Guardian AI-ngels: Artificial Intelligence Will Predict Patient Outcomes In Future Pandemics
An artificial intelligence algorithm has been used to analyze patient data from past viral pandemics including SARS, MERS and swine flu to find patterns in which genes turned on or off during infection, and identified two sets of genes that can be used to predict patient outcomes in future pandemics.
The multidisciplinary study reveals how gene expressions in patients with viral infections — including the novel coronavirus — can predict severity of illness and immune response, and how the model can be used to predict the outcomes of test therapies.
The research, which bridges a gap between medicine and computer science, was led by Pradipta Ghosh, Debashis Sahoo and Soumita Das at the University of California-San Diego.
Researchers identified two sets of genes involved in the process. A set of 166 genes reveals how a person’s immune system responds to infection, while a set of 20 signature genes can predict disease severity, including whether the patient will need to be hospitalized or put on a ventilator.
To test their algorithm against COVID-19, researchers used lung tissue collected from autopsies of infected patients along with animal test subjects.
“These viral pandemic-associated signatures tell us how a person’s immune system responds to a viral infection and how severe it might get, and that gives us a map for this and future pandemics,” said Ghosh, a professor of cellular and molecular medicine.
Ghosh, who is also executive director of the university’s HUMANOID Center of Research Excellence, said she and Sahoo “work on using innovative computational solutions to extract meaningful actionable insights from ‘big data.’
“We have used these approaches to identify novel biomarkers for diagnosis and prognostication, and even to pick high-value targets in the human genome/proteome that are more likely to successfully pass Phase III clinical trials [the typical bar a drug must meet for U.S. Food and Drug Administration clearance for use in the clinic, with enough efficacy and not too much toxicity].”
“While we had tested these approaches in the setting of other diseases, we never tested them on viral pandemics,” she said. “When the pandemic shut down all campuses around the world, we naturally asked each other why not use these to build a computational framework that will not only help dissect the current pandemic, but also maintain relevance in the pandemics of the future?”
Analyzing the novel coronavirus was not really different from analyzing other viruses, in terms of the challenges, Ghosh said.
“The human body reacts to any external threat by mounting a response. There are only so many ways how our cells respond to any pathogen,” she said. “One can safely assume that the cell will be stressed and will mount its defense system (inflammatory cytokines) and skew its intracellular signals and processes to try to control such inflammation and clear the pathogen.”
The team looked for the common, fundamental response of a human host to any viral respiratory pathogen, and identified the core gene expression changes using data from past pandemics.
“Once the signatures were fixed, then we started to test each and every emerging COVID-19 dataset,” she said, noting that a total of about 900 COVID-19 datasets have been analyzed to date.
“Any sample — including whole blood, immune cells, nasal swabs, airway lavage, lungs of deceased subjects — from these patients were analyzed, and the signatures worked. They distinguished healthy from [infected], mild from severe COVID-19, [and] predicted who might need prolonged hospitalization and/or ICU [intensive care unit] and ventilator support.”
A good surprise
The only unexpected outcome was that the signatures worked in pandemics that are not respiratory in nature, but are deadly, Ghosh said.
“For example, HIV, Hepatitis C, Zika, Ebola — the signatures worked in all of them. They also worked in diverse samples,” she said.
“Why? Because our algorithms were geared to ignore the superficial differences between sample types and/or the virus itself, and instead, home in on the fundamental core genes that are invariably changing in all those situations.”
Analysis of the genes involved revealed the source — and consequences — of cytokine storms, an immune response that causes inflammation and damages organs.
“We could see and show the world that the alveolar cells in our lungs that are normally designed to allow gas exchange and oxygenation of our blood, are one of the major sources of the cytokine storm, and hence, serve as the eye of the cytokine storm,” Das said. “Next, our HUMANOID Center team is modeling human lungs in the context of COVID-19 infection in order to examine both acute and post-COVID-19 effects.”
Researchers believe this information can help guide treatment for patients experiencing a cytokine storm.
To test the possibility, the team pre-treated rodents with either a precursor version of Molnupiravir, a drug being tested in clinical trials for the treatment of COVID-19 patients, or SARS-CoV-2-neutralizing antibodies, and then exposed them to SARS-CoV-2.
After exposure, the lung cells of control-treated rodents showed the pandemic-associated gene expression signatures. The treated rodents did not, suggesting that the treatments were effective in blunting the cytokine storm.
Sahoo, founding director of the university’s Center for Precision Computational Systems Network that sifts through big data to find meaningful information, said the endeavor “develops novel machine learning algorithms to drive precision drug discovery.”
Their research was published June 11 in the journal EBioMedicine.
(Edited by Judith Isacoff and Matthew B. Hall)