Comparing Methods to Detect Somatic Mutations

A nice comparison of four methods (VarScan, SomaticSniper, JointSNVMix and Strelka) to detect somatic single nucleotide variants (SNVs) was published this week in Bioinformatics.  This is clearly a very hot topic in the analysis of cancer genomes because as I was reading through the article, I was thinking to myself, I can think of at least three other methods which have been published in 2013 alone. This particular article was received by Bioinformatics in March 2013, accepted in June and published first online in July, therefore the more recent methods could not have made it in into this comparison paper, but I would have loved to see the authors take on the three methods discussed in this post.

For those less familiar with the topic, you might ask why is this such an important topic?  One of the first steps in analyzing cancer sequencing data, is to accurately detect the genetic variants with some level of sensitivity and specificity.  The second step would be to understand the genetic (or epigenetic) effect followed by associating the mutational profiles with outcome.  Two genetic variants of interest in matched normal and tumor samples are called somatic and germline single nucleotide polymorphisms (SNPs) where

• 'Somatic' refers to a variant observed in tumor sample and not observed in the normal sample and the normal allele matches the reference allele
• 'Germline' refers to the same variant observed in both the normal and tumor samples but the normal allele does not match the reference allele

One straightforward approach is the "subtraction method" in which variants are called separately from normal and tumor samples using a tool such as GATK.  Next, you just subtract the high quality variant calls from the tumor and the high quality 'no-calls' in normal leaving the somatic calls.

The main problem with this approach is when detecting somatic variants from the next-generation sequencing data the "sensitive and accurate detection of somatic singlenucleotide variants (sSNVs) from next-generation sequencing data is still a challenging informatic problem because of differing levels of purity and numbers of subclonal populations represented in tumor samples" [Hansen et al. (2013)].  This implies the 'subtraction method' may not be the best approach to detect true somatic and true germline variants because the somatic variants called from a given method may vary depending on tumor purity or clonality.

• Carter et al. (2012), Koboldt et al. (2012) and Hansen et al. (2013) agree "the necessary read depth coverage from the NGS data to detect a mutation accurately depends on 
• the prevalence of the mutation in that sample which depends on 
• levels of copy number variation, 
• levels of stromal contamination and 
• prevalence of mutation among subcones within the sample".
• The authors note: "Current software does not account for the how deep coverage should be to attain a specific sensitivity and accuracy".
• Cibulskis et al. (2013) writes "the sensitivity and specificity of any somatic mutation–calling method varies along the genome and depends on several factors, including 
• the depth of sequence coverage in the tumor and a patient-matched normal sample, 
• the local sequencing error rate, 
• the allelic fraction of the mutation and
• the evidence thresholds used to declare a mutation". 
• The authors note: "Characterizing how sensitivity and specificity depend on these factors is necessary for designing experiments with adequate power to detect mutations at a given allelic fraction, as well as for inferring the mutation frequency along the genome, which is a key parameter for understanding mutational processes and significance analysis".

Recently, several groups have developed methods specifically to deal with this problem (detecting somatic variants using next-generation sequencing data).  Rather than calling variants separately from the normal and tumor samples, one idea is to simultaneously analyze the reads from the matched normal and tumor samples together resulting in increased confidence in the germline and somatic variants.  Some studies have shown this can lead to better discrimination of germline variants from sequencing errors.  I refer you to the comparison paper published in Bioinformatics this week for a detailed discussion on VarScan, SomaticSniper, JointSNVMix and Strelka. In this post, I will discuss three more recently published tools to detect somatic mutations Shimmer, Cake and MuTect.

Shimmer [Hansen et al. (2013)]

• Method: Uses Fisher's exact test at each genomic position covered by both the tumor and normal sample and corrects for multiple testing.  
• Documentation: https://github.com/nhansen/Shimmer/blob/master/DOCUMENTATION
• Input: samtools BAM files of the matched normal and tumor samples and reference genome
• Output: Shimmer will produce a list of somatic variants and somatic indels.  Additionally, Shimmer will provide the output in the ANNOVAR format (and will annotate the file if ANNOVAR is locally installed on the system).
• Reference paper shows Shimmer is superior to other methods in large amounts of stromal contamination, heterogenity and copy number alterations, but comparable accuracy in high tumor purity.  

