PairHMMSyntheticBenchmark and PairHMMEmpirical benchmark were written to test the banded pairHMM, and were archived along with it. I returned them to the test directory for use in benchmarking the ArrayLoglessPairHMM. I commented out references to the banded pairHMM (which was left in archive), rather than removing those references entirely.
Renamed PairHMMEmpiricalBenchmark to PairHMMBandedEmpiricalBenchmark and returned it to the archive. It has a few problems for use as a general benchmark, including initializing the HMM too frequently and doing too much setup work in the 'time' method. However, since the size selection and debug printing are useful for testing the banded implementation, I decided to keep it as-is and archive it alongside with the other banded pairHMM classes. I did fix one bug that was causing the selectWorkingData function to return prematurely. As a result, the benchmark was only evaluating 4-40 pairHMM calls instead of the desired "maxRecords".
I wrote a new PairHMMEmpiricalBenchmark that simply works through a list of data, with setup work and hmm-initialization moved to its own function. This involved writing a new data read-in function in PairHMMTestData. The original was not maintaining the input data in order, the end result of which would be an over-estimate of how much caching we are able to do. The new benchmark class more closely mirrors real-world operation over large data.
It might be cleaner to fix some of the issues with the BandedEmpiricalBenchmark and use one read-in function. However, this would involve more extensive changes to:
PairHMMBandedEmpiricalBenchmark
PairHMMTestData
BandedLoglessPairHMMUnitTest
I decided against this as the banded benchmark and unit test are archived.
A new PairHMM implementation for read/haplotype likelihood calculations. Output is the same as the LOGLESS_CACHING version.
Instead of allocating an entire (read x haplotype) matrix for each HMM state, this version stores sub-computations in 1D arrays. It also accesses intersections of the (read x haplotype) alignment in a different order, proceeding over "diagonals" if we think of the alignment as a matrix.
This implementation makes use of haplotype caching. Because arrays are overwritten, it has to explicitly store mid-process information. Knowing where to capture this info requires us to look ahead at the subsequent haplotype to be analyzed. This necessitated a signature change in the primary method for all pairHMM implementations.
We also had to adjust the classes that employ the pairHMM:
LikelihoodCalculationEngine (used by HaplotypeCaller)
PairHMMIndelErrorModel (used by indel genotyping classes)
Made the array version the default in the HaplotypeCaller and the UnifiedArgumentCollection.
The latter affects classes:
ErrorModel
GeneralPloidyIndelGenotypeLikelihoodsCalculationModel
IndelGenotypeLikelihoodsCalculationModel
... all of which use the pairHMM via PairHMMIndelErrorModel
1. Some minor refactorings and claenup (e.g. removing unused imports) throughout.
2. Updates to the KB assessment functionality:
a. Exclude duplicate reads when checking to see whether there's enough coverage to make a call.
b. Lower the threshold on FS for FPs that would easily be filtered since it's only single sample calling.
3. Make the HC consistent in how it treats the pruning factor. As part of this I removed and archived
the DeBruijn assembler.
4. Improvements to the likelihoods for the HC
a. We now include a "tristate" correction in the PairHMM (just like we do with UG). Basically, we need
to divide e by 3 because the observed base could have come from any of the non-observed alleles.
b. We now correct overlapping read pairs. Note that the fragments are not merged (which we know is
dangerous). Rather, the overlapping bases are just down-weighted so that their quals are not more
than Q20 (or more specifically, half of the phred-scaled PCR error rate); mismatching bases are
turned into Q0s for now.
c. We no longer run contamination removal by default in the UG or HC. The exome tends to have real
sites with off kilter allele balances and we occasionally lose them to contamination removal.
5. Improved the dangling tail merging implementation.
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]
The current implementation of the PairHMM had issues with the probabilities and the state machines. Probabilities were not adding up to one because:
# Initial conditions were not being set properly
# Emission probabilities in the last row were not adding up to 1
The following commit fixes both by
# averaging all potential start locations (giving an equal prior to the state machine in it's first iteration -- allowing the read to start it's alignment anywhere in the haplotype with equal probability)
# discounting all paths that end in deletions by not adding the last row of the deletion matrix and summing over all paths ending in matches and insertions (this saves us from a fourth matrix to represent the end state)
Summarized changes:
* Fix LoglessCachingPairHMM and Log10PairHMM according to the new algorithm
* Refactor probabilities check to throw exception if we ever encounter probabilities greater than 1.
