Skip to main content
Premium Trial:

Request an Annual Quote

GNS Healthcare's Machine Learning Platform Leads to Novel Parkinson's Gene Discovery


CHICAGO (GenomeWeb) – Recently published research powered by GNS Healthcare, maker of a causal machine-learning platform for precision medicine, promises to unravel some of the mysteries of Parkinson's disease, specifically the rate of progression of the neurodegenerative condition.

Using data from the Michael J. Fox Foundation-sponsored Parkinson's Progression Markers Initiative, researchers from Cambridge, Massachusetts-based GNS Healthcare and from the University of Rochester were able to identify genetic and molecular markers of rapid motor progression of Parkinson's. The work, described in an article that appeared in Lancet Neurology last week, confirmed the role of one biomarker that had a known association with Parkinson's disease and also discovered a novel predictor.

"What's interesting is that it's now the two of them in combination that work" in predicting the pace of progression, said Iya Khalil, cofounder and chief commercial officer of GNS Healthcare and one of 10 listed authors of the Lancet Neurology article.

The researchers came to their discoveries by running PPMI data and other clinical information through GNS Healthcare's causal machine-learning platform called REFS, which stands for Reverse Engineering and Forward Simulation.

REFS takes in many types of data, including whole-genome sequences, proteomic data, laboratory results, patient records, imaging readouts, and motor progression scores, over time. "It reverse-engineers or learns from the data it reconstructs that causal mechanism that gave rise to that data, the causal biological mechanism that explains why we're seeing patients progress, [which] patients would be better off with other treatments, observing causality. It's not just extracting a pattern of the data," Khalil said.

"Once we've learned these models, we can run simulations and run different scenarios" to test different drugs or different gene perturbations, Khalil explained.

"These analyses around the simulation that can happen on the computer very quickly can lead to much more powerful hypotheses that we can then give to clinicians to inform their decisions, or to researchers on the biological side," she said.

In the case of Parkinson's, nobody knows the cause of the ailment or why the disease progresses more quickly in some patients than in others.

Seeking answers, the Michael J. Fox Foundation in 2010 established the PPMI, a collaboration between academic, government, and industry aimed at verifying progression markers for Parkinson's disease. The PPMI includes a longitudinal, standardized repository for Parkinson's researchers, featuring data on motor skills, sleep patterns, and other phenotypic measures. Some records have been combined with imaging analysis, genome/exome sequencing, and targeted, SNP-array data, according to Khalil.

"We thought these are exactly the kinds of analyses and data that we want to now push through our platform and see what insights we can learn," Khalil said. "It's a really great match because they had this rich data and we had a platform that could learn from that data new insights and apply it directly to that very specific patient that they have a mandate for helping," Khalil said.

Indeed, the MJFF provided funding for the GNS Healthcare study, with additional financial support from the National Institute for Neurological Disorders and Stroke, one of the US National Institutes of Health.

GNS Healthcare took the PPMI data and adapted it to a series of measures to train the machine-learning algorithm. The company then did the same with genotypes, imaging reports, outcome measures, and other clinical data points for patients in the new study's cohort of 312 Parkinson's patients.

"We then let REFS learn all the possible combinations of things that could explain the rate of motor progression in these patients," Khalil said. This included genotypes in combination with baseline clinical measures.

"It's a massive computational effort to go through all of those possibilities. There are billions and billions of models now that could explain outcomes, and we run those until we come up with the best set of models that could explain the outcomes," Khalil explained.

"It turns out that the best explainer for the rate of motor progression wasn't a single SNP. It was a combination of SNPs, along with a number of clinical variables at baseline. It's a complex set of interactions that explains the outcomes," she said.

Khalil said that a standard genome-wide association study would have missed the genomic markers the GNS team found. "It comes down to a combination of using vast amounts of compute power to search through many possibilities, as well as the mathematics that helps you pull out that unique insight."

The University of Rochester independently validated the findings, which prompted publication in Lancet Neurology, a journal that does not often feature studies of machine learning.

It took "three pillars" to make this study work, according to Khalil. "It's not just about the genes. It's not just about the clinical experience or the clinical disease trajectory. The machine learning is what allows you to get insights from all of those things," she said.

The causal machine learning also is what GNS Healthcare is counting on to advance its work.

"Just being able to go full circle between the data to the algorithm to the clinical validation and now thinking through the clinical implementations have been hugely useful," Khalil said.

"With this core piece of technology that we've built up that learns from lots of different data parts, learns causality in a probabilistic way from that data, [and] allows you to run these different simulations on a per-patient level, we're able to apply this across the healthcare ecosystem, from the discovery of novel biology to the translation of that to patient and clinical care to how whole health systems operate and can be optimized to benefit the patient," Khalil continued.

"At the end of the day, for this to become real, for us to go from these large datasets to actionable insights quickly, we need to get to something actionable," she said.

Until recently, some researchers and clinicians were not ready to adopt this approach because they did not understand the technology, while others were not collecting the right kind of data. "But that's changing," Khalil said. "People are understanding the value of these things and there's the data to support the ability to learn really powerful insights from that data."

GNS Healthcare has collaborations in place with pharmaceutical companies including Celgene as well as health insurers. The company is just starting to work with care providers as well. Last month, GNS Healthcare announced a deal to aggregate patient data from the breast cancer registry and personalized medicine research program at Seattle's Swedish Cancer Institute, then build computer models to simulate potential treatments and their effect on patient outcomes.

Khalil also said that GNS was "in discussions" about the next steps in the Parkinson's research.

Parkinson's can take years to manifest itself, so it's difficult to diagnose early and to measure progression in the immediate aftermath of a diagnosis. "You could be giving somebody a drug and you could think it's working, but it may not work. They may just be a slow progressor," Khalil said.

"By showing which patients are going to progress slowly vs. quickly, we can actually optimize the trial and test drugs faster," Khalil said. Such a trial would require about 20 percent fewer patients, she noted. Such testing could also help get fast-progressing patients on more aggressive treatments sooner.

"Our goal remains to replicate this, do more within Parkinson's, but also replicate this within other disease areas," Khalil said. She noted that GNS is working with several pharma companies trying to replicate the MJFF study for Huntington's disease, multiple myeloma, and other conditions.