First, I'm sure there's a better way to do this, but I wanted to create a single commit summarizing the changes from my branch SamRecordFactory. What's the best way to do this? Rebase?
Now, on to the changes here:
-- Picard added a SamRecordFactory that is used to create instances the subclass SamRecord or BAMRecord. This factory allows us to have low-level picard readers (SamFileReader) create objects of type GATKSamRecord. The abomination of the extends and contains GATKSamRecord is now gone. GATKSamRecords are now produced by this factory, the GATK provides this factory to our SamFileReaders, and everything works with GATKSamRecord just extending BAMRecord. This results in up to a 2x performance improvement in writing BAMs and a ~10% improvement when reading BAMs files.
-- As a consequence of this, we no longer officially support SAM records. Attempting to create SAMRecord objects with the factory will throw a user exception.
-- Created a standard NGSPlatform enum, and GATKSamRecords support efficiently obtaining this value. The real BQSR (not the copy indel version) got the efficient code to use this. Please add all future platforms to this enum.
-- GATKSamRecord no longer supports using the OQ or defaultBaseQuality. This is performed in a wrapper iterator that's only added when these command line options are used.
-- ReducedRead code has been moved from ReadUtils until efficiency caching assessors in GATKSamRecord.
-- ArtificialSamUtils creates GATKSamRecords now, just SAMRecords. Added code here to create artifical pairs and using that code to create artificial ReadBackedPileups with specific properties
-- New smarter algorithm for FragmentPileup. This new code is up to 3x faster than the previous version, and is lazy so is more efficient when no overlapping pairs are actually in the pileup. Created extensive DataProvider driven UnitTest. Added Caliper-based benchmarking system to characterize the performance differences between the old and new algorithms. TODO still remains to make a efficient version that works for non-pileups for the HaplotypeCaller
Moved gsalib and queueJobReport.R to embeddable namespaced locations.
Updated packager dependencies/dir to add an @includes which filters the embedded fileset.
RScriptExecutor can now JIT compiles the gsalib.
RScriptExecutor uses ProcessController and sends the Rscript output to java's stdout when run under -l DEBUG.
Refactored ProcessController and IOUtils from Queue to Sting Utils.
Added more unit tests to ProcessController along with a utility class to hard stop OutputStreams at a specified byte count.
Replaced uses of some IOUtils with Apache Commons IO.
ShellJobRunner refactored to use direct ProcessController and now kills jobs on shutdown.
Better QGraph responsiveness on shutdown by using Object.wait() instead of Thread.sleep().
-- removed intermiate functions. Now only original version and best optimized new version remain
-- Moved general artificial read backed pileup creation code into ArtificialSamUtils
- Sites with missing genotypes in pairs/trios are handled as follows:
-- Missing child -> Homozygous parents are phased, no transmission probability is emitted
-- Two individuals missing -> Phase if homozygous, no transmission probability is emitted
-- One parent missing -> Phased / transmission probability emitted
- Mutation prior set as argument
-- Uses mayOverlapRoutine in ReadUtils
-- Attempts to be smart when doing overlap calculation, to avoid unnecessary allocations
-- PileupElement now comparable (sorts on offset than on start)
-- Caliper microbenchmark to assess performance
Moved the validation of the GATKSamRecord to the MalformedReadFilter with the intent to make the read filter the ultimate validation location for sam records. This way we can opt to filter out malformed reads if we know what we are doing or blow up otherwise.
- Adapted to get the trio information from the SampleDB (i.e. from Pedigree file (ped)) => Multiple trios can be passed as argument
- Mendelian violations and trio phasing possibilities are pre-calculated and stored in Maps. => Runtime is ~3x faster
- Genotype combinations possible only given two MVs are now given a squared MV prior (e.g. 0/0+0/0=>1/1 is given 10^-16 prior if the MV prior is 10^-8)
- Corrected bug: In case the best genotype combination is Het/Het/Het, the genotypes are now set appropriately (before original genotypes were left even if they weren't Het/Het/Het)
- Basic reporting added:
-- mvf argument let the user specify a file to report remaining MVs
-- When the walker ends, some basic stats about the genotype reconfiguration and phasing are output
Known problems:
- GQ is not recalculated even if the genotype changes
Possible improvements:
- Phase partially typed trios
- Use standard Allele/Genotype Classes for the storage of the pre-calculated phase
-- Creates all combinatinos of overlapping and non-overlapping read pair pileups in all orientations and first/second pairings to validate fragment detection.
