AGBT 2010 - Kristian Cibulskis - Broad Institute

ITector: Accurate Somatic Mutation Detection in Whole Genome and Exome Capture

Mutation Detection - the goal

* Somatic Point Mutations: SNV in the tumour DNA that are not present in the normal

One challenge: Sensitivity

* tumour purity : Normal tissue gets into the sample - you may be testing normal tissue in high quantities

* ploidy: often there are multiple copies of the DNA

60% tumour is common - with 3x ploidy and min allele fraction: 0.23

Challenge 2: Specificity

* Signal: 1 somatic mutation per Mb

* Noise: 1000 common germline varients per Mb (in dbsnp)

* Mutations are not recurrent. (Constant discovery mode)

* 1000s mutations per sample, 100s of samples

* Too expensive to validate every mutation - would cost more than to discover.

MuTector:

* Core detection algorithm and practical artifact filters

* Under dev since Nov 2008

* Built upon GATK

Some artifacts can be cleaned up globally

* Remove molecular Duplicates

* Recalibrate Quality Scores (make Q values match)

* Locally Realign [Gapped - uses SW - I saw the poster]

Core Statistical Test

* Prior genotype probablities enforce variant expectation rate..

* first calculate score for non-reference (for tumor)

* then calculate scover for it being reference (for normal)

* Controling sequencing error

* Controlling missing a germline ref in the normal.

Running: you get more somatic mutations

* expected 30 somatic mutations, ended up with 133 in 30mb of coding sequence

* Error processes not captured by the core statistic produce high confidence mistakes

* Information about reference alleles and mutatn alleles should come from similar distributions

* linked mutations, library errors... etc

Filters:

* Sequence context causes base hallucinations

* Fisher's exact test to check distribution of strand of reads containing reference allele versus alternate allele

* Bigger effect in capture than whole genome

Misalignment:

* Sequencers/Aligners tend to make reproduceable errors, which then show up in alignments

Small changes to filters have big effects

* Very sensitive!

Filtering goes from 133 to 35.

Validation:

* 26/29, 30/35, 31/36, 92/100

* Around 95%

How Sensitive?

* use core statistics.

* depends on coverage! [of course]

* use theoretical prediction data and ultra deep coverage as "control"

* Both seem to give the same/similar results

* Average 60-80% power to detect

Beta Testing going on

* Release of the software will be soon!