Mega Doctor News
Newswise — The secrets of idiopathic pulmonary fibrosis (IPF) are written in its very name. Idiopathic refers to a disease of unknown cause, and the condition, which turns healthy lung tissue into fibrous scar tissue, still raises many questions.
IPF originates at the periphery of the lung and progresses inward, compromising more and more tissue and, eventually, making it difficult for a person to breathe. There is no cure for IPF, and neither of the two drugs that are approved as treatments can reverse the scarring—they only slow it down.
In a new study published June 20 in Nature Biomedical Engineering, researchers at Yale School of Medicine and collaborators took a significant step toward understanding IPF—and numerous other complex diseases—with an algorithm that interprets disease data and proposes treatments.
The research team developed a deep generative neural network called UNAGI (unified in-silico cellular dynamics and drug screening framework) that can identify patterns in disease-specific data. In a matter of hours or a couple of days, depending on the computer, UNAGI can learn how to glean insights from hundreds of thousands of cells, differentiating between cell types, picking out genes that are involved in disease progression, and identifying relevant regulatory networks. Then, it tries out different drugs, pulling from a long list of approved compounds to see if any of them work against whichever disease is being studied.
Although UNAGI was developed with IPF data, it can also be applied to changing physical states, such as aging, and other diseases, which the researchers demonstrated using COVID data.
“The model looks for regulation—what characterizes and regulates changes—and then, using known drug databases, also suggests treatments,” says co-senior author Naftali Kaminski, MD, Boehringer Ingelheim Pharmaceuticals, Inc. Professor of Medicine (Pulmonary) at Yale School of Medicine.
The research was done in collaboration with scientists at McGill University in Canada, KU Leuven in Belgium, and several other institutions.
AI reads between the lines
In IPF, disease distribution is variable, with some areas becoming “sicker” than others. Several years ago, Kaminski and collaborators at KU Leuven developed a method for tracking IPF progression that doesn’t require repeated follow-up with the same patients—which can be a challenge for researchers—by grading disease progression within a single sample.
The KU Leuven team collected diseased lungs during transplant surgeries and cut them into slices, selecting small pieces to represent different stages of the disease. The Yale team cataloged gene expression patterns in individual cells from these samples, yielding the pulmonary fibrosis single-cell atlas.
“We made some key discoveries, including novel cell types and populations,” says Kaminski. But the researchers wanted to understand disease trajectory better and lacked the analytical tools to do so with their data. “So, we started thinking about a way to apply AI to this problem.”
Kaminski teamed up with Jun Ding, PhD, an assistant professor at McGill University School of Medicine, who leads a lab that specializes in computational biology. The researchers developed UNAGI using sequencing data from 230,000 cells, provided by Kaminski.
“Our model is designed to represent virtual cells and virtual disease progression,” says Ding. Unlike most existing models, which are generic, UNAGI is disease-informed, meaning that it models the disease in question by identifying associated genes and regulatory networks. It then loops that information back into the model, refining its representation of cells and disease progression to add nuance. UNAGI also requires minimal researcher oversight, learning autonomously through an embedded iterative refinement process. By contrast, other models must be manually re-trained to interpret new datasets or test different drugs, which can be expensive and time-consuming.
UNAGI bypasses this need by integrating new information along the way, going deeper and deeper into the data until it can say with confidence which cells, genes, and pathways are involved in disease progression. “The model evolves to understand more and more about the disease,” says Ding. “It’s a bidirectional exchange of information.”
Then, pulling from a database of thousands of drugs with known mechanisms of action, UNAGI can test thousands of compounds and spit out a short list of potential therapeutics. In this study, it identified eight possible drugs, one of which is already used for IPF. The researchers selected one from the list that seemed out of place for their validation studies. Their pick—a calcium channel blocker called nifedipine—is used to treat hypertension, but UNAGI thought it might also have anti-fibrotic effects.
When the researchers applied the drug to slices of human lung tissue that were designed to model IPF, the drug blocked the formation of scar tissue, as UNAGI predicted it would. Even if nifedipine doesn’t end up being a good fit for fibrosis, says Kaminski, “UNAGI is hitting pathways we did not think about before.”
Merging technologies like single-cell sequencing with AI will shape the future of this field, he adds. “We have a convergence of both very sophisticated AI-based analytical methods and the ability to generate data that is high-resolution enough to actually make these observations.”
Researchers at Hannover Medical School in Germany, Pacific Northwest National Laboratory, Baylor College of Medicine, and University of Pittsburgh also collaborated on this work.
The research reported in this news article was supported by the National Institutes of Health (awards R01HL127349, R01HL141852, U01HL145567, R21HL161723, P01HL11450, and U01HL148860) and Yale University. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. This work was also supported by the U.S. Department of Defense, the Else Kröner-Fresenius Foundation, CORE100Pilot (Advanced) Clinician Scientist Program of Hannover Medical School, and the German Research Foundation.
