John West
`Richard Chen
`Accuracy next steps, spreadsheet to organize the discussion
`Thursday, June 7, 2012 6:24:18 PM
`Error mech Detect n Poss solutions JW 6Jun2012.xls
`
`From:
`To:
`Subject:
`Date:
`Attachments:
`
`Hi Rich,
`
`I put this together earlier this week to begin organizing my thoughts on how we move towards
`accuracy solutions, somewhat in parallel with ongoing analysis. I would be happy to go
`through it when you have time.
`
`John
`
`Personalis EX2061
`
`

`

`MECHANISM OF ERROR GENERATION
`
`Mis-alignment due to:
`
`Repeats / degeneracy
`
`CAN WE DETECT & COUNT EACH
`ERROR FROM THIS MECHANISM ?
`HOW ?
`ARE WE SET UP TO LOOK FOR THIS ?
`
`MIE due to false positive het SNP's, clustering
`in known degenerate regions
`Reproducibly abnormal coverage in known
`degenerate regions
`Ratio of reads at het loci far from 50/50
`Tri-allelic in a single genome (fairly easy to see)
`Triple-haplotype (more work to make QC tool)
`Systematically different result from paired-end
`reads of different insert lengths
`
`HOW WE COULD ELIMINATE / REDUCE THIS ERROR TYPE
`BIOINFORMATICALY
`LAB METHODS
`REFERENCE
`
`ALGORITHMS
`
`Chain-link SNP alleles by shared raw
`reads and PE-reads across a degen
`region
`
`Special focus on junction sequences
`
`Variable alignment stringency across
`the genome, tailored to known
`differences in degeneracy, likely as
`a 2nd stage re-alignment
`
`Solve degenerate regions
`in "gold standard" genome
`by "spare no expense"
`
`Develop MAF data in
`degenerate regions to guide
`later alignment of others
`
`SNP/InDel cluster
`
`Low raw sequence quality:
`
`Reproducible coverage drop around SNP clusters ?
`Ratio of reads at het loci far from 50/50
`InDel MIE due to failure to detect InDel due to cluster
`Apparent failure of LD - Alleles at nearby loci
`don't "travel together" as expected
`
`New version of GATK is better ?
`
`Use BreakSeq approach for SNP/InDel
`clusters
`
`Do SNP clusters near 50% MAF
`split 49/51 in non-real
`combinations in our ethnically
`specific major allele refs ?
`
`Check for allele combinations which
`are unlikely from known LD
`
`Add InDels to our ref where
`major allele
`
`Raw Q-score depression downstream from
`specific sequence motifs
`
`Directional Q-score average dips, particularly
`in standard locations
`
`Motif-specific pattern recognition
`including length, not just substitutions
`
`Clusters systematically too faint in certain
`sequence contexts
`
`Phasing worse than normal in certain
`sequence contexts
`
`Low raw Q-scores near the start of a read, e.g.
`bases 2-20. We don't currently look
`for this, but could.
`We could confirm with access to fluorescence
`intensity data - an option on the
`system but not usually on
`
`We could confirm with access to fluorescence
`intensity data - an option on the
`system but not usually on
`
`Look specifically for these & if they
`are reproducible. Should we exclude
`them ? What motifs cause them ?
`
`Motif-specific algorithms
`
`Motif-based sequencing primers
`
`Pile-up of low raw Q-scores by random
`near-registration of the ends of reads,
`which are inherently low-Q
`
`We don't currently look for this, but could
`May be a good refinement of the
`directional Q-score metric
`
`Look for this and flag it (at least)
`
`Some shorter reads, which are
`highly accurate all the way to
`the ends
`
`Longer reads
`ILMN 2x250
`Ion Proton
`Oxford Nanopore
`Moleculo
`Longer PE inserts
`Multiple PE insert lengths
`
`Custom pullout for long
`insert paired ends or other
`assays
`
`Custom SNP array or pullout
`incorporating the allele sets
`known to be in LD
`
`If GC or hairpin driven, use
`nucleotide analog kits
`
`Lower cluster density, longer
`imaging exposure times ?
`Different cluster growth, e.g.
`# cycles
`
`Low coverage:
`
`Leading to a lack of calls,
`missed heterozygocity, or missed SV
`The low coverage being due to:
`
`Extremes of %GC
`
`Hairpin
`
`Poisson
`
`Look for directional coverage reproducible
`dropouts
`
`Look for where this is unusual, i.e.
`not seen in test genomes sequenced
`the same way
`
`Need to find best scale and lateral offset for GC
`correlation with coverage
`
`Need to understand mechanism of
`GC bias to know how to fix it
`
`Nucleotide analogs
`Extreme GC pullout(s)
`
`JW spreadsheet-based model can find these but
`is too slow for genome scale.
