sic example: The blood-thinner drug Warfarin is widely prescribed to prevent blood clots. Dosage is critical; too high and the patient can bleed to death, too low and the drug might not prevent life-threatening blood clots. Often, the right dosage is established through multiple visits to the clinic and regular testing. However, recent reports16 suggest that knowledge of the patient’s genetic program can help establish the right dosage. We outline this approach (genetic association and discovery workflow) in three steps:

Collect samples. Collect a sample of affected and “genetically matched” control individuals; then sample DNA and catalog variations;

Identify variations. Identify and report variations that co-segregate, or correlate, with the affected/control status of the individual; and

Follow up with studies. Follow up on the genetic basis of the correlation through costly studies and experiments in animal models and clinical trials; then transfer knowledge to the clinic.

Even with its success, the discov-
ery approach involves complications:
First, studies are resource-intensive,
requiring identifying and sequenc-
ing large cohorts of individuals with
and without a disease. Second, it is
unclear how to apply study results to
a specific individual, especially one
genetically different from the inves-
tigated cohort. Finally, data reuse is
difficult; significant computation is
needed to dig out data from a previous
study, and much care is required to
reuse it. We contrast “discovery work
flow” with “personalized medicine.”
Here, a physician treating individual
A may query a database for treatments
suitable for patients with genetic vari-
ations similar to those of A or query
for patients genetically similar to A for
treatments and dosages that worked
well for these patients.

figure 1. universal sequencing, discovery, and personalized medicine.

Assume every individual is sequenced at birth. in discovery, clinical geneticists logically select a subset of individuals with a specific phenotype (such as disease) and another without the phenotype, then identify genetic determinants for the phenotype. By contrast, in personalized medicine medical professionals retrieve the medical records of all patients genetically similar to a sick patient S.

gs, Ms
g1

g2

gn
M
Personalized
Medicine
M1

M2

Mn discovery workflow

for a patient, a medical team might identify a collection of individuals genetically similar to the patient and on a Warfarin regimen; query their genomes and EMRs for genetic variation in candidate genes and Warfarin dosage, respectively; and choose the appropriate dosage based on the patient’s specific genetic program. The ability to logically select from a very large database of individuals using the phenotype as a key removes the first problem with the discovery workflow. Using genetic variations specific to individual A as a key to return treatments (that work well for such variations) addresses the second problem. Finally, if the accompanying software system has good abstractions, then the third problem (computational burden to reuse data) is greatly eased. Here we focus on key software abstractions for genomics, suggesting that like other CS areas (such as VLSI/sys-tems), software abstractions will enable genomic medicine.

We start with basic genetics using programming metaphors, then describe trends in sequencing and how genetic variations are called today and outline our vision for a vast genomic database built in layers; the key idea is the separation of “evidence” and “ inference.” We then propose a language for specifying genome queries and end by outlining research directions for other areas of computer science to further this vision.

We limit our scope to genomics, ignoring dynamic aspects of genomic analysis (such as transcriptomics, pro-teomics expression, and networks). Genomic information is traditionally analyzed using two complementary paradigms: First, in comparative genomics, where different species are compared, most regions are dissimilar, and the conserved regions are functionally interesting. 6, 7 The second is population genomics, where genomes from a single population are compared under the baseline hypothesis that the genomes are identical, and it is the variations that define phenotypes and are functionally interesting. We focus on population genomics and its application to personalized medicine and do not discuss specific sequencing technologies (such as strobe sequencing vs. color space encoding).