New research demonstrates advanced AI algorithms successfully and rapidly modeling clinical trials to find new uses for existing drugs and therapeutics.
When a new disease strikes, drug discovery is crucial — but so is drug repurposing.
In the US alone, there are over 20,000 pharmaceuticals approved for marketing, and finding more uses for them could help save lives. We’ve turned to AI to help us do just that.
Our research, “Emulated Clinical Trials from Longitudinal Real-World Data Efficiently Identify Candidates for Neurological Disease Modification: Examples from Parkinson’s Disease,” published in the journal Frontiers,1 demonstrates advanced AI algorithms successfully and rapidly modeling clinical trials to find new uses for existing drugs and therapeutics.
In the paper, we describe how we found two new potential treatments for dementia that typically accompanies Parkinson’s disease, known as PDD. This type of dementia is defined as acquired objective cognitive impairment in multiple domains, including attention, memory, executive and visuospatial ability.
One treatment is an existing Parkinson’s drug, rasagiline. The other is based on an entirely new use of a well-known insomnia medication, zolpidem.
First, we applied scalable causal inference technologies on retrospective data pulled in from our IBM Watson Health network. The data included volumes of de-identified – produced in a way so that to prevent someone’s personal identity from being revealed — electronic medical records and claims sourced from our industry partners.
The data allowed us to model and emulate clinical trials by generating simulated cohorts of individuals who either received or did not receive certain drugs. Our AI models were able to take into account the imbalances of different groups and cohorts, such as gender, comorbidities (the simultaneous presence of two or more diseases or medical conditions in a patient), Charlson’s Comorbidity Index,2 age at diagnosis with Parkinson’s and other factors and simulate a scenario of a randomized clinical trial.
In the study, we corrected for potential confounding and selection biases using two different causal inference approaches. The first one relies on balancing weight through the Inverse Probability Weighting (IPW) process. It involves reweighing patients to emulate random treatment assignment and uninformative censoring. We also used the method of outcome model, using standardization to predict counterfactual outcomes. More details about the method and code for inferring causal effect from observational data is here.
Our AI network also relied on medical domain knowledge. It organized the drugs using the Anatomical Therapeutic Chemical (ATC) classification system and assessed their effectiveness according to the outcomes detailed in the Electronic Health Records (EHRs) and insurance claims data. We tested the effect of hundreds of concomitant drugs on the emergence of Parkinson’s-associated dementia as one of the more common hallmarks of the disease’s progression.
Only those drugs for which a statistically significant effect was found independently in both EHR and claims data were further considered for repurposing potential. Our analysis unraveled therapeutic benefits of two drugs in decreasing the population-level incidence of dementia associated with Parkinson’s.
The first therapeutic our AI identified is an existing drug used to treat Parkinson’s, rasagiline, intended to slow down motor symptoms. This drug showed potential to also be useful for the reduction of symptoms of dementia.
In previous clinical trials, the use of rasagiline in patients diagnosed with Parkinson’s implied possible disease-modifying effects, albeit inconclusively. In fact, none of the previous studies showed the statistical power to support or refute the slowing the progression of the disease. The largest study, called ADAGIO, took place in 2009.3 It was a double-blind, delayed-start trial of rasagiline in patients suffering from Parkinson’s. But it also failed to demonstrate a dose-dependent effect on the Unified Parkinson’s disease Rating Scale (UPDRS) scores.
The second, perhaps more surprising finding, concerns a widely used insomnia medication called zolpidem that was not developed to treat Parkinson’s disease. A single prior report published more than two decades ago speculated that zolpidem would not be effective against Parkinson’s.4 That research was based on limited clinical experience with the drug and without specific consideration for cognition.
However, recent publications show zolpidem’s ability to treat a large variety of neurologic disorders, most often related to movement disorders and those of consciousness. Recent research also suggests that zolpidem induces transient effects on Unified Parkinson’s disease Rating Scale (UPDRS).5 Another very recent clinical study details the assessment of the benefits of low-dose zolpidem in late-stage Parkinson’s disease.6
While our research is an important use case, there is tremendous potential to repurpose other drugs for a range of neurodegenerative and infectious diseases. And AI can be of huge help.
With the current challenges of addressing the COVID-19 pandemic, the need to find new uses for existing drugs is crucial. We are now looking into novel ways to identify patients who were diagnosed with COVID-19. Using the Explorys and MartketScan data, our aim is to understand if and what drugs might have positive effects on long-term COVID.
Laifenfeld, D. et al. Emulated Clinical Trials from Longitudinal Real-World Data Efficiently Identify Candidates for Neurological Disease Modification: Examples from Parkinson’s Disease. Front. Pharmacol. 12, (2021). ↩
Charlson, M. E., Pompei, P., Ales, K. L. & MacKenzie, C. R. A new method of classifying prognostic comorbidity in longitudinal studies: Development and validation. Journal of Chronic Diseases 40, 373–383 (1987). ↩
Olanow, C. W. et al. A Double-Blind, Delayed-Start Trial of Rasagiline in Parkinson’s Disease. N Engl J Med 361, 1268–1278 (2009). ↩
Claflin, E. S., Townsend, W. & Peterson, M. D. Zolpidem for the Treatment of Neurologic Disorders. JAMA Neurol 74, 1130 (2017). ↩