This commit contains:
- IntronLossGenotyper is brought into its current incarnation
- A couple of simple new filters (ReadName is super useful for debugging, MateUnmapped is useful for selecting out reads that may have a relevant unaligned mate)
- RFA now matches my current local repository. It's in flux since I'm transitioning to the new traversal type.
+ the triggering read stash pilot required me to change the scope of some of the variables in the ReadClipping code, private -> protected. Those are all the changes there.
- MendelianViolation restored to its former glory (and an annotator module that uses the likelihood calculation has been added)
+ use this rather than a hard GQ threshold if you're doing MV analyses.
- Some miscellaneous QScripts
-- all of the methods in the class must be synchronized or the internal state can be inconsistent with the contract invariant when entering the class in a non-synchronized method, even when that method doesn't care about the object's internal state
We can't have a public test that depends on both public and private
code/data -- the new release system needs to do public-only tests,
and will catch this sort of thing.
The public integration test VariantContextIntegrationTest was dependent on the
private walker TestVariantContextWalker. Moved this walker to public/java/test
(NOT public/java/src, since this walker is only used by the test suite) to avoid
errors during public-only tests.
-- MD5 db had spelling error; fixed
-- Bug in AlignmentUtils resulted in some bases not being color space corrected. The integration test caught the change, and it's clear that the new version is correct, as the prev. version was not considering the last the N qualities for reads with a ND operation.
-- SimpleTimer is now threadsafe using synchronized method keywords
-- Bug fix for alignmentToByteArray() where the N case was refPos++ not the now correct refPos += elementLength
Having SnpEff grouped with the Experimental annotations was proving problematic, since it
requires a rod. Placing it in its own group should improve the situation somewhat, making it
easier to request "all annotations except for SnpEff".
This allows the annotation classes to perform any necessary initialization/validation.
For example, it allows the SnpEff annotator to (among other things) validate its rod binding.
This will prevent a NullPointerException when SnpEff annotation is requested but no rod binding
is present.
Added an integration test to cover this case so that it doesn't break again.
-- Changes associated code throughout the codebase
-- Updated necessary (but minimal) UnitTests to reflect new behavior
-- Much better makealleles() function in VC.java that enforces a lot of key constraints in VC
-- For no apparent reason we were using a HashSet to store the ReadFilters, so the order of operations was really arbitrarily applied. The order now is
(1) the order of the walker intrinsic filters
(2) read group black list (if provided)
(3) command line filters (if provided)
classloading of bcel*.jar/ant-apache-bcel*.jar. Switching instead to manually
specifying a minimal set of packages/classes to include in the vcf.jar via
build.xml, and adding a unit test which creates a limited classloader
only aware of vcf.jar and tribble.jar and tries to use it to load the core
classes in the vcf jar.
Hopefully third time's the charm.
-- The GATK will now throw a user exception if it opens a SAM/BAM file that doesn't have at least one RG defined
-- LIBS again throws an error if the complete list of samples isn't provided
-- Updating ExmpleCountLociPipeline test to use the well-formated versions of the exampleBAM and exampleFASTA files in testdata, instead of the old broken ones in validation_data.
-- Convenience constructors for UserExceptions.MalformedBAM
-- Right now we cannot process BAM files without read groups because we enforce the samples list to not be empty when there's a SAM record. Now if there are reads and there are no samples we add the "null" sample so that LIBS walks the reads properly
-- the underlying data structure is still present, but until I decide what to do for the extensible system I've completely disabled the subsystem
-- Added code to merge Samples, so that a mostly full record can be merged with a consistent empty record. If the two records are inconsistent, an error is thrown
-- addSample() in Sample.class now invokes mergeSample() when appropriate
-- Validation types are now only STRICT or SILENT
-- Validation code implemented in SampleDBBuilder
-- Extensive unit tests for SampleDBBuilder
-- Passes significiant unit tests
-- Implicit sample creation for mom / dad when you create single samples
-- Continuing cleanup of Sample and SampleDataSource
-- These could be simplied in their downstream uses
-- Or they could be replaced with a generic getSAMFileHeaders() function and then apply the getSamples(header) as desired downstream
-- A nearly identical piece of code already lived in SampleUtils. Now there are two functions, one taking a regular header and another grabbing the merged header from the GATK engine itself. Much cleaner
If both ends of the interval falls within a deletion in the read then hardClipBothEnds would cut the right tail first including the entire deletion, then fail to cut the left tail because there would not be any bases there anymore. Fixed.
