Here are a few more, a summary of the other answers, and updated links:
For a much more general discussion of variant calling (not necessarily somatic or limited to SNVs/InDels) check out this thread: What Methods Do You Use For In/Del/Snp Calling?
Some papers describing comparisons of these callers:
The ICGC-TCGA DREAM Mutation Calling challenge has a component on somatic SNV calling.
This paper used validation data to compare popular somatic SNV callers:
You'll need to update the link to MuTect. Broad Institute has begun to put portable versions of their tools on Github, like thelatest release of MuTect. The Genome Institute at WashU has been using Github for a while, but portable versions of their tools can be found here and here.
To rehash/expand on what Dan said, if you're sequencing normal tissue, you generally expect to see single-nucleotide variant sites fall into one of three bins: 0%, 50%, or 100%, depending on whether they're heterozygous or homozygous.
With tumors, you have to deal with a whole host of other factors:
These, and other factors, make calling somatic variants difficult and still an area that is being heavily researched. If someone tells you that somatic variant calling is a solved problem, they probably have never tried to call somatic variants.
Sounds like somatic / tumor variant calling is something that will be solved by improvements at the wet lab side ( single cell selection / amplification / sequencing ) . Rather than at the computational side.
Well, single cell has a role to play (and would have more of one if WGA wasn't so lossy), but realistically, you can't sequence billions of cells from a tumor individually. Bulk sequencing still is going to have a role for quite a while.
Hell germ line calling isn't even a solved problem. Still get lots of false positives (and false negatives). It just tends to work so well that it is hard to improve it much except by making it faster, less memory intensive, etc
Solved was the wrong word. I just meant improved. There is only so much you can do at the computational side. Wet lab also has its part to play.
A germline variant caller generally has a ploidy-based genotyping algorithm built in to part of the algorithm/pipeline. I believe, IIRC, the GATK UnifiedGenotyper for instance does both variant calling and then genotype calling. So to call a genotype for a variant it is expecting a certain number of reads to support the alternative allele. When working with somatic variants all of the assumptions about how many reads you expect with a variant at a position to distinguish between true and false positives are no longer valid. Except for fixed mutations throughout the tumor population only some proportion of cells will hold a somatic variation. You also typically have some contamination from normal non-cancerous cells. Add in complications from significant genomic instability with lots of copy number variations and such and you have a need for a major change in your model for calling variation while minimizing artifactual calls. So you have a host of other programs that have been developed specifically for looking at somatic variation in tumor samples.
High-throughput sequencing is rapidly becoming common practice in clinical diagnosis and cancer research. Many algorithms have been developed for somatic single nucleotide variant (SNV) detection in matched tumor-normal DNA sequencing. Although numerous studies have compared the performance of various algorithms on exome data, there has not yet been a systematic evaluation using PCR-enriched amplicon data with a range of variant allele fractions. The recently developed gold standard variant set for the reference individual NA12878 by the NIST-led “Genome in a Bottle” Consortium (NIST-GIAB) provides a good resource to evaluate admixtures with various SNV fractions.
Using the NIST-GIAB gold standard, we compared the performance of five popular somatic SNV calling algorithms (GATK UnifiedGenotyper followed by simple subtraction, MuTect, Strelka, SomaticSniper and VarScan2) for matched tumor-normal amplicon and exome sequencing data.
Nevertheless, detecting somatic mutations is still challenging, especially for low-allelic-fraction variants caused by tumor heterogeneity, copy number alteration, and sample degradation
We used QIAGEN’s GeneRead DNAseq Comprehensive Cancer Gene Panel (CCP, Version 1) for enrichment and library construction in triplicate。
QIAGEN’s GeneRead DNAseq Comprehensive Cancer Gene Panel (Version 1) was used to amplify the target region of interest (124 genes, 800 Kb).
When analyzing different types of data, use of different algorithms may be appropriate.
DNA samples of NA12878 and NA19129 were purchased from Coriell Institute. Sample mixtures were created based on the actual amplifiable DNA in each sample, resulting in 0%, 8%, 16%, 36%, and 100% of NA12878 sample mixed in the NA19129 sample, respectively.We treated the mixed samples at 8%, 16%, 36%, and 100% as the virtual tumor samples and the 0% as the virtual normal sample.
1. NaiveSubtract — SNVs were called separately from virtual tumor and normal samples using GATK UnifiedGenotyper . For exome sequencing data, reads were already mapped, locally realigned and recalibrated by the 1,000 Genomes Project. So SNVs were directly called on the BAM files using GATK Unified Genotyper. Then, SNVs detected in the virtual normal sample were removed from the list of SNVs detected in the virtual tumor sample, leaving the “somatic” SNVs.
2. MuTect — MuTect is a method developed for detecting the most likely somatic point mutations in NGS data using a Bayesian classifier approach. The method includes pre-processing aligned reads separately in tumor and normal samples and post-processing resulting variants by applying an additional set of filters. We ran MuTect under the High-Confidence mode with its default parameter settings. We disabled the “Clustered position” filter and the “dbSNP filter” for the amplicon sequencing reads, and we disabled the “dbSNP filter” for the exome sequencing.
3. SomaticSniper — SomaticSniper calculates the Bayesian posterior probability of each possible joint genotype across the normal and cancer samples. We tuned the software’s parameters to increase sensitivity and then filtered raw results using a Somatic Score cut-off of 20 to improve specificity.
4. Strelka — Strelka reports the most likely genotype for tumor and normal samples based on a Bayesian probability model. Post-calling filters built into the software are based on factors such as read depth, mismatches, and overlap with indels. We skipped depth filtration for exome and amplicon sequencing data as recommended by the Strelka authors. For the amplicon sequencing reads, we set the minimum MAPQ score at 17 for consistency with the defaults in GATK UnifiedGenotyper. We used variants passing Strelka post-calling filters for analysis.
5. VarScan2 — VarScan2 performs analyses independently on pileup files from the tumor and normal samples to heuristically call a genotype at positions achieving certain thresholds of coverage and quality. Then, sites of the genotypes not matched in tumor and normal samples are classified into somatic, germline, or ambiguous groups using Fisher’s exact test. We generated the pileup files using SAMtools mpileup command.
The compatibility of the output VCF files between different methods as well as the NIST-GIAB gold standard was examined using bcbio.variation tools and manual inspection. The reported SNP call representations between files are comparable to each other.