Bioinformatics for the 100,000 Genomes Project

Transcription

1 Bioinformatics for the 100,000 Genomes Project Augusto Rendón Director of Bioinformatics Genomics England Principal Research Associate University of Cambridge Barcelona,

2 Outline Introduction to the UK s 100,000 genomes project Analyses in rare diseases Analyses in cancer Bioinformatics Platform Data models and flows Databases Interpretation

3 Inception of the 100,000 genomes project (2012, 2014) If we get this right, we could transform how we diagnose and treat our most complex diseases not only here but across the world (December 2012) I am determined to do all I can to support the health and scientific sector to unlock the power of DNA, turning an important scientific breakthrough into something that will help deliver better tests, better drugs and above all better care for patients. (August 2014)

4 Goals of the Genomics England project Sequence 100,000 genomes Cancer and rare genetic disease Capture data delivered electronically, store it securely and analyse it within an English data centre (reading library) Combine genomes with extracted clinical information for analysis, interpretation, and aggregation Create capacity, capability and legacy in personalised medicine for the UK 1. To bring benefit to NHS patients 3. To kickstart the development of a UK genomics industry 2. To enable new scientific discovery and medical insights 4. To create an ethical and transparent programme based on consent

5 Genomics England project

6 Recruitment and clinical interface via 13 GMCs, Scotland and Northern Ireland + Genomic Medicine Centres Networks of NHS hospitals including genomics labs 13 Lead organisation plus 71 Local Delivery Partners Contracted by NHS England Cover recruitment, data and return of results Scotland Doing own sequencing Northern Ireland Similar to a GMC Contracted by NI payer

7 Feedback to participants 7

8 Additional finding Genes Requirements: A treatable or preventable condition. Reliably detected by next generation sequencing. Each gene will have a curated list of high confidence, high penetrance variants. Other conditions may be added if clinically appropriate and technically feasible.

9 Participants recruited in RD About 400 RD participants currently recruited per week 5,000 participants recruited to the RD pilot Family Size *Data from Main Programme

10 Recruitment by tumour type Sarcoma, 44, 4% Renal, 65, 6% Testicular Germ Cell Tumours, 1, 0% Adult Glioma, 19, 2% Bladder, 28, 3% Prostate, 95, 9% Breast, 321, 29% Ovarian, 91, 8% Malignant Melanoma, 3, 0% Lung, 139, 13% Childhood, 1, 0% Endometrial Carcinoma, 20, 2% Colorectal, 264, 24% 10

11 >14,896 genomes sequenced (Nov 1) N Bases x 10 9 (Q30-nodup) Median %Autosomal coverage >= 15X = 97.4% About 1.4 PB of data % Autosomal coverage >=15x(Q30-nodup) Germline data only

12 Analyses in Rare Diseases Genetic based test

13 Checks of reported data vs genetics Sex checks Coverage-based (WGS) X chromosome heterozygosity and Y chromosome genotyping rate (array) Predicted minor karyotypes include XO, XXY, XYY Relatedness checks Mendelian inconsistency rate (where at least one parent sequenced) Estimated identity by descent sharing for all pairs in cohort and working on a family only workflow - PLINK and PC-Relate Can identify rare phenomena, e.g. large-scale uniparental isodisomy

14 Coverage based sex checks

15 Relatedness checking 15

16 Analyses in Cancer Assessing the quality of sample preparation protocols

17 Fresh Frozen (FF) vs Formalin Fixed Paraffin Embedded (FFPE) FF Costly and not widely available Difficult to capture tumour High quality DNA FFPE Routinely used Digital pathology for tumour selection Low quality and quantity of DNA

18 AT/CG dropout effect on copy number variant calling FFPE GC dropout FFPE AT dropout FF AT dropout FF GC dropout AT rich GC rich AT dropout GC dropout FF sample low coverage for GC-rich regions FFPE sample trend is reversed with poor coverage of AT-rich regions

19 Fresh frozen and FFPE paired samples: ability to call CNVs FF AT drop Pur ity RMSD cov FFPE AT drop Pur ity RMSD cov NA

20 FFPE also affects small variant calling Proportion of variants Overlapping SNVs in FF and FFPE samples from paired VAF < 5% filtered out

