A Lawrence Livermore National Laboratory (LLNL)-led team of scientists and computational engineers has identified several existing medications that may be associated with longer survival in people with amyotrophic lateral sclerosis (ALS), using one of the largest electronic health record datasets ever assembled for ALS.
Published in The Lancet Digital Health, the study analyzed health records from more than 11,000 U.S. military veterans diagnosed with ALS between 2009 and 2019 and treated within the Veterans Health Administration. By combining causal-inference methods with machine learning (ML), researchers evaluated 162 medications to identify drugs prescribed for other conditions that were associated with meaningful differences in survival. The work was performed in collaboration with the Stanford University School of Medicine, the Veterans Affairs Palo Alto Health Care System (VA Palo Alto) and the University of California, Los Angeles (UCLA).
The timing of the study was driven by a rare convergence of data access and funding, said LLNL principal investigator Priyadip Ray, a research scientist in LLNL's Computational Engineering Division (CED), Beginning in 2009, when ALS was formally recognized as a service-connected disease, the VA saw a sharp increase in veterans receiving ALS care, creating a decade-plus record of detailed treatment data within a single healthcare system. At the same time, new targeted-funding programs made it possible to pursue ALS research at a scale that had historically been difficult for a rare disease.
The work was also motivated by recent setbacks in ALS drug development. Early enthusiasm around the drug Relyvrio, approved by the U.S. Food and Drug Administration in 2022, faded after a larger follow-up trial failed to show benefit and the drug was withdrawn from the market in 2024. That outcome underscored how difficult ALS clinical trials can be and helped spur interest in other discovery pathways.
"We realized that the large amount of experience with treating ALS in the VA system could provide an alternative approach to identifying medications for the disease," Ray said.
Rather than relying on traditional ML approaches, the team focused on causal inference - a more demanding framework that aims to isolate potential treatment effects while accounting for bias, confounding factors and uneven treatment patterns in real-world data, researchers explained.
"Our team developed a set of methods that combine rigorous statistical techniques with modern machine learning to isolate causal effects at the population level, even when data aren't collected in a controlled way," said co-author Braden Soper, an LLNL data scientist.
The analysis identified 27 medications associated with statistically significant changes in mortality risk. Notably, multiple drugs within the same therapeutic classes - including statins, phosphodiesterase type 5 inhibitors and alpha-adrenergic antagonists - showed similar associations with prolonged survival.
"Scientifically, what stood out was that there were multiple drugs in each of these groups that had the same positive effect," Ray said. "This gave us a great deal of confidence in the relationship between slowing ALS progression and these drugs."
To explore why these drugs might influence disease progression, the team used PathFX, a protein-protein interaction modeling tool developed by collaborators at UCLA and Stanford. The network analysis suggested that several of the identified drugs converge on shared downstream protein pathways, pointing to possible common mechanisms and new molecular targets for ALS research. The work builds on LLNL's broader investments in causal modeling, ML and computational tools for biomedical and national security applications.
Ray emphasized that the findings do not prove clinical benefit but provide a strong foundation for next steps, including deeper modeling that accounts for time-varying health factors and validation in independent datasets that include more diverse civilian populations.
"Because sensitive medical data are hard to share due to privacy and access restrictions, we're working to release our software pipeline as open source so researchers anywhere can apply these tools to their own datasets, diseases and interventions," added Soper.
Co-authors on the study include Andre Goncalves and Jose Cadena Pico from LLNL's CED; Amy Gryshuk (formerly of LLNL - now at the University of California, San Francisco's Innovation Ventures); Richard Reimer and Thomas Osborne of the Stanford University School of Medicine and VA Palo Alto; Jennifer Wilson of UCLA; Paola Suarez of VA Palo Alto; and Kevin Grimes of the Stanford University School of Medicine.
