Frequently Asked Questions¶
What type of variants does pVACseq support?
pVACseq makes predictions for all transcripts of a variant that were annotated
frameshift_variant by VEP as long as the transcript was not also annotated
start_lost. In addition, pVACseq only includes variants that were
called as homozygous or heterozygous variant. Variants that were not called
in the sample specified are skipped (determined by examining the
field in the VCF).
My pVACseq command has been running for a long time. Why is that?
The rate-limiting factor in running pVACseq is the number of calls that are made to the IEDB software for binding score predictions.
It is generally faster to make IEDB calls using a local install of IEDB than using the IEDB web API. It is, therefore, recommended to use a local IEDB install for any in-depth analysis. You should either install IEDB locally yourself or use the pvactools docker image that includes it.
There are a number of factors that determine the number of IEDB calls to be made:
Number of variants in your VCF
pVACseq will make predictions for each missense, inframe insertion, inframe deletion, protein altering, and frameshift variant in your VCF.
Speedup suggestion: Split the VCF into smaller subsets and process each one individually, in parallel.
Number of transcripts for each variant
pVACseq will make predictions for each transcript of a supported variant individually. The number of transcripts for each variant depends on how VEP was run when the VCF was annotated.
Speedup suggestion: Use the
--pickoption when running VEP to annotate each variant with the top transcript only.
pVACseq takes an input VCF and creates a wildtype and a mutant FASTA for each transcript. The number of FASTA entries that get submitted to IEDB at a time is limited by the
--fasta-sizeparameter in order to reduce the load on the IEDB servers. The smaller the FASTA size, the more calls have to be made to IEDB.
Speedup suggestion: When using a local IEDB install, increase the size of this parameter.
Number of prediction algorithms, epitope lengths, and HLA-alleles
One call to IEDB is made for each combination of these parameters for each chunk of FASTA sequences. That means, for example, when 8 prediction algorithms, 4 epitope lengths (8-11), and 6 HLA-alleles are chosen, 7*4*6=192 calls to IEDB have to be made for each chunk of FASTA.
Speedup suggestion: Reduce the number of prediction algorithms, epitope lengths, and/or HLA-alleles to the ones that will be the most meaningful for your analysis. For example, the NetMHCcons method is already a consensus method between NetMHC, NetMHCpan, and PickPocket. If NetMHCcons is chosen, you may want to omit the underlying prediction methods. Likewise, if you want to run NetMHC, NetMHCpan, and PickPocket individually, you may want to skip NetMHCcons.
This parameter determines how many amino acids of the downstream sequence after a frameshift mutation will be included in the wildtype FASTA sequence. The shorter the downstream sequence length, the lower the number of epitopes that IEDB needs to make binding predictions for.
Speedup suggestion: Reduce the value of this parameter.
This parameter determines the number of threads pvacseq will use for parallel processing.
Speedup suggestion: Use a host with multiple cores and sufficient memory and use a larger number of threads.
My pVACseq output file does not contain entries for all of the alleles I chose. Why is that?
There could be a few reasons why the pVACseq output does not contain predictions for alleles:
The alleles you picked might have not been compatible with the prediction algorithm and/or epitope lengths chosen. In that case no calls for that allele would’ve been made and a status message would’ve printed to the screen.
It could be that all epitope predictions for some alleles got filtered out. You can check the
<sample_name>.all_epitopes.tsvfile to see all called epitopes before filtering.
Why are some values in the WT Epitope Seq column NA ?
Not all mutant epitope sequences will have a corresponding wildtype epitope sequence. This occurs when the mutant epitope sequence is novel and a comparison is therefore not meaningful. For example:
An epitope in the downstream portion of a frameshift might not have a corresponding wildtype epitope at the same position at all. The epitope is completely novel.
An epitope that overlaps an inframe indel or multinucleotide polymorphism (MNP) might have a large number of amino acids that are different from the wildtype epitope at the corresponding position. If less than half of the amino acids between the mutant epitope sequence and the corresponding wildtype sequence match, the corresponding wildtype sequence in the report is set to
What filters are applied during a pVACseq run?
By default we filter the neoepitopes on their binding score. If readcount
and/or expression annotations are available in the VCF we also filter on the depth, VAF,
and gene/trancript FPKM. In addition, candidates where the mutant epitope sequence is the
same as the wildtype epitope sequence will also be filtered out (i.e., they
don’t overlap the mutation). pVACseq also filters on the transcript support
level, if the
--tsl option was chosen during VEP annotation. Lastly, the
top score filter will pick the best epitope for each variant.
How can I see all of the candidate epitopes without any filters applied?
<sample_name>.all_epitopes.tsv will contain all of the epitopes predicted
before filters are applied.
Why have some of my epitopes been filtered out even though the Best MT Score is below 500?
By default, the binding filter will be applied to the
Median MT Score
column. This is the median score value among all chosen prediction algorithms.
Best MT Score column shows the lowest score among all
chosen prediction algorithms. To change this behavior and apply the binding
filter to the
Best MT Score column you may set the
Why are entries with NA in the VAF and depth columns not filtered?
We do not filter out
NA entries for depth and VAF since there is not
enough information to determine whether the cutoff has been met one way or another.
Why do some of my epitopes have no score predictions for certain prediction methods?
Not all prediction methods support all epitope lengths or all alleles. To see
a list of supported alleles for a prediction method you may use the
pvacseq valid_alleles command. For more details on
each algorithm refer to the IEDB MHC Class I
and Class II documentation.
How is pVACseq licensed?
pVACseq is licensed under the open source license NBSD 3-Clause Clear License.
How do I cite pVACseq?
Jasreet Hundal+, Susanna Kiwala+, Joshua McMichael, Christopher A Miller, Alexander T Wollam, Huiming Xia, Connor J Liu, Sidi Zhao, Yang-Yang Feng, Aaron P Graubert, Amber Z Wollam, Jonas Neichin, Megan Neveau, Jason Walker, William E Gillanders, Elaine R Mardis, Obi L Griffith, Malachi Griffith. pVACtools: a computational toolkit to select and visualize cancer neoantigens. (+)equal contribution. bioRxiv 501817; doi: https://doi.org/10.1101/501817
Jasreet Hundal, Susanna Kiwala, Yang-Yang Feng, Connor J. Liu, Ramaswamy Govindan, William C. Chapman, Ravindra Uppaluri, S. Joshua Swamidass, Obi L. Griffith, Elaine R. Mardis, and Malachi Griffith. Accounting for proximal variants improves neoantigen prediction. Nature Genetics. 2018, DOI: 10.1038/s41588-018-0283-9. PMID: 30510237.
Jasreet Hundal, Beatriz M. Carreno, Allegra A. Petti, Gerald P. Linette, Obi L. Griffith, Elaine R. Mardis, and Malachi Griffith. pVACseq: A genome-guided in silico approach to identifying tumor neoantigens. Genome Medicine. 2016, 8:11, DOI: 10.1186/s13073-016-0264-5. PMID: 26825632.