Exploring the Intersection of Genomic Data and AI in Healthcare

AI and genomics: A tailor-made match

To make the promise of precision medicine treatments a reality, the first step is translating genomic data from an incomprehensible resource to a meaningful medical asset -- and AI will be key in that undertaking, Denny noted.

“AI can enhance the interpretability of genomic data, and we're trying to convert genomic data to actionable clinical information. We want to improve disease diagnosis, understand what medications we should give and to whom, minimize side effects and maximize efficacy. All of those things require us to go from raw data to interpretability,” he said. 

“AI will hopefully help us shorten that process because there are so many variables involved and AI can analyze data faster. Additionally, if humans try to find the right variables by purposeful investigation, we can impute our biases onto the data. The hope is that we move from data to actionability faster and in ways that we wouldn't have considered otherwise.”

The All of Us Research Program is leveraging AI and cloud computing technologies to make biomedical data available to COVID-19 researchers across the country. In a partnership with Intel and Google Cloud, the program will make genomic information, EHR data, and socioeconomic factors broadly accessible through the Researcher Workbench.

The collaboration will help speed data analysis and research for COVID-19, as well as a wide range of other diseases. 

“Intel will provide funding credits for use of the Google Cloud platform, which will host all our data to help accelerate COVID-19 research. All of the computing on the platform will help us increase access to biomedical data and accelerate research nationwide,” said Denny.

Elsewhere in the research sector, entities are partnering with vendors to implement data-driven analytics tools and identify genetic biomarkers of disease. The University of Utah recently announced a collaboration with Renalytix AI to improve kidney health and reduce the risk of kidney failure for large-scale populations in the earliest stages of disease.

“Partnerships like these allow companies to bring their strength in research and development as well as actual commercial application together with data and faculty expertise,” said Maia Hightower, MD, MPH, chief medial information officer at University of Utah Health.

“There are all sorts of data sources that are housed within a provider organization that industry doesn’t have direct access to. We’re hoping to provide cutting-edge clinical decision support that's personalized to an individual kidney patient’s journey based on their Renalytix results, their molecular data, and their lab data.” 

Ultimately, the partnership will aim to accelerate the availability of precision medicine therapies for diabetes and kidney disease patients in Utah.

“Combining our data in benchmarking and best practices is going to assist in the diagnosis and treatment protocols in these areas,” said Donna Roach, MS, chief information officer at University of Utah Health. 

Researchers are also using AI and genomic data to gather new insights about incurable, prevalent diseases. Understanding how these conditions vary between individuals can open the door to novel, personalized treatments, and analytics tools can help the industry get there.

“Many diseases are much more heterogeneous than originally thought. For example, what people typically refer to as Alzheimer's is probably a whole spectrum of different pathologies that present with similar symptoms, and the underlying mechanisms of how the disease occurs might be fairly different between patients,” said Clemens Hug, computational biologist at the Laboratory of Systems Pharmacology at Harvard Medical School.

“There's definitely potential to develop targeted therapies for specific subtypes of patients, and machine learning models and AI hold much promise for discriminating between different patterns of disease.”

In addition to discerning varying forms of disease, AI and machine learning can examine the impact of drugs on disease progression and human biology. In a recent study, Hug and his team used a machine learning-based method to screen currently available medications for their potential to treat Alzheimer’s disease.

While there are currently a handful of approved Alzheimer’s drugs, none of these medications actually stop or slow the progression of the disease, Hug said. The machine learning method could offer a rapid and inexpensive way to propose new therapies for the condition.

“A big challenge in trying to identify new drugs is that clinical trials are extremely costly, often tens to hundreds of millions of dollars. It would be great if you could effectively pre-screen the drugs that go to clinical trials, and one way to do that is to use machine learning models to prioritize drugs to go into trials,” said Hug. 

“We don't claim that we can use machine learning to definitively say that a drug is going to work, but we can pre-select a set of drugs that we believe have a greater chance of succeeding in a clinical trial.”

After identifying medications that can treat Alzheimer’s, AI tools can help support precision medicine strategies by determining which drugs will perform optimally in specific patients.

“If our drugs work in patients, then we expect them to only work in a subset of patients, not in all of them. We would need to use machine learning and genomics to identify which patients are most likely to benefit from certain treatments,” said Hug.  

Oversight, biases, and errors – oh my

While decoding genomic data seems like a task perfectly suited for AI, some familiar challenges can come with using the technology to support precision medicine strategies. 

Issues with underrepresentation, bias, and potential inaccuracies are prevalent in algorithm development and use, and researchers need to consider these obstacles when applying analytics models to genomic information.

