• What factors complicate somatic SNV calling and why is additional filtering necessary?

• CaVEMan filters

• Metric-based approach to filtering false positives

• Other approaches to filtering (panel of normals, ensemble calling)

• Benchmarks for assessing SNV calling and filtering

• Low cellularity (tumour DNA content)

• Intra-tumour heterogeneity in which multiple tumour cell populations (subclones) exist

• Aneuploidy

• Unbalanced structural variation (deletions, duplications, etc.)

• Matched normal contaminated with cancer DNA

  • adjacent normal tissue may contain residual disease or early tumour-initiating somatic mutations

  • circulating tumour DNA in blood normals

• Sequencing errors

• Alignment artefacts

Mwenifumbo & Marra, Nat Rev Genet. 2013

• Base qualities drop off toward ends of reads on Illumina sequencing platform, errors in base calls more likely

Filter: minimum base quality for variant alleles [CaVEMan --min-base-qual parameter]

Filter: no variant alleles found in first 2/3 of a read [CaVEMan RP filter]

• Duplicate reads

  • PCR amplification during library construction introduces duplicate reads

  • Bioinformatic tools identify likely PCR duplicates based on aligned start positions of both ends of a read pair

Post-alignment: mark duplicate reads and exclude from SNV calling

• Strand bias

  • Calls supported mainly by reads aligning to one strand may also be PCR artefacts

Filter: variant alleles in reads aligning in one direction only [SE filter]

• GC bias

  • Stretches of low GC content tend to be under-represented leading to uneven coverage across the genome

  • Poor confidence in calls at these low coverage regions as germline SNVs more likely to be mistaken for somatic SNVs due to inadequate sampling in the matched normal

Filter: minimum read depth at variant position particularly in the normal

Filter: minimum number of reads supporting variant call

Alignment issues are common source of false positive SNV calls

• Missing sequence in the reference genome causes misalignments, usually with mismatches

Alignment: use decoy sequence included in reference genome used by 1000 Genomes project

Filter: variants supported by reads with many mismatches

• Assembly in region around variant may differ from the the reference sequence causing incorrect alignments, e.g. indels, rearrangements

Post-alignment: Perform local realignment around indels [GATK IndelRealigner]

Filter: variants within or close to germline indels [GI filter]

Filter: variant position always toward beginning/end of alignment [RP filter]

• Repetitive regions in the genome cause difficulties for alignment tools

  • Aligners assign a mapping quality of 0 if they cannot uniquely place a read

  • Regions of low mappability are usually out of bounds for short read sequencing

Filter: minimum mapping quality of variant reads [MQ filter]

Filter: calls from low-complexity and low-mappability regions [SR, CR filters]

HCC1143 dataset: 138042 of 159001 (87%) of initial SNV calls made by CaVEMan are filtered

Hard filters

• Based on summary statistics, or metrics, computed from the sequence reads covering the variant position, e.g. average mapping or base quality scores

Ensemble calling

• Consensus call set from majority voting on candidate SNVs called by multiple callers

Blacklisting

• Exclude list of problematic genomic positions and/or substitutions, e.g. based on panel of normals

Machine learning techniques

• Computers can potentially be trained to recognize characteristics that distinguish true variants from erroneous calls [GATK Variant Quality Score Recalibration, mutationSeq]

• Benchmark datasets can be used to tune and assess filters

• Ideally need to test filters on separate dataset to that used for training

• Danger of overfitting

• Approach:

  • Plot distribution of true and false positive variants for variety of metrics

  • Choose threshold to best distinguish between TP and FP

Here we use the ICGC medulloblastoma benchmark dataset to derive filters and the ICGC-TCGA DREAM Challenge sythentic dataset 4 to test these filters.

ICGC benchmarking exercise

• Medulloblastoma tumour/normal pair sequenced in 6 different centres to combined 300-fold coverage used to establish 'truth'

• 16 ICGC project teams ran their pipelines on data from one centre (40x)

Alioto et al., Nat Commun. 2015

ICGC-TCGA DREAM Somatic Mutation Calling challenge

• 6 synthetic datasets based on cell line sequenced to 80x, BAM randomnly split into 2 ('tumour' and 'normal'), mutations added to one computationally

• Synthetic dataset 4: 80% cellularity; 50% and 35% subclone VAF (effectively 30% and 15%)

Ewing et al., Nat Methods 2015 [leaderboards]

• CalculateSNVMetrics tool available in CRUK-CI gatk-tools package

• VariantFiltration tool in the Genome Analysis Toolkit (GATK)

# filter set 1 (complements MuTect2's own in-built hard filters)

java -jar GenomeAnalysisTK.jar \
  --analysis_type VariantFiltration \
  --reference_sequence reference.fasta \
  --variant input.vcf \
  --out output.vcf \
  --filterName VariantAlleleCount    --filterExpression "VariantAlleleCount < 3" \
  --filterName VariantCountControl   --filterExpression "VariantAlleleCountControl > 1" \
  --filterName VariantBaseQualMedian --filterExpression "VariantBaseQualMedian < 25.0" \
  --filterName VariantMapQualMedian  --filterExpression "VariantMapQualMedian < 40.0" \
  --filterName MapQualDiffMedian     --filterExpression "MapQualDiffMedian < -5.0 || MapQualDiffMedian > 5.0" \
  --filterName LowMapQual            --filterExpression "LowMapQual > 0.05"

SNVs called by MuTect2 (12432 true, 665 false, 3836 not called)

• Artefacts usually cancel out in the tumour normal comparison but depends on adequate sampling

  • Low depth in normal can cause germline variants to appear as somatic

• An approach to detecting likely artefacts is to look for the variant in a panel of unrelated normal samples

  • Filters both polymorphisms and locations prone to aberrant mapping or systematic sequencing artefacts

cgpCaVEManWrapper [Jones et al., 2016]

CRUK-CI blacklist

• Based on 50x sequence data for 149 blood normal samples from the UK oesophageal cancer ICGC project

• Variants appearing in at least 5 normals (minimum 3 reads, 5% allele fraction in each sample)

Can combining the results from multiple callers improve accuracy?

Ensemble calling

• Majority voting on candidate SNVs from multiple callers to produce consensus call set

• ICGC PanCancer project using SNV calls made by at least 2 out of 4 callers (CaVEMan, MuTect2, MuSE, samtools)

• bcbio cancer variant calling pipeline – also see Brad Chapman's blog

• SomaticSeq [Fang et al, 2014]

• SNV calling in cancer genomes is difficult for many reasons

• Using an SNV caller out-of-the-box may give a reasonable set of calls but is likely to result in call sets with higher sensitivity at the expense of precision

• Simple filtering strategies can improve precision but there is a trade-off between sensitivity and accuracy

• The cancer genome sequencing community has been active in establishing benchmark datasets that can be used to assess and improve somatic mutation calling pipelines