-- old algorithm was O(kmerSize * readLen) for each read. New algorithm is O(readLen)
-- Added real unit tests for the addKmersFromReads to the graph. Using a builder is great because we can create a MockBuilder that captures all of the calls, and then verify that all of the added kmers are the ones we'd expect.
-- The previous creation algorithm used the following algorithm:
for each kmer1 -> kmer2 in each read
add kmers 1 and 2 to the graph
add edge kmer1 -> kmer2 in the graph, if it's not present (does check)
update edge count by 1 if kmer1 -> kmer2 already existed in the graph
-- This algorithm had O(reads * kmers / read * (getEdge cost + addEdge cost)). This is actually pretty expensive because get and add edges is expensive in jgrapht.
-- The new approach uses the following algorithm:
for each kmer1 -> kmer2 in each read
add kmers 1 and 2 to a kmer counter, that counts kmer1+kmer2 in a fast hashmap
for each kmer pair 1 and 2 in the hash counter
add edge kmer1 -> kmer2 in the graph, if it's not present (does check) with multiplicity count from map
update edge count by count from map if kmer1 -> kmer2 already existed in the graph
-- This algorithm ensures that we add very much fewer edges
-- Additionally, created a fast kmer class that lets us create kmers from larger byte[]s of bases without cutting up the byte[] itself.
-- Overall runtimes are greatly reduced using this algorith
-- When the alignments are sufficiently apart from each other all the scores in the sw matrix could be negative which screwed up the max score calculation since it started at zero.
-Bamboo now uses the local mirrors for git checkouts, which should hopefully
resolve the intermittent long delays we've been seeing in Bamboo during git
operations
-Use a daemon process rather than a cron job to guarantee strict serial
execution of the git updates (since the time git update operations
require can vary). The spawn script for the daemon runs as a cron job
instead to make sure the daemon is always running and restart it if
necessary.
-- The previous version would enter into an infinite loop in the case where we have a graph that looks like:
X -> A -> B
Y -> A -> B
So that the incoming vertices of B all have the same sequence. This would cause us to remodel the graph endless by extracting the common sequence A and rebuilding exactly the same graph. Fixed and unit tested
-- Additionally add a max to the number of simplification cycles that are run (100), which will throw an error and write out the graph for future debugging. So the GATK will always error out, rather than just go on forever
-- After 5 rounds of simplification we start keeping a copy of the previous graph, and then check if the current graph is actually different from the previous graph. Equals here means that all vertices have equivalents in both graphs, as do all edges. If the two graphs are equal we stop simplifying. It can be a bit expensive but it only happens when we end up cycling due to the structure of the graph.
-- Added a unittest that goes into an infinite loop (found empirically in running the CEU trio) and confirmed that the new approach aborts out correctly
-- #resolves GSA-924
-- See https://jira.broadinstitute.org/browse/GSA-924 for more details
-- Update MD5s due to change in assembly graph construction
-- HC now throws a UserException if this model is provided. Documented this option as not being supported in the HC in the docs for EXACT_GENERAL_PLOIDY
Problem:
--------
Print Reads was running out of disk space when using the -BQSR option even for small bam files
Solution:
---------
Configure setupWriter to expect pre sorted reads
-- Add a maximum per sample and overall maximum number of reads held in memory by the ART at any one time. Does this in a new TAROrderedReadCache data structure that uses a reservior downsampler to limit the total number of reads to a constant amount. This constant is set to be by default 3000 reads * nSamples to a global maximum of 1M reads, all controlled via the ActiveRegionTraversalParameters annotation.
-- Added an integration test and associated excessively covered BAM excessiveCoverage.1.121484835.bam (private/testdata) that checks that the system is operating correctly.
-- #resolves GSA-921
-- This method provides client with the current number of elements, without having to retreive the underlying list<T>. Added unit tests for LevelingDownsampler and ReservoirDownsampler as these are the only two complex ones. All of the others are trivially obviously correct.
-- The function getReducedCounts() was returning the undecoded reduced read tag, which looks like [10, 5, -1, -5] when the depths were [10, 15, 9, 5]. The only function that actually gave the real counts was getReducedCount(int i) which did the proper decoding. Now GATKSAMRecord decodes the tag into the proper depths vector so that getReduceCounts() returns what one reasonably expects it to, and getReduceCount(i) merely looks up the value at i. Added unit test to ensure this behavior going forward.