The base qualities of a consensus reads are now the average quality of the bases forming the consensus base (most common base) and the consensus quality tag now carry an array with the counts of each base in the consensus. This should increase file size but improve calling sensitivity/specificity.
* Includes tests that include HardClip to Read and Reference Coords.
* Changed ReadUtils.HardClipByReferenceCoordinates from private to protected to allow for testing
when hard clipping the right tail of a read falls inside a deletion, clipping should fall back to the last base before the deletion to follow the ReadClipper's contract.
* RR will now compress reads that span across multiple intervals correctly and output them in the correct order.
* Fixed bug in getReadCoordinateForReferenceCoordinate where if the requested reference coordinate fell inside a deletion in the read the read would be clipped up to one element past the deletion.
-- resulted in massive code cleanup
-- GdA will integrate his new banded algorithm here
-- Removed: DO_CONTEXT_DEPENDENT_PENALTIES, GET_GAP_PENALTIES_FROM_DATA, INDEL_RECAL_FILE, dovit, GSA_PRODUCTION_ONLY
-- UnitTests for key functions on reduced reads
-- PileupElement calls static functions in ReadUtils
-- Simple routine that takes a reduced read and fills in its quals with its reduced qual
b) Change md5 to reflect records that are now merged correctly.
c) Change unit merge alleles test to reflect the fact that a null non-variant vc object is not valid and not supported because there's no way to codify such object in a vcf. The code correctly converts this to a non-variant single-base event with whatever the reference is at that location.
With the current implementation, a read cannot start with a deletion or an insertion. Maybe this will change in the future, but for now, chop the leading insertion off.
b) First reimplementation of new vc merger of different types. Previous version did it in two steps, first merging all vc's per type and then trying to see if resulting vc's would be merged if alleles of one type were a subset of another, but this won't work when uniquifying genotypes since sample names would be messed up and GT sample names wouldn't match VC sample names. Now, it's actually simpler: when splitting vc's by type before merging, we check for alleles of one vc being a subset of alleles of vc of another type and if so we put them together in same list.
* Deletions now count as hard clipped bases in order to recover the original alignment start of a clipped read.
* Insertions do not count as hard clipped bases for the same reason.
* This created a bug in the previous cigar cleaning function. Fixed.
-We now assign a functional class (nonsense, missense, silent, or none) to each SnpEff effect, and add a
SNPEFF_FUNCTIONAL_CLASS annotation to the INFO field of the output VCF.
-Effects are now prioritized according to both biological impact and functional class, instead of impact only.
-Many of SnpEff's "low-impact" effects are now classified as "modifiers" with lower priority than every
other effect. This includes such "effects" as DOWNSTREAM, UPSTREAM, INTRON, GENE, EXON, and others that
really describe the location of the variant rather than its biological effect.
This code will be short-lived (likely 1.2-only), as the next version of SnpEff will include most of these
features directly.
Checking this change into Stable+Unstable instead of Unstable because the current functional class stratification
in VariantEval is basically broken and urgently needs to be fixed for production purposes.
if soft clipped bases were after a hard clipped section of the read, the hard clip was clipping the left soft clip tail as if it were a right tail. Mayhem.
* Hard clipped Cigar now includes all insertions that were hard clipped and not the deletions.
* The alignment start is now recalculated according to the new hard clipped cigar representation
Big (but not major) cleanup of code in ILG - mostly excising the old likelihood model
Activated the early-abort check for ILG. I think it should be better this way.
Be careful when using this - if you're writing a bam file it will be potentially written out of order (since the previous alignment start was at the M, not the S).
