In Virtual Trials, Two Alzheimer’s Drugs Do Poor Job Of Slowing Cognitive Decline

By Deborah Borfitz 

December 14, 2022 | Researchers from Duke, Penn State, and the University of Tennessee (UT) collaboratively built a mathematical model of Alzheimer’s disease that used real, de-identified patient data to simulate health outcomes from a pair of anti-amyloid-beta drugs. Notably, the virtual (aka in silico) clinical trials employed a computational approach that provides “individualized prescriptions” based on optimal timing and dosing of treatment while minimizing possible side effects. 

The pharmaceutical industry has been interested in developing and employing computer-simulated clinical trials to identify their most promising compounds to advance into preclinical and early human trials, says Jeffrey Petrella, M.D., professor of radiology and director of the Alzheimer Imaging Research Laboratory at Duke University. Traditionally, companies have poured hundreds of millions of dollars and a decade or more time into a preclinical and clinical trials pipeline for drugs that originally showed promise based on having a druggable target and an identified market need. 

A virtual clinical trial, in contrast, can quickly evaluate the potential of a variety of drugs in humans in a short period of time, “to separate potential winners from losers,” Petrella says. As recently described in PLOS Computational Biology (DOI: 10.1371/journal.pcbi.1010481), the personalized mathematical modeling approach was used for the first time to directly compared the Alzheimer’s drugs aducanumab (an FDA-approved drug developed by Biogen) and donanemab (a compound of Eli Lilly given breakthrough therapy designation by the agency last year) to confirm what was found in actual clinical trials—without the advantage of having study data on either of the two drugs in hand.  

“What we found aligns almost exactly with findings in prior clinical trials of 18 months duration, but because we were using a virtual simulation, we had the added benefit of directly comparing the efficacy of different drugs over longer trial periods,” says Wenrui Hao, Ph.D., associate professor of mathematics at Penn State. Treatment with either of the two medicines, when studied over a simulated 10-year period, had a minimal effect on slowing cognitive decline in patients. 

Up until last year, the U.S. Food and Drug Administration had approved six medications for Alzheimer’s disease, all of them symptomatic treatments. “These Band-Aid approaches are merely jump-starting the brain cells that you have left rather than preventing further loss of brain tissue,” says Petrella. Aducanumab and donanemab are among the first of a class of agents that aim to modify the disease process itself.  

Defying Intuition 

With the virtual trials for the two Alzheimer’s drugs, the goal was to maintain cognitive function for as long as possible over the trial—importantly, while minimizing side effects such as brain swelling, bleeding, headaches, dizziness, nausea, confusion and vision problems, explains Petrella. Suzanne Lenhart, professor of mathematics at the UT, Knoxville, used her expertise in “optimal control theory” to arrive at the optimal dosing scenario for administering each of the medicines to individual patients.  

“Computational scientists are limited in their ability to develop models for personalized medicine by the quality of available datasets to validate what happens to a drug in the body and where,” says Lenhart. “Having a physician on the team made all the difference,” she adds, noting that her prior simulation models sometimes factored in side effects but only in a vague way. It’s a step forward to consider important, and sometimes life-threatening, side-effects of a drug prior to in silico dosing, just as it would be before administering it to real patients in a clinical trial.  

For the virtual trials, researchers set the trial timeframe for both medium-term (18-month) and long-term (10-year) periods with low-dose (6 mg/kg) and high-dose (10 mg/kg) regimens for aducanumab, and a single-dose regimen (1400 mg) for donanemab. These are the same doses used in human trials for the drugs.  

“The ability to model a disease computationally, especially using biomarkers, represents a new era for clinical trials where mathematics can be used to recapitulate disease mechanisms,” says Petrella. “Given the complex multifactorial mechanisms of certain diseases, such as Alzheimer’s disease, combined with the varied therapeutic targets of different drugs, knowing which drug or combination of drugs to invest in often defies intuition.”  

‘Black Box’ 

“The platform used in the study could easily test combinations of different drugs in addition to individual drugs separately, which is where an accurate depiction of side effects would potentially be even more critical,” Petrella says. Predicting the effect of a drug combination is difficult without the benefit of detailed clinical trial data on the drugs being modeled. 

