1. Class meetings: TR 3:30-4:50 MCGIL 2315
CSE291: Personal genomics for bioinformaticians
Class meetings: TR 3:30-4:50 MCGIL 2315
Office hours: M 3:00-5:00, W 4:00-5:00 CSE 4216
Contact:
Today’s schedule:
3:30-4:10 Prioritization and Filtering
4:10-4:15 Break
4:15-4:50 Time to work on PS5 (+command line tips and info for final presentations)
Announcements:
PS5 due Thursday
Reading posted for Thursday

2. Prioritizing and Filtering variants
CSE291: Personal Genomics for Bioinformaticians
03/07/17

3. The challenge: needles in haystacks Annotation annotation annotation
Outline
The challenge: needles in haystacks
Annotation annotation annotation
Family information
Prior knowledge of gene function
Selection
Remaining challenges

4. The challenge: needles in haystacks

5. The challenge: sifting through piles of variants…
From the exome alone, THOUSANDS of candidate variants

6. The challenge: sifting through piles of variants
2. Is this variant in affected family members?
1. Have I seen this variant before?
3. Does this mutation affect gene function?
GAAAATATCATGTGGTGTTTCC
GAAAATATCATATGGTGTTTCC
6. Is the gene likely relevant to this disease?
4. Is this position conserved across species?
5. Gene expression pattern of this gene makes sense?
Common approach: progressively apply filters from highest to least confidence
Caveat: some truly pathogenic variants will fail these filters!

7. Annotation annotation annotation

8. Annotating impact of mutations on genes

9. Loss of function variants
LOFTEE ( VEP plug-in to annotate LoF
Assesses suspected LoF variants:
Stop gain (nonsense)
Splice site disrupting
Frameshift variants
KEY: filters to identify true vs. false positive LoF annotations, e.g.:
Nonsense variants in last 5% of the gene unlikely to be that damaging (why?)
Nonsense variants in an exon without canonical splice sites around it likely false positive (why?)
Splice sites in very small introns (e.g. <15bp) likely not that critical
If the LoF allele matches the ancestral allele, likely not really LoF (why?)

10. Are missense variants important?
Polyphen2: predict impact of an amino acid substitution on gene structure and function
8 sequence based features, 3 structure based features
Classify variants as: probably damaging, possibly damaging, benign
See also: SIFT, MutationTaster, SNAP, and more. Most people use multiple methods and e.g. require more than one method to call the mutation damaging
Adzhubei et al. Nature Methods 2010

11. Ensemble annotations Ensemble: combination of different methods
Idea: there’s lots of annotations out there. Some combination of which are probably important. Let’s combine them into a single classifier
CADD: Combined Annotation-Dependent Depletion
Kircher, et al. Nature Genetics 2014 (Shendure Lab)
Features (63 annotations total):
VEP annotations (e.g. nonsense, missense)
SIFT, PolyPhen2
Mappability
Conservation (PhastCons, PhyloP, GERP)
Segmental duplications
Expression
Histone modifications
SVM Classifier
Train on simulated data to determine:
Observed (likely benign) vs.
Simulated de novos (likely pathogenic)

12. Family information

13. Pedigrees help narrow down the disease location
Thousands of candidate variants
Dozens of candidate variants

14. Pedigrees help narrow down the disease location
Causal variant homozygous in affecteds, missing or heterozygous in unaffecteds
Affected siblings almost always share the region IBD=2
Autosomal recessive
Causal variant het (or hom) in affecteds, missing in unaffecteds
Affected siblings likely share the region IBD=1, both inherited from affected parent
Autosomal dominant
Bigger pedigree=better. Why?
Example of different types
Dominant
Heterozygous
De novo
De novo
Mutation not present in parents or affected siblings

15. Prior knowledge of gene function

16. Databases of clinical consequences of variants
Has my candidate gene been previously implicated in a human disease?
If yes, is it related to the current disease I’m trying to solve?