-- Changed the name of setReducedCounts() to setReducedCountsTag as this function assumes that counts have already been encoded in the tag way.
-- Bugfix to puts all files in the subdirectory, regardless of whether the outputDir is provided with a ending / or not
-- UG now runs single threaded in GeneralCallingPipeline
-- GCP HC only needs 2 GB now
-- Extension increased to 200 bp
-- Min prune factor defaults to 0
-- LD merging enabled by default for complex variants, only when there are 10+ samples for SNP + SNP merging
-- Active region trimming enabled by default
-- The kbest paths algorithm now takes an explicit set of starting and ending vertices, which is conceptually cleaner and works for either the cycle or no-cycle models. Allowing cycles can be re-enabled with an HC command line switch.
-- The previous likelihood calculation proceeds as normal, but after each read has been evaluated against each haplotype we go through the read / allele / likelihoods map and eliminate all reads that have poor fit to any of the haplotypes. This functionality stops us from making a particular type of error in the HC, where we have a haplotype that's very far from the reference allele but not the right true haplotype. All of the reads that are slightly closer to this FP haplotype than the reference previously generated enormous likelihoods in favor of this FP haplotype because they were closer to it than the reference, even if each read had many mismatches w.r.t. the FP haplotype (and so the FP haplotype was a bad model for the true underlying haplotype).
-- Trims down active regions and associated reads and haplotypes to a smaller interval based on the events actually in the haplotypes within the original active region (without extension). Radically speeds up calculations when using large active region extensions. The ActiveRegion.trim algorithm does the best job it can of trimming an active region down to a requested interval while ensuring the resulting active region has a region (and extension) no bigger than the original while spanning as much of the requested extend as possible. The trimming results in an active region that is a subset of the previous active region based on the position and types of variants found among the haplotypes
-- Retire error corrector, archive old code and repurpose subsystem into a general kmer counter. The previous error corrector was just broken (conceptually) and was disabled by default in the engine. Now turning on error correction throws a UserException. Old part of the error corrector that counts kmers was extracted and put into KMerCounter.java
-- Add final simplify graph call after we prune away the non-reference paths in DeBruijnAssembler
-- outgoingVerticesOf and incomingVerticesOf return a list not a set now, as the corresponding values must be unique since our super directed graph doesn't allow multiple edges between vertices
-- Make DeBruijnGraph, SeqGraph, SeqVertex, and DeBruijnVertex all final
-- Cache HashCode calculation in BaseVertex
-- Better docs before the pruneGraph call
-- The previous version of the head merging (and tail merging to a lesser degree) would inappropriately merge source and sinks without sufficient evidence to do so. This would introduce large deletion events at the start / end of the assemblies. Refcatored code to require 20 bp of overlap in the head or tail nodes, as well as unit tested functions to support this.
-- Goes through the graph looking for chains to zip, accumulates the vertices of the chains, and then finally go through and updates the graph in one big go. Vastly more efficient than the previous version, but unfortunately doesn't actually work now
-- Also incorporate edge weight propagation into SeqGraph zipLinearChains. The edge weights for all incoming and outgoing edges are now their previous value, plus the sum of the internal chain edges / n such edges
-- These events always occur on the very edge of the haplotypes, and are intrinsically dodgy. So instead of emitting them and then potentially having to deal with merging real basepair events into them we just no longer emit those events.
-- Moved R^2 LD haplotype merging system to the utils.haplotype package
-- New LD merging only enabled with HC argument.
-- EventExtractor and EventExtractorUnitTest refactors so we can test the block substitution code without having to enabled it via a static variable
-- A few misc. bug fixes in LDMerger itself
-- Refactoring of Haplotype event splitting and merging code
-- Renamed EventExtractor to EventMap
-- EventMap has a static method that computes the event maps among n haplotypes
-- Refactor Haplotype score and base comparators into their own classes and unit tested them
-- Refactored R^2 based LD merging code into its own class HaplotypeR2Calculator and unit tested much of it.
-- LDMerger now uses the HaplotypeR2Calculator, which cleans up the code a bunch and allowed me to easily test that code with a MockHaplotypeR2Calculator. For those who haven't seen this testing idiom, have a look, and very useful
-- New algorithm uses a likelihood-ratio test to compute the probability that only the phased haplotypes exist in the population.
