Harvard grad student devises an efficient method to sort and organize billions of genomic matches

A DOE graduate fellow has developed an algorithm that will dramatically slash the time it takes to sort and catalog billions of genome sequences from the Joint Genome Institute and other research centers.

The algorithm, developed by Ben Campbell Smith from Harvard University, can search and organize billions of genomic sequence comparisons in a day instead of a month. The efficiency will enable staff at the Biological Data Management and Technology Center (BDMTC) at Berkeley Lab to massage raw data into materials that scientists can easily use for genomic analyses.

BDMTC develops informatics tools and provides data management for the Joint Genome Institute (JGI), UC San Francisco, Berkeley Lab’s Life Sciences and Physical Biosciences divisions and the California Institute of Quantitative Biomedical Research (QB3). The Integrated Microbial Genomes (IMG) system, created by BDMTC, integrates microbial data from the JGI and other public sources and enables comparative analyses across species, something researchers look for in hunting for clues about evolution, for example.

“Ben is an outstanding worker. Bioinformatics is now flooded with a huge amount of data to be analyzed and cross compared. Scalability is a major issue especially for a tightly integrated system such as IMG,” said Ernest Szeto, a BDMTC researcher who works closely with Smith. “Ben has applied solid computer science skills to help deal with some of these most pressing disk oriented processing scalability issues.”

Smith is working in BDMTC this summer as a DOE Computational Science Graduate Fellow. The fellowship, funded by the DOE Office of Science and the National Nuclear Security Administration, not only pays for each fellow’s tuition and other school fees, it also provides an annual stipend of $31,200 and other funds for research-related expenses.

Bioinformatics is not Smith’s research focus in school. In fact, the Harvard graduate student is partial to high-energy physics. His Ph.D. work involves hunting for the elusive Higgs boson particle, whose existence can validate a theory on how fundamental particles such as electrons and quarks acquire mass.

But working with biological data isn’t new for Smith. He recalled fondly the time he worked in his father’s bioinformatics lab at the University of British Columbia, where the elder Smith is a hematologist/oncologist.

“Before I started graduate school, my dad said, ‘Come work for me and write some code.’ I had a lot of fun doing that,” said Smith, who searched Berkeley Lab’s Computing Sciences web site for research ideas and learned about the work by Markowitz and his group. “I thought it would be cool to work on something that people use all the time.”

Genomics and high-energy physics share one similarity — they both generate an incredible amount of research data that must be culled to obtain useful information for research. With that in mind, Smith said he was able to immerse himself quickly in the informatics project at Berkeley Lab.

Every time a microbe’s genome is sequenced, that information goes to the Pacific Northwest National Laboratory (PNNL), which uses a supercomputer and a software called Basic Local Alignment Search Tool (BLAST) to look for matching sequences among the roughly 3.2 million microbial sequences in the database.

Instead of looking for matches only between the newly sequenced genome and those already in the database, however, the PNNL computer carries out the “all versus all” BLAST search, spitting out results that show all the matches among various microbes’ genomes. When each microbe’s genome can produce thousands of gene sequences, the process of matching them with each other will produce an enormous set of data. As a result, BDMTC staff aren’t able to update the IMG system frequently.

Smith’s task is to organize and format those results so that researchers can quickly find specific comparisons among the two sequences or microbes they are studying. The dataset he is working with contains 20 billion lines, each corresponding with a match.

The 20 billion matches aren’t in any particular order, making it even more difficult to sort them by taxons and then “score,” which refers to a statistical analysis of the quality of the matches (some matches could have been made in error).

Before Smith devised the new method, BDMTC staff used a brute force algorithm that read the output a single line at a time and wrote the match to a file based upon the two taxons involved. Because of the inefficiency inherent in accessing a different file for each of the nearly 20 billion matches, this process would take approximately 30 days.

Smith’s algorithm, on the other hand, first breaks down the data into thousands of smaller files. Using a cluster with 35 dual core CPUs, the smaller chunks of data are catalogued by the genomes they contain. This allows a sorting program to focus only on very small subsets of the data corresponding to the genome of interest. The process is further sped up through the use of a binary search tree, which allows the sorting to remain computationally efficient, even for very large datasets.

“The process now takes a day. You take all the data and run and sort it. Then anyone who needs it again can quickly look up the results,” Smith said.

With the new technique, the IGM system can be updated four times a year instead of two. The algorithm will be used in the next release of IMG/M, the metagenomics version of IMG, which is accompanied with a big batch of computational results from PNNL. The next release is scheduled for December or January.

Learn more about the IMG system at http://crd.lbl.gov/html/BDMTC. Information about the Computational Science Graduate Fellowship program can be found at http://www.krellinst.org/csgf/index.shtml.

—–

Source: Lawrence Berkeley National Laboratory, Computational Research Division. This article was originally published in the August 2007 issue of the Computational Research Division Report which can be found at http://crd.lbl.gov/html/news/CRDreport.html.