Let's say I ran forty cancer samples, twenty responders and twenty non-responders to some treatment. I performed alignment and annotation of SNVs and indels in each sample. Now I want to know if there are any deferentially affected genes in the two groups that may justify the different response I observed in those tumors. Problem is, simple chi squares won't do, because a single gene can be affected by multiple potentially deleterious SNVs and indels, and I may have to bin multiple alterations to have enough power. I may therefore tag each gene as affected/unaffected by a potentially deleterious variation, and proceed this way with my analyses. But how do I know if a "potentially" deleterious variation is really so? It's impossible to validate biologically all of them. Or I may restrict my analysis to variations that recur more than once in two different samples, but then again, some genes can be altered in many places, and in one place just once out of several samples, and I would overlook them.
Finally, I may consider not genes but pathways, and see whether some pathways are involved by deleterious alterations in responder tumors and not in non-responders, or vice-verse, but pathways are not as well defined as one may want to think; there is a risk of including deleterious variations that have nothing to do with the underlying biology of my samples. It's more philosophy than statistics here. What would you do? What do you you recommend?
You can get a lot of information about your mutations from the CRAVAT web server (www.cravat.us) . The server hosts some machine learning-based tools for the analysis of missense mutations detected in tumor sequencing. This includes functional analysis (is the protein affected?) as well as a method that is cancer-specific ( does the mutation look more like a driver or a passenger?). These methods return a continuous score between 0 and 1 that can be used to prioritize mutations, as well as p-value and FDR estimates to help you select a cutoff for deciding which missense mutations to keep. CRAVAT annotates all of your variants with dbSNP ids, allele frequencies for 1000 genomes and ESP6500 populations, overlap with the COSMIC database, provides gene functional annotations from the GeneCards database, and the results of a PubMed search for each mutated gene.
If you want to look for more fundamental differences between the groups, you might want to read "Mutational processes molding the genomes of 21 breast cancers" from Mike Stratton's group (Cell 2012). They look at differences in the processes underlying mutation among 21 individuals with Breast Cancer.
to be honest there`s a much bigger problem in your setting. If you do not have normal tissue samples to compare with, how are you going to correct your data for possible population biasses? If you do not have them, it will be impossible o identify the correct significantly different variations.
Ingenuity has just started a tool for genotyping and pathway-related analysis of variants. Played with it today - seems to work nicely. They plan to charge quite a bit of money for it, but chances are for an open trial limited version.
Benedetta why should I have population biases if all my patients come from the same region and are from the same ethnicity? I assume you mean the issue of having healthy matched tissues is related to filtering germline variants from somatic ones.
The first question is are you investigating germ-line variation or somatically acquired variation? Either or both is good, but somatic variation will be easier as you can exclude germ-line variation from your matched non-cancer tissue. Analysis of variation in your germ-line material will require alignment to some kind of reference genome, and you will get a huge number of potential variants from this analysis.
This is a massive field, but briefly.
Assuming you have suitable read depth then you can filter your variants based on some type of depth filter and that the variant occurs on both DNA strands. It is worth eye-balling your variants in software such as IGV to see how the variants look in the context of you read depth. If you are looking at somatic variants, you can calculate a somatic p value based on the read depth in your germ-line and somatic tissue, this can be useful at prioritising candidate genes.
Then you should define whether the variants are within genes and if they are non-synonymous, giving an insight into whether they are effecting gene function. Furthermore, whether the variants result in a frame-shift or the creation of a stop codon. Again, insight into functional consequences. SAMTools can do this kind of stuff, but there are plenty of other packages.
You can filter your data based on the 1000 genome project (or dbSNP), if your looking at cancer genes you can also look for ones in COSMIC.
There are also packages that predict the potential impact of a non-synonymous variant, such as SIFT, POLYPHEN and GRANTHAM. Certain genes are also highly mutable in the context of normal individuals, and you could consider removing them, there is a paper by Fuentes (cannot remember the reference) that defined these genes.
Then you can look at recurrently targeted genes and pathways.
This might help you prepare a list of genes that are likely to be real based on your NGS data, and then prioritise them based on potential effects. There are many, many things that you can do, but hopefully this helps.
Yes, I did mean that and I think Jon Strefford gave you a complete explanation. I`m not a cancer person, but there`s been a lot of discussion in my lab and in general in the cancer community here at Harvard about the need of including healthy tissues in cancer profiling-studies. I`m epigenetist, and in this field this is a must for each of our studies in the lab on cancer. I suppose it is not exactly the same in genetics, but I would try to get as much as I can and Jon suggested some interesting approaches to do that. Thank you for replying I`m always interested in learning new stuff especially from people working in different fields.
Good luck!
You can get a lot of information about your mutations from the CRAVAT web server (www.cravat.us) . The server hosts some machine learning-based tools for the analysis of missense mutations detected in tumor sequencing. This includes functional analysis (is the protein affected?) as well as a method that is cancer-specific ( does the mutation look more like a driver or a passenger?). These methods return a continuous score between 0 and 1 that can be used to prioritize mutations, as well as p-value and FDR estimates to help you select a cutoff for deciding which missense mutations to keep. CRAVAT annotates all of your variants with dbSNP ids, allele frequencies for 1000 genomes and ESP6500 populations, overlap with the COSMIC database, provides gene functional annotations from the GeneCards database, and the results of a PubMed search for each mutated gene.
If you want to look for more fundamental differences between the groups, you might want to read "Mutational processes molding the genomes of 21 breast cancers" from Mike Stratton's group (Cell 2012). They look at differences in the processes underlying mutation among 21 individuals with Breast Cancer.
Hannah, CRAVAT looks great and really powerful, thank you for the introduction. I have passed the details on to my group.
I would suggest to have a look also to PriVar, a toolkit for prioritizing SNVs and indels from next generation sequencing data. Executable jar package is available at http://paed.hku.hk/uploadarea/yangwl/html/software.html
You might find FATHMM useful (http://supfam2.cs.bris.ac.uk/FATHMM/), it even includes cancer-specific options. Also it is connected to dcGO (http://supfam.org/SUPERFAMILY/dcGO/) which can give you pathway and other useful ontology information.
- Badges
- Science topic