAR  summarizes  alignment  structure  

CIGAR  =  Concise  Idiosyncra%c  Gapped  Alignment  Report  

read1 99 ref 2 30 3M1D2M1I1M = 14 20 CATCTAG *

 At  Broad:  Unmapped  BAM  instead  of  FASTQ  

Special  workflow  using  Picard  tools  for  improved  data  management    

Reference  genome  

BWA  MEM  
Raw  mapped  BAM   Mapped,  cleaned,    
sorted  BAM  

Unmapped  BAM   MergeBamAlignment  

Step  2:  Mark  duplicates  to  mi)gate  duplica)on  ar)facts  

Duplicates  =  non-­‐independent  measurements    

             of  a  sequence  fragment  
 
-­‐>  Must  be  removed  to  assess  support  for  alleles  correctly  

Reference  

Mapped  
reads  

Picard  MarkDuplicates  

=  sequencing  error  propagated  in  duplicates  

 Step  3:  Local  realignment  around  indels  corrects  mapping  errors  

Several  consecu%ve  SNPs  

BEFORE   only  found  on  reads  ending  
on  the  right  of  the  
homopolymer  

Several  consecu%ve  SNPs  

only  found  on  reads  ending  
on  the  le^  of  the  
homopolymer   7bp  homopolymer  run  

AFTER  

Adding  a  1-­‐bp  inser%on    

brings  sanity  to  the    
en%re  alignment  
Op%onal  with  assembly-­‐
based  variant  callers    
Step  4:  Base  Recalibra)on  (BQSR)  corrects  for  machine  errors

• Sequencers  make  systema)c  errors  in  base  quality  scores  

• BQSR  corrects  the  quality  scores  (not  the  bases)  

Example  of  bias:  quali)es  reported  depending  on  nucleo)de  context      

RMSE = 4.188 RMSE = 0.281

10

10
5

5
Empirical − Reported Quality

Empirical − Reported Quality

0

0
−5

−5
original   recalibrated  
−10

−10
AA AG CA CG GA GG TA TG AA AG CA CG GA GG TA TG

Dinuc Dinuc
Special handling for RNAseq splice junctions
 Variant  Discovery  

Data   Variant   Callset    

Pre-­‐Processing   Discovery   Refinement  

FASTQ  -­‐>  BAM   BAM  -­‐>  VCF  

 Best  Prac)ces  workflows  by  variant  type  

Germline  SNPs  &  Indels   Soma%c  SNVs  &  Indels   Soma%c  CNVs  
Discovery  of  germline  short  variants  is  done  on  cohorts  

• Single  genome  in  isola)on:  almost  never  useful  

• Family  or  popula)on  data    
add  valuable  informa)on  
– rarity  of  variants  
– de  novo  muta)ons  
– ethnic  background  
Visualiza)on  of  reads  at  a  probable  SNP  site  in  IGV  

Depth  of  coverage   Probable  C/T  SNP  

First  and  second  read  from  

the  same  fragment  

Non-­‐reference  bases  are  

colored;  reference  bases    
are  grey  
Individual  reads  
aligned  to  the  genome  
Reference  genome  
 Short  variants  are  reported  in  VCF:  Variant  Call  Format    

##fileformat=VCFv4.1
##reference=1000GenomesPilot-NCBI36
##INFO=<ID=DP,Number=1,Type=Integer,Description="Total Depth">
##INFO=<ID=AF,Number=A,Type=Float,Description="Allele Frequency">

Header  
##INFO=<ID=DB,Number=0,Type=Flag,Description="dbSNP membership">
##FILTER=<ID=s50,Description="Less than 50% of samples have data">
##FORMAT=<ID=GT,Number=1,Type=String,Description="Genotype">
##FORMAT=<ID=GQ,Number=1,Type=Integer,Description="Genotype Quality">
##FORMAT=<ID=DP,Number=1,Type=Integer,Description="Read Depth">
#CHROM POS ID REF ALT QUAL FILTER INFO FORMAT NA00001 NA00002 NA00003
20 14370 rs6054257 G A 29 PASS DP=14;AF=0.5 GT:GQ:DP 0/0:48:1 1/0:48:8 1/1:43:5

Records  
20 1230237 . T . 47 PASS DP=13 GT:GQ:DP 0/0:54:7 0/0:48:4 0/0:61:2
20 1234567 . GT G 50 PASS DP=9 GT:GQ:DP 0/1:35:4 0/2:17:2 1/1:40:3

