U-M Data Science Annual Symposium 2020

November 11 at 9am

Novel Tools to Increase the Reliability and Reproducibility of Population Genetics Research (9:00am-9:20am)

Addressing selection bias in population data

Yajuan Si – Research Assistant Professor, Survey Research Center, Institute for Social Research
Advances in population genetic research have the potential to create numerous important advances in the science of population dynamics. The interplay of micro-level biology and macro-level social sciences documents gene–environment–phenotype interactions and allows us to examine how genetics relates to child health and wellbeing. However, traditional genetics research is based on nonrepresentative samples that deviate from the target population, such as convenience and volunteer samples. This lack of representativeness may distort association studies. Recent findings have provoked concern about misinterpretation, irreproducibility and lack of generalizability, exemplifying the need to leverage survey research with genetics for population-based research. This project is motivated by the research team’s collaborative work on the Fragile Family and Child Wellbeing Study and the Adolescent Brain Cognitive Development Study, which present these common problems in population genetics studies, to advance the integration of genetic science into population dynamics research. The project will evaluate sample selection effects, identify population heterogeneity in polygenic score analysis, and develop strategies to adjust for selection bias in the association studies of educational attainment, cognition status and substance use for child health and wellbeing. This interdisciplinary project will strengthen the validity and generalizability of population genetics research, deepen new understandings of human behavior and facilitate advances in population science.

View Slides

An end-to-end deep learning system for rapid analysis of the breath metabolome with applications in critical care illness and beyond (9:20am-9:40am)

Deep learning for biomedical research

Christopher Gillies – Assistant Research Scientist, Emergency Medicine
The metabolome is the set of low-molecular-weight metabolites and its quantification represents a summary of the physiological state of an organism. Metabolite concentration levels in biospecimens are important for many critical care health illnesses like sepsis and acute respiratory distress syndrome (ARDS). Sepsis is responsible for 35% of patients who die in the hospital and ARDS has a mortality rate of 40%. Missing data is a common challenge in metabolomics datasets. Many metabolomics investigators impute fixed values for missing metabolite concentrations and this imputation approach leads to lower statistical power, biased parameter estimates, and reduced prediction accuracy. Certain applications of metabolomics data, like breath analysis by gas chromatography, used for the prediction or detection of ARDS, can be done without the quantification of individual metabolites. This would circumvent the quantification step of individual metabolites, eliminating the missing data problem. Our team has developed a rapid gas chromatography breath analyzer, which has been challenged by missing data, a time-consuming process of breath signature alignment, and the following quantification of metabolites across patients. Analyzing the breath signal directly could eliminate these challenges. End-to-end deep learning systems are neural networks that operate directly on a raw data source and make a prediction directly for the target application. These systems have been successful in diverse fields from speech recognition to medicine. We envision an end-to-end deep learning that leverages transfer learning, from the collection of many healthy samples, that could rapidly multiply the applications of our breath analyzer. The end-to-end deep learning system will enhance our breath analyzer so it could be used more efficiently in areas of the intensive care unit to the battlefield to identity patients or soldiers with critical illnesses like sepsis and ARDS and monitor longitudinal changes in breath metabolites.

View Slides

Machine learning-guided equations for the on-demand prediction of natural gas storage capacities of materials for vehicular applications (9:40am-10:00am)

