The Genes Have Spoken: UMD Researchers Turn to AI to Decode Stress in Fish

Scientists know that stress can take a toll on all living things, but a doctor-patient heart-to-heart about letting off steam to live a healthier life isn’t going to happen if your subject lives its whole life underwater.

That’s why University of Maryland researcher Mohamed Salem and his team set out to find a more practical way to measure the effect of the things that stress out fish, whether heat, low oxygen, overcrowding or improper diet. It turns out that answer, which impacts much of the world’s ability to feed itself, was arriving in the language of genetics. Now Salem’s lab is working to decode it.

"We are essentially developing a translator for the fish’s biological experience,” said Salem, a professor of animal and avian sciences. “By decoding these genetic signals, we can hear what the fish are telling us about their environment long before they show outward signs of distress, allowing us to intervene and ensure their well-being." 

This is more than just feel-good animal husbandry for Salem. Globally, more than 3.3 billion people get at least 20% of their daily animal protein from fish, and with rising water temperatures threatening commercial fish stocks, and ever-increasing demands on modern fish farming, understanding and more importantly, preventing fish stress is critical for both conservation and the world’s food supply.

Traditionally, scientists test levels of the hormone cortisol in fish blood to determine stress levels. But that approach can yield variable results, so Salem’s team has been looking for something more precise and definitive. 

Every cell in an animal’s body is constantly reacting to the world around it on a genomic level. And those reactions leave behind a trail of evidence. For example, when we go out into the sun, we’re exposed to damaging UV rays, so our skin cells turn on genes involved in DNA repair. They also turn off genes involved in cell division to prevent damage from being copied. So, reading the genes to see what’s been turned on and off can tell someone quite a bit about an animal’s environment.

The challenge for Salem’s team was in the sheer number of genes in a fish and the possible responses they could have to different stressors. When he and the students working in his lab looked at rainbow trout exposed to five different kinds of stressors, things like high heat, low oxygen and bad water, they found that over 21,000 different genes were either switched on or switched off—far too many to evaluate if the objective is a simple test for stress. What’s more, each type of stress created its own unique genetic signature.

The team needed to sift through all that data to identify patterns and pare it down to something more usable, so they turned to machine learning with the help of freshman animal science student Youssef Ali, who was a high school intern in the lab at the time. Youssef had always been interested in computer science, and he recognized that AI tools could help make sense of all the data Salem had been collecting.

Under the guidance of Ph.D. student Guglielmo Raymo, Youssef took the lead in applying AI to the problem, first looking for a handful of genes that could tell them when a fish was experiencing any kind of stress, like a single key to decode all possible stresses. But that effort turned out to be too broad. The AI algorithm identified 39 core genes associated with stress in the test data the team provided, but it was unable to use them to accurately predict general stress in the genomes of fish it hadn’t seen before.

So the team took another approach and guided its machine learning algorithms to look for genetic markers of just one type of stress: heat stress. The algorithm searched the 12,000-plus genes related specifically to that, narrowing it from thousands to hundreds to dozens, until finally it was able to identify just two genes that, taken together, could act as a nearly perfect biomarker, predicting heat stress in rainbow trout with up to 98.6% accuracy. 

“It could be a game changer,” Salem said, “a simple, elegant and incredibly powerful tool that gives us a real-world, practical framework for actually improving stress resilience in fish.”

The discovery, which was published in the journal Nature Scientific Reports in December provides fish farmers and conservationists alike with a tool to know when fish are in danger. So far, the work has focused on rainbow trout, but the genetic markers Salem and his team identified are present across fish species. The researchers expect further studies to show similar success in other commercially important fish.

“If they prove transferable, it can help conservationists monitor wild fish populations, and fish farmers monitor health and environmental stress in aquaculture systems,” Salem said.  “And it can even be used in breeding programs to select for fish that are naturally more resilient to heat.”

The team is also expanding its work to look at additional stressors, and it recently received a grant from the U.S. Department of Agriculture’s National Institute of Food and Agriculture to identify similar markers for stress across a broader range of environmental conditions.