…ATG…GAGTA… Reference Assembly
…ACG…GAGTA… Maternal chromo-
some 1
…ATG…GAGCA… Paternal chromo-
some
Individual A is bi-allelic, or heterozy-
gous, at the two SNV sites and has the
genotype …C/T…C/T…, and the geno-
types are resolved into two haplotypes
…C…T… , … T…C…
Sites containing SNVs that are
prevalent in a population demarcate
chromosomal positions as varying,
or polymorphic. Consequently, these
locations are called single nucleotide
polymorphisms (SNPs). In discovery
workflows, geneticists test popula-
tions to see if the occurrence of varia-
tion correlates, or associates, with the
phenotype status of the individual.

So far we have discussed simple variations involving one or a small number of changes at a location. By contrast, geneticists also investigate structural variations in which large (1kbp up to several million bases) genomic fragments are deleted, inserted, translocated, duplicated, or inverted, relative to the reference. 19

(and decreasing more slowly) than the cost of sequencing.

We begin with exemplar queries on genomic data that illustrate the difficulty of genomic analysis and lack of consensus as to a best method. Abstractions must be flexible enough to handle a variety of methods.

sequencing trends

Four technological trends are relevant for designing a genomic software architecture:

Reduced cost. While the Human
Genome Project ( http://www.genome.
gov/) cost $100 million, human re-se-
quencing for redundant (15x) cover-
age now costs less than $5,000 in the
U.S., projected to fall below $1,000.
This implies universal sequencing
may be realizable, and archiving and
analysis, not sequencing, will domi-
nate cost;

figure 2. evidence for variation in the donor.

ACCGTACACTCAT
CCT-AC
GTAGACT
GTACAC
TAGACTCA
TACACTCAC

Snv deleted region Reference

50

cb
a
(a) evidence for Snvs is provided by aligning donor reads against the reference sequence; the g/t
variation might be a sequencing error, as the variant reads maps with too many errors, though the
g/C variation appears to be a true Snv. (b) Paired-end sequencing and mapping provides evidence
for deletion in the genome; the dotted rectangle demarcates the region in the reference deleted
in one of the two donor chromosomes. Read “a” samples the region around the deletion (marked
with the lightning bolt), mapping “discordantly” in the reference; read “b” maps concordantly,
but with coverage of only about half of neighboring regions; and read “c” is sampled from the
breakpoint, mapping at only one end.

evidence
1. Paired-end mapping
2. Depth of coverage
3. Loss of heterozygosity
4. Split reads

0

c
a
(a)

(b)

b donor genome

Variation Calling

The key to efficiency in genomics is the premise that an individual’s genetic record can be summarized succinctly by a much smaller list of individual genetic variations. While we develop this premise further in our layering proposal, we provide insight as to how variants are called today; the expert should skip this section and proceed to our layering proposal. We start with querying for SNVs in the donor genome, the simplest form of variation:

Calling SNVs. Figure 2 outlines how a mutation may be called. Consider the reference allele C. We see two copies of the donor genome with a G allele and some copies with a C, indicating a heterozygous SNV. If the variation is homozygous, all overlapping reads would be expected to have a G in that position, though even this simple call can be confounded. Some reads may have been mapped to the wrong place on the reference (such as the top donor read in the figure). The G/T mutation may not be correct, and the alignment of the incorrectly mapped read might present many variations. Even if the read is mapped correctly, sequencing errors could incorrectly appear as heterozygous mutations.

Mutation callers use statistical methods informed by mapping the quality of the read (such as number of potential places in the genome a read can map to), the quality score of a base call, and the distribution of bases or alleles in the reads for that location. Some mutation callers use evidence based on the surrounding locations (such as an excess of insertion/dele-tion events nearby suggesting alignment problems). The decision itself could be based on frequentist, Bayes-ian inference, or other machine-learning techniques. While SNP callers use various inference techniques, all refer to the same evidence—the set