Pre-softclipped reads (with high qual) are a complicated event to deal with in the Reduced Reads environment. I chose to hard clip them out for now and added a todo item to bring them back on in the future, perhaps as a variant region.
-- Old code required qual to be <64, which isn't strictly necessary. Now uses the Picard SAMUtils.MAX_PHRED_SCORE constant
-- Unittest to enforce this behavior
-- Now handles multiple records at a site, so that you don't see records like set=dbsnp-dbsnp-dbsnp when combining something with dbsnp
-- Proper handling of ids. If you are merging files with multiple ids for the same record, the ids are merged into a comma separated list
This change is urgently required for production, which is why it's going into Stable+Unstable
instead of just Unstable.
The keys for the SnpEff version and command header lines in the VCF file output by
VariantAnnotator (OriginalSnpEffVersion and OriginalSnpEffCmd) are intentionally
different from the keys for those same lines in the SnpEff output file (SnpEffVersion
and SnpEffCmd), so that output files from VariantAnnotator won't be confused
with output files from SnpEff itself.
-- Previously, on the fly indices didn't have dictionary set on the fly, so the GATK would read, add dictionary, and rewrite the index. This is now fixed, so that the on the fly index contains the reference dictionary when first written, avoiding the unnecessary read and write
-- Added a GenomeAnalysisEngine and Walker function called getMasterSequenceDictionary() that fetches the reference sequence dictionary. This can be used conveniently everywhere, and is what's written into the Tribble index
-- Refactored tribble index utilities from RMDTrackBuilder into IndexDictionaryUtils
-- VCFWriter now requires the master sequence dictionary
-- Updated walkers that create VCFWriters to provide the master sequence dictionary
Now the functions getRefCoordSoftUnclippedStart and getRefCoordSoftUnclippedEnd will return getUnclippedStart if the read is all contained within an insertion. Updated the contracts accordingly. This should give the same behavior as the GenomeLocParser now.
-- No functional changes (my algorithm wouldn't work)
-- Major structural cleanup (returning more basic data structures that allow us to development new algorithm)
-- Unit tests for the efficiency of interval partitioning
The ClippingOp clip cigar function would run into a endless loop if the parameter were out of the reads range, I stopped the bug.
* There is no check to make sure the read coordinate are covered by the read though
When Hard clipping to interval, I added a check for deletions.
NOTE: method works for NA12878 WEx but needs to be more thoroughly tested/optimized
After discussing this with Mark, it seems clear that the old version of the
VariantEval FunctionalClass stratification is preferable to this version.
By reverting, we maintain backwards compatibility with legacy output files
from the old GenomicAnnotator, and can add SnpEff support later without
breaking that backwards compatibility.
This reverts commit b44acd1abd9ab6eec37111a19fa797f9e2ca3326.
This is a temporary and hopefully short-lived solution. I've modified
the FunctionalClass stratification to stratify by effect impact as
defined by SnpEff annotations (high, moderate, and low impact) rather
than by the silent/missense/nonsense categories.
If we want to bring back the silent/missense/nonsense stratification,
we should probably take the approach of asking the SnpEff author
to add it as a feature to SnpEff rather than coding it ourselves,
since the whole point of moving to SnpEff was to outsource genomic
annotation.
-Rewrote SnpEff support in VariantAnnotator to support the latest SnpEff release (version 2.0.2)
-Removed support for SnpEff 1.9.6 (and associated tribble codec)
-Will refuse to parse SnpEff output files produced by unsupported versions (or without a version tag)
-Correctly matches ref/alt alleles before annotating a record, unlike the previous version
-Correctly handles indels (again, unlike the previous version
-- Cannot reproduce the very long waits reported by some users.
-- Fixed problem that exception might result in an undeleted file, which is now fixed with deleteOnExit()
All VariantAnnotator annotation classes may now have an (optional) initialize() method
that gets called by the VariantAnnotatorEngine ONCE before annotation starts.
As an example of how this can be used, the SnpEff annotation class will use the initialize()
method to check whether the SnpEff version number stored in the vcf header is a supported
version, and also to verify that its required RodBinding is present.
b) Added (but left commented out since it may affect integration tests and to isolate commits) fix to per-sample DP reporting, so that deletions are included in count.
c) Bug fix to avoid having non-reference genotypes assigned to samples with PL=0,0,0. Correct behavior should be to no-call these samples, and to ignore these samples when computing AC distribution since their likelihoods are not informative.
