VariantEval is overly abusive of the GATKReport (lack of) spec.
1. It converts numeric values (longs, integers and doubles) to string before sending to the Report, then expects it to decipher that those were actually numbers.
2. Worse, the stratification modules somehow instead of sending the actual values to the report table, sends a string with the value "unknown" and then abuses the GATKReport spec to convert those "unknown" placeholder values with numbers. Then again, it expects the report to know those are numbers, not strings.
Now that the GATKReport HAS specs, VariantEval needs to be overhauled to conform with that. In the meantime, I have added special ad-hoc treatment to these wrong contracts. It works, and the integration tests all passed without changing any MD5's, but right after Mark and Ryan commit their VariantEval refactors, I will step in to change the way it interacts with the GATKReport, so we can clean up the GATKReport.
No wonder, the printing needed to be O(n^2).
* when gathering, be aware that some keys will be missing from some tables.
* when a gatktable has no elements, it should still output the header so we know it had no records
- The Integer column type now accepts byte and shorts
- Updated Unit Tests and added a new testParse() test
Signed-off-by: Mauricio Carneiro <carneiro@broadinstitute.org>
* restructured the hash tables into one class (RecalibrationReport) that has all the functionality for the different tables and key managers
* optmized empirical qual calculation when merging recalibration reports
* centralized the quality score quantization functionalities
* unified the creating/loading of all the key manager/hash table structures.
* added unit tests for the gatherer (disabled because gatk report needs to be sorted for automated testing)
* added integration tests for BQSR and on-the-fly recalibration
-- Minor refactoring of state key iteration in VEW.map to make the dependencies more clear
-- Long discussion about the performance problems with StateKey, and how to fix it, which I have run out of time to address before ESP meeting.
-- The previous approach (requiring > 5 copies among all reads) is breaking down in many samples (>1000) just from sequencing errors.
-- This breakdown is producing spurious clustered indels (lots of these!) around real common indels
-- The new approach requires >X% of reads in a sample to carry an indel of any type (no allele matching) to be including in the counting towards 5. This actually makes sense in that if you have enough data we expect most reads to have the indel, but the allele might be wrong because of alignment, etc. If you have very few reads, then the threshold is crossed with any indel containing read, and it's counted.
-- As far as I can tell this is the right thing to do in general. We'll make another call set in ESP and see how it works at scale.
-- Added integration tests to ensure that the system is behaving as I expect on the site I developed the code on from ESP
-- StateKey no longer extends TreeMap. It's now a final immutable data structure that caches it's toString and hashcode values. TODO optimizations to entirely remove the TreeMap and just store the HashMap for performance and use the tree for the sorted tostring function.
-- NewEvaluationContext has a method makeStateKey() that contains all of the functionality that once was spread around VEUtils
-- AnalysisModuleScanner uses an annotationCache to speed up the reflections getAnnotations() call when invoked over and over on the same objects. Still expensive to convert each field to a string for the cache, but the only way around that is a complete refactoring of the toTransversalDone of VE
-- VariantEvaluator base class has a cached getSimpleName() function
-- VEUtils: general cleanup due to refactoring of StateKey
-- VEWalker: much better iteration of map data structures. If you need access to iterate over all key/value pairs use the Map.Entry construct with entrySet. This is far better than iterating over the keys and calling get() on each key.
-- Now the only use for update0, calculating the number of processed loci, is centrally tracked in the walker itself not the evaluations.
-- This allows us to avoid calling update0 are every genomic base in 100ks of evaluates when there are a lot of stratifications.
-- No need to modify the integration tests, this optimization doesn't change the result of the calculation
* added empirical quality counts to allow quantization during on-the-fly recalibration to any level
* added number of observations and errors to all tables to enable plotting of all covariates
* restructured BQSR to report recalibrated tables.
* implemented empirical quality calculation to the BQSR stage (instead of on-the-fly recalibration)
* linked quality score quantization to the BQSR stage, outputting a quantization histogram
* included the arguments used in BQSR to the GATK Report
* included all three tables (RG, QUAL and COVARIATES) to the GATK Report with empirical qualities
On-the-fly recalibration with GATK Report
* loads all tables from the GATKReport using existing infrastructure (with minor updates)
* implemented initialiazation of the covariates using BQSR's argument list
* reduced memory usage significantly by loading only the empirical quality and estimated quality reported for each bit set key
* applied quality quantization to the base recalibration
* excluded low quality bases from on-the-fly recalibration for mismatches, insertions or deletions
-- This behavior, which isn't obviously valuable at all, continued to grab and rethrow exceptions in the HMS that, if run without NT, would show up as more meaningful errors. Now HMS simply checks whether the throwable it received on error was a RuntimeException. If so, it is stored and rethrow without wrapping later. If it isn't, only in this case is the exception wrapped in a ReviewedStingException.
-- Added a QC walker ErrorThrowingWalker that will throw a UserException, ReviewedStingException, and NullPointerException from map as specified on the command line
-- Added IT that ensures that all three types are thrown properly (i.e., you catch a NullPointerException when you ask for one to be thrown) with and without threading enabled.
-- I believe this will finally put to rest all of these annoying HMS captures.
-- Use a LinkedHashMap not a TreeMap so iteration is faster.
-- Note that with a lot of stratifications the update0 is taking up a lot of time. For example, with 822 samples and functional class and sample on there are 100K contexts and 30% of the runtime is just in the update0 call
-- Now you always get SNP and indel metrics with VariantEval!