-- Fixed fundamental bug in the way the previous R^2 implementation worked
-- Optimizations for HaplotypeLDCalculator: only compute the per sample per haplotype summed likelihoods once, regardless of how many calls there are
-- Previous version would enter infinite loop if it merged two events but the second event had other low likelihood events in other haplotypes that didn't get removed. Now when events are removed they are removed from all event maps, regardless of whether the haplotypes carry both events
-- Bugfixes for EventMap in the HaplotypeCaller as well. Previous version was overly restrictive, requiring that the first event to make into a block substitution was a snp. In some cases we need to merge an insertion with a deletion, such as when the cigar is 10M2I3D4M. The new code supports this. UnitTested and documented as well. LDMerger handles case where merging two alleles results in a no-op event. Merging CA/C + A/AA -> CAA/CAA -> no op. Handles this case by removing the two events. UnitTested
-- Turn off debugging output for the LDMerger in the HaplotypeCaller unless -debug was enabled
-- This new version does a much more specific test (that's actually right). Here's the new algorithm:
* Compute probability that two variants are in phase with each other and that no
* compound hets exist in the population.
*
* Implemented as a likelihood ratio test of the hypothesis:
*
* x11 and x22 are the only haplotypes in the populations
*
* vs.
*
* all four haplotype combinations (x11, x12, x21, and x22) all exist in the population.
*
* Now, since we have to have both variants in the population, we exclude the x11 & x11 state. So the
* p of having just x11 and x22 is P(x11 & x22) + p(x22 & x22).
*
* Alternatively, we might have any configuration that gives us both 1 and 2 alts, which are:
*
* - P(x11 & x12 & x21) -- we have hom-ref and both hets
* - P(x22 & x12 & x21) -- we have hom-alt and both hets
* - P(x22 & x12) -- one haplotype is 22 and the other is het 12
* - P(x22 & x21) -- one haplotype is 22 and the other is het 21
-- This fixes edge base bugs where non-consolidated cigars are causing problems in users of the Haplotype object. Input arguments are now checks (let's see if we blow up)
-- Picard extension so Queue scripts can use FastqToSam
-- Single-sample BAM processing: merge/trim reads + BWA + IR + MD + BQSR. Mostly identical to standard pipeline,
except for the adaptor trimming/merging which is critical for short-insert libraries.
-- Single-sample calling (experimental, work in progress): standard UG run but outputting at all sites, meant for
deep whole genomes.
New scripts
Problem:
--------
PairHMM was generating positive likelihoods (even after the re-work of the model)
Solution:
---------
The caching idices were never re-initializing the initial conditions in the first position of the deletion matrix. Also the match matrix was being wrongly initialized (there is not necessarily a match in the first position). This commit fixes both issues on both the Logless and the Log10 versions of the PairHMM.
Summarized Changes:
------------------
* Redesign the matrices to have only 1 col/row of padding instead of 2.
* PairHMM class now owns the caching of the haplotype (keeps track of last haplotypes, and decides where the caching should start)
* Initial condition (in the deletionMatrix) is now updated every time the haplotypes differ in length (this was wrong in the previous version)
* Adjust the prior and probability matrices to be one based (logless)
* Update Log10PairHMM to work with prior and probability matrices as well
* Move prior and probability matrices to parent class
* Move and rename padded lengths to parent class to simplify interface and prevent off by one errors in new implementations
* Simple cleanup of PairHMMUnitTest class for a little speedup
* Updated HC and UG integration test MD5's because of the new initialization (without enforcing match on first base).
* Create static indices for the transition probabilities (for better readability)
[fixes#47399227]
* As reported here: http://gatkforums.broadinstitute.org/discussion/comment/4270#Comment_4270
* This was a commit into the variant.jar; the changes here are a rev of that jar and handling of errors in VF
* Added integration test to confirm failure with User Error
* Removed illegal header line in KB test VCF that was causing related tests to fail.
* Very trivial, but I happened to see this code and it drove me nuts so I felt compelled to refactor it.
* Instead of iterating over keys in map to get the values, just iterate over the values...
Mark DePristo
Ryan Poplin
Eric Banks
droazen
Mauricio Carneiro
Guillermo del Angel