Artificial intelligence tools help scientists decode genomic disorders and communicate genomic risks
Three researchers explain how they use artificial intelligence for their genomic studies.
As researchers continue to unravel the many mysteries of genomics, they require more and more sophisticated technologies to diagnose, monitor and treat genetic conditions. Artificial intelligence tools, which mimic human intelligence to solve problems, are well-suited to tackle these complex tasks.
Machine learning, a type of artificial intelligence, has the unique ability to learn and improve itself. These clever methods have already been used to predict how a certain type of cancer will progress, find disease-causing genomic variants and identify genetic disorders by examining people’s faces.
Researchers at the National Human Genome Research Institute (NHGRI) are increasingly using artificial intelligence tools to answer compelling questions in genomics, such as predicting rare genetic disorders and their severity, and to understand how genomic information influences decision-making.
Diagnosis is intentionally only skin deep
Machine learning and other artificial intelligence tools are already improving the detection of relatively common conditions, such as breast cancer through mammography.
Benjamin Solomon, M.D., NHGRI clinical director and senior clinician in the NHGRI Medical Genetics Branch, wants to know if we can find a way to use these tools at surface level -- to diagnose genetic conditions that affect the skin.
Genetic disorders are often rare and notoriously difficult to diagnose. On average, it takes between five and 10 years from the onset of symptoms to pinpoint the exact genetic cause of a rare condition. The long and arduous diagnostic journey often delays treatment, and it typically ends up being costly and isolating.
But that timescale is shrinking. To shorten this process even more, Dr. Solomon and his colleagues are exploring how artificial intelligence tools can help clinicians identify genetic conditions that involve characteristic marks or patterns on the skin, similar to birthmarks.
If we can create a consistent system to recognize genetic skin conditions across the board, we can help clinicians diagnose disorders right at the bedside. The sooner we can make an accurate diagnosis, the faster we can treat the patient.
In a recent study, they used computing systems to recognize and classify rare genetic skin disorders based on photographs that illustrated different abnormal skin characteristics. The underlying algorithm, which is a set of programmed instructions for solving tasks, could classify genetic skin conditions more accurately than pediatricians or medical geneticists.
These computing systems, called neural networks, mimic how the brain processes information. Researchers must train the neural networks to associate specific skin characteristics with the corresponding genetic conditions, much in the same way a person would show photographs of apples and oranges to teach children the difference between these fruits.
So far, the study has taught the algorithm to recognize six genetic skin conditions, and Dr. Solomon hopes to expand this tool to diagnose more rare genetic disorders.
“If we can create a consistent system to recognize genetic skin conditions across the board, we can help clinicians diagnose disorders right at the bedside,” Solomon says. “The sooner we can make an accurate diagnosis, the faster we can treat the patient.”
Marking the severity of rare disorders
Oleg Shchelochkov, M.D., NHGRI director of residency and fellowship programs, is also harnessing the power of artificial intelligence to help diagnose rare genetic disorders more accurately.
Specifically, Dr. Shchelochkov is interested in a rare metabolic disorder called propionic acidemia, which affects one in 20,000 to 500,000 people worldwide. Patients with propionic acidemia have higher levels of a chemical called propionic acid in their bodies, which can cause organ damage and frequent hospitalizations. In some cases, a liver transplant is necessary.
For decades, researchers and clinicians have discussed the possibility of two types of propionic acidemia — mild and severe — which could have an impact on the type of treatment that a patient receives. But because of the limited number of people with this condition, researchers have found it difficult to predict which patients might benefit from the different treatment approaches.
Recently, Dr. Shchelochkov published a study with Charles Venditti, M.D., Ph.D., chief of the NHGRI Metabolic Medicine Branch, that used machine learning to find biological markers, also called biomarkers, associated with mild and severe forms of the condition.
The researchers collected nearly 500 types of genetic, laboratory and imaging data. After working with propionic acidemia disease experts to create a system to classify patients into mild and severe categories, the researchers trained the algorithm to determine which pieces of the data are uniquely associated with the two forms of the disease. After training, the researchers gave the algorithm new patient information. The algorithm was very successful at establishing which data types were associated with the mild versus severe form of propionic acidemia.
If we can use machine learning to make these kinds of useful predictions about rare diseases, even with such little data, it would be a boon for more common conditions like cancer, hypertension and diabetes.
The results of this study support a decades-long intuition held by experienced clinicians that there are distinct versions of propionic acidemia. With early insights into the severity of a given case, clinicians can better design the treatment plan for that patient.
“It would have been very difficult for humans to distill so much data into what really matters for the severity of the disorder,” says Dr. Shchelochkov. “This is the kind of predictive power we want to continue harnessing for future efforts.”
With information about which biomarkers are most closely associated with the severity of propionic acidemia, clinicians can focus on identifying severe patients more rapidly and provide them with the help they need as early as possible.
“If we can use machine learning to make these kinds of useful predictions about rare diseases, even with such little data, it would be a boon for more common conditions like cancer, hypertension and diabetes,” says Dr. Venditti.
Understanding how to communicate genomic information through a simulated food buffet
Instead of using artificial intelligence to understand and recognize rare genetic disorders, Susan Persky, Ph.D., associate investigator in the NHGRI Social and Behavioral Research Branch, is using such tools to understand how different ways of communicating genomic information can influence behavior.
Dr. Persky heads the Immersive Simulation Program at NHGRI, in which her research group uses virtual reality to simulate realistic interactions to assess how delivering individual genomic information can affect someone’s decisions and behavior.
Because both genomics and environmental factors, such as food choices, influence health, these researchers are interested in studying how parents choose food for their children after learning about their child’s genomic risk for health conditions.
Using immersive virtual reality technology and machine learning analysis, we want to develop ways for researchers and clinicians to improve how we communicate genomic risk to individuals.
To examine this dynamic, the researchers created a simulated food buffet where parents can select lunch for their child, just like they would in real life.
"We want to know what ends up on the plate, but that's just step one," says Dr. Persky. "Thanks to this virtual reality approach, we can collect a lot of information about the parents’ behavior, like where the parents are walking, what they are looking at, the order in which their choices are made and how long they took at each step."
Dr. Persky believes that the parsing out of such complex data will be a key step for using machine learning tools. Her research group studies physical movements in the virtual reality food buffet and how they might relate to participant’s emotional and cognitive state. Through machine learning algorithms, they hope to create models that predict behavioral and emotional responses.
"Using immersive virtual reality technology and machine learning analysis, we want to develop ways for researchers and clinicians to improve how we communicate genomic risk to individuals," says Dr. Persky.
On the horizon of genomics and artificial intelligence
Already, these artificial intelligence tools have yielded many applications in science and medicine that in some cases surpassed human expectations and abilities. Yet, the intersection of genomics and artificial intelligence is a complex area that also involves considerations of many ethical, legal and social implications of the work. With the recent NIH investment in the Bridge to Artificial Intelligence (Bridge2AI) program, researchers must ensure that these tools do not contribute to perpetuating health inequities or ethical problems during the collection and analysis of genomic data.
Last updated: December 14, 2022