quirements and can be recomputed on demand.

EMRs (information retrieval). The phenotypes associated with each sequenced individual are already in patient medical records. Initial results from the eMERGE network indicate that, for a limited set of diseases, EMRs can be used for phenotype characterization in genome-wide association studies within a reasonable margin of error. 9, 13 We anticipate that most health-care institutions will be using EMRs by 2014, given incentives provided by the Health Information Technology for Economic and Clinical Health Act of 2009.17 Increasing adherence to interoperability standards5 and advances in biomedical natural language processing12 make efficient querying possible. However, there is no integration of genotype and phenotype data today. GQL should be useful for both interrogating a single genome and interrogating multiple genomes across groups of individuals but will need to integrate with existing EMR systems so phenotype data can be queried together with genomes.

Privacy (computer security). The genome is the ultimate unique identifier. All privacy is lost once the public has access to the genome of an individual, but current regulation, based on the Health Information Portabil-ity and Accountability Act, is silent about it. 2, 3, 11 Though the Genetic Information Nondiscrimination Act addresses accountability for the use of genetic information8 privacy laws must change to ensure sensitive information is available only to the appropriate agents. Checking that a given study satisfies a specific privacy definition requires formal reasoning about the data manipulations that generated the disclosed data—impossible without a declarative specification (such as GQL) of such manipulations.

Provenance (software engineering). GQL is an ideal way to record the provenance of genomic study conclusions. Current scripts (such as GATK) often consist of code that is too ad hoc for human readability and span various programming languages too low level for automatic analysis. By contrast, publishing the set of declarative GQL queries along with their results would

While the work
is a challenge,
making genetics
interactive is
potentially as
transformative
as the move
from batch
processing to
time sharing.

significantly enhance the clarity and reproducibility of a study’s claims.

Provenance queries also enable scientists to reuse the data of previously published computing-intensive studies. Rather than run their costly queries directly on the original input databases, these scientists would prefer to launch an automatic search for previously published studies in which provenance queries correspond to (parts of) the computation needed by their own queries. The results of provenance queries can be directly imported and used as partial results of a new study’s queries, skipping re-computa-tion. This scenario corresponds in relational database practice to rewriting queries using views.

Scaling (probabilistic inference). Learning the correlation between diseases and variations can be tackled differently if there are a large number of genomes. It may be less critical to accurately evaluate individual variations for such a discovery problem, as erroneous variations are unlikely to occur over a large group of randomly selected individuals. More generally, do other inference techniques leverage the presence of data at scale? As an example, Google leverages the big-data collections it has to find common misspellings. Note that accurately screening individual variations is still needed for personalized medicine.

Crowd sourcing (data mining). Crowdsourcing might be able to address difficult challenges like cancer, 14 but the query system must first have mechanisms to allow a group to work coherently on a problem. Imagine that a group of talented high-school science students are looking for genetic associations from cases and controls for a disease. A potentially useful GQL mechanism would be to select a random subset of cases and controls that are nevertheless genetically matched (arising from a single mixing population). Researchers could then query for a random subset of 100 individuals with a fraction of network bandwidth while still providing similar statistical power for detecting associations.

Reducing costs (computer systems). Personalized medicine must be commoditized to be successful so requires computer systems research; for example, since most genomes are