Installation and Usage: 
Shimmer is available in a perl script on github:
$ git clone git://github.com/nhansen/Shimmer.git
$ cd Shimmer
$ perl Build.PL
$ ./Build
$ ./Build test
$ ./Build install

Basic usage:
$ shimmer.pl [normal_bamfile.bam] [tumor_bamfile.bam] --ref reference.fa 


Cake [Rashid et al. (2013)]

• Method: Uses set of mutations called from at least two callers (VarScan2, CaVEMan, Bambino and SAMtools mpileup). Applies post-processing filters. 
• Documentation: http://cakesomatic.sourceforge.net/Cake_User_Guide.pdf
• Input: samtools BAM files of the matched normal and tumor samples and reference genome
• Output: Cake generates several output files/directory depending on what the user defined, but it will report the somatic variants (from each individual caller) in a VCF format. 
• The Cake_User_Guide.pdf provides extensive documentation on how to download and install Cake, VarScan2, CaVEMan and Bambino. It also provides documentation on how to install all the prerequistes such as samtools, vcftools, tabix, and additional perl modules.  

Installation and Usage: 
Cake is available in a perl script here: http://sourceforge.net/projects/cakesomatic/files/
$ mkdir cake
$ cd cake
$ wget http://sourceforget.net/projects/cakesomatic/files/Cake_1.0.tar.gz
$ tar xfz Cake_1.0.tar.gz

Basic Usage:
$ perl run_cake_somatic_pipeline.pl -s [Sample_File] -species human -callers pileup,varscan,bambino, somatic sniper -o [output_directory] -separator [,|\t] -mode [FULL|CALLING|FILTERING] 


MuTect [Cibulskis et al. (2013)]

• Method: Pre-processes the aligned reads from the matched normal and tumor samples separately.  Calculates two log odds ratios (LODs) - see below.  Applies an additional set of post-processing filters.   
• 1) LODs of the tumor containing a non-reference allele for a given position 
• 2) If the position is called non-reference, the LODs of the normal sample containing the reference allele
• Documentation: http://www.broadinstitute.org/cancer/cga/mutect_run
• Input: samtools BAM files of the matched normal and tumor samples and wiggle file containing the coverage output
• Output: A list of somatic calls (use grep to filter the calls not considered confident by MuTect)
• Reference paper compares MuTect to SomaticSniper, JointSNVMix and Strelka  
• Can use Cosmic & dbSNP databases for annotation

Installation and Usage: 
To run, must have java installed.  MuTect is available from here: http://www.broadinstitute.org/cancer/cga/mutect_download and should be run in the command line.

Basic usage:
$ java -Xmx2g -jar muTect-XXXX-XX-XX.jar --analysis_type MuTect --reference_sequence <reference> --input_file: normal [normal_bamfile.bam] --input_file:tumor [tumor_bamfile.bam] --out [output_file.out] --coverage_file [coverages.wig.txt]

To only see the 'confident' calls from MuTect, use grep to filter for lines not containing REJECT
$ grep -v REJECT [my.output_file.out]



Three additional tools (from 2011 and 2012):

• deepSNV [Gerstung et al. (2012)] 
• Based on a likelihood ratio test at each genomic position and corrects for multiple testing.
• Uses samtools with heavy physical memory requirements
• CaVEMan [Stephens et al. 2012]
• Rashid et al. (2013) says this method is similar to SAMtools mpileup which "computes the genotype likelihood of nucleotide positions in tumor and normal genome sequences by use of an expectation-maximization method"
• Bambino [Edmonson et al. 2011]
• Bayesian scoring model (as opposed to Fisher's exact testing) which compares alternate allele frequency between tumor and normal samples
• Similar to VarScan2 (except Bayesian)

Beyond the Genome 2012: Highlights and Discussion Points

This past week I attended the Beyond the Genome conference Sept 27-29 held in Boston at Harvard Medical School this year.  Genome Medicine and Genome Biology have hosted this conference for three years now and it was my first time attending.  To see more details on individual talks, check out the blogposts from Oliver Hofmann tagged with btg2012 or you can search for tweets with the hashtag #BtG2012. I want to give a few highlights from the conference, but first I'll begin with a cartoon which was a very fitting way to describe the conference!