Format  specifica)on  in  

h`ps://samtools.github.io/hts-­‐specs/VCFv4.2.pdf  
Joint  analysis  from  early  stages  empowers  discovery  

Individual  callsets   Joint  callset  

Underpowered  analysis   Empowered  analysis  

Discovery  is  empowered  at  difficult  sites  

• Sample  #1  or  Sample  #N  alone:  

• weak  evidence  for  variant  
• may  miss  calling  the  variant    

• Both  samples  seen  together:  

• unlikely  to  be  ar%fact  
• call  the  variant  more  
confidently  
 
 And  we  get  full  informa)on  at  all  sites  of  interest  

• Analyzed  individually:  
– No  call  for  either  sample  
– Very  different  reasons!  

SAMPLE  A  
• In  joint  analysis  with  other  
samples:  
– Hom-­‐ref  call  and  no-­‐call  
genotypes  emi`ed  

SAMPLE  B  
Tradi)onal  mul%-­‐sample  calling  approach  :  very  inefficient  

It  gives  us  the  right  answers,  but…  

Compute  requirements  
scale  very  badly  with  
number  of  samples!!!  
Want  to  add  new  samples?    
 
Got  to  re-­‐run  pipeline  from    
scratch!  The  N+1  problem!  
GVCF  workflow  solves  both  problems,  yields  same  results  

Compute  requirements  
scale  linearly  with    
number  of  samples   Want  to  add  a  new  sample?    
 
Just  call  it  by  itself  then  re-­‐genotype  the  
cohort  at  will!  
 Best  Prac)ces  workflows  by  variant  type  

Germline  SNPs  &  Indels   Soma%c  SNVs  &  Indels   Soma%c  CNVs  
Key  challenges  :  tumor  heterogeneity  and  contamina)on  

normal  
     
dysplas%c   *      
cancer  
**"
**  
     
*  
     
*    

Tumor  Sample  
Tissue-­‐adjacent  Normal   Blood  Normal  

Adapted  from  h`ps://science.educa)on.nih.gov/supplements/nih1/cancer/guide/understanding1.html  

Amount  of  signal  may  be  comparable  to  noise  

Expecta%on  for     Expecta%on  for    

germline  variants   soma%c  variants  

signal  

signal  
heterogeneity  

contamina%on  

noise   noise  

+  AF  expected  to  follow  ploidy   +  no  reliance  on  ploidy  for  AF  
 Best  Prac)ces  workflows  by  variant  type  

Germline  SNPs  &  Indels   Soma%c  SNVs  &  Indels   Soma%c  CNVs  
Copy  number:  it’s  all  about  coverage  

Collect  propor%onal  coverage     Normalize  to  remove  noise  

Iden%fy  segment  boundaries  

 Callset  Refinement  

Data   Variant   Callset    

Pre-­‐Processing   Discovery   Refinement  

FASTQ  -­‐>  BAM   BAM  -­‐>  VCF  

PROJECT    
DEPENDENT  
h`ps://sowware.broadins)tute.org/gatk/best-­‐prac)ces/  
Part  3  

PIPELINES  AT  BROAD  

 Pipelines  implemented  on  local  systems  at  Broad  

“The  Picard  Pipeline”  

Reads  

Implementa)on  

BAMs  

GATK  Best  Prac)ces  

Pre-­‐processing  

Germline  
SNPs  &  Indels  
“FireHose”  
 Pipelines  implemented  on  Google  Cloud  

“Genomes  On  The  Cloud”   “FireCloud  /  Workbench”  

Reads  

BAMs  

WGS  Germline  
SNPs  &  Indels   hkps://so^ware.broadins%tute.org/firecloud/  
GOTC  WDL  script  shared  at    
h`ps://github.com/broadins)tute/wdl/tree/develop/scripts/broad_pipelines  
 Write  your  own  pipelines  in  WDL!  

Tired  of  these  op=ons  for  wri=ng  pipelines?   Meet  the  Workflow  Descrip=on  Language  

Finally  a  workflow  language  meant  to  be  read    

(low-­‐tech)   and  wriken  by  humans  
Plain  Old  Scripts  

WDL  makes  it  easy  to:    

-­‐ Describe  analysis  tasks  
-­‐ Chain  tasks  into  workflows  
-­‐ Specify  advanced  behaviors  like  
paralleliza)on  
(high-­‐tech)  
Domain-­‐Specific  Languages  
hkps://so^ware.broadins%tute.org/wdl/