Mining the Big Data Mountain

Posted on May 6th, 2014 by Dr. Francis Collins

Biomedical researchers and clinicians are generating an enormous, ever-expanding trove of digital data through DNA sequencing, biomedical imaging, and by replacing a patient’s medical chart with a lifelong electronic medical record. What can be done with all of this “Big Data”?

Besides being handy for patients and doctors, Big Data may provide priceless raw material for the next era of biomedical research. Today, I want to share an example of research that is leveraging the power of Big Data.

NIH-funded researcher Atul Butte of Stanford University recently mined mountains of existing data to find new links among genes, diseases, and traits. In this instance, traits are defined as any detectable physical or behavioral characteristic, such as cholesterol levels or other blood chemistries; bone density; or body weight. Butte reasoned that a trait that was closely linked to a disease through specific genes might be useful as a predictive marker of disease risk.

To discover these new links, he tapped into the VARiants Informing MEDicine (VARIMED) database, a resource that he began building in 2008 to interpret the clinical consequences of DNA variation in patients [1]. To create VARIMED, Butte and his colleagues read scientific papers on human genetics—including many genome-wide association studies (GWAS), which identify common genetic variants that are associated with disease risk—and noted the genes, variations, and traits mentioned in each paper and the connections between them. Over the years, the privately funded database grew; today it contains findings from more than 9,000 studies.

In their most recent study, Butte’s team examined the genetic architecture of each disease—all of the genetic variations that influence disease risk—and made a list of the gene-disease pairs. They found 801 genes that were reliably linked to 69 diseases [2]. Next, the researchers made a list of genes associated with particular traits; 796 genes were reliably linked to 85 traits. Finally, they searched for overlaps between the two lists: were there any genes that influenced both a disease and a trait?

Butte’s team found 120 diseases and traits that were linked by the activity of just a few genes. Many of these disease-trait associations were already known to the biomedical community, but about 20 percent of these connections were novel.

The next question was whether these traits can be used to predict whether an individual would develop a particular disease. Butte tested his hypothesis by taking five of the new disease-trait connections and examining patient data from Stanford Hospital and Clinics and New York’s Mount Sinai Medical Center and Columbia University Medical Center, all of which had at least a decade’s worth of electronic medical records (EMRs).

For example, one of the disease-trait pairs was gastric cancer-magnesium, which is a trace mineral found in blood serum. This pair was connected by three genes—MUC1, THBS3, and TRIM46—previously implicated in gastric cancer and also known to influence magnesium levels in blood serum.

However, it wasn’t clear whether serum magnesium levels were actually predictive for gastric cancer. To find out, Butte’s team analyzed the EMRs of 704 patients who had a magnesium measurement one year before they were diagnosed with gastric cancer. For controls, he examined the EMRs of more than 324,000 patients who had magnesium measurements, but no diagnosis of gastric cancer. The comparison revealed that patients with elevated magnesium levels were significantly more likely to be diagnosed with gastric cancer than patients whose levels were normal.

Another surprising connection was made through the CLPTM1L and TERT genes that linked lung cancer with prostate specific antigen (PSA), a protein that can be detected through blood tests and is a known risk biomarker for prostate cancer. Butte’s team studied 240 men who had their PSA measured one year before their diagnosis of lung cancer and compared them with men who had PSA measurements taken, but did not develop lung cancer. The team found that men with abnormally high PSA levels were more likely to be diagnosed with lung cancer a year later than those with normal PSA levels. One important caveat about this finding is that the study did not control for smoking, which is a known risk factor for lung cancer and has also been implicated in prostate cancer.

Using the same approach, the team also validated links between: the average volume of a red blood cell and the risk of developing acute lymphoblastic leukemia (through shared variants in theIKZF1 gene); levels of the alkaline phosphatase enzyme and the risk of developing a blood clot in a leg, liver, or lung vein (the ABO and TERT genes); and low platelet counts and likelihood that someone is alcohol dependent (the C12orf51 gene).

The authors emphasize that it is premature to develop predictive blood tests for any of these five diseases because it’s not yet clear whether the trait precedes the disease, or is the result of it. While finding a connection between a disease and trait may someday lead to new blood, genetic, imaging, and other predictive or diagnostic tests, this is not the only clinical implication of the new study. If different diseases are found to be connected by common traits, it’s possible there may be a common biological mechanism underlying those diseases—which could mean that a drug that works for one disease might also be effective for the others. Getting answers to these complicated questions clearly will require much further study in cells, animal models, and humans.

What I find most noteworthy about this work is not the specific findings, but how the researchers demonstrate the feasibility of mining vast troves of existing data—genetic, phenotypic, and clinical—to test new hypotheses.

Indeed, we are at a point in history where Big Data should not intimidate, but inspire us. We are in the midst of a revolution that is transforming the way we do biomedical research. In some cases, rather than posing a question, designing experiments to answer that question, and then gathering data, we already have the needed data in hand—we just have to devise creative ways to sift through this mountain of data and make sense of it.

At NIH, a bold new initiative called Big Data to Knowledge (BD2K) is providing a focal point for catalyzing this historic research opportunity. I’ve recently recruited the distinguished informatics expert Dr. Philip Bourne from the University of California at San Diego, and he now serves as my Associate Director for Data Science (ADDS, isn’t that an appropriate acronym?). If you’d like to learn more about what NIH is doing to enable researchers to capitalize on this wealth of data, check out our BD2K web site.

References:

[1] Clinical assessment incorporating a personal genome. Ashley EA, Butte AJ, Wheeler MT, Chen R, Klein TE, Dewey FE, Dudley JT, Ormond KE, Pavlovic A, Morgan AA, Pushkarev D, Neff NF, Hudgins L, Gong L, Hodges LM, Berlin DS, Thorn CF, Sangkuhl K, Hebert JM, Woon M, Sagreiya H, Whaley R, Knowles JW, Chou MF, Thakuria JV, Rosenbaum AM, Zaranek AW, Church GM, Greely HT, Quake SR, Altman RB. Lancet. 2010 May 1;375(9725):1525-35.

[2] Disease Risk Factors Identified Through Shared Genetic Architecture and Electronic Medical Records. Li L, Ruau DJ, Patel CJ, Weber SC, Chen R, Tatonetti NP, Dudley JT, Butte AJ. Sci Transl Med. 2014 Apr 30;6(234):234ra57.

Links:

NIH Big Data to Knowledge (BD2K)

Atul Butte at TEDMED 2012

NIH support: National Center for Advancing Translational Sciences