Analyses of Indels

From Genome Analysis Wiki
Jump to navigationJump to search


This wiki page details some standard Indel analyses which hopefully can help the group in understanding the issues and perform the analyses quickly without reinventing the wheel.


This walkthrough requires vt.


The file generated from the indel calling is a binary version [BCFv2.1] of the Variant Call Format (VCF). BCFv2.1 is more efficient to process as the data is already stored in computer readable format on the hard disk. It is however not necessarily more compact than VCF4.2 especially when the format fields are rich in details.

File Preparation

The VCF file you work with should preferably be BCF2.1 compatible. Here we provide an example in /net/fantasia/home/atks/indel_analysis_tutorial. \\

To convert to BCF format which will work fast with vt:

  vt view mills.vcf -o mills.bcf

You will encounter an error as the header does not contain contigs. To fix this, you should construct a complete header for mills.vcf. This is done for you in mills.with.alt.hdr.*

  vt view mills.with.alt.hdr.vcf -o mills.genotypes.bcf

To index:

 vt index mills.genotypes.bcf

To extract just the site list which is convenient for working with if you are not analysing the genotypes of the individuals

 vt view -s mills.genotypes.bcf -o mills.sites.bcf

To index:

 vt index mills.sites.bcf

You may also work with vcf.gz, just name the output as *.vcf.gz. But it will be slower with vt.


You can see what you have in the file with:

 vt peek mills.genotypes.bcf

You can also focus on a chromosome:

 vt peek mills.genotypes.bcf -i 20

Or with just passed variants:

 vt peek mills.genotypes.bcf -i 20 -f PASS

Or with failed variants:

 vt peek mills.genotypes.bcf -i 20 -f ~PASS

Or with just 1bp indels:

 vt peek mills.genotypes.bcf -i 20 -f "PASS&&DLEN==1"

Or with just 1bp deletions:

 vt peek mills.genotypes.bcf -i 20 -f "PASS&&LEN==-1"

Or with just biallelic 1bp indels:

 vt peek mills.genotypes.bcf -i 20 -f "PASS&&N_ALLELE==2&&LEN==1"

Or with just biallelic 1bp indels that are somewhat rare:

 vt peek mills.sites.bcf -f "PASS&&N_ALLELE==2&&LEN==1&&INFO.AF<0.03"

Or with just biallelic 1bp indels that are somewhat rare with sanity checking:

 vt peek mills.sites.bcf -f "PASS&&N_ALLELE==2&&LEN==1&&INFO.AC/INFO.AN<0.03"

which you will observe discrepancies due to rounding off in AF. So you should probably use INFO.AC/INFO.AN.


Indel representation is not unique, you should normalize them and remove duplicates.

  Variant normalization is implemented in vt and this page explains the algorithm 
  and also provides a simple proof of correctness - Variant Normalization

The following table shows the number of variants that had to be normalized and the corresponding type of normalization performed and the ensuing number of duplicate variants found for some of the 1000 Genomes Trio High Coverage call sets. Although left alignment seems to be a trivial concept, it is easily overlooked and remain a common mistake.

Dataset Freebayes Haplotyecaller PINDEL Platypus RTG Samtools SGA
Left trim 27069 1 0 0 0 0 15047
Left aligned 3 1 1 0 12262 2 1892
Left trim 40782 0 0 0 374
Left aligned 1892 0 0 0 1329 1 0
Right trimmed 0 0 0 25393 0 11 0
Duplicate variants 0 1 155 3143 286 8 7541

Another example is the Mills et al. data set which followed up with 10004 Indels for validation. Out of 9996 passed variants, it was found that after normalization, only 8904 distinct Indels remain - about a loss of 11% of variant thought distinct.

