Search Magazine April 26, 2022

An eye on healthy function and disease

An animation depicting a mouse scurrying across the screen

JAX scientist Vivek Kumar, Ph.D., is using machine learning to automate and improve analysis of mouse behavior and movement. The new methods have already yielded important biological insights, and they may have even more profound implications for research into human diseases such as ALS, Alzheimer’s disease, and even cancer.

What is disease?

It’s not a difficult question on the surface. Everyone can think of a quick answer and provide a good example—cancer, the flu, diabetes, Alzheimer’s, COVID, and many more. There are tens of thousands of diseases that have a huge variety of causes, from inherited genetics to pathogenic infection to too much time in the sun. What they all involve, however, is a disruption in healthy function in one way or another. The disruption may be subtle or severe, fleeting or chronic, but it has a negative effect in one way or another.

To really understand disease, however, we first need to thoroughly understand healthy function. As in, very thoroughly. That way, it’s possible to detect the changes associated with disease, even if they’re very subtle, and figure out what’s causing them. But while we’ve learned a huge amount about our biology in recent years, both in normal function and in disease states, we still have a long way to go. That’s where the research of Jackson Laboratory (JAX) Associate Professor  Vivek Kumar, Ph.D.Understand the genetic and neurobiological basis of complex behaviors that are important in psychiatric conditions such as addiction, ADHD, and depression using genomic, neural circuit, and computational tools.Vivek Kumar , Ph.D., comes in.

Beyond observing

At first, Kumar’s work may not seem to connect directly with such broad objectives for human health. He is a behavioral researcher who primarily uses mice to study the genetics underlying addiction. His findings depend on obtaining accurate data regarding mouse susceptibility to addictive behaviors with substances such as cocaine. Acquiring that data is onerous, however, as it is in all behavioral research.

Behavioral assays—experimental measurements within specific parameters—provide a snapshot of responses to a situation or stimulus. High-resolution video can track mice movements 24/7. But current assays have limited accuracy and reproducibility. Indeed, researchers have found that data disparities can result from whether the person administering an assay is a man or a woman. And all that video data needs to be watched by experts to actually characterize what the mouse is doing, a labor- and time-intensive process.

“We need to be able to accurately and holistically analyze animal behavior,” says Kumar, “and be able to link that behavior with physiology instead of looking at the two as separate. To accomplish that, our tools need to be more sensitive, higher throughput and require less human involvement throughout the process.”

To achieve his goals, Kumar has turned to machine learning (ML). ML and the larger field of  artificial intelligence involves “teaching” a computational tool to perform tasks like humans. Once taught, an ML algorithm, known as a neural network, can process and analyze large amounts of data unsupervised. For example, Kumar had experts categorize millions of frames of mouse footage to identify whether a mouse was performing a particular behavior (grooming) or not, then trained a neural network to detect grooming based on the experts’ input. Once trained, the network could categorize mouse grooming with an accuracy of well over 90 percent, the same level at which the experts agreed with one another. The trained computational network can then analyze thousands of hours of data and tell Kumar whether the animal is carrying out the behavior, something that’s impossible for humans to do. Furthermore, with more input the network can be used to identify behaviors in addition to grooming, providing a general solution for animal behavior detection.

“This kind of ‘behavior extraction’ can provide insight into disease states,” says Kumar. “Grooming is associated with anxiety, stress and stereotyped behaviors seen in psychiatric conditions such as autism spectrum disorder. But to determine which genetic variants and molecular pathways are associated with grooming frequency, we need to observe enough mice and acquire enough data to start making functional associations. Before ML, that wasn’t possible.”

An infographic describing the difference between machine learning and human observation

Grooming, sleep and gait

Kumar’s work with ML has led to a series of papers over the past nine months, beginning in the summer of 2021 with the grooming analysis. That paper went far beyond simply presenting the ML tool, as Kumar took full advantage of the mouse resources at JAX. They characterized the grooming behaviors of 62 different mouse strains and thousands of mice. The large dataset from different strains displayed a continuum of grooming behaviors and frequencies, with recently wild-derived strains grooming far more than their long-domesticated laboratory counterparts. The researchers were also able to parse out the genetic differences underlying the variable grooming behaviors. The genes highlighted by the study are known to regulate nervous system function and development and have been implicated in neurodegenerative diseases. Furthermore, Kumar was able to link the underlying genetics of grooming behavior in the mouse with human psychiatric traits, providing new cross-species insights.

Using other ML approaches, subsequent papers from Kumar’s lab have analyzed mouse movement (gait and pose) and sleep patterns. As with grooming, both are closely tied with human health. Sleep disruption has been associated with a growing list of health consequences, including increased risk for Alzheimer’s disease, and altered movement can be extremely important for insight into multiple diseases. Again, they used artificial intelligence methods to train ML algorithms, then analyzed large datasets to gain insight into healthy and disease states.

