Communications of the ACM

Since the conference, further prog-
ress has been made in yet another
collaboration: the public-private part-
nership for Accelerating Therapeutic
Opportunities in Medicine (ATOM)
involving GlaxoSmithKline, the DOE,
and the NCI, he says. Additionally,
“most significantly, the 21st Century
Cures Act was just signed into law,
setting the stage for a very promising
future at the intersection of predictive
oncology and computing.”
Several universities also are actively
researching ways to tackle big data,
which is a big challenge given the tre-
mendous amount of information col-
lected in the life sciences, notes Sunita
Chandrasekaran, an assistant profes-
sor in the Center for Bioinformatics
and Computational Biology at the
University of Delaware, and one of the
meeting’s organizers.

“Efforts are under way in universi-
ties that collaborate with medical re-
search institutes or facilities in order
to accelerate such large-scale compu-
tations like sequence alignment using
accelerators like GPUs (graphics-pro-
cessing units),” she says. “Efforts are
also under way to build suitable and
portable software algorithms that can
adapt to varying input and generate
results dynamically adapting to evolv-
ing hardware.”
Stevens says what makes it possible
now to use data more effectively than
several years ago is that researchers have
found ways to accelerate deep learning
through things like GPUs. “This, cou-
pled with breakthroughs in some meth-
ods like convolutional neural networks,
has suddenly made deep learning effec-
tive on many problems where we have
large amounts of training data.”
When the single model has been
put into effect, researchers will be able
to add more information about cancer
cells as well as more information about
drugs, “and we would have many more
instances of ‘this drug worked this well
on a given tumor,’ so many more train-
ing pairs between cancers and drugs,”
says Stevens.

While acknowledging he hates to
make predictions, Kibbe feels con-
fident that “in the next 10 years we
should see that many of what are very
hard-to-treat cancers will be treated,”
and that regardless of where someone
lives and what their socioeconomic
status is, they will have access to the
same level of care.

“I think that’s what will come out of
these collaborations and use of com-
puting; as sensors and instrumenta-
tion get cheaper and cheaper to im-
plement and become more and more
ubiquitous, the hope is there will be
a leveling effect on cancer treatment
across the country, and perhaps the
whole world.”
Perhaps working in collaboration,
combined with deep learning and
highly advanced computing, will prove
to be that holy grail. Kibbe calls the
DOE/NCI partnership unique in that
two very different cultures are working
together as a team. While everyone is
excited about their individual projects,
he says, they are also excited about
their joint mission of creating a work-
force that has both biomedical knowl-
edge and computational expertise.

“That side of the collaboration is going to continue to pay dividends for as long as we have computation in biomedical research, which I hope is forever.”

Further Reading

Davis, J.

Can Big Data Help Cure Cancer?

Information Week, July 17, 2016. http://www.informationweek.com/big-data/ big-data-analytics/can-big-data-help-cure- cancer-/d/d-id/1326295

Agus, D.B.

Giving Up Your Data to Cure Disease,

The New York Times, Feb. 6, 2016. https://www.nytimes.com/2016/02/07/ opinion/sunday/give-up-your-data-to-cure- disease.html?_r=0

Panda, B.

Big Data and Cancer Research. Springer, Oct. 13, 2016. http://link.springer.com/chapter/10.1007%

2F978-81-322-3628-3_ 14

Cho, W.C.

Big Data for Cancer Research, Clinical Medicine Insights: Oncology. v. 9; 2015

PMC4697768. https://www.ncbi.nlm.nih.gov/pmc/articles/ PMC4697768/

Reed, D.A., and Dongarra, J.

Exascale Computing and Big Data,

Communications of the ACM, Volume 58, Issue 7, July 2015, pp. 56-68. http://dl.acm.org/citation.cfm?id=2699414

Esther Shein is a freelance technology and business writer based in the Boston area.

cation of gene expression,” he says. And there are several other problems as well, including trying to predict the response to an individual drug based on its formula and profile, and the auto encoder problem, in which a network is trained to learn the compressed representation of a drug structure, for example, and then has to be trained to accurately reproduce the input so the team can build an improved algorithm.

The benchmarks will change over time, but they are a way to develop a common language among the vendors and the teams working on the pilots, Stevens says.

Once the first iteration of the model has been built and validated, it should be able to analyze tumor information from a newly diagnosed cancer patient and predict which drug will be the most effective at attacking the tumor.

Meanwhile, to help foster existing collaborations and pursue new ones, the first of a series of meetings was held in July 2016. The Frontiers of Predictive Oncology and Computing meeting focused on predictive oncology and computing in a few areas of interest in NCI/DOE collaboration: basic biology, pre-clinical, clinical applications and computing, says Eric Stahlberg, a contractor working on the high-performance computing strategy within the Data Science and Information Technology Program at the Frederick National Laboratory for Cancer Research in Rockville, MD.

“Efforts at the frontier of pre-clinical predictive oncology … included developing new models using patient-derived xenografts and predicting drug efficacy through regulatory networks,” Stahlberg says. Other areas of focus were how to gain better insights into Ras-related cancers, gathering quality data for use in predictive model development, and improving the SEER database.

“The meeting attendees were very enthusiastic about the prospects for improving cancer patient outcomes with increased use of computing,” Stahlberg says. That said, “One of the largest challenges exists in developing interoperability among solutions used in predictive oncology.” Others include gathering consistent data and having enough data to understand the complexity of individual cells, he says.