21 Comparing sequence quality metrics across labs GMC1 Other GMCs After standardising on optimised FFPE protocol

22 Bioinformatics platform

23 GEL bioinformatics platform Design Goals Scalability: able to operate on several hundred whole genomes per day Traceability: able to keep the provenance of every artefact produced in the process Knowledge accumulation: able to capture and aggregate the knowledge, decisions captured during the interpretation in order to generate better knowledge bases Service oriented: components talk to each other via well defined APIs and data formats

24 Hospitals Genomics England Interp. providers Genome intake service Workflow management Clinical Data intake service Tracking GxP associations Interpretation Metadata Reference Knowledge Variants Interpretation platform services

25 Same data model, many manifestations How to ensure that all the data is coherently stored and easily retrievable? Inspired by Model-Driven Architecture approaches Models(schemas) controlled in github including boiler plate functions to validate data against model Documentation auto-generated out of the model Services communicate using JSON derived from the model Data written against the schema auto-generated from the model in the metadata store using document stores

26 Data models in the platform Use of avro for its interface definition language, JSON out of the box, automatic code generation of classes to handle these data Models (and auxiliary libraries) available here: s/tree/releases/schemas/idls Documentation for the master branch here: ndex.html Bioinformatics models here: For reads and variants we use protocol buffers compatible with GA4GH standards

27

28 InterpretedGenomeRD

29 Bertha: Distributed workflow management system (really an enterprise service bus for genomic data) DeliveryAPI Auditor Message Broker Job Scheduler Grid Consumer Restarts Scatter-gather Single and group processes Multiple concurrent workflows (work in progress) Orchestrator Producer publishes Exchange routes Queue consumes Consumer Tracking DB Dashboard

30 Workflow diagramme bertha_default Intake QC Delivery API Integrity Check Sequence received IntakeAPI MD5 Check Filtered Bamstats Unfiltered Bamstats Q30 Bamstats VCF QC Data intake Validate BAM Picard Plot Filtered Bamstats Generate Filtered Metrics Bamstats Plot Unfiltered Bamstats Generate Q30 Metrics Bamstats Fix Permissions QC Stats Post-processing Intake QC Check Point Merge Array Genotypes Single Sample QC & Processing Single Sample QC & Processing Sex Check Cross Sample Contamination Somatic VCF re-headering Cross Species Contamination Tumour Cross Sample Contamination Depth of Coverage Concordance check Single-Sample QC Check Point Multi Sample QC Multi-sample QC Mendelian Inconsistency Rate Identity by Decent Multi-Sample QC Check Point Consent Check Point Analysis Inbreeding Coefficient Mutation Burden SNV & Indel Refinement Actionable Mutation Coverage Viral Insertions Mutation Signature Tumour Clonality Tumour Purity Tumour Ploidy Variant Calling Homozygosity Runs Analysis Variant Normalisation Variant Annotation Interpretation Request Dispatched Interpretation Dispatch Variant Tiering Exomiser

31

32 Interpretation approach Virtual Gene Panels Initially assigned by a clinician Working on automated panel suggestions Variant filtering Allele Frequency: variant is rare Segregation: variant segregates with condition in family Panel membership (including mode of inheritance) Different for cancer Interpretation Automated pathogenicity scoring Manual review Several manual QC points

33 Panelapp: Crowdsourcing curation of gene disease associations

34 Status of Panels 190 panels 97 >=v1 panels 3,512 genes 435 registered reviewers 15,149 gene level reviews Recognised by the UK genetic testing network Curation reaching a point of diminishing returns Number of reviews Reviewers

35 Automated panel suggestion Disease X HP1 HP2 HP3 HP4 Diseases, core HP terms and panels HP1 HP2 HP3 HP4 HP5 X Y Z Panel X G1,G2,G3 Panel Y G1,G4,G5 Panel Z G6,G7,G8 Diseases, HP annotation and genes X Y Z G1,G2,G3 G1,G4,G5 G6,G7,G8 Also get a QC score for how phenotypically similar patient is to recruited disease HP2 HP3 HP4 G6 G7

36 RD pilot benchmarking 1831 participants with HPO terms, assigned panels (2674 total) and core disease 847/1831 (46%) have exactly same panels 728/1831 (40%) have same panels plus 1 or 2 extra 256/1831 (14%) are missing some of medical review 1200 panels Gene gains or losses Frequency November More 36 Bin

