CHICAGO (GenomeWeb) – A new partnership between the Scripps Research Translational Institute (SRTI) in San Diego and computational technology company Nvidia seeks to marry artificial intelligence and deep learning with genomics and digital health sensors to advance predictive medicine.
SRTI and Santa Clara, California-based Nvidia recently announced that they would work together to create a center of excellence in genomics and digital sensors, developing best practices for AI in these two domains. In the process, they will attempt to bring the genomics, sensor, and medical imaging worlds together.
Nvidia is best known as a hardware company, including graphics processing units, but it also is involved in software algorithm development for hardware acceleration and high-performance computing. The company, which has been working in life sciences for a dozen years, has been partnering with clinical organizations since about 2015, with many of its early users being in radiology.
"It is great to utilize all of the digital data we create with images, but I think, to tell the full story of what is happening at an individual level, we need to start getting a lot smarter at being able to operate on data that isn't images," said Kimberly Powell, head of healthcare operations at Nvidia. "That's what triggered this interest in this partnership with Scripps."
It just so happens that genomics and sensors — smartwatches, blood-pressure cuffs, continuous glucose monitors, smart weight scales, and the like — are the two main focuses of SRTI in pursuit of "digitizing human beings," in the words of Director Eric Topol.
"There are really just three ways to do that. One is sensors and one is images and one is [with] biologic layers like DNA omics," Topol said. "The rest of it is just descriptive stuff, like your electronic health record."
He noted that EHRs are fairly ubiquitous in US health systems now, but not many institutions have the thousands of whole-genome sequences and thousands of records of continuous sensor data from humans, as Scripps does. With this data, SRTI and Nvidia will focus on deep learning for better mutation detection and correlation of mutations with wearable data to predict diseases.
"They have some very unique datasets. We have our own applied research team in genomics that's going to be working side-by-side with their data scientists as well as their clinical specialists," Powell said.
"Of course we're going to be working on research, but in the process of doing that research, we're going to understand what kind of computational tools and infrastructure are needed to be able to do this kind of research," she added.
Nvidia has partnered with other genomic research institutions, but with deep learning and AI, SRTI is trying to harness the power of the Scripps databases. "We're trying to take it to the next level," Topol said. One of the main goals of the center of excellence is to publish research, he added.
For the last six years or so, Nvidia has been investing in high-performance AI computing platforms to help clients create neural networks to support deep learning.
"We're now in this phase of making AI very accessible to the domain experts of various industries," Powell said. "We want to make sure that it's not just the computer-science student or PhD who can develop these deep-learning networks, but rather it's the domain expert."
Since the computer essentially is writing code on the fly, domain experts — perhaps geneticists, radiologists, or pathologists — need to teach the computer how to learn. "It's really a completely new approach for developing software," Powell said.
The SRTI-Nvidia center of excellence will be setting ground rules and developing best practices for AI in both genomics and digital health, including sensors and wearable devices. "We're starting with whole-genome sequences, but we'll probably get into biologic layers over time, beyond just DNA sequencing," Topol said.
The work will start with atrial fibrillation, a longtime research interest of Topol. "If you can detect that you might be going into that state, what can you do to intervene?" he wondered.
"I think we will be jointly publishing research against the datasets that Scripps has and the problems that we're trying to solve," Powell said. Topol, a prominent cardiologist, has years of data on atrial fibrillation, for example. "We're going to do our first work along that disease state, and then we will move on to other disease states," she said.
Researchers will take independent looks at data from genomes and data from digital wearables. The latter may be just as resource-intensive as the former.
"Doing sensor data is tricky because it requires recurrent neural networks," Topol said. "When you have someone with continuous, every-minute or every-second data, analyzing that data across large numbers of people is not easy to eke out … signals that a human being couldn't see."
A third run will go "multimodal," according to Topol, addressing how to integrate the two datasets and draw in from EHRs and other sources of information to advance predictive capabilities. "No one has really done that yet, by the way," Topol said.
"Being able to use several data sources is really going to be the Holy Grail," Powell added.
One key difference between the two domains is that sensor data changes frequently, while genomic data is mostly static, with the notable exception of tumor genomes.
"I think [wearables] are going to give a very interesting perspective and change the way we think about how we deliver medicine, how we predict disease, and how we can prevent disease from happening in the first place," Powell said.
"With imaging, if you're going in for an MRI, you're largely symptomatic already. Something is already wrong," she noted. "But we also want to get into the prevention stage to help people live healthier, happier, longer."
While there is no formal roadmap for the collaboration, Powell said that Nvidia and Scripps will try to show progress in increments of six months. "This isn't meant to be a partnership where you wait five years to see [results]," she said.
Topol said that the partnership will be moving into diabetes before long. "Today, the algorithm [for diabetes care] is a dumb algorithm. I mean, all you have is glucose that goes up or down," he said.
"We know that things like physical activity, what you eat and drink, your sleep, your gut microbiome, your level of stress, and your genomics — all these things have a very potent impact on your glucose. Both the genomics and your gut microbiome are biologic layers," Topol explained.
"If you want to develop a smart algorithm, you are going to need to fold in all of these data elements. These are critical determinants. There may be others," he said.
"What we intend to do is start to get multimodal over time," Topol said. "We're starting out with a couple of projects that are well-delineated, and then over time, [we will] start to get into the world of what I would label as smart algorithms, not the ones we have today, which are rules-based, heuristic, and very primitive."