From:
`To:
`Subject:
`Date:
`Attachments:
`
`John West
`Richard Chen; Hugo Lam; Mark Pratt; Scott Kirk
`Spreadsheet for 2pm discussion of Accuracy Differentiation
`Tuesday, May 22, 2012 1:17:12 PM
`Accuracy differentiation JW 22May2012.xls
`
`All,
`
`For our meeting this afternoon on accuracy differentiation, I have put together some notes to
`serve as a starting point.
`They are attached in Excel. Comments are welcome.
`
`John
`
`Personalis EX2054.1
`
`

`

`ACCURACY DIFFERENTIATION
`Draft for discussion, JW, 22 May, 2012
`
`ORGANIZATION OF THIS FILE
`Company differentiation around accuracy
`Better sequencing in the laboratory
`Better variant detection & reporting
`More accurate databases, for interpretation
`
`COMPANY DIFFERENTIATION AROUND ACCURACY
`Focus on accuracy relevant to medical interpretation
`Better understanding of the issues than anyone
`Best track record of publications on the issue
`Unbiased from a platform standpoint & able to combine platforms
`More comprehensive data on accuracy than anyone
`World's only collection of genomes sequenced on both ILMN and CG platforms, plus arrays and karyotyping
`Largest family pedigree sequenced to high coverage
`Only genome sequenced on Sanger, ILMN and CGI
`Databases more accurate than those publicly available
`Able to provide a detailed quantitative view of mechanisms underlying errors
`Deep understanding of accuracy issues used to create better results
`Not just flagging errors, or filtering out those loci, but fixing the problems
`More insightful approaches delivery accuracy affordably
`
`BETTER SEQUENCING IN THE LABORATORY (WHEN MANAGED BY PERSONALIS)
`
`Focus on getting the whole medically-interpretable genome, accurately, even if more expensive
`Use insight into error types & medical content to keep this affordable
`
`Combine data from multiple different runs of a single platform
`Combine paired-end libraries made with multiple insert lengths
`Use longer read lengths (e.g. 2 x 250 bases, when available later in 2012)
`More expensive because only MiSeq, but clearly better
`Combine with bulk shorter-read data from HiSeq
`Substantially more efficient at split-read & junction-sequence SV detection
`Key to single-base breakpoint determination
`
`Combine data from multiple platforms
`More experience with Illumina / Complete Genomics than anyone
`May add Ion Torrent, Oxford Nanopore
`Guided by deep understanding of differential error mechanisms in each platform
`Not tied to any one platform
`Use whatever it takes to get the best possible combination
`
`Combine data from outside next-gen sequencing
`
`Several major areas of medical genetics are not well assayed by next gen sequencing
`Example 1 : Diseases caused by STR-expansion (e.g. Huntinton's)
`Example 2 : Robertsonian translocations
`
`Personalis will combine NGS data with Non-NGS technologies to create a complete assessment
`Add karyotyping
`Add electrophoresis where appropriate (TBD)
`Others
`
`Orthogonal technologies also provide validation of NGS results
`Integrate NGS with array (fluorescence for SV's, in addition to genotypes)
`Sanger follow-up to findings of specific genomes (option TBD)
`
`Ability to create semi-custom products focused on medically interpretable parts of the genome
`Leverages Personalis advantages in content
`Custom hybridization array
`Custom pullout set
`Other assays to fix specific error types
`
`Question : Should there be a "Personalis exome" option ?
`More comprehensive / accurate at exome price level ?
`
`What is proprietary about this approach ?
`
`Personalis EX2054.2
`
`

