Potential treatments for amyotrophic lateral sclerosis (ALS) and other neurodegenerative diseases may already be out there in the form of drugs prescribed for other conditions. A team of researchers from Lawrence Livermore National Laboratory (LLNL), Stanford University and the University of California, Los Angeles (UCLA) are using artificial intelligence and machine learning (AI/ML) to try to find them.
Clinical trials for new drugs can take up to 5-7 years, so repurposing existing drugs is one of the best ways to deliver treatments quickly. AI/ML can make it even faster. By analyzing long-term electronic health records (EHRs) of patients with ALS, the team can identify drugs - or combinations of drugs - prescribed for other conditions that may influence the progression of the disease. The drugs' "off-target" effects may not only affect patient survival but also provide insight into how neurodegenerative diseases work and inform better therapies.
"If you talk to any ALS caregiver, you will be moved because the disease has such a grim prognosis, so being able to do something is tremendously motivating," said Priyadip Ray, a staff scientist in LLNL's Computational Engineering Division (CED) who leads the effort.
Computers to clinics
The Center for Disease Control estimates that as many as 31,000 Americans suffer from ALS (also known as Lou Gehrig's disease), with veterans being diagnosed at higher rates than the average population. The disease attacks motor neurons in the spinal cord and brain, causing increasing mobility loss until the body shuts down, usually within 2-5 years of onset. Its cause is unknown, there is no cure, and the only three FDA-approved drugs have a minor impact.
However, the emergence of EHRs - digital files with patients' medical history, prescriptions, demographic information and more - has opened the door for unprecedented research opportunities.
"ALS is a relatively rare disease, and it has a rapid onset, so we really don't have the numbers or the time to run large clinical trials," said Ray. "The [EHR] data is critical, because now we can use advanced AI/ML tools to create good, high-confidence hypotheses, and we can do 1-3 targeted clinical trials that have a much higher rate of success."
In a clinical trial, a group of similar patients are randomly given either a treatment or a placebo. If the half that received the treatment has a better outcome, it proves that the treatment works. With EHR data, Ray and his team use a technique called causal machine learning.
"Causal machine learning creates a sort of synthetic clinical trial," he said. "We looked for patients who were given a particular drug and matched them with a group of patients who are very similar and who were likely to be given that drug but were not."
Moving with (re)purpose
Ray, his CED colleagues, Braden Soper, Andre Goncalves and Jose Cadena Pico, and their collaborators began by creating a surrogate model (a mathematical approximation) of ALS progression with a small publicly available EHR dataset. Through seed funding from the ALS CURE Project - established by LLNL employee Mike Piscotty in memory of his wife - the team was able to access more than 20,000 EHRs of veterans with ALS from Veterans Affairs (VA). After the EHRs were scrubbed of individualized information, the team investigated risk factors for ALS and received funding from the Department of Defense for further analysis.
The team studied 162 drugs that patients were regularly taking around the onset of ALS and identified three classes that had a significant positive effect on survival: statins (which reduce cholesterol), alpha-blockers (which reduce blood pressure and relax muscles) and PDE5-inhibitors (which treat erectile dysfunction). They also found that combining statins and alpha-blockers had a synergistic effect.
The team found a few early-stage studies on these drugs and ALS that backed up their results, suggesting they could all be good repurposing candidates. Collaborators at Stanford and UCLA collaborators also ran protein-protein interaction studies on each of the drug types and found a few common downstream protein targets - what the drugs ultimately affect.
"We are pretty excited about these initial findings," said Ray. "If we can also identify these shared downstream protein targets, we can make drugs that specifically target those proteins and work even better."
Since the VA data skews heavily toward men with military backgrounds - both risk factors for ALS - the team aims to corroborate and generalize their results. To do this, they plan to analyze millions of patient files from the Optum EHR dataset, which they gained access to thanks to new funding from the ALS network, the ALS CURE Project, the Livermore Lab Foundation, RDM Positive Impact Foundation and Stanford University. They also plan to apply their AI/ML approaches to study Parkinson's disease, which Ray hopes will shed light on treating all neurodegenerative diseases.
Meanwhile, the team seeks funding to validate their findings in a clinical setting, which would not only be one of the final steps of getting the drugs approved to treat ALS but also confirm that their approach works.
Ray feels grateful for the opportunity to use AI/ML to make a difference in medical research and the Lab's unique infrastructure and connections with academia, industry and government that make it possible.
"The Lab recognizes that building these tools and working with patient data can have a tremendous impact," he said. "The opportunity to make a difference on healthcare as well as national security motivates me to work on this high-impact research."
-Noah Plueger-Peters