“A lot of the challenges of AI are very similar to challenges of traditionally analyzing data,” said Denny.  

“There’s been a lot of discussion about how some clinical analyses and scoring mechanisms are based on non-diverse, non-representative datasets. This is especially true for genomics, where the vast majority of genetic data is derived from people of European ancestry. A minority of about four percent would represent African or Latinx ancestries. With that kind of dataset, you can't necessarily derive accurate conclusions for diverse populations.” 

Underrepresentation of certain groups can translate to an algorithm that doesn’t benefit all patients – and that’s a chance that researchers can’t afford to take. 

“Some of the conclusions may hold through and some models may be useful in all populations, but some may not. The problem is you don't know which ones work well and which ones don't if you don't have datasets to test them,” said Denny. 

“You need lots of representation and diversity across the data sources and data seeds. Machine learning algorithms analyze so many data inputs that they can be much more sensitive to biases. Having diverse data and rigorously testing your algorithms is important, especially when it comes to the healthcare industry.” 

Hightower also emphasized the gravity of building unbiased models. 

“There's been a lot of advancement in using AI and machine learning for clinical forecasting. But we've seen some examples where these models can perpetuate certain biases,” she said. 

“This has stalled some of the enthusiasm for AI when it comes to using algorithms for prediction in clinical care. It’s a little different when you talk about automating processes, because there’s less risk associated with automating your billing cycle. There is a lot of potential around augmented clinical decision support, where a clinician is interpreting the information, but we still need to understand what’s going on inside the algorithm.”

With AI and machine learning increasingly edging into clinical settings, algorithm development is just one factor to consider. Training medical professionals on how to best utilize these tools presents another set of critical challenges, Roach added.

“As we educate medical students and residents, the question is how should we incorporate this kind of training in medical schools? How do we show students and residents the best way to use AI in healthcare?” she said.

As the technology becomes more complex, comprehensive training and education will only grow more essential.

“Right now, we're facing a challenge with our algorithms that use race, because we’ve recognized that race has been used as a surrogate for social determinants and other factors. Even now, we’re just learning that it’s important to make our clinicians aware of how race can affect the output of an algorithm. And this is just in straightforward regression models, let alone more complex algorithms,” said Hightower.

“So, how do we translate that into clinical medicine, a field that relies heavily on human decision-making? It will be interesting moving into the future and training our students and residents.”

Considering the potential of genomics and AI in healthcare

With so many industries working to provide personalized consumer experiences, it only makes sense that healthcare would follow suit. The health of each individual depends on so many different factors that clinical treatments can’t simply be one-size-fits-all. 

Genomic data analysis will provide the basis for precision medicine and targeted, effective therapies. 

“We are very excited about the resource we're building. We’ve started to return genetic results to participants, and we're providing information on their genetic ancestry. We're looking forward to returning health-related traits – probably in about a year's time – that represents a real foray for a program of our size. We want to be able to return information that might have actionability,” said Denny.

“It's really important to try to return the value to our participants, and this is one of the ways in which we can do that. It's part of that loop of creating a research program that engages participants and researchers, and it creates the foundation for artificial intelligence and other kinds of approaches.”

With AI and genomic data, researchers and providers can make new connections and gather new insights about diseases they may have never discovered had they not had these resources.

“We're very excited because a lot of the drugs that have been tested for Alzheimer's were focused on a specific target – that is, plaques that form in the brain, which are a hallmark of the disease,” said Hug. 

“There are a lot of drugs that were developed that are very good at dissolving these plaques, but they didn’t substantially help the patients in the end. With our approach, we found anti-inflammatory drugs that seem to be very promising. These anti-inflammatory drugs operate via a distinct mechanism from what people have previously investigated.” 

Additionally, through cross-sector partnerships, the industry can get closer to leveraging genomic information and designing personalized treatment strategies.

“We have so many strengths as an organization, whether it’s our wealth of clinical data or our faculty expertise. However, we may not be in a position to commercialize or advance a product that’s already in the market,” said Hightower. 

“Through this partnership, we’re hoping to enhance patient diagnosis, monitor treatments, and individualize care delivery based on patients’ genomic data, molecular data, as well as their environmental and social determinants data. That’s the pathway to precision medicine.”

The advancement of AI and genomic data in clinical care could usher in a new era of medicine tailored to the individual. The healthcare industry just has a few hurdles to clear before this vision can become a reality. 

“AI needs to make actionable data available to the clinician sooner. That’s the next step the technology has to take,” Roach concluded.