`

`Personalis can focus it's efforts based on the world's best content (re medical interpretation)
`Personalis will develop proprietary understanding about how best to combine multiple technologies
`Personalis does not face the competitive & anti-trust barriers that platform companies do
`Personalis' people combine deep experience with both platforms and interpretation & can leverage the two against each other
`Personalis can combine work in the lab and in bioinformatics, in a way that pure informatics companies can't
`
`BETTER VARIANT DETECTION & REPORTING
`
`Fewer false positive SNP's due to method of generating laboratory data
`Combination of paired-end insert lengths covers more of the genome uniquely
`
`Fewer false negative SNP's due to method of generating laboratory data
`More uniform coverage by combining library prep methods & platforms
`
`Orthogonal validation of millions of SNP genotypes by array
`Integrated with next gen sequencing data, not just another separate report
`
`Better alignment, due to better reference sequence
`SNP major allele ref by ethnicity (in our first product)
`InDel major allele ref by ethnicity (later)
`Other advances as R&D develops them:
`Itterative alignment
`
`# TBD changes inSNP alleles called with Personalis reference vs public standard
`Likely more improvement in non-European ethnicities
`Include (eventually) changes in InDels called as well
`
`We provide the only support available specifically for admixed genomes
`Major-allele non-ethnic reference, or even more advanced options
`
`Focused effort to align in the presence of SNP / InDel clusters, MNP's
`May leverage Hugo's BreakSeq approach
`May need time in the development plan
`
`Better SNP reporting due to better reference
`
`We report variants when sequence is a homozygous match to the public ref but that's the minor allele
`Entirely missed by systems which use the public reference
`Example : Factor V Leiden
`Rong had a whole paper on all the disease variants in the public ref
`> 1M loci where we can be different
`We should calculate the average # actual loci / genome, by ethnicity
`
`At het loci, we report both alleles, but we report the minor allele as the variant
`Not the allele which is different from the public reference
`
`Better detection of SV's
`
`Better lab data for SV detection:
`Longer reads (better for approaches based on split-read & junction-sequences)
`MiSeq 2x250 or other platform
`Multiple insert lengths
`Electrophoretic assay of STR-expansions
`Karyotyping for Robertsonian translocations
`
`Orthogonal technologies for validation of SV's:
`Fluorescence intensity data from hybridization arrays
`
`We combine the results from five different algorithmic approaches
`
`We test our SV algorithms by Mendelian Inheritance in high coverage whole genome family data sets
`one which was sequenced with ten different paired-end libraries spanning 200 - 40,000 bases
`and validate them using fluorescent intensities from high density hybridization arrays
`
`We don't treat all SV's as novel - we have the world's best database of known SV's and their junction sequences
`Detection is better when you know exactly what you are looking for
`We should have a meeting to discuss how we can (easily ?) build this
`Start with 1,000 Genomes result Hugo has helped create
`Large data set but low coverage may make detection less certain in low MAF SV's
`Others will be able to access this eventually, potentially catching up, or claiming to
`Augment this with (more confident ?) SV's from:
`
`Personalis EX2054.3
`
`

`

`Full coverage (30-40x) genomes (West, Altman, 40 Koreans, others we can download)
`High coverage (>60x) genomes (Snyder, CEPH1463, Venter, others ?)
`
`Better reporting of SV's
`We determine the zygocity of deletions and report it
`Deletions integrated with SNP report, e.g. "A-" vs "AA" inside a het deletion
`We report SV's with their allele frequencies in the ethnicity matching the sample
`
`Flagging of potential errors
`Many subtle error types not recognized by others
`Error mechanisms underlying differences when the same person is sequenced twice (it's not just Poisson !)
`Error loci determined from deep & multi-platform sequencing of large families
`Error loci determined by extensive platform comparison, both NGS/NGS and NGS/Non-NGS
`Detailed understanding of compressions, and large unpublished catalog of them
`
`MORE ACCURATE DATABASES, FOR INTERPRETATION
`
`Cleaner databases:
`Well financed, systematic manual curation to industrial QC standard
`Standardized medical language hierarchy
`Extensive cross checking of databases developed independently
`VariMed vs HGMD
`MendelDB vs OMIM
`Personalis PharmGKB vs public PGKB (need to be careful in this positioning)
`
`Databases others will not have:
`Regulome
`BreakSeq (esp if augmented with private Personalis data)
`Compression list (described in publications but not released)
`Variant data derived from a broad collection of genomes
`Multiple public data sets, some processed in proprietary ways by Personalis
`Access to private data sets, sequenced by others
`Access to private data sets, sequenced by Personalis
`
`Personalis EX2054.4
`
`