A technique known as deep learning may transform a variety of fields of biomedical research, according to a review article co-authored by an NIEHS scientist and published April 4 in the Journal of the Royal Society Interface.
Deep learning, a powerful subset of machine learning algorithms, can extract meaningful patterns from complex data sets, enabling new discoveries that may improve treatments for human diseases.
“Deep learning approaches are being widely applied to difficult problems in life and environmental health science, ranging from foundational biology to drug development and clinical trials,” said co-author Christopher Lavender, Ph.D., a former postdoctoral fellow at NIEHS and now a contractor for the institute’s Integrative Bioinformatics Support Group.
“We hope that this review will facilitate the development and application of new technologies to address longstanding questions in environmental health science,” he said.
Visualizing protein structure
Lavender is particularly excited about applying deep learning methods to problems in cryo-electron microscopy (cryo-EM). This cutting-edge technique involves flash-freezing a biological sample, such as a thin layer of a protein-containing solution, and then visualizing the sample using transmission electron microscopy. Cryo-EM allows scientists to determine the structure of proteins with near-atomic resolution.
“Proteins play fundamental roles in almost all biological processes, and understanding their structure is critical for basic biology and for drug development,” Lavender said. “In the paper, I described how approaches used by companies like Google for image recognition can be used in cryo-EM to find protein particles in microscope images.”
This application of deep learning is being expedited by the Molecular Microscopy Consortium, a collaboration among NIEHS, Duke University, and the University of North Carolina at Chapel Hill. The consortium allows scientists to use cryo-EM and other microscopy tools to solve molecular structures at the atomic level.
As part of the initiative, NIEHS opened the first cryo-EM facility in the Carolinas in June 2017. “Advances in cryo-EM were awarded the 2017 Nobel Prize in Chemistry, and we are proud that the NIEHS is a local leader in bringing this technology to the area,” said Mario Borgnia, Ph.D., director of the Molecular Microscopy Consortium.
Applications in environmental health sciences
Other research areas at NIEHS are incorporating deep learning. For example, the Integrative Bioinformatics Support Group is developing new deep learning approaches for analyzing large genomic and epigenomic data sets. The institute’s Office of Data Science recently sponsored a deep learning training course.
The National Toxicology Program (NTP) is finding uses for the technology as well, according to Nicole Kleinstreuer, Ph.D., deputy director of the NTP Interagency Center for the Evaluation of Alternative Toxicological Methods.
“Deep learning is being leveraged in the field of toxicology to predict complex endpoints that involve multiple factors, by using information such as chemical structural features and physicochemical properties,” she said. “Several of the top performing models in a recent workshop on acute oral systemic toxicity applied deep learning to successfully predict this endpoint for tens of thousands of chemicals.”
Meanwhile, the NIEHS Office of Scientific Computing provides the critical infrastructure for running computationally demanding deep learning programs. “We see deep learning contributing to advances in the diversity of fields represented here at NIEHS,” said David Fargo, Ph.D., who is the institute’s first scientific information officer. “Cryo-EM is one such example, but deep learning is also being actively applied to genomics, epidemiology, and neuroscience.”
For example, deep learning could be used to identify genetic variants that influence susceptibility to environmental exposures, Fargo explained. It could also reveal previously unnoticed epidemiological patterns in large air pollution exposure data sets. “The application of this new technology to difficult longstanding problems is an area of active development,” said Fargo.
Citation: Ching T, Himmelstein DS, Beaulieu-Jones BK, Kalinin AA, Do BT, Way GP, Ferrero E, Agapow PM, Zietz M, Hoffman MM, Xie W, Rosen GL, Lengerich BJ, Israeli J, Lanchantin J, Woloszynek S, Carpenter AE, Shrikumar A, Xu J, Cofer EM, Lavender CA, Turaga SC, Alexandari AM, Lu Z, Harris DJ, DeCaprio D, Qi Y, Kundaje A, Peng Y, Wiley LK, Segler MHS, Boca SM, Swamidass SJ, Huang A, Gitter A, Greene CS. 2018. Opportunities and obstacles for deep learning in biology and medicine. J R Soc Interface 15(141):20170387.
(Janelle Weaver, Ph.D., is a contract writer for the NIEHS Office of Communications and Public Liaison.)