Lecture 2.2

Signal or Noise?

Tue Sep 22SlidesAssignment

Learning objectives

After this lecture, you should be able to do the following.

Interpret the confidence reported with a variant call, and judge whether a call should be trusted and reported.

1. Interpret the GQ score as a Phred-scaled genotype error probability, \(GQ = -10\log_{10}(P(\text{Error}))\), and state the confidence a stated GQ carries, as GQ30 carries 99.9%.
2. Describe why a called variant is filtered for artifacts, reviewed by eye, and confirmed by an independent method before it is reported.
3. Given the GQ values reported for several calls and a stated confidence requirement, determine which calls clear it.
4. Given an unfamiliar site whose evidence pulls both ways, such as a high GQ inside a duplicated region or deep coverage by poorly aligned reads, judge whether the call should be trusted and defend the verdict.

Separate a biological mutation from a technical error, and account for why sequencing the same position many times is what makes the separation possible.

5. Distinguish a biological mutation (a physical change in the sample DNA) from a sequencing error (introduced during library prep, PCR, or the sequencer).
6. Explain why sequencing the same region many times gives the statistical redundancy needed to separate a consistent true-variant signal from sporadic sequencing error.
7. Given a described false positive, attribute it either to low base quality at that position or to poor alignment at the ends of the reads covering it.
8. Given the two allele counts at a position, determine whether the share of reads carrying the alternate base sits nearer one half or nearer the error rate, and say which reading of the site that supports.

Account for how a caller weighs two competing genotypes against the reads, and identify what each term of Bayes theorem contributes.

9. Given a reference base and a pileup holding two different bases, state the genotype hypotheses a caller has to weigh against each other.
10. Explain what the likelihood \(P(R\mid G)\) asks about a candidate genotype, and why the base quality scores at that position are what set it.
11. Explain why the numerator of Bayes theorem alone is a score rather than a probability, and what dividing by the summed scores of all three genotypes gives you.
12. Explain why a position with 19 reads carrying one base and 9 carrying another fits both a heterozygote sampled unevenly and a homozygote with nine errors.
13. Given a variant already known to be common in the population, predict which way the prior moves the posterior for the heterozygous genotype.
14. Given a drop in the base quality of the reads carrying the alternate base, predict the effect on the likelihood of the heterozygous genotype and on the call.

Describe how a haplotype-aware caller rebuilds a messy region instead of trusting its alignment, and judge when that is worth doing.

15. State what marks a stretch of the genome as an active region to a haplotype-aware caller.
16. Define a haplotype as a set of alleles inherited together from one parent.
17. Describe how a haplotype-aware caller sets the existing alignment aside in an active region and assembles the reads there on their own, resolving indels the pileup grid gets wrong.
18. Given a described variant site, determine whether a pileup caller is enough or a haplotype-aware caller is needed.
19. Given the steps of local reassembly out of order, put them in sequence: flag the active region, discard the alignment in it, assemble the reads from that window alone, and pick the haplotypes best supported by them.

Account for how PCR duplication manufactures a variant that is not there, and for what removing duplicates does to the evidence.

20. Explain how a deduplication tool picks out duplicates from identical mapping coordinates, and why a unique molecular identifier separates a duplicate from an independent molecule that happens to start at the same coordinate.
21. Explain how amplifying a library by PCR leads to the same original molecule being sequenced many times.
22. Explain why a candidate variant that disappears after duplicate filtering was amplified noise rather than biological signal.
23. Explain how PCR duplication violates the independence-of-reads assumption, so that one early amplification error is copied into every duplicate.
24. Given a position whose alternate reads are mostly duplicates of one molecule, predict what happens to the apparent allele share and to the call once duplicates are flagged and set aside.

Account for what a pileup can and cannot see, and for the cluster of false positives an indel leaves in one.

25. Define a pileup as the column of bases covering one reference position, and state what that column is inspected for.
26. State the assumption a pileup makes about the aligner, and explain why a length-changing variant breaks the fixed coordinate grid that assumption rests on.
27. Explain why a gap placed at one position in one read and a neighboring position in another turns real matches into apparent mismatches, leaving a cluster of ghost SNVs that hides the true indel.
28. Explain why a substitution leaves every surrounding base on the reference grid, so that the whole signal for it sits in one column.

Explain variant calling as inferring the true sequence from noisy reads, and account for what makes it an inference rather than a readout.

29. State that the reference genome is a composite map rather than a universal standard, and that a variant is a difference between a sample and that reference.
30. State that two human genomes are about 99.9% identical, and that the remaining fraction of \(3 \times 10^9\) base pairs leaves roughly 3 million variants per person.
31. Distinguish the four routes from reads to variants, reads or an assembly compared against a linear reference.
32. Explain why most of a persons variants are harmless and the one that matters is still hard to find.
33. Explain why one linear reference cannot stand for a whole population, and what becomes of reads from a segment the reference does not carry.
34. Explain variant calling as inferring the true sequence from noisy reads, distinguishing biological signal from technical noise.
35. Explain why a genome is inferred probabilistically rather than read off, given that the sequencer makes physical errors and the reference itself holds gaps and errors.