`Also, not all large hairpins lead to
`low coverage - needs work
`
`Refine model to understand correlation
`between hairpin properties and
`error occurrence, so we can think
`about how to fix the problem
`
`Genome-wide database of
`hairpins, their properties &
`their impact
`
`Nucleotide analogs
`Pullouts or PCR sets which have
`just one side of the hairpin
`
`The part of coverage variability which is not
`reproducible run to run on same library
`
`Exome pullout imperfection
`
`Coverage vs position wil show this
`
`Maybe room to add SV detection to
`exomes, esp via BreakSeq
`
`Failure to detect small InDel's, due to:
`
`Algorithm weakness
`
`InDel MIE's may identify some of these ?
`
`New GATK may improve this quite a bit
`
`Can BAM crawler identify InDels in alignment of
`raw reads which were not called in the
`variant lists (possible false negatives) ?
`
`If BAM crawler can do this (see left)
`then local re-allignment may correct
`miss
`
`Difficult seq contexts
`Seq with InDel maps elsewhere
`InDel creates degeneracy
`InDel creates hairpin
`InDel near reference gap
`InDel close to, but not in, STR
`InDel close to, but not in, SV
`
`Many reads too close to either end
`at location of InDel to be able to
`detect
`
`Special case not currently handled
`Special case not currently handled
`Special case not currently handled
`Special case not currently handled
`Special case not currently handled
`Special case not currently handled
`
`Similar to issue of pile-up of low raw Q scores
`by random near-registration of
`ends of reads, but either end is a
`problem here.
`
`InDels should be in LD with nearby SNP's
`so we could look for those and use that
`
`Could create dedicated algorithm if priority
`Could create dedicated algorithm if priority
`Could create dedicated algorithm if priority
`Could create dedicated algorithm if priority
`Could create dedicated algorithm if priority
`Could create dedicated algorithm if priority
`
`Extend the InDel detection algorithms for
`loci with known high-allele-frequency
`InDels, to support detection even near
`either end of a read. Different from
`detection of novel InDels.
`
`STR errors due to:
`
`Biochemistry fails
`
`Shows up as a drop in coverage and/or a reproducible
`drop in raw Q-scores downstream
`
`Failure is generally for very long STR's
`which are generally heterozygous.
`This may be detectable as an allelic read
`depth imbalance at nearby het SNP's
`
`Probably worst problem where
`combined with other sources
`low coverage, so fix those
`
`Supplementary pullout
`
`Ethnically-specific major InDel
`allele reference, on Hugo's
`plan.
`
`Longer Illumina reads should
`help: Minimal false positive
`but a much bigger % of the
`read will be usable
`
`See other hairpin options, above
`
`Detectable by electrophoresis
`Standard technology we could
` incorporate
`
`Long STR's can also be strong
`hairpins, so hairpin-reducing
`nucleotide analogs could help
`
`Seq OK but algorithm fails, esp if het
`Worse (?) if both alleles
`don't match reference
`
`Would this show up as a low alignment Q-score ?
`Would we notice a pile-up of those ?
`Should we add this to our BAM crawler ?
`
`STR MIE's, if we look for these
`
`Since all STR loci should already be known
`we could write a dedicated algorithm for
`this. lobSTR may already be this.
`Mark has suggested trying it.
`
`Current reference almost
`certainly does not have the
`most common # STR units at
`each position, let alone by
`ethnicity. This should be easy
`to change in the reference,
`though it requires mapping
`coordinate systems.
`
`Homopolymer errors due to:
`
`Biochemistry fails
`
`Shows up as a drop in coverage and/or a reproducible
`drop in raw Q-scores downstream
`
`A simple dedicated algorithm could
`prioritize getting the sequence right
`downstream of the homopolymer, vs the
`length of the homopolymer, by working
`with reads which begin near the end of the
`
`Dedicated sample prep which
` intentionally and reproducibly
`spikes different bases into these
`regions, to break them up and
`make them more sequenceable.
`
`Personalis EX2061
`
`

`

`Seq OK but algorithm fails, esp if het
`Worse (?) if both alleles
`don't match reference
`
`Would this show up as a low alignment Q-score ?
`Would we notice a pile-up of those ?
`Should we add this to our BAM crawler ?
`
`Homopolymer MIE's, if we look for these
`
`Translocation errors:
`
`False positive due to repeat seq
`
`Translocation MIE's, if we look for them
`
`Missed reads which span junction
`
`Translocation detected by paired-end algorithm but
`not split-read algorithm
`
`Robertsonian translocations missed
`Not detectable with next gen seq.
`
`Larger (>50bp) deletion errors:
`
`Zygocity not reported, or wrong
`
`MIE in large deletion detection
`
`Del detected but not junction seq so
`length & position inexact
`
`HugeSeq's SV algorithm requires detection by at
`least two algorithms, but doesn't
`require junction seq detection
`Should it ?
`
`homopolymer, or end just into it.
`Almost all homopolymer loci in the genome
`should be known now, so we can use that info
`
`Does lobSTR potentially help here ?
`
`Current reference likely does
`not have the most common
`# homopolymer base units
`in all of these regions, let alone
`by ethnicity
`
`Paper circulated fairly recently (by Gemma ?)