* Rename LoglessCachingPairHMM to LoglessPairHMM (this is the default implementation in the HC now)
* Rename matrices to matchMatrix, insertionMatrix and deletionMatrix for clarity
* Rename metric lengths to read and haplotype lengths for clarity
* Rename private methods to initializePriors (distance) and initializeProbabilities (constants) for clarity
* Eliminate first row constants (because they're not used anyway!) and directly assign initial conditions in the deletionMatrix
* Remove unnecessary parameters from updateCell()
* Fix the expected probabilities coming from the exact model in PairHMMUnitTest
* Neatify PairHMM class (removed unused methods) and PairHMMUnitTest (removed unused variables)
* Update MD5s: Probabilities have changed according to the new PairHMM model and as expected HC and UG integration tests have new MD5s.
[fix 47164949]
The issue here is that the OptimizedLikelihoodTestProvider uses the same basic underlying class as the
BasicLikelihoodTestProvider and we were using the BasicTestProvider functionality to pull out tests of
that class; so if the optimized tests were run first we were unintentionally running those same tests
again with the basic ones (but expecting different results).
-- Uses 1/N for N potential start sites as the probability of starting at any one of the potential start sites
-- Add flag that says to use the original edge condition, respected by all subclasses. This brings the new code back to the original state, but with all of the cleanup I've done
-- Only test configurations where the read length <= haplotype length. I think this is actually the contract, but we'll talk about this tomorrow
-- Fix egregious bug with the myLog10SumLog10 function doing the exact opposite of the requested arguments, so that doExact really meant don't do exact
-- PairHMM now exposes computeReadLikelihoodGivenHaplotypeLog10 but subclasses must overload subComputeReadLikelihoodGivenHaplotypeLog10. This protected function does the work, and the public function will do argument and result QC
-- Have to be more tolerant of reference (approximate) HMM. All unit tests from the original HMM implementations pass now
-- Added locs of docs
-- Generalize unit tests with multiple equivalent matches of read to haplotype
-- Added runtime argument checking for initial and computeReadLikelihoodGivenHaplotypeLog10
-- Functions to dumpMatrices for debugging
-- Fix nasty bug (without original unit tests) in LoglessPairHMM
-- Max read and haplotype lengths only worked in previous code if they were exactly equal to the provided read and haplotype sizes. Fixed bug. Added unit test to ensure this doesn't break again.
-- Added dupString(string, n) method to Utils
-- Added TODOs for next commit. Need to compute number of potential start sites not in initialize but in the calc routine since this number depends not on the max sizes but the actual read sizes
-- Unit tests for the hapStartIndex functionality of PairHMM
-- Moved computeFirstDifferingPosition to PairHMM, and added unit tests
-- Added extensive unit tests for the hapStartIndex functionality of computeReadLikelihoodGivenHaplotypeLog10
-- Still TODOs left in the code that I'll fix up
-- Logless now compute constants, if they haven't been yet initialized, even if you forgot to say so
-- General: the likelihood penalty for potential start sites is now properly computed against the actual read and reference bases, not the maximum. This involved moving some initialize() code into the computeLikelihoods function. That's ok because all of the potential log10 functions are actually going to cached versions, so the slowdown is minimal
-- Added some unit tests to ensure that common errors (providing haplotypes too long, reads too long, not initializing the HMM) are captured as errors
-- Would have been squashed but could not because of subsequent deletion of Caching and Exact/Original PairHMMs
-- Actual working unit tests for PairHMMUnitTest
-- Fixed incorrect logic in how I compared hmm results to the theoretical and exact results
-- PairHMM has protected variables used throughout the subclasses
This is an intermediate commit so that there is a record of these changes in our
commit history. Next step is to isolate the test classes as well, and then move
the entire package to the Picard repository and replace it with a jar in our repo.
-Removed all dependencies on org.broadinstitute.sting (still need to do the test classes,
though)
-Had to split some of the utility classes into "GATK-specific" vs generic methods
(eg., GATKVCFUtils vs. VCFUtils)
-Placement of some methods and choice of exception classes to replace the StingExceptions
and UserExceptions may need to be tweaked until everyone is happy, but this can be
done after the move.
bradtaylor
Mauricio Carneiro
sathibault
Eric Banks
Mark DePristo
Ryan Poplin
David Roazen