AI aids analysis of lifetime environmental exposures

The idea that biology is not destiny is hardly new. Studies in twins have shown that even among identical pairs—those sharing 100% of their DNA—the same disease genes do not turn into full-blown illness in both individuals.

Most human diseases are fueled by both genetic and environmental factors. Genes account for 30 to 70% of the risk for developing conditions such as type 1 diabetes, hypertension, stroke, or Parkinson’s disease. The rest likely stems from external, non-biologic factors.

This poorly understood jumble of environmental variables, unique to each one of us, forms our individual exposomes.

Untangling these external influences and how they intersect with our genes could provide critical clues to the puzzle of complex diseases arising from the interplay between genes and environment. Such knowledge can optimize the way the way scientists forecast a patient’s disease risk and individualize treatments.

But the scope of our environmental exposures is so vast and the interactions so intricate, that to understand the human exposome scientists will increasingly need to turn to artificial intelligence for help.

“Some of our causal models of disease are failing, and there is a growing interest to understand how diseases arise so that we can develop therapies and interventions that factor environmental and behavioral sources of variation,” said exposome researcher Chirag Patel, associate professor of biomedical informatics in the Blavatnik Institute at Harvard Medical School. Patel said.

“We need to evaluate AI approaches for reclassifying disease or stitching together exposure histories to piece together the causal picture.”

Case in point—Alzheimer’s disease, a highly heritable condition that tends to emerge later in life.

“This pattern implies that despite the significant role of genetics, many decades of environmental exposures also play an important role in the disease onset and progression,” said Hyun-Sik Yang, HMS assistant professor of neurology at Brigham and Women’s Hospital.

But the exact nature of the interaction between genes and environment in Alzheimer’s disease remains unclear. “It’s a complex problem, given that environmental exposure occurs many decades before the disease onset,” Yang said.

Understanding the exposome better can go a long way toward filling these knowledge gaps in Alzheimer’s risk and disease progression. Knowing a patient’s genetic and exposome profile can also inform tailored treatments. Currently, there is only one FDA-approved disease-modifying therapy, an antibody that blocks the accumulation of amyloid plaque to slow down disease progression by 25 to 30%, Yang said. This underscores the importance of understanding both additional biologic pathways that fuel the disease and modifiable risk factors, including the exposome.

Evolution in understanding of nature and nurture

The question of how our environment shapes us has tantalized the curiosity of scientists and philosophers since antiquity. Nearly 2,500 years ago, the Greek physician Hippocrates attempted to explain how environmental conditions affect health in his treatise “On Airs, Waters, and Places.” Aristotle and Plato famously disagreed on the role of nature versus nurture in human development.

But it was not until the mid-20th century that modern exposure research emerged in earnest with landmark studies from England linking smoking to lung cancer and levels of daily physical activity to cardiovascular risk.

Since then, science has made serious progress toward defining some of the mechanisms by which various exposures alter physiology and affect health. Studies have shown that these exposures can modify our DNA, a major driver of cancer, change the way our genes are activated or deactivated, and affect our biology in other ways.

Some attempts to capture the exposome have focused on measuring molecular changes in the blood caused by various exposures—classic epigenetics research. But exposome science should go beyond that, Patel added, because some of our most important exposures may not be readily captured in molecules floating in the blood.

So, the full story of the interplay between nature and nurture in shaping disease and health remains maddeningly elusive. Now a group of academic, government, and industry researchers are hoping to change that.

Time for the Human Exposome Project

The group, of which Patel is a member, recently convened a brainstorm to define the scope of exposomic science and thus pave the way for a study of the Human Exposome Project. Conceived in the vein of the Human Genome Project, which in 2003 generated a blueprint of human genes, the Human Exposome Project aims to discover the role of various environmental exposures in human disease and health.

In December of 2023, the group held a national workshop to clarify what exposome science is and how it can be integrated into other biomedical research.

The need for a field definition is more than an exercise in semantics, said Patel. A clear, uniform definition charts the course of study, maps the areas of interest, enumerates the disciplines involved in the research, and spells out the ultimate goals for the field.

Workshop attendees converged on three complementary definitions of “exposomics,” geared toward a different set of experts.

• The study of the integrated compilation of all physical, chemical, biological, and psychosocial influences that impact biology.
• The study of the comprehensive and cumulative effects of physical, chemical, biological, and psychosocial mediators that impact biological systems by integrating data from a variety of interdisciplinary methodologies and streams to enable discovery-based analyses of environmental influences on health.
• A transdisciplinary field aimed at discovery-based understanding of how the exposome influences biology and health.

As a biomedical informatics expert, Patel favors the second definition.