Machine learning for energy research

Alauddin Ahmed – Assistant Research Scientist, Mechanical Engineering
Transportation is responsible for nearly one-third of the world’s carbon dioxide (CO2) emission because of burning fossil fuel. While we dream for zero-carbon vehicles, future projections suggest little decline in fossil fuel consumption by the transportation sector until 2050. Therefore, ‘bending the curve’ of CO2 emission prompts the adoption of low-cost and reduced-emission alternative fuels. Natural gas (NG), the most abundant fossil fuel on earth, is such an alternative with nearly 25% lower carbon footprint and lower price compared to its gasoline counterpart. However, the widespread adoption of natural gas as a vehicular fuel is hindered by the scarcity of high-capacity, light-weight, low-cost, and safe storage systems. Recently, materials-based natural gas storage for vehicular applications have become one of the most viable options. Especially, nanoporous materials (NPMs) are in the spotlight of the U.S. Department of Energy (DOE) because of their exceptional energy storage capacities. However, the number of such NPMs is nearly infinite. It is unknown, a priori, which materials would have the expected natural gas storage capacity. Therefore, searching a high-performing material is like ‘finding a needle in a haystack’ that slows down the speed of materials discovery against growing technological demand. Here we present a novel approach of developing machine learning-guided equations for the on-demand prediction of energy storage capacities of NPMs using a few physically meaningful structural properties. These equations provide users the ability to calculate energy storage capacity of an arbitrary NPM rapidly using only paper and pencil. We show the utility of these equations by predicting NG storage of over 500,000 covalent-organic frameworks (COFs), a class of NPMs. We discovered a COF with record-setting NG storage capacity, surpassing the unmet target set by DOE. In principle, the data-driven approach presented here might be relevant to other disciplines including science, engineering, and health care.

View Slides

Fusing Computer Vision And Space Weather Modeling (10:00am-10:20am)

Deep learning and computer vision for space science

David Fouhey – Assistant Professor, UM EECS
Space weather has impacts on Earth ranging from rare, immensely disruptive events (e.g., electrical blackouts caused by solar flares and coronal mass ejections) to more frequent impacts (e.g., satellite GPS interference from fluctuations in the Earth’s ionosphere caused by rapid variations in the solar extreme UV emission). Earth-impacting events are driven by changes in the Sun’s magnetic field; we now have myriad instruments capturing petabytes worth of images of the Sun at a variety of wavelengths, resolutions, and vantage points. These data present opportunities for learning-based computer vision since the massive, well-calibrated image archive is often accompanied by physical models. This talk will describe some of the work that we have been doing to start integrating computer vision and space physics by learning mappings from one image or representation of the Sun to another. I will center the talk on a new system we have developed that emulates parts of the data processing pipeline of the Solar Dynamics Observatory’s Helioseismic and Magnetic Imager (SDO/HMI). This pipeline produces data products that help study and serve as boundary conditions for solar models of the energetic events alluded to above. Our deep-learning-based system emulates a key component hundreds of times faster than the current method, potentially opening doors to new applications in near-real-time space weather modeling. In keeping with the goals of the symposium, however, I will focus on some of the benefits close collaboration has enabled in terms of understanding how to frame the problem, measure success of the model, and even set up the deep network.

View Slides

Decoding the Environment of Most Energetic Sources in the Universe (10:20am-10:40am)

Machine learning for astronomy

Oleg Gnedin – Professor, Department of Astronomy, LSA
Astrophysics has always been at the forefront of data analysis. It has led to advancements in image processing and numerical simulations. The coming decade is bringing qualitatively new and larger datasets than ever before. The next generation of observational facilities will produce an explosion in the quantity and quality of data for the most distant sources, such as the first galaxies and first quasars. Quasars are the most energetic objects in the universe, reaching luminosity up to 10^14 that of the Sun. Their emission is powered by giant black holes that convert matter into energy according to the famous Einstein’s equation E = mc^2. The largest progress will occur in quasar spectroscopy. Detailed measurements of spectrum of quasar light, as it is being emitted near the central black hole and partially absorbed by clouds of gas on the way to the observer on Earth, allows for a particularly powerful probe of quasar environment. Because spectra of different chemical elements are unique, spectroscopy allows to study not only the overall properties of matter such as density and temperature, but also the detailed chemical composition of the intervening matter. However, the interpretation of these spectra is made very challenging by the many sources contributing to the absorption of light. In order to take a full advantage of this new window into the nature of supermassive black holes we need detailed theoretical understanding of the origin of quasar spectral features. In a MIDAS PODS project we are applying machine learning to model and extract such features. We are training the models using data from the state-of-the-art numerical simulations of the early universe. This approach is fundamentally different from traditional astronomical data analysis. We have only started learning what information can be extracted and still looking for a new framework to interpret these data.

View Slides