GenomicConsensus is the current PacBio consensus and variant
calling suite. It contains a main driver program, variantCaller
,
which provides two consensus/variant calling algorithms: Arrow and
Quiver. These algorithms can be run by calling
variantCaller.py --algorithm=[arrow|quiver|plurality]
or by going
through the convenience wrapper scripts quiver
and arrow
.
- EviCons was the previous generation PacBio variant caller (removed
- in software release v1.3.1).
Separate packages called ConsensusCore and ConsensusCore2 are C++ libraries where all the computation behind Quiver and Arrow are done, respectively. This is transparent to the user after installation.
Plurality is a very simple variant calling algorithm: it stacks up the aligned reads (alignment as produced by BLASR, or alternate mapping tool), and for each column under a reference base, calls the most abundant (i.e., the plurality) read base (or bases, or deletion) as the consensus at that reference position.
Plurality does not perform any local realignment. This means it is heavily biased by the alignment produced by the mapper (BLASR, typically). It also means that it is insensitive at detecting indels. Consider this example:
Reference AAAA ---- Aligned A-AA reads AA-A -AAA ---- Plurality AAAA consensus
Note here that every read has a deletion and the correct consensus call would be "AAA", but due to the mapper's freedom in gap-placement at the single-read level, the plurality sequence is "AAAA"---so the deletion is missed. Local realignment, which plurality does not do, but which could be considered as implicit in the Quiver algorithm, essentially pushes the gaps here to the same column, thus identifying the deletion. While plurality could be adjusted to use a simple "gap normalizing" realignment, in practice noncognate extras (spurious non-homopolymer base calls) in the midst of homopolymer runs pose challenges.
Quiver is a more sophisticated algorithm that finds the maximum quasi-likelihood template sequence given PacBio reads of the template. PacBio reads are modeled using a conditional random field approach that scores the quasi-likelihood of a read given a template sequence. In addition to the base sequence of each read, Quiver uses several additional QV covariates that the basecaller provides. Using these covariates provides additional information about each read, allowing more accurate consensus calls.
Quiver does not use the alignment provided by the mapper (BLASR, typically), except for determining how to group reads together at a macro level. It implicitly performs its own realignment, so it is highly sensitive to all variant types, including indels---for example, it resolves the example above with ease.
The name Quiver reflects a consensus-calling algorithm that is QV-aware.
We use the lowercase "quiver" to denote the quiver tool in GenomicConsensus, which applies the Quiver algorithm to mapped reads to derive sequence consensus and variants.
Quiver is described in detail in the supplementary material to the HGAP paper.
Arrow is a newer model intended to supercede Quiver in the near future. The key differences from Quiver are that it uses an HMM model instead of a CRF, it computes true likelihoods, and it uses a smaller set of covariates. We expect a whitepaper on Arrow to be available soon.
We use the lowercase "arrow" to denote the arrow tool, which applies the Arrow algorithm to mapped reads to derive sequence consensus and variants.
Quiver supports all PacBio RSII data since the C2 chemistry release (late 2012).
Arrow supports all data from the PacBio Sequel instrument, as well as data from the PacBio RSII using the P6-C4 chemistry.
For general instructions on installing and running, see the HowTo document.
There are three output files from the GenomicConsensus tools:
- A consensus FASTA file containing the consensus sequence
- A consensus FASTQ file containing the consensus sequence with quality annotations
- A variants GFF file containing a filtered, annotated list of variants identified
It is important to note that the variants included in the output variants GFF file are filtered by coverage and quality, so not all variants that are apparent in comparing the reference to the consensus FASTA output will correspond to variants in the output variants GFF file.
To enable all output files, the following can be run (for example):
% quiver -j16 aligned_reads.cmp.h5 -r ref.fa \ -o consensus.fa \ -o consensus.fq \ -o variants.gff
The extension is used to determine the output file format.
Quiver's consensus is de novo in the sense that the reference and the reference alignment are not used to inform the consensus output. Only the reads factor into the determination of the consensus.