To normalize and remove duplicate variants:

 vt normalize  mills.genotypes.bcf -r ~/ref/vt/grch37/hs37d5.fa  | vt mergedups - -o mills.normalized.genotypes.bcf 

and you will observe that 3994 variants had to be left aligned and 1092 variants were removed.

 stats: biallelic
         no. left trimmed                      : 0
         no. left trimmed and left aligned     : 0
         no. left trimmed and right trimmed    : 0
         no. left aligned                      : 3994
         no. right trimmed                     : 0 
multiallelic no. left trimmed : 0 no. left trimmed and left aligned : 0 no. left trimmed and right trimmed : 0 no. left aligned : 0 no. right trimmed : 0
no. variants observed : 9996
Time elapsed: 0.14s

stats: Total number of observed variants 9996 Total number of unique variants 8904
Time elapsed: 0.13s

The following will be slight faster: + denotes using of uncompressed bcf stream.

 vt normalize  mills.genotypes.bcf -r ~/ref/vt/grch37/hs37d5.fa -o + | vt mergedups + -o mills.normalized.genotypes.bcf

Also remember to index this file and extract the sites.

Insertion/Deletion ratios, Coding Regions and Overlap analysis

You can obtain measure of insertion deletion ratios, coding region indels and sensitivity analysis by using the profile_indels analysis.

 vt profile_indels -g indel.reference.txt -r ~/ref/vt/grch37/hs37d5.fa mills.normalized.sites.bcf

The indel.reference.txt file contains the required reference to perform the overlap analysis.

 data set
   No Indels :       8904 [0.93]   //#variants in your data set [ins/del ratio]
      FS/NFS :       0.66 (67/35)  //Proportion of frameshift Indels. (#Frameshift Indels/#Nonframeshift Indels)
dbsnp //A represents the data set you input, B represents dbsnp A-B 2975 [1.06] //#variants in A only [ins/del ratio] A&B 5929 [0.86] //#variants in A and B B-A 2059845 [1.51] Precision 66.6% //A&B/A this represents how novel your data set is in the variants represented. Sensitivity 0.3% //A&B/B this represents sensitivity somewhat if dbsnp is considered a high quality Indel //set and the sample are the same in both data sets. (which they usually are not, this is still //nonetheless a useful indicator)
mills A-B 5705 [0.81] A&B 3199 [1.18] B-A 203819 [0.98] Precision 35.9% Sensitivity 1.5%
mills.chip A-B 0 [-nan] A&B 8904 [0.93] B-A 0 [-nan] Precision 100.0% Sensitivity 100.0%
affy.exome.chip A-B 8821 [0.93] A&B 83 [0.69] B-A 34011 [0.47] Precision 0.9% Sensitivity 0.2%

Ins/Del ratios: Reference alignment based methods tend to be biased towards the detection of deletions. This provides a useful measure for discovery Indel sets to show the varying degree of biasness. It also appears that as coverage increases, the ins/del ratio tends to 1.

Coding region analysis: Coding region Indels may be categorised as Frame shift Indels and Non frameshift Indels. A lower proportion of Frameshift Indels may indicate a better quality data set but this depends also on the individuals sequenced.

Overlap analysis: overlap analysis with other data sets is an indicator of sensitivity.

  • dbsnp: contains Indels submitted from everywhere, I am not sure what does this represent exactly. But assuming most are real, then precision is a useful estimated quantity from this reference data set.
  • Mills: contains doublehit common indels from the Mills. et al paper and is a relatively good measure of sensitivity for common variants. Because not all Indels in this set is expected to be present in your sample, this actually gives you an underestimate of sensitivity.
  • Mills chip: This is a subset of the Mills data set. There are genotypes here that are useful for subsetting polymophic subsets of variants that are present in samples common with your data set, this can potentially provide a better estimate of sensitivity. In general not very useful unless you happen to be working on 1000 Genomes data or any data set who's individuals are commonly studied.
  • Affy Exome Chip: This contains somewhat rare variants in exonic regions and is useful for exome chip analysis. You should subset your exome data to exome region Indels before comparing against this data set.

This analysis supports filters too.

to document

  • Annotation of STRs is really important. Show example of a deceptive single base pair variant
  • Mendelian analysis
  • AFS
  • Can check concordance of genotypes between callers - partitiion
  • Type of Indels - homopolymer types and STR types and isolated, Adjacent SNPs ,Adjacent MNPs,Clumping variants
  • genotype likelihood concordance
  • concordance stratified by indel length or tract length
  • mendelian concordance by tract length