AI Blood Test Predicts Neurodegenerative Disease Progression
A novel machine-learning algorithm that uses genetic data from patient blood samples can significantly predict and explain the progression of neurodegenerative diseases such as Alzheimer’s and Huntington’s.
The trial of nearly 2000 patients with late-onset Alzheimer’s and Huntington’s diseases shows that when applied to in vivo blood samples at baseline, the algorithm significantly predicted clinical deterioration and conversion to advanced disease stages.
Furthermore, the technique also allowed for discovery of genes and molecular pathways in peripheral tissues and brain tissues that are highly predictive of disease evolution.
“We found that if you reorder the data using the algorithm, you will be able to in some way give a molecular score to each patient that is highly reflective of the neuropathology and the clinical decline in that patient,” study investigator Yasser Iturria-Medina, PhD, told Medscape Medical News.
“A second finding is that we can make these predictions with data coming from the brain and from the periphery,” added Iturria-Medina, assistant professor of neurology and neurosurgery at The Neuro (Montreal Neurological Institute-Hospital), McGill University, Montreal, Canada.
“That’s very important because there has been a lot of emphasis on finding brain biomarkers of neurodegenerative diseases like Alzheimer’s. But what if we could get similar information from something simpler and less invasive than that?”
The study was published online January 28 in the journal Brain.
Most neurodegenerative disorders take decades to develop. Early detection of these disorders is made even more challenging by confounding nonpathological aging processes.
Not surprisingly, there is a considerable lack of longitudinal gene expression data covering the complete evolution of neurodegenerative conditions. In fact, the overwhelming majority of neurodegenerative studies are based on either cross-sectional or short-term longitudinal data. This further hinders the understanding of the underlying dynamic molecular mechanisms involved in the pathogenesis of neurodegenerative conditions.
“All of the data we have in this area come from a few time points in cross-sectional studies,” said Iturria-Medina. “So we don’t have a real picture of how these diseases evolve over a number of years or decades. It would be helpful if we had a way to analyze the data in a way that will reflect the disease’s evolution.”
Genomic and proteomic analyses may serve that purpose, particularly when coupled with machine-learning techniques that can tease out temporal components from cross-sectional data.
Previous research has demonstrated the ability of machine-learning algorithms to use data from cross-sectional research to reconstruct temporal paths of various disease states.
Such techniques create a relative ordering of individuals within a population to accurately identify a series of molecular states that constitute a longitudinal trajectory for a process of interest.
In the current study, Iturria-Medina and his team used these processes to analyze postmortem and in vivo gene expression neurodegenerative samples.
“We don’t have good biomarkers for neurodegenerative disease,” he said. “So a second aim of the study was to see if we can take a minimally invasive test such as a blood test and use it to give us an idea about which stage of disease patients are at, or how close they are to developing a particular disorder.”
Three Large Datasets
The investigators used data from 1969 patients with late-onset Alzheimer’s and Huntington’s disease from three large-scale databases, each of which was processed and analyzed independently.
Dataset 1 comprised RNA expression data from the prefrontal cortex of a subset of 489 autopsied individuals. Dataset 2 included 736 individual postmortem tissue samples, also from the prefrontal cortex. Finally, the third dataset used data from 744 patients with blood-gene expression information from the Alzheimer’s Disease Neuroimaging Initiative (ADNI).
Data were analyzed through a novel gene expression contrastive trajectory inference (GE-cTI) method, which reveals enriched temporal patterns in the diseased population. As part of the analysis, gene expression, neuropathology, and cognitive/clinical deterioration were assessed in the representative blood and tissue samples.
With an eye toward determining the molecular reconfigurations underlying neurodegenerative evolution, the investigators then reordered the gene expression patterns using a novel algorithm for detecting enriched trajectories in a diseased population relative to a background dataset of normal controls. Finally, the cTI was independently applied to the three populations, providing population-specific trajectories.