Researchers were unable to get access to proprietary trial data on aducanumab and donanemab to validate their model, so they played forward certain assumptions about the drugs based on published study findings, he notes. But what they really need is individual-level data showing patient-to-patient variation in drug responses and side effects. 

“Given the clinical trial data for the two drugs, the research team could revisit the hypothesis arising from the modeling exercise that donanemab’s efficacy is higher than that of aducanumab,” points out Hao.  

Creation of the virtual patients was based on biomarker data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohort, a freely available longitudinal natural history database of patients across the spectrum from cognitively normal to dementia. The same data was used to incorporate time-varying treatment controls and side effects into the model. “The observations of interest were the rate of movement between biomarker-related compartments,” says Lenhart.  

“In simple terms, the model can be thought of as a black box where the input is real patient data and the output is a prediction of how that patient is going to respond to a drug [i.e., the outcome of a clinical trial],” says Petrella. “You can turn the dials on the box and up the dose of the drug, you can change the side effects of the drug—that is, you can do all kinds of the things that may affect the response—then assess the outcome. 

“These include scenarios such as administering a drug earlier and earlier, testing its limits... [which] would never be done in the context of a real clinical trial,” he adds. In the latest virtual trials, for example, aducanumab had a barely detectable effect on preserving cognition in later years, even when the drug was administered to asymptomatic people in their 20s who were at elevated risk of developing Alzheimer’s disease.  

With the optimization, the simulated results indicated that the “personal optimum treatment” is to gradually work up to a maximum dosage and then hold steady, says Hao. Treatment with aducanumab was shown to have only a modest mitigation effect over both the short and long term, leading to the conclusion that the drug is “not very effective for slowing the cognitive decline associated with Alzheimer’s disease.” 

Balanced Approach 

Moving forward, Petrella says he foresees the computational causal model being refined for use in virtual trials of other drugs under development for Alzheimer’s disease—including more than 10 new therapies targeting different types of amyloid. Among the targets, alone or in combination, are soluble amyloid-beta protein that can be present in blood and cerebrospinal fluid, and tau protein, a marker of the downstream neurodegenerative process of Alzheimer’s disease.  

The model will most immediately be useful for predicting the results of the next wave of clinical trials over the next one to two years, primarily for drugs that aim to reduce amyloid plaque build-up in the brain, says Petrella. Most recently, Genentech and Roche announced disappointing results from their Phase 3 GRADUATE trials of gantenerumab, an anti-amyloid antibody that showed smaller effects on removing amyloid brain plaques than expected and failed to slow cognitive decline. Trials for other classes of Alzheimer’s disease treatments, including neuroprotective and cognitive-enhancing drugs, have had a high failure rate on cognitive endpoints and could likewise benefit from personalized treatment optimization.  

The idea of “tailoring” administration of a drug for individual patients is being increasingly embraced by the clinical trials community, Petrella says. “The idea is that one size doesn’t fit all, because that approach has been shown to fail.” In cancer, notably, patients’ response to therapy can differ based on their genetic profile, age, and hormonal status.  

Lenhart says she has built treatment models for many diseases to help with decision-making about the trajectory of treatment for a given patient. Her latest one for arriving at the optimal anti-amyloid-beta therapy for Alzheimer’s disease is just “a bit more realistic.” 

“The model employs a common mathematical method known as ordinary differential equations, making it straightforward to implement, in theory,” notes Hao. “Which isn’t to say this cascade of Alzheimer’s pathobiology was easy to build in the first place, or to optimize treatment effects, for that matter.” 

The model represents a balanced approach between two extremes, remarks Petrella, “one where you would make all these assumptions about the disease and then play that forward for your prediction, and the other where you assume nothing about the disease and let the data speak to you.” Though objectively appealing, the main disadvantages of the data-only approach are that the output is less interpretable, it doesn’t take advantage of decades of knowledge gained about the disease from other sources, and, finally, it requires lots of data that we might not yet have.”