-- Support for streaming VCF writing via the VCFWriter interface
-- GCF now has a header and a footer. The header is minimal, and contains a forward pointer to the position of the footer in the file.
-- Readers now read the header, and then jump to the footer to get the rest of the "header" information
-- Version now a field in GCF
-- per sample stratification was not being calculated correctly. The alt allele was always remaining, even if the genotype of the sample was hom-ref. Although conceptually fine, this breaks the assumptions of all of the eval modules, so per sample stratifications actually included all variants for everything. Eric is going to fix the system in general, so this commit may break the build.
-- ArrayList are List where possible
-- states refactored into VariantStratifier base class (reduces many lines of duplicate code)
-- Added VariantType stratification that partitions report by VariantContext.Type
- Instead of using readLength, the ReadUtil function are used to get a proper read coordinate
- Added debug info in interval clipping ( with -dl)
NOTE: method might not be safe for production and checks need to be added to the ClippingOp code
Updates for HybridSelectionPipeline:
- Use VQSR on SNPs for projects using bait set whole_exome_agilent_1 and applying cut at 98.5.
- If a whole_exome_agilent_1 project has less than 50 samples also mixing in 1000G samples to reach VQSR thresholds.
- Updated SNP hard filters based on analysis done with ebanks to approximate VQSR results on small target batches.
- Removed GSA_PRODUCTION_ONLY flag from indel caller.
- Updated indel hard filters based on delangel's analysis.
- Updated HybridSelectionPipelineTest to use HARD SNP filters only, for now.
-- Very minimal working version that can read / write binary VCFs with genotypes
-- Already 10x faster for sites, 5x for fully parsed genotypes, and 1000x for skipping genotypes when reading
-- Removed versions getAttribriteAsX(key) that except on not having the value.
-- Removed version that getAttributeAsXNoException(key)
-- The only available assessors are now getAttributeAsX(key, default).
-- This single accessors properly handle their argument types, so if the value is a double it is returned directly for getAttributeAsDouble(), or if it's a string it's converted to a double. If the key isn't found, default is returned.
-- Don't create an empty LinkedHashSet() for PASS fields. Just return Collections.emptySet() instead.
-- For filter fields with actual values, returns an unmodifiableSet instead of one that can be changed
Found this neat little walker Kiran wrote stashed in the private tree. Very useful. Generalized it a bit, added GATKDocs and moved it to public. I might include it as a QC step on the pacbio processing pipeline.
* generalize it so it works with non pair ended reads.
* generalize it to work with no read group information
* getRefCoordSoftUnclippedEnd was not resetting the shift when hitting insertions. Fixed.
* getReadCoordinateForReferenceCoordinateBeforeAlignmentEnd was returning the wrong read coordinate position. Fixed.
The clipper could leave an insertion or deletion as the start or end of a read after hardclipping a read if the element adjacent to the clipping point was an indel. Fixed.
Read clipper now identifies and clips even if the requested coordinate is outside the alignment but the read contains soft clipped bases in that region.
* When hard clipping a read that had insertions in it, the insertion was being added to the cigar string's hard clip element. This way, the old UnclippedStart() was being modified and so was the calculation of the new AlignmentStart(). Fixed it by subtracting the number of insertions clipped from the total number of hard clipped bases.
* Walker was sending read instead of filtered read when deleting a read that contains only Q2 bases
* Sliding the window was causing reads that started on the new start position to be entirely clipped.
-- General purpose RScript executor in java (please use when invoking RScripts)
-- Removed groupName. This is now analysisName
-- Explicitly added capability to enable/disable individual QFunction
b) More useful AC,AF logging in VariantsToTable with multiallelic sites: instead of logging comma-separated values, log max value by default. Hidden, experimental argument -logACSum to log sum of ACs instead. This is due to extreme slowness of R in parsing strings to tokens and computing max/sum itself (~100x slower than gatk).
c) Added integrationtest for new SelectVariants commands
-- Useful if you want to have a parameter like MAX_RECORDS that wants the walker to stop after some number of map calls without having to resort to the old System.exit() call directly.