Here are a few key discussion points I took away from the conference:

1. Genomic or bioinformatic tools used to analyze next-generation sequencing data need to be reproducible, accessible, fast, interactive and web-based. James Taylor from Galaxy advocated for making bioinformatics more reproducible and accessible and even gave an example of in a review only 7 out of 50 papers using BWA provided all the parameters necessary to be able to reproduce their research. Gabor Marth argued bioinformatic tools shouldn't be useable by only informaticians, but also should be useable by biologists.
2. Bioinformaticians are re-inventing the wheel. I found two recent papers giving a review of batch annotation tools for variants obtained using next-gernation sequencing (Sifrim et al. 2012; Lyon and Wang 2012) with a list of over 19 tools published just since 2010!  At the conference, several of the speakers noted in their talks, "As bioinformaticians we apparently like to keep reinventing the wheel".  I would agree with this statement.  In addition to genomic tools needing to be reproducible, accessible, fast, interactive and web-based, I would argue we need to create a standardized tool or format for annotating variants.  
3. With mutations, context matters. Why do some of the "right mutations" not respond to treatment? Josh Stuart advocated for using pathway-based analyses to assess the impact of mutations. Should focus on recurrent, rare variants (most likely to be impactful), but also need to make sure the background mutation rate is correct and determine if mutations are in key domains such as DNA binding or conserved domains. Daniel MacArthur said we need to analyze variants in the context of tens of thousands of genomes because the human genomic landscape is dominated by ultra-rare variation and we need to require consistent variant calls across studies. He also gave a set of recommendations on establishing causality from the NHGRI (see picture below).  Lynda Chin argued even if we know the mutation changes the function of the protein, we don't necessarily know the biological consequence. Therefore she argues we need to use a model systems approach to characterize and interpret a complete catalogue of driver mutations with functional validation.  
4. The idea of using whole exome sequencing as a diagnostic test in clinic has arrived (especially for rare genetic disorders), but we are just now starting to deal with all the challenges that come along with it.  Sharon Plon discussed her first year experience with clinical whole exome sequencing and said almost 25% of sequenced patients have "medically actionable" findings (mostly cardiac, but some cancer). Elizabeth Worthey said we need to be very careful because she explained "medically actionable" doesn't always mean "can be treated". With secondary variants and findings in clinic, Leslie Biesecker argued context matters. In his experience, the response of the patient varies depending on prior family history: with previous diagnosis - mundane response, with prior family history - mild surprise, without family history - dazed and confused.  Amy McGuire gave a beautiful talk from the legal perspective on the reasons to disclose or not disclose incidental findings and included survey results from actual GWAS researchers. In terms of drug discovery, Stuart Schreiber gave many examples of relating the genomic alterations in cancer to the small-molecule sensitivity. 

Finally, I thought the talk by Richard Gibbs deserved it's own paragraph. His talk focused on who should we be sequencing when it comes to human genetic diseases.  Though next-generation sequencing technology has come a long way, it's still not perfect.  Performing whole exome sequencing as a service can be effectively done by experts, but this is inefficient and not scalable.  Automating the process of interpretation often ends with dumb results. Sequencing healthy individuals should be a low priority especially when not tying phenotypes to the control individuals (crazy!).  Sequencing individuals with complex diseases is still tricky because of the debate regarding CDCV v. CDRV hypothesis.  He argued that if complex diseases are caused by rare variants, then we should focus within families and not broader populations.  Sequencing individuals with mendelian disease should be high on the priority list because there is a huge value in obtaining a molecular diagnosis even without a treatment.  Finally recreational sequencing is low on the priority list.  He has even coined the term 'narcciss-ome'!  Overall, he predicts sequencing will become standard of care and soon all the excitement will be passé. 

Final thoughts:  The conference was filled with fantastic talks given by world-renowned speakers.  The level of genetic complexity of diseases never ceases to amaze me, but at the same time seeing such great research being produced to answer some tough genetic questions is always exciting.  I learned a great deal and would highly recommend Beyond the Genome to future participants!  

Antioxidants May Lower Risk of Pancreatic Cancer

The journal Gut (impact factor 10.11) published a prospective study yesterday investigating whether eating antioxidants such as vitamin C (found in fruits and vegetables) and E (found in nuts, seeds, egg yolk), selenium (found in cereals, nuts, fish and meat) and zinc decreases the risk of pancreatic cancer.  They looked at 7-day food diaries of 23,658 participants over 10 years in the European Prospective Investigation of Cancer (EPIC-Norfolk) Study.  The estimates in this study were adjusted for other variables such as age at recruitment, gender, smoking, diabetes, BMI and total energy (smoking and diabetes are two known risk factors for pancreatic cancer).   Interestingly, those with the highest intake of selenium, vitamin C and E reduced the risk of pancreatic cancer by two-thirds. Those with high intakes of selenium and vitamin E had 40% reduced risk.  Those are huge reductions in risk based on these micronutrients.