37 Filtering in the rare diseases programme

38 Filtering in the cancer programme Frequency filters: exclude common variants (1000G, ExAC, GEL) Consequence filters: exclude synonymous variants Domain 1 Domain 2 Domain 3 Variants in a virtual panel of actionable genes (between 20 and 40). Actionable genes are defined as genes with short variants associated with therapeutic, prognostic or diagnostic actions by GenomOncology (My Cancer Genome) Matching at the variant level Variants in the genes from Cancer Gene Census genes. Variants in all other genes Two part reports: Actionable and Interesting

39 Structural variants Mutational density Coverage and copy number Mutational signatures Supplementary analysis Mutation context Hypermutation rain plots

40 Bioinformatics platform components Cellbase Reference data store/annotation Engine OpenCGA Catalog: metadata and clinical data store Storage: variant database InterpretationPlatform Interpretation service: manage various producers and consumers Interpretation warehouse (under construction): stores and serves interpretation data

41 OpenCB family of applications Genome Browser Variant Analysis Data Discovery Interface Layer Cellbase OpenCGA Catalog OpenCGA Storage MongoDB MongoDB MongoDB HBASE PosixFS

42 Cellbase Knowledge base management Uses Ensembl, Uniprot, IntAct, ClinVar, etc. Current database engine is MongoDB JSON outputs against well defined model Supports annotation against local DBs Annotates about 10,000 variants/second per instance Python and R APIs

43 Annotation against Cellbase

44 CellBase VEP 82 Consequence type benchmark (1kG phase 3, 83M variants) VEP annotations: 346M CellBase annotations: 346M Coincidence at SO term level (346M annotations) Annotations provided by VEP and not provided by CellBase: 3364 (99.999% coincidence) Annotations provided by CellBase and not provided by VEP: 4918 (99.999% coincidence) 60% Due to differences on mirna data sources 39% Difficulties with VEP output format parsing Coincidence at variant level (83M variants) Variants with conflicting annotation: 4990 (99.994% coincidence)

45 Annotation for phased MNVs and CNVs Support for CNVs new in CellBase 4.5 Beta Main challenge: support imprecise calling - match against already reported CNVs (population frequencies, clinical variants) Same annotation data as for the rest of variants: consequence type, population frequencies, etc. Example CNV Support for MNVs and phased variants from CellBase 4.0 Consequence type depends on variants affecting the same codon Variants are assigned a phase set (phased VCFs include the PS tag) - all variants on the same phase set shall be processed together

46 Annotation of MNVs Example: 17:270550: AA CA G : TG CA A Example MNV Decompose into single phased variants members of the same phase set: { { "id": "17:270550:A:T", "result": [ { "codon": "caa/ctg", "proteinvariantannotation": { "reference": "GLN", "alternate": "LEU" }, "sequenceontologyterms": [ { "accession": "SO: ", "name": "missense_variant" } "id": "17:270551:A:G", "result": [ { "codon": "caa/ctg", "proteinvariantannotation": { "reference": "GLN", "alternate": "LEU" }, "sequenceontologyterms": [ { "accession": "SO: ", "name": "missense_variant" } { "id": "17:270554:G:A", "result": [ { "codon": "cag/caa", "proteinvariantannotation": { "reference": "GLN", "alternate": "GLN" }, "sequenceontologyterms": [ { "accession": "SO: ", "name": "synonymous_variant" }

47 OpenCGA - Catalog Metadata store and A&A for OpenCGA Manages roles, groups, acls Audit log LDAP integration Arbitrary schemas (annotation sets)

48 OpenCGA - Storage 6 node Hadoop cluster: Transform: 97 min Load: 80 sec Merge: 84 sec Millisecond response times for regional queries Whole genome filtering queries for all individuals within seconds Extensive capabilities to query across genotype and phenotype relationships

49 Aspiration to be fully GA4GH compatible from v1.0

50 Platform for interpretation (under construction)

51

52 Key (personal) learnings There is great strength in multidisciplinary teams with specialisation, but those individuals that can span both biology/genetics and software engineer are pivotal the connect the specialist Good software engineering practices also apply to bioinformatics, to name a few: designing, documenting, Testing, support and service. Skipping them don t really save you time I have become a big fan of using well established technologies with rich ecosystems (e.g. hadoop) rather than inventing new formats, data structures, toolchains

53 Final thoughts The future in human genetics will be underpinned by academic/industrial partnerships; both the task and the benefits are too big to go at it alone Genomic Medicine is just one of the pilots of a digital revolution in health care where artificial intelligence will complement/replace the diagnostic journey But genomics is the easy part, clinical data is the real challenge