Reads that were not hard clipped for the variable site no longer show up in output file
Walker now uses unclippedStart of Read to determine position in the sliding Window
It is time to bring the ReadClipper class to the main repo. Read Clipper has tested functionality for soft and hard clipping reads. I will prepare thorough documentation for it as it will be very useful for the assembler and the GATK in general.
b) More R-friendly VariantsToTable printing of AC in case of multiple alt alleles
c) Rename FixPLOrderingWalker to FixGenotypesWalker and rewrote: no longer need older code, replaced with code to replace genotypes with all-zero PL's with a no-call.
- Ability to pass a different resident memory reservation and limits. Useful for large pileups of low pass genome data that sometimes need high -Xmx6g but usually don't exceed 2-3g in actual heap size.
- Fixed jobPriority to work for all job runners. Now must be a integer between 0 and 100- even for GridEngine- and will be mapped to the correct values.
- Passing parallel environment and job resource requests to LSF and GridEngine. Useful for passing tokens like iodine_io=1 and -pe pe_slots 8
- Refactored GridEngine JobRunner to also provide basic support for other job dispatchers with DRMAA implementations such as Torque/PBS. Should work for basic running but advanced users must pass their own jobNativeArgs from the command line or in customized QScripts until someone maps properties like jobQueue, jobPriority, residentRequest, etc. into a Torque/PBS/etc. dispatcher.
Misc updates to WholeGenomeIndelCalling.scala
Bug fix in VariantEval (may be temporary, need more investigation): if -disc option is used in sites-only vcf's then a null pointer exception is produced, caused by recent introduction of -xl_sf options.
-- Sorting of ArgumentSources now done in GATKDoclet, not in the ParsingEngine, as the system depends on the LinkedTreeMap
-- Fixed broken exception throwing in the case where a file's type could not be determined
-- ArgumentSources are now sorted by case insensitive names, so arguments are shown in alphabetical order (Ryan)
-- @Advanced annotation can be used to indicate that an argument is an advanced option and should be visually deemphasized in the GATKs. There's now an advanced section. Mauricio or Ryan -- could you figure out how to make this section less prominent in the style.css?
-- Allowed values for RodBinding<T> are displayed in the GATKDocs
-- Longest name up to 30 characters is chosen for main argument list (suggested by Ryan/Mauricio)
-- Features are listed in alphabetical order
-- Moved useful getParameterizedType() function to JVMUtils
-- Tests of these features in the Documentation Test
VariantEval module CountVariants is corrected and an additional column is added so that we log mixed events and complex indels separately (before they were being conflated).
VariantEval module IndelStatistics is considerably simplified as the sample stratification was wrong and redundant, now it should work with the VE-generic Sample stratification. Several columns are renamed or removed since they're not really useful
-- Now supports a static list of root classes / interfaces that should receive docs. A complementary approach to documenting features to the DocumentedGATKFeature annotation
-- Tribble codecs are now documented!
-- No longer displayed sub and super classes
-- refdata/features now in utils/codecs with the other codecs
-- Deleted dbsnpHelper. rsID function now in VCFutils. Remaining code either deleted or put into VariantContextAdaptors
-- Many associated import updates due to code move
-- RefSeqCodec bug: getFeatureClass() returned RefSeqCodec.class, not RefSeqFeature.class. Really should change this in Tribble to require Class<T extends Feature> to get compile time type checking
-- Better error messages that actually list the available tribble types, when there's a type error
can produce data loss in a narrow corner case where the BGZF block(s) locations
and offsets in the last index bucket of contig n overlap exactly with the BGZF
block locations and offset in the last index bucket of contig n+1.
A proper fix that keeps the optimization has already been introduced into
unstable, but disabling the optimization is a low risk way to make sure that
users of stable experience no data loss.
Mark DePristo
Laurent Francioli
Ryan Poplin
David Roazen
Guillermo del Angel
Khalid Shakir
Mauricio Carneiro
Christopher Hartl
Eric Banks
Matt Hanna
Roger Zurawicki
Andrey Sivachenko
Christopher Hartl
Menachem Fromer