The study results showed that the algorithm strongly predicted neuropathological severity — as expressed by Braak, amyloid, and Vonsattel stages — in each of the three individual datasets, as well as the combined study cohort.
Next, the researchers examined whether unsupervised ordering of gene expression patterns in the blood samples reflect neuropathological severity and, if so, might be used as a marker of future clinical deterioration.
This analysis revealed that when applied to in vivo blood samples at baseline, the algorithm significantly predicted both clinical deterioration and conversion to advanced disease stages. Nevertheless, the algorithm was considerably less powerful when it came to predicting detailed alterations in memory and executive function.
Finally, the investigators set out to identify the genes, molecular functions, and pathways responsible for the accurate prediction of neurodegenerative progression in late-onset Alzheimer’s disease.
Interestingly, this analysis showed that as many as 90% of the highly predictive molecular pathways in the neurodegenerating brain were also among the most relevant pathways detected in the blood data.
The common blood–brain functional pathways relevant for progression to late-onset Alzheimer’s included blood coagulation, angiogenesis, p53, B-cell activation, and Wnt signaling.
Broad Clinical, Research Implications
The findings have broad implications for patient stratification in the clinic as well as monitoring response to personalized treatments in neurodegeneration, said Iturria-Medina.
“This test could one day be used by doctors to evaluate patients and prescribe therapies tailored to their needs. It could also be used in clinical trials to categorize patients and better determine how experimental drugs impact their predicted disease progression,” he noted.
In addition to shedding light on disease dynamics, cTI may enable the data-driven identification of novel subpopulations within an otherwise heterogeneous neurodegenerative population. This, the researchers explained, may open the door to precision medical approaches in these patients.
While the current study focused on neurodegenerative evolution, Iturria-Medina believes cTI can be applied to multiple neurological and neuropsychiatric conditions.
“We’re working on validating it more and testing it in different scenarios,” he said. “But assuming it all works out, there are two potential applications I see here. One is a complementary technique for patient diagnosis. This gives you the ability to stratify the patient and see at which point they are in their disease course.
“The second one is that since this test is minimally invasive, it gives us a direct way to monitor how someone is responding to treatment. If their trajectory is not changing or perhaps headed in the wrong direction, then maybe it’s time to change the treatment. If the trajectory progresses or is getting better, then it’s an indication that the person should continue the treatment.”
“A Good First Step”
Commenting on the findings for Medscape Medical News, David A. Bennett, MD, explained that while he was encouraged by the research, treating such disorders as Alzheimer’s disease is hamstrung by a lack of pharmaceutical options.
“These kinds of algorithms are affecting clinical care in other spaces,” said Bennett, director of the Rush Alzheimer’s Disease Center at Rush University Medical Center in Chicago, Illinois.
“The real problem that AI can’t solve, however, is that Alzheimer’s disease is simply not actionable, no matter how good you are at predicting it,” continued Bennett, who was not involved with the study. “The drugs that we have are poor, and there has not been a new drug approved by the FDA in more than 15 years.
“That said, you want the science to proceed in parallel,” he added. “You don’t want to wait until you have a drug to start doing this kind of research. When the drug comes out, you want to be prepared already with all kinds of decision-making around it. There’s still work to be done, but this is a good first step.”
For his part, Iturria-Medina believes the research opens the door to patient-specific treatment strategies.
“I see this as having strong potential implications for what is called dynamic personalized treatments, which allow us to continuously adjust the treatment based on how the patient responds,” he explained. “Because how a patients responds to treatment is not always reflected in the clinical data; that’s why molecular indicators are so promising.”
Iturria-Medina and Bennett have disclosed no relevant financial relationships. The study was funded by McGill University’s Healthy Brains for Healthy Lives Initiative, the Ludmer Centre for Neuroinformatics and Mental Health, the Brain Canada Foundation, and Health Canada.
Brain. Published online January 28, 2020. Abstract