The only time the reference sequence is used to make consensus calls -
when the --noEvidenceConsensusCall
flag is set to reference
or
lowercasereference
(the default)- is when there is no effective
coverage in a genomic window, so Quiver has no evidence for computing
consensus. One can set --noEvidenceConsensusCall=nocall
to
avoid using the reference even in zero coverage regions.
Quiver's expected accuracy is a function of coverage and chemistry. The C2 chemistry (no longer available), P6-C4 and P4-C2 chemistries provide the most accuracy. Nominal consensus accuracy levels are as follows:
Coverage | Expected consensus accuracy | |
---|---|---|
C2, P4-C2, P6-C4 | P5-C3 | |
10x | > Q30 | > Q30 |
20x | > Q40 | > Q40 |
40x | > Q50 | > Q45 |
60-80x | ~ Q60 | > Q55 |
The "Q" values referred to are Phred-scaled quality values:
q = -10 \log_{10} p_{error}
for instance, Q50 corresponds to a p_error of 0.00001---an accuracy of 99.999%. These accuracy expectations are based on routine validations performed on multiple bacterial genomes before each chemistry release.
arrow achieves similar accuracy to quiver. Numbers will be published soon.
If there are errors remaining applying Quiver, they will almost invariably be homopolymer run-length errors (insertions or deletions).
At present, the Quiver model is trained per-chemistry, so it is very important that Quiver knows the sequencing chemistries used.
If SMRT Analysis software was used to build the cmp.h5 or BAM input file, the cmp.h5 will be loaded with information about the sequencing chemistry used for each SMRT Cell, and GenomicConsensus will automatically identify the right parameters to use.
If custom software was used to build the cmp.h5, or an override of Quiver's autodetection is desired, then the chemistry or model must be explicity entered. For example:
% quiver -p P4-C2 ... % quiver -p P4-C2.AllQVsMergingByChannelModel ...
Yes! GenomicConsensus tools automatically see the chemistry per-SMRT Cell, so it can figure out the right parameters for each read and model them appropriately.
For Quiver: all PacBio RS chemistries are supported. Chemistry mixtures of P6-C4, P4-C2, P5-C3, and C2 are supported.
For Arrow: the RS chemistry P6-C4, and all PacBio Sequel chemistries are supported. Mixes of these chemistries are supported.
Quiver uses additional QV tracks provided by the basecaller. These QVs may be looked at as little breadcrumbs that are left behind by the basecaller to help identify positions where it was likely that errors of a given type occurred. Formally, the QVs for a given read are vectors of the same length as the number of bases called; the QVs used are as follows:
- DeletionQV
- InsertionQV
- MergeQV
- SubstitutionQV
- DeletionTag
To find out if your cmp.h5 file is loaded with these QV tracks, run the command
% h5ls -rv aligned_reads.cmp.h5
and look for the QV track names in the output. If your cmp.h5 file is lacking some of these tracks, Quiver will still run, though it will issue a warning that its performance will be suboptimal.
The most likely cause for true errors made by these tools is that the coverage in the region was low. If there is 5x coverage over a 1000-base region, then 10 errors in that region can be expected.
It is important to understand that the effective coverage available to quiver/arrow is not the full coverage apparent in plots---the tools filter out ambiguously mapped reads by default. The remaining coverage after filtering is called the /effective coverage/. See the next section for discussion of MapQV.
If you have verified that there is high effective coverage in the region in question, it is highly possible---given the high accuracy quiver and arrow can achieve---that the apparent errors actually reflect true sequence variants. Inspect the FASTQ output file to ensure that the region was called at high confidence; if an erroneous sequence variant is being called at high confidence, please report a bug to us.
For regions with no effective coverage, no variants are outputted, and the FASTQ confidence is 0.
The output in the FASTA and FASTQ consensus sequence tracks is
dependent on the setting of the --noEvidenceConsensusCall
flag.
