Using Big Data technologies to uncover genetic causes of Amyotrophic lateral sclerosis

Transcription

1 Using Big Data technologies to uncover genetic causes of Amyotrophic lateral sclerosis Dr Natalie Twine Transformational Bioinformatics 11 October 2017 HEATH & BIOSECURITY

2 Astronomy Twitter YouTube Genomics Genomics will outpace other BigData disciplines Stephens et al. PLOS Biology 2015

3 Population-scale genomic data analysis requires BigData solutions Desktop compute High-performance compute cluster Hadoop/Spark compute cluster Focus small data Compute-intensive Data-intensive Fault tolerant No No Yes Node-bound Yes Yes No Parallelization 10 CPU 100 CPU 1000 CPU Parallelization procedure bespoke bespoke standardized CSIRO solution

4 Amyotrophic lateral sclerosis (ALS) ALS is a devastating motor neurone disease Leads to death within 3 years Affects more than 200,000 people worldwide Causes largely unknown genetic component Project MinE - sequencing 15,000 ALS genomes worldwide

5 What is our hypothesis? Sporadic cases are potentially related but separated over generations ALS is reported to be 5% familial and 95% sporadic Familial component is potentially higher than 5% Australia is a small population and disease is late onset What are our aims? Uncover hidden patient relationships to increase detection power Identify novel disease causing variants Datasets available: Exomes (Familial, n=137) WGS (Sporadic, n=800) Project MinE WGS (Sporadic, n=15,000) GOI SNP A SNP B SNP C

6 Existing methods for measuring relatedness PLINK (Chang CC et al. GigaScience, 2015) KING (Manichaikul A et al. Bioinformatics, 2010) SNPduo, ERSA, GRAB, XIBD (etc.) Limitations of these tools Designed to identify and remove relatives as part of GWAS workflow Identifying more distant relatives is challenging Tools effective at distant relationship detection are SLOW We want to expand existing family structures Identify more distant relationships with confidence

7 Relatedness between ALS patients using KING KING identifies close relationships 172 Familial and Sporadic ALS Exomes Each blue dot represents a relationship between a pair of ALS patients.

8 Relatedness between ALS patients using KING n=172 ( 137 Familial and 35 Sporadic) Degree of relationship Number True positives False positives Unknown Duplicates 6 6 (100%) 0 0 1st degree (100%) 0 0 2nd degree (100%) 0 3rd degree (44%) 9 (33%) 6 (22%) 4th degree 1310 n/a n/a n/a 5th degree 7852 n/a n/a n/a KING identified 8 novel relationships, at 3 rd degree 3 we have ruled out as false positives. 5 are potentially REAL as can t be classified as FP (no mutation status known).

9 How can we improve on this result? True positive rate - 44% - this needs to improve Whole Genome Sequencing >> Exome (SNP density) WGS (n=800) cohort has 42 million variants High density data - > more informative BUT - Existing tools struggle/fail with Big Data volumes We are implementing relatedness testing in VariantSpark to identify novel relationships in 800 WGS Sporadic and Familial ALS 15,000 WGS samples (Project MinE)

10 Speed BMC Genomics 2015, 16:1052 PMID: (IF=4) Bringing BigLearning to genomics applications. VariantSpark learns from 3000 individuals and 80 million mutations in under 30 minutes z Association testing Clustering Classification Accuracy

11 Using VariantSpark to identify relatedness Data-driven rather than model-driven approach VariantSpark can handle 80 million variants x 3000 individuals What is the genetic distance between samples? (allele sharing) Euclidean distance Identity by descent (IBD) (as in PLINK) Sliding window for IBD segments (as in ERSA) Include data from 1000 Genomes as controls (family and ancestry known) Approaches currently being tested for feasibility

12 distance accuracy Testing using different distance measures Euclidean distance 1 (IBD) UR 1 5 Degree of relationship Degree of relationship Exomes (n= 137 Familial ALS) 10 Euclidean distance performs well until 4 th degree Plink (IBD) performs better than other distances

13 Effective methods are compute intensive Ramstetter et al., Genetics 2017 IBD segment based methods most accurate for more distant relatives BUT They are also are most compute intensive (a.k.a SLOW)

14 Next steps in tool development Simulate a large pedigree using whole genome data Calculate genetic distance between each sample Measure performance of different distance measures sensitivity and specificity (AUC) Generate relationship degree metrics from simulated cohort Implement these methods in VariantSpark speed and scalability proof of principle cohort Familial ALS WGS (n=89) Identify novel relationships in Sporadic ALS WGS (n=800)

15 VariantSpark application genetic association Bone Mineral Density (BMD) as the phenotype; 1,936 individuals with 7.2 Million variants (imputed from array). Joint-loci analysis (machine learning - random forest) Replicate known BMD genes identified by traditional GWAS (single loci regression). Amplify signal over traditional methods so smaller cohorts give robust insights Random forests identifies interaction of 2 or more loci We will use this methodology to identify novel & modulating ALS variants

16 In summary: BigLearning to understand ALS Identify related individuals Novel diseasecausing variants Preventative measures Personalised treatment

17 Transformational Bioinformatics Team Denis Bauer Oscar Luo Laurence Wilson Aidan O Brien Natalie Twine Rob Dunne Piotr Szul Kaitao Lai Collaborators News Ian Blair Kelly Williams Emily McCann Jenn Fifita Adrian White Mia Champion Software