What Gets Animals Fired Up? An Interview With Eppendorf & Science Prize Winner Ann Kennedy

Trapped in a car jam? Burnt your toast? Received your 50^th spam email of the morning? You might be feeling a little … angry. You’re not alone. Aggression is everywhere in the animal kingdom. But many species use aggression to fulfill functions rather more important to survival than lobbing a Breville out of the window.

Aggressive behavior is a routine behavior for species as diverse as mammals, birds and even insects. As an example, aggression can help establish hierarchies that are often key to reproduction and the passing on of genes. Rodents, for example, can be shown to settle intra-group squabbles with aggressive fighting.

But these actions are usually not all-or-nothing. Instead, there are many steps separating small infractions between animals and all-out war.

Instead, animals typically engage in actions that ramp in aggression – from display behavior to non-violent contact to full-on fighting. This ensures that combat – which can leave even the victorious animal damaged or dying – is avoided unless necessary.

Northwestern University assistant professor Ann Kennedy’s research focuses on the brain activity underlying this complex behavior. Kennedy recently won the Eppendorf & Science Prize for Neurobiology in recognition of her work. The Prize is awarded for the best 1000-word essay summing up an area of research. Kennedy’s essay, Boiling Over, can be read here.

At the award ceremony for the Prize, held alongside the Society for Neuroscience’s annual meeting in San Diego, Technology Networks spoke with Kennedy to find out more about her research and what gets our brains fired up.

Ruairi J Mackenzie (RJM): What aspects of animal behavior are you interested in?

Ann Kennedy (AK): There's a set of behaviors that are necessary for survival of an animal and of a species – social interactions, feeding, defensive behaviors – and they're so critical that you need to get them right, and you can't just learn by trial and error. So, they're genetically baked into the brain. And as a theorist, I'm interested in how you genetically wire up a brain so that it is going to be afraid of a predator the first time we encounter one. The work that I wrote about for this essay concerned a brain region that's involved in control of territorial aggression, which is another one of these evolutionarily ancient survival behaviors. I went into this wanting to understand how the brain does these things that it has evolved to do and how neural populations are interacting with each other to drive these behaviors. I wanted to understand how they produce them in a way that is adaptive, so that you're not going to forage if there's a predator nearby; that you balance different competing needs and behave in a way that is appropriate.

RJM: You've got a computational background. When you talk about adaptive learning, it makes me think of artificial neural networks. How does that feed into your behavioral research?

AK: My background was in engineering before I went to grad school. There aren’t single neurons that drive these behaviors. It's entire populations of cells, and you can't really make sense of it if you look at things one neuron at a time; you have to look at what all of the cells are doing and describe their dynamics as a group. You have to try to come up with interpretations for that population signal and relate it to the animal's actions. You really need computational methods to do this. You can't just stare at the raw data and get a good intuition for it without having these kinds of approaches.

RJM: Could you tell us about the research that went into the Eppendorf Prize essay?

AK: The core of the essay for Eppendorf was based on a preprint, which will be published early next year. That preprint was about making sense of activity in a brain region that is optogenetically associated with aggressive behavior called the ventrolateral portion of the ventromedial hypothalamus (VMHvl). You don't see individual neurons in this region that are activated specifically during attack—in fact, you see more attack-specific neurons in a mating-associated region tightly coupled to the VMHvl, the medial preoptic area (MPOA), than you do in the VMHvl.

But if you stimulate the VMHvl, you get aggressive behavior. We had this paradox in the lab for years, not understanding what made VMHvl the attack area, given that it didn't have attack-tuned neurons. In the preprint, we found that you don't have attack-tuned neurons, but you have this low-dimensional population code for aggressive motivation that ramps up over tens to hundreds of seconds.

If you look at the population level, how much activity there is