Assuming the reference in the window is "ACGT", the options are:
--noEvidenceConsensusCall=... |
Consensus output |
---|---|
nocall (default in 1.4) |
NNNN |
reference |
ACGT |
lowercasereference (new post 1.4, and the
default) |
acgt |
MapQV is a single scalar Phred-scaled QV per aligned read that reflects the mapper's degree of certainty that the read aligned to this part of the reference and not some other. Unambigously mapped reads will have a high MapQV (typically 255), while a read that was equally likely to have come from two parts of the reference would have a MapQV of 3.
MapQV is pretty important when you want highly accurate variant calls. Quiver and Plurality both filter out aligned reads with a MapQV below 20 (by default), so as not to call a variant using data of uncertain genomic origin.
This can be problematic if using quiver/arrow to get a consensus sequence. If the genome of interest contains long (relative to the library insert size) highly-similar repeats, the effective coverage (after MapQV filtering) may be reduced in the repeat regions---this is termed these MapQV dropouts. If the coverage is sufficiently reduced in these regions, quiver/arrow will not call consensus in these regions---see `What do quiver/arrow do for genomic regions with no effective coverage?`_.
If you want to use ambiguously mapped reads in computing a consensus for a denovo assembly, the MapQV filter can be turned off entirely. In this case, the consensus for each instance of a genomic repeat will be calculated using reads that may actually be from other instances of the repeat, so the exact trustworthiness of the consensus in that region may be suspect. The next section describes how to disable the MapQV filter.
The MapQV filter can be disabled using the flag
--mapQvThreshold=0
(shorthand: -m=0
). If running a
quiver/arrow job via SMRT Portal, this can be done by unchecking the "Use
only unambiguously mapped reads" option. Consider this in
de novo assembly projects, but it is not recommended for variant
calling applications.
When in doubt, it is easiest to inspect the region in a tool like SMRT View, which enables you to view the reads aligned to the region. Deletions and substitutions should be fairly easy to spot; to view insertions, right-click on the reference base and select "View Insertions Before...".
The available options limit read coverage, filters reads by MapQV, and filters variants by quality and coverage.
- The overall read coverage used to call consensus in every window is
100x by default, but can be changed using
-X=value
. - The MapQV filter, by default, removes reads with MapQV < 20. This
is configured using
--mapQvThreshold=value
/-m=value
- Variants are only called if the read coverage of the site exceeds
5x, by default---this is configurable using
-x=value
. Further, they will not be called if the confidence (Phred-scaled) does not exceed 40---configurable using-q=value
.
At present, quiver/arrow assume a haploid sample, and the behavior of on sample mixtures or diploid/polyploid samples is undefined. The program will not crash, but the output results are not guaranteed to accord with any one of the haplotypes in the sample, as opposed to a potential patchwork.
Some customers using quiver for polishing highly repetitive genomes have found that if they take the consensus FASTA output of quiver, use it as a new reference, and then perform mapping and Quiver again to get a new consensus, they get improved results from the second round of quiver.
This can be explained by noting that the output of the first round of quiver is more accurate than the initial draft consensus output by the assembler, so the second round's mapping to the quiver consensus can be more sensitive in mapping reads from repetitive regions. This can then result in improved consensus in those repetitive regions, because the reads have been assigned more correctly to their true genomic loci. However there is also a possibility that the potential shifting of reads around from one rounds' mapping to the next might alter borderline (low confidence) consensus calls even away from repetitive regions.
We recommend the (mapping+quiver) iteration for customers polishing repetitive genomes, and it could also prove useful for resequencing applications. However we caution that this is very much an exploratory procedure and we make no guarantees about its performance. In particular, borderline consensus calls can change when the procedure is iterated, and the procedure is not guaranteed to be convergent.
We have seen many examples where (mapping+quiver), repeated many times, is evidently not a convergent procedure. For example, a variant call may be present in iteration n, absent in n+1, and then present again in n+2. It is possible for subtle changes in mapping to change the set of reads examined upon inspecting a genomic window, and therefore result in a different consensus sequence there. We expect this to be the case primarily for "borderline" (low confidence) base calls.