Upfront Imputation Improves Read Alignment

Taher Mun

Joint Lab Meeting 11/22/2019

Reference bias

1

1

2

3

2

3

Reference bias affects mapping accuracy

30x 100bp reads simulated from chr21 of 5 1KG samples

Ref bias = total # reads w REF allele / total # reads w ALT allele

Reference bias affects alignments over het sites

50x paired-end reads from NA12878 provided by GIAB (Zook et al, 2016, doi: 10.1038/sdata.2016.25)

​

(chr21)

Reference bias affects low-coverage variant calling

• ​

Bobo et. al. 2018 https://doi.org/10.1101/066043

How can we improve alignment in the face of reference bias?

Solution: change the reference

• Major allele reference

​

• Pros: contains most common alleles
• Cons: represents just one haplotype, ignores rare variants
• Graph genome

�

• Pros: can contain any number of alleles, represents any possible haplotype
• Cons: blow-up negatively affects accuracy and resources (Pitt et al, 2018 https://doi.org/10.1186/s13059-018-1595-x)

Personal is “best case scenario”

30x 100bp reads simulated from chr21 of 5 1KG samples

Ref bias = total # reads w REF allele / total # reads w ALT allele

Solution 2: build a personalized reference

• Build a genome personalized towards sample
• Methods: assembly, genotyping, variant calling
• Pros:
• Genome will contain variation specific to individual
• Can align to a diploid genome rather than one haploid or a graph
• Cons:
• All of these methods are time-consuming, resource-heavy
• Requires deep coverage reads from sample
• Might require extra technology (microarrays, long reads)

Proposed Solution: Create a personalized reference with sparse data

Goals:

• Generate a reference specific to study sample
• Minimal resource usage
• Minimal time
• Closest possible approximation to personal

Li Stephens model helps impute unknown genotypes

• Hidden Markov Model
• Built from
• Haplotype reference panel
• Recombination map
• Existing implementations available
• Beagle (Browning, et al 2018)
• IMPUTE5 (Rubinacci et al 2019)
• Use Viterbi algorithm to impute�values in incomplete genotype

Transition prob = recombination rate

= error rate

Li and Stephens, 2003

1000 Genomes Project

Imputation is dominant step

Future work

• Resource-efficient imputation
• Speed up lift-over step
• Evaluate on downstream results
• Variant Calls
• Allele specific expression
• etc.

Acknowledgements

Langmead Lab

Dr. Ben Langmead

Dr. Brad Solomon

Christopher Wilks

Daniel Baker

Charlotte Darby

Nae-Chyun Chen

Rone Charles

​