17. Gene ontologies Does the annotated function of my gene make sense?
e.g. for Marfan Syndrome, FBN1
Biological process: skeletal/heart/kidney development
Cellular component: basement membrane, extracellular, microfibril
Molecular function: Calcium ion binding, structural, protein binding

18. Incorporating gene expression data - tissue
Is this gene expressed in tissues that make sense for this disease?
e.g. if disease is primarily liver related, the causal gene is probably expressed in liver
We now have resources (e.g. GTeX) reporting expression across tissues
CFTR Expression by tissue

19. Incorporating gene expression data from patients
Identified novel exon in COL6A1 formed by dominantly acting splice gain event, causes external collagen-VI-like dystrophy
Overall, diagnosis rate of 35% in patients with undiagnosed neuromuscular diseases
Can incorporate RNA-seq from family members to leverage traditional pedigree approaches
Cummings et al. 2016

20. Selection

21. Types of selection
Positive selection: a new mutation confers a selective advantage, and rises to frequency quickly. OR a new environmental factor makes an existing mutation suddenly more advantageous.
Examples: LCT (lactase persistence), EDAR1
Tests: Long haplotypes, high derived allele frequency
Purifying selection: mutations in critical regions of the genome are often deleterious and quickly eliminated
Examples: protein coding sequence vs. introns, ultra-conserved regions
Tests: all of these compare observed vs. expected variation
Tajima’s D, Fu and Li Test, many others
Genetic constraint (Tuesday)
Implication: deleterious mutations are rare!

22. Selection says disease-causing mutations are rare
severe
e.g. Tay-Sachs
Severe Mendelian disorders
Nonexistent
(removed by selection)
(well actually… AD APO e4. why?)
Effect size
e.g. high cholesterol, Crohn’s Disease, Type II Diabetes
(many common alleles with small effect sizes)
Likely many examples, but low power to detect these
mild
rare
common
Allele Frequency

23. Metrics of purifying selection
Site frequency spectrum: the distribution of allele frequencies of a given set of SNPs in a population or sample
Site frequency spectrum define it
Use the three different metrics:
% variable sites (e.g. for de novos we’ll talk about thurs)
% singleton
Mean MAF
Summarizing the SFS:
% Singletons (seen 1 time in the population). Higher=rarer=stronger selection
Mean MAF (higher=more common=weaker selection)
% variable sites (how many positions were never variable. Higher=weaker selection)

24. Purifying selection for variant classes
MAPS: mutability adjusted % singletons
Missense, nonsense, syn, splice
Show intron exon plot from daly
Help us prioritize variation, but can only be applied to classes of vars, not single
Generally, the more variants that are singletons, the more deleterious that mutation class is
Caveat: these metrics describe categories of variants, and may or may not be useful for predicting impact of a specific mutation
Lek et al. 2016

25. Conservation across species informative
Phastcons: compute conservation based on phylogenetic tree + HMM
phyloP: compute conservation using sequence alignment and model of neutral evolution
GERP: identify constrained elements in multiple sequence alignment by quantifying “substitution deficits”
Davydov et al. 2010

26. Remaining challenges

27. Some examples defy our annotation pipelines
Cystic fibrosis (deltaF508 in CFTR) I I F G V
In-frame deletion!
(Usually would prioritize frameshifts) rs
GAAAATATCATCTTTGGTGTTTCC
GAAAATATCAT---TGGTGTTTCC I I F G V
Synonymous variants can be pathogenic! (e.g. in MDR1, multidrug resistance. UBE1 spinal muscular atrophy
Deep intronic mutation

28. The non-coding genome…
With exome sequencing, we analyze 2% of the genome but still have too many variants.
Helped by the fact that we have a decent idea of how to analyze coding variants.
How will we deal with the overwhelming number of false positives from WGS?
Requires… annotation and prioritization! More on non-coding regions next Tuesday.

29. Final projects + PS5