-- Includes Number of SNPs, Number of singleton SNPs, Number of Indels, Number of singleton Indels, Percent of indel sites that are multi-allelic, SNP to indel ratio, Singleton SNP to indel ratio, Indel novelty rate, 1 to 2 bp indel ratio, 1 to 3 bp indel ratio, 2 to 3 bp indel ratio, 1 and 2 to 3 bp indel ratio, Frameshift percent, Insertion to deletion ratio, Insertion to deletion ratio for 1 bp events, Number of indels in protein-coding regions labeled as frameshift, Number of indels in protein-coding regions not labeled as frameshift, Het to hom ratio for SNPs, Het to hom ratio for indels, a Histogram of indel lengths, Number of large (>10 bp) deletions, Number of large (>10 bp) insertions, Ratio of large (>10 bp) insertions to deletions
-- Updated VE integration tests as appropriate
-- Moved a variety of useful formatting routines for ratios, percentages, etc, into VariantEvalator.java so everyone can share. Code updated to use these routines where appropriate
-- Added variantWasSingleton() to VariantEvaluator, which can be used to determine if a site, even after subsetting to specific samples, was a singleton in the original full VCF
-- TableType, which used to be an interface, is now an abstract class, allowing us to implement some generally functionality and avoid duplication.
-- This included creating a getRowName() function that used to be hardcoded as "row" but how can be overridden.
-- #### This allows us implement molten tables, which are vastly easier to use than multi-row data sets. See IndelHistogram class (in later commit) for example of molten VE output
-- No more IndelLengthHistogram (superceded by IndelSummary in subsequent commit)
-- No more SamplePreviousGenotypes or PhaseStats
-- No more MultiallelicAFs
* fixed BadCigarFilter to filter out reads starting/ending in deletion and that have adjacent I/D events.
* added Unit tests for BadCigarFilter
* updated all exceptions in LocusIteratorByState to tell the user that he can instead run with -rf BadCigar
* added the BadCigar filter to ReduceReads and RealignTargetCreator (if your walker blows up with these malformed reads, you may want to add it too)
- Updated the documentation on the code
- Made the table.write() method private and updated necessary files.
- Added a constructor to GATKReport that takes GATKReportTables
- Optimized my code
Signed-off-by: Mauricio Carneiro <carneiro@broadinstitute.org>
This is important for quick turnaround in the analysis cycle of the new covariates. Also added a dummy unit test that doesn't really test anything (disabled), but helps in debugging.
Pulled out the functionality from Indel Realigner and Table Recalibrator into Utils.setupWriter to make everyone else's life's easier if they want to include the PG tag in their walkers.
Infrastructure:
* Added static interface to all different clipping algorithms of low quality tail clipping
* Added reverse direction pileup element event lookup (indels) to the PileupElement and LocusIteratorByState
* Complete refactor of the KeyManager. Much cleaner implementation that handles keys with no optional covariates (necessary for on-the-fly recalibration)
* EventType is now an independent enum with added capabilities. All functionality is now centralized.
BQSR and RecalibrateBases:
* On-the-fly recalibration is now generic and uses the same bit set structure as BQSR for a reduced memory footprint
* Refactored the object creation to take advantage of the compact key structure
* Replaced nested hash maps with single hash maps indexed by bitsets
* Eliminated low quality tails from the context covariate (using ReadClipper's write N's algorithm).
* Excluded contexts with N's from the output file.
* Fixed cycle covariate for discrete platforms (need to check flow cycle platforms now!)
* Redfined error for indels to look at the previous base in negative strand reads (using new PE functionality)
* Added the covariate ID (for optional covariates) to the output for disambiguation purposes
* Refactored CovariateKeySet -- eventType functionality is now handled by the EventType enum.
* Reduced memory usage of the BQSR script to 4
Tests:
* Refactored BQSRKeyManagerUnitTest to handle the new implementation of the key manager
* Added tests for keys without optional covariates
* Added tests for on-the-fly recalibration (but more tests are necessary)
Infrastructure:
* Generic BitSet implementation with any precision (up to long)
* Two's complement implementation of the bit set handles negative numbers (cycle covariate)
* Memoized implementation of the BitSet utils for better performance.
* All exponents are now calculated with bit shifts, fixing numerical precision issues with the double Math.pow.
* Replace log/sqrt with bitwise logic to get rid of numerical issues
BQSR:
* All covariates output BitSets and have the functionality to decode them back into Object values.
* Covariates are responsible for determining the size of the key they will use (number of bits).
* Generalized KeyManager implementation combines any arbitrary number of covariates into one bitset key with event type
* No more NestedHashMaps. Single key system now fits in one hash to reduce hash table objects overhead
Tests:
* Unit tests added to every method of BitSetUtils
* Unit tests added to the generalized key system infrastructure of BQSRv2 (KeyManager)
* Unit tests added to the cycle and context covariates (will add unit tests to all covariates)
-- TODO for ryan -- there are bugs in ActivityProfile code that I cannot fix right now :-(
-- UnitTesting framework for ActivityProfile -- needs to be expanded
-- Minor helper functions for ActiveRegion to help with unit tests
Eric Banks
Mauricio Carneiro
Ryan Poplin
Guillermo del Angel
Roger Zurawicki
Mark DePristo
Menachem Fromer