The American Institute for Cancer Research (AICR) wrote "37,000 Americans will die of pancreatic cancer this year. According to the National Cancer Institute, pancreatic cancer is the fourth leading cause of cancer-related death. It is often not diagnosed until the advanced stages, when treatment is challenging".  With the recent death of Sally Ride, I thought this article discussing pancreatic cancer was particularly relevant.  

War on Cancer

As a statistician applying my research to cancer genetics it is always encouraging to see stories like this that make me feel sometimes all this work isn't for nothing.

The article published last week in the NY Times describes the story of a woman who was diagnosed with a rare form of lymphoma in which white blood cells (T cells) become cancerous and move to the skin.  Faced with a disease with no cure and no standard treatment, she was able to keep the cancer at bay using chemotherapy for five years.  At that point, her health took a turn for the worse and her son who worked at Illumina, which produces some of the latest sequencing technology, decided to quit his job and help his mom full-time by sequencing her genome.  After sequencing her normal and tumor DNA, they found 18K differences (or mutations) with no known significance for disease.  In the analysis of her DNA, the researchers found two genes fused together forcing the growth signals in cancer cells to be reversed: the signal to stop was forcing the cells to grow and the signal to grow was forcing the cells to stop growth.  The researchers decided to give her a new melanoma drug ipilimumab which forces normal T cells to grow, in hopes of stopping the growth of the tumor cells. The drug performed beautifully keeping the cancerous T cells at bay for 8 weeks before the cancer came back and ultimately she passed away a few weeks later.

41 years ago Richard Nixon signed the National Cancer Act of 1971 which started a "war on cancer". The goal was to find a "cure for cancer" by increasing the funding toward cancer research and find more effective treatments.  As defined by PubMed "cancer is the uncontrolled growth of abnormal cells in the body".  Cancer is not just one disease, in fact it is the word we have for hundreds of diseases we label as 'cancer'.  When people talk about "curing cancer", it suggests that once a treatment/cure for one type of cancer is found, it should theoretically be applied to another form of cancer.  In my experience this is definitely not the case.  Every type of cancer is unique in its etiology and treatment.  It is true that some drugs today developed for one type of cancer can be applied to another type of cancer because the target molecule may happen be the same for two different cancers, but this is a separate idea than the finding a "cure for cancer".

I work with several groups of researchers in the Texas Medical Center at University of Texas MD Anderson, Baylor College of Medicine and Texas Children's Hospital who are all trying to exactly this.  They are sequencing the genomes of individuals affected by a particular disease and trying to find the causal mutation or reason for the disease.  My small contribution comes in helping to analyze the results that come out of the sequencing.  Bioinformatics attempts to take in the large amount of sequencing data and make sense of it.  This is by no means easy and will take much longer to properly analyze the data than to just sequence it.  But, even with these small steps I look forward to more success stories like these and seeing further progress being made in the war on cancer.

Eliminating Highly Polymorphic Genes from Whole Exome Sequencing

Whole exome sequencing yields thousands of variants for each individual sequenced.  Many of these variants are in genes that are highly polymorphic or in regions that do not sequence well and therefore would not be of interest when searching for a disease-causing variant.  This is because any frequently mutated gene containing many deleterious variants will have a low probability of containing the disease-causing mutation.  Also, if the gene is in a region that does not sequence well then a high number of variants will be often reported.  When you find such variants, this would be considered a false positive.   There is a big interest in detecting these false positive signals and eliminating these variants from the list reported back from whole exome sequencing.  

A paper was just published this month in Human Mutation describing a way to do this.  They published several lists of genes for researchers to use in their own projects.  The researchers hypothesize these genes will not contain the disease-causing mutation.   This list of genes could be an incredibly useful tool to filter out highly polymorphic genes or genes that simply do not sequence well.

Nearly 800,000 lung cancer death averted by decline in smoking

The results from a major study on lung cancer conducted by six institutions (including Rice) was released yesterday in the Journal of National Cancer Institute.  The goal of the study was to ask  how many deaths have been prevented after the release of the US Surgeon General's report on Smoking and Health in 1964. This is the first time the NCI is publishing a paper of this magnitude on estimates of lung cancer deaths using all model-based approaches.  It was estimated nearly 800,000 lung cancer deaths have been averted from 1975-2000 with the decline in smoking.  What's the most impressive is that if smoking had been completely eliminated, then it was estimated that 2.5 million deaths would have been prevented.