In one striking example, Kumar analyzed the movement of a Down syndrome mice that was created by another JAX faculty member, Muriel Davisson, Ph.D., almost 20 years ago. Children with Down syndrome often have motor changes and are miscoordinated. These motor changes often present themselves earlier than cognitive deficits. “Our data shows that Down syndrome mice move very differently than control mice. Our methods are very simple and scalable because we just watch the animal move around naturally, and these approaches could be used for screening for therapeutic drugs,” says Kumar. His team also found changes in mouse models of autism spectrum disorder, Rhett syndrome, and ALS. “Gait and movement issues are also one of the first signs of ASD in very young children, and in neurodegenerative diseases in adults. But in research, how do you characterize how a mouse walks? How can you tell if it’s just how that strain typically moves or if there’s a deficit? With ML we’re able to detect minute differences and quantitate exactly how animals move normally and in disease states.”

The sleep analyses also represent an important step forward for the field. It’s possible to get accurate sleep data, particularly using electroencephalogram/electromyogram (EEG/EMG) outputs to determine sleep states. But EEG/EMG requires invasive physical implantation of electrodes in the mice as well as expert analysis of EEG/EMG readouts, both of which introduce unwanted variables and severely limit the scalability of the research. Kumar collaborated with Allan Pack, MBChB, Ph.D., at the University of Pennsylvania to create a sleep state classification system that is noninvasive and high throughput. It uses only video data to detect small changes in the area and shape of the mice to identify transitions from non-rapid eye movement (NREM) to rapid eye movement (REM) sleep. As a result, it can accurately score wake, NREM, and REM sleep states. While only one strain of mouse was used for the study, it can now be applied to genetically diverse mouse strains to map, for the first time, the genetic architecture of sleep. It can also be used for interventional studies that test the efficacy of sleep therapeutics.

AI gait analysis example

An example of what a computer "sees" when being trained to analyze video.

Surpassing human capability

Much of what Kumar has accomplished with ML to this point has been to expand human expertise, capturing more data and, in turn, insight, with equivalent accuracy. With gait and sleep, however, he has moved the tools beyond real-time human capability, and they are able to detect differences more subtle than can be seen with the naked eye. Moving forward, he is looking to take the technology further still, so that it can make determinations about mouse function and health states that even human experts cannot. His ultimate goal is to use advanced phenotyping methods [the process of measuring biological traits] to understand how altered genetics and neurobiology leads to a wide array of diseases, and to screen for better drugs for human diseases that are modeled in the mouse.

His work has even expanded to study aging with JAX Professor Gary Churchill, Ph.D.Employs a systems approach to investigate the genetics of health and disease and complex disease-related traits in the mouse.Gary Churchill , Ph.D., and the  JAX Center for Aging Research We focus our diverse expertise in biology and genomics on the problems and disorders associated with aging, employing a systems-wide approach to understand aging processes.JAX Aging Center . “We’re working to visualize ‘frailty’ for aging research, to accurately determine the biological—not calendar—age of an animal,” says Kumar. “Even an experienced mouse technician is poor at assessing biological age, so it’s more abstract and more difficult than what we’ve tackled so far. The video methods we’ve used to this point have associated advantages and disadvantages, so we need to merge them so that we can really fine tune the analyses and movement detection sensitivity.”

What’s the payoff? The methods remove human bias and provide a scalable platform for interventional studies of aging. “Imagine being able to use the mouse to screen for genetic or pharmacological factors that enhance healthy aging as oppose to unhealthy aging,” says Kumar. “Or being able to detect signs of aging earlier than current methods for better preclinical animal studies.” The potential benefits are enormous.

Toward human health

Of course, Kumar is working with mice, and the ultimate goal for any kind of disease research is to translate experimental findings to clinical progress. Mice have proven to be invaluable for learning about fundamental mammalian biology, but bridging the gap to human medicine, with our highly variable genetics, environments and behaviors, is far more difficult. In order to improve the mouse-human interface and, in turn, improve medical safety and efficacy, it’s essential to obtain better preclinical data. That’s the whole point of Kumar’s research, and he returns to ASD as an example.

“In autism, the earliest symptoms are motor symptoms, and the cognitive changes develop later,” he says. “Cognitive function related to ASD is very hard to measure in mice, so what if we use the motor symptoms for preclinical research instead? If you’re screening compounds for early ASD intervention, looking at mouse movement with the ML tool could be a great surrogate.”

Kumar originally conceived of the tools in the context of behavior and addiction, but they can apply to diseases of all kinds beyond those already mentioned, including complex diseases that need better treatments. Identifying and quantifying the small changes associated with diabetes, cancer, Alzheimer’s disease and more will narrow the focus ever more precisely on the genetics and molecular processes involved. A solid foundation of preclinical data obtained in concert with human patient data will provide a much better starting point for identifying effective therapeutic targets as well as weeding out those that are unlikely to work. The ability to identify changes in the mice before overt symptoms emerge also holds promise for improving diagnostics and opening avenues for early intervention.

The broader concept is still relatively new, however. “NIDA (National Institute on Drug Abuse) is familiar with using ML and computer vision to map behavior and brain circuitry, but I’m looking for a broader impact,” says Kumar. He laughs. “Some ‘preaching’ will need to happen. But automated ML-based mouse phenotyping is the future for the field. In 10 years every cage will have video capture.”

It will be exciting to see what the next 10 years brings.