To untangle the dark matter of the exposome, scientists look to AI

The exposome encompasses various external influences an individual encounters from the day they’re born until the day they die—some fleeting, some persistent. These include pollution, microplastics, medications, viral and bacterial infections, food, exercise, climate, and physical and psychological stressors.

In a further twist to an already complicated plot, some research even hints at transgenerational effects of exposures—for example, the role of our mothers’ and grandmothers’ diets in modulating our health.

Untangling it all is akin to finding thousands of needles lost in a huge haystack of disease, Patel said in a 2019 TEDx talk. All these needles need to be identified and analyzed, not one variable at time, but in combination, for their synergistic effects.

“Studying exposures in isolation is reductive and can lead to false associations because you might not be getting the totality of the picture,” Patel said.

Such a multilayered analysis is beyond the capacity of the human brain. This is where AI can help.

In fact, AI might be the only way to understand the exposome, Patel said, because AI can take in and simultaneously look at disparate types of data—genomic profiles and food intake or place of residence and biologic tissue samples, such as blood or urine or skin cells, for example.

“I totally believe that using these new multimodal approaches in a way that stitches all the pieces together might be the only way to analyze a person’s complex history and go beyond the usual do you smoke, do you drink,” Patel said.

AI tools could one day help physicians individualize disease risk prediction based on a patient’s unique exposome and genome, Patel said. The results could lead to tailored interventions, such as more frequent monitoring or earlier treatment.

Some of this work is already taking place at HMS.

For example, a 2019 study led by Patel teased out how the environment affects the activity of various genes. The research was based on analysis of medical records of 56, 000 U.S. fraternal and identical twin pairs who grew up in different zip codes.

In another project, still in its nascent stages, Patel is hoping to use genomics and exposomics data to design a poly-exposure risk score that captures the myriad non-genetic influences that give rise to different forms of diabetes across the world among various populations. The team also includes Broad Institute genomics specialist Josep Mercader, and AI and health disparities expert Raj Manrai, assistant professor of biomedical informatics at HMS.

In another planned study, Patel and collaborators—including Deborah Blacker, HMS professor of psychiatry at Mass General—will analyze the constellation of exposures in older people to define who is at risk for developing Alzheimer’s disease and other forms of dementia. To do so, they will integrate metabolomics, genomics, and epigenetics data along with information from brain and whole-body imaging.

“We will shove all this stuff into a model and extract predictors for dementia,” Patel said. “I’m really excited about that, because for the first time, we’ll have what has been missing in the field, which is measuring everything all at once within the same individuals.”

What will the future look like?

Some scientists see a not-too-distant future in which one way of capturing a patient’s exposome would be done with a simple blood test. Andrea Baccarelli, dean of the Harvard T. H. Chan School of Public Health, forecasts that this might happen in as soon as 10 years.

Such a test would hinge on detecting a process called methylation, which occurs when molecules called methyl groups get tacked onto specific segments of DNA. These molecules can awaken certain genes and drive others into dormancy without changing the underlying genetic code itself.

Various exposures, including chemicals, food, chronic stress, and viruses, can drive methylation changes that have profound effects on our cells and organs. In that sense, methylation changes could serve as the footprints of past exposures.

With AI by their side, physicians will be able to individualize risk prediction based on a patient’s individual exposome—their unique set of exposures over their lifetime. Based on a person’s exposome and genome, AI could forecast long-term risks and recommend more aggressive monitoring and earlier treatment where necessary.

And, Patel says, AI could help in another way—by extracting patient exposure clues already lurking in existing data that are currently overlooked. For example, if a patient shows up in the emergency room with a set of symptoms, a physician might feed into an AI model the patient’s medical, social, and family history; their chief complaint; blood work results; and various images. The physician would then receive a report on the patient’s likely exposures based on elements that human clinicians might easily miss.

To be sure, AI models are not there yet. Researchers need to eliminate bugs, build and refine algorithms, and conduct rigorous testing and validation to ensure that models are extracting and predicting accurately. But being able to do all that, Patel said, is only a matter of time.

Patel also cautions that by creating new knowledge, untangling the exposome might pose new challenges and require new ways of thinking. Frontline clinicians and policymakers would need to decide how to apply newly generated insights to disease prevention, for example. Today, disease prevention and regulatory science remain reductive, focusing on regulating one variable at a time—air pollution, water quality, smoking.

“Once you have a full exposome predictor, how do you regulate for multiple exposures?” Patel said. “This is the next challenge for us as a community: If we believe that these things are causal, how do we start to think about new paradigms of either regulation or intervention that consider all this stuff simultaneously rather than one at a time?”