PHILADELPHIA— A powerful new machine-learning technique can be applied to large datasets in the biological sciences to uncover previously unknown features of organisms and their genes, according to a team led by researchers from the Perelman School of Medicine at the University of Pennsylvania. For example, the technique learned the characteristic gene-expression patterns that appear when the bacterium is exposed to low-oxygen conditions and robustly identified changes that occur in response to antibiotics.
The technique employs a recently developed algorithm called a “denoising autoencoder,” which learns to identify recurrent features or patterns in large datasets without being told what specific features to look for. In 2012, for instance, when Google-sponsored researchers applied a similar method to randomly selected YouTube images, their system successfully learned to recognize major recurrent features of those images—including cats.
In the new study, published in the online journal mSystems this week, Casey Greene, PhD, an assistant professor of Systems Pharmacology and Translational Therapeutics, in collaboration with Deborah Hogan, PhD at Dartmouth College, used a system of denoising autoencoders to analyze many large datasets that measure how genes in the bacteria are expressed in different conditions. Greene is also a senior fellow at the Penn Institute for Biomedical Informatics.
“The system learned fundamental principles of bacterial genomics just from these data,” Greene said. “We expect that this approach will be particularly useful to microbiologists researching bacterial species that lack a decades-long history of study in the lab. Microbiologists can use these models to identify where the data agree with their own knowledge and where the data seem to be pointing in a different direction.” Greene thinks that these are cases where the data may suggest new biological mechanisms.
Last year, Greene and his team published the first demonstration of the new method in a biological context: an analysis of two gene-expression datasets of breast cancers. The new study was considerably more ambitious -- it covered all 950 gene-expression arrays publicly available at the time for the bacterium Pseudomonas aeruginosa, from 109 distinct datasets. This bacterium is a notorious pathogen in the hospital and in individuals with cystic fibrosis and other chronic lung conditions and is often difficult to treat due to its high resistance to standard antibiotic therapies.
First author Jie Tan, a graduate student at Dartmouth, where Greene, until recently, had his laboratory, developed ADAGE (Analysis using Denoising Autoencoders of Gene Expression) and applied it to the P. aeruginosa datasets. The data included only the identities of the roughly 5,000 P. aeruginosa genes, their measured expression levels in each published experiment. The goal was to show that this “unsupervised” learning system could uncover important patterns in P. aeruginosa gene expression and clarify how those patterns change when the bacterium’s environment changes, for example when in the presence of an antibiotic.
Even though the model built with ADAGE was relatively simple—roughly equivalent to a brain with only a few dozen neurons—it had no trouble learning which sets of P. aeruginosa genes tend to work together or in opposition. To the researchers’ surprise, the ADAGE system also detected differences between the main laboratory strain of P. aeruginosa and strains isolated from infected patients. “That turned out to be one of the strongest features of the data,” Greene said.
“We were struck by the similarities between P. aeruginosa grown in association with cultured lung epithelial cells and these bacteriataken directly from the lungs of individuals with cystic fibrosis,” said John H. Hammond, a graduate student in the Hogan Lab who collaborated on this project. “We are excited to continue to use ADAGE in combination with data from patient samples and experiments using laboratory models to discover better ways to find therapies to treat cystic fibrosis lung infections.”
“We think that the proliferation of ‘big data’ provides an opportunity, through the use of unsupervised machine-learning, to find completely new things in biology that we didn’t even know to look for,” Greene said.
Support for the research came from the Gordon and Betty Moore Foundation (GBMF4552), the William H. Neukom Institute for Computational Science, the National Institutes of Health (RO1AI091702, T32DK007301, P30GM106394), and the Cystic Fibrosis Foundation (STANTO07R0, STANTO15R0).
Penn Medicine is one of the world’s leading academic medical centers, dedicated to the related missions of medical education, biomedical research, excellence in patient care, and community service. The organization consists of the University of Pennsylvania Health System and Penn’s Raymond and Ruth Perelman School of Medicine, founded in 1765 as the nation’s first medical school.
The Perelman School of Medicine is consistently among the nation's top recipients of funding from the National Institutes of Health, with $550 million awarded in the 2022 fiscal year. Home to a proud history of “firsts” in medicine, Penn Medicine teams have pioneered discoveries and innovations that have shaped modern medicine, including recent breakthroughs such as CAR T cell therapy for cancer and the mRNA technology used in COVID-19 vaccines.
The University of Pennsylvania Health System’s patient care facilities stretch from the Susquehanna River in Pennsylvania to the New Jersey shore. These include the Hospital of the University of Pennsylvania, Penn Presbyterian Medical Center, Chester County Hospital, Lancaster General Health, Penn Medicine Princeton Health, and Pennsylvania Hospital—the nation’s first hospital, founded in 1751. Additional facilities and enterprises include Good Shepherd Penn Partners, Penn Medicine at Home, Lancaster Behavioral Health Hospital, and Princeton House Behavioral Health, among others.
Penn Medicine is an $11.1 billion enterprise powered by more than 49,000 talented faculty and staff.