`showed success ealing with this by filtering
`out repeat regions & using long mate pairs
`
`HugeSeq has two algorithms which look
`for junction sequences & could filter out
`potential translocations for which it did not
`find them
`
`Map out regions of systematically low
`coverage (in the absence of a deletion)
`Potentially useful in distinguishing deletions
`from normal coverage variation
`
`Junction sequences are most likely to be
`missed where coverage is systematically low
`and a deletion is heterozygous.
`Also where deletion removes one unit of
`a multiple unit tandem repeat or STR
`because junction seq can be same with or
`without deletion. Potentially solvable with
`a dedicated algorithm if priorotized
`
`Could use a targeted mega-
`nuclease.
`What priority / value ?
`
`Supplement sequencing with
` long mate pairs
`
`Supplement next gen seq with
`karyotyping
`
`Longer reads should be propor-
`tionately better at detecting
`junction sequences (with & w/o
`deletion, in case of het del).
`If we can put more weight on
`junction-based algorithms, we
`will be less subject to false
`positives from coverage variation.
`
`Potential for pull-out in junction
`sequence regions, to increase
`ability to detect these known
` junctions
`
`False positive from read-depth algorithm
`due to coverage variation caused
`by other mechanisms
`
`MIE in large deletion detection, and HugeSeq
`requires two algorithms to detect
`SV's confidently
`
`See above
`
`Larger insertion errors:
`
`Haven't investigated enough to categorize
`
`MIE's and reproducibility with multiple PE insert
`lengths
`
`Haven't investigated enough to
`propose approaches
`
`Can gold standard genomes
`and 1,000 Genomes' SV data
`help produce a better ref
`sequence re insertions ?
`
`Longer insert read-pairs might
`be quite helpful.
`
`CNV errors:
`
`Copy number wrong due to mis-alignment
`esp in repeats
`
`Total copy # OK, but heterozygocity wrong
`e.g. 4x as 2+2 vs 3+1
`
`Junction sequences of CNV's not detected
`
`CNV MIE's, if we look for these
`
`CNV MIE's, if we look for these
`
`Algorithms to improve alignment in
`degenerate & compression regions
`might also help with CNV's - another
`variation on the same type of problem
`
`Gold standard genomes, esp
`using molecules 10-100kbp,
`might help solve some of this
`and create a framework on
`which to organize the
`population statistics
`
`HugeSeq's SV algorithm requires detection by at
`least two algorithms, but doesn't
`require junction seq detection
`Should it ?
`
`Complex CNV's: differing, overlapping blocks
`Also re-arrangements, deletions, etc
`
`MIE's and reproducibility with multiple PE insert
`lengths
`
`Reference sequence errors:
`
`Incorporates rare SNP alleles,
`leading to mis-alignment
`and missed variant calls
`
`Incorporates rare InDel alleles,
`leading to mis-alignment
`
`Incorporates rare SV alleles, leading to SV
`inverse (e.g. Insert for Del)
`
`Compressions (large duplications not
`recognized in the reference)
`
`Gaps in the reference
`
`Reference is not 1,000 Genomes major allele
`Done by Rick Dewey for PLoS paper
`
`Hugo plans to look for InDels with >50% Alt
`Allele Frequency in 1,000 Genomes
`result he calculated
`
`Hugo plans to look for SV's with >50% Alt
`Allele Frequency in 1,000 Genomes
`result he calculated
`
`All quartet family members het where inheritance
`state would not allow that
`Split of reads at het loci is far from 50/50
`esp in very high coverage genome
`Blocks of elevated coverage seen identically in
`other ethnicities
`Loci tri-allelic in a single individual
`Three haplotypes in one region of a single genome
`
`These are well documented in the reference
`itself.
`
`More complex cases could likely be
`addressed with more sophisticated
`and dedicated algorithms.
`Development of these would benefit from
`interaction with an effort to build gold
`standard genomes using maximal
`technology.
`
`Update with newer 1,000 Genomes data ?
`Cross-check against NHLBI exome major alleles ?
`Call all interpretable loci regardless of
`variance vs reference sequence
`
`Addressing this is on Hugo's project plan
`(right ? When ?)
`
`Addressing this is on Hugo's project plan
`(right ? When ?)
`
`Make an effort to vallidate more compressions
`from Rick Dewey' list: Top 36 was not a limit -
`We just ran out of time manually validating from
`the original "top 50".
`Alignment algorithm proposed based on
`population data, supplemented by gold
`standard genomes
`
`Look for reads which align at edge of gap
`and extend into the "unknown" to build
`our own extensions to narrow the gaps.
`Others have done some of this, which we
`might leverage.
` If we wanted to make significant progress
`here, it could be very hard and we would
`certainly want to understand what kept the
`original human genome project from
`closing these gaps. They certainly tried.
`
`Reads of very long molecules
`would help. Not just mate pairs,
`but possibly very long insert
`clones, Moleculo approach or
`Shendure approach or CGI "long
`fragment read" approach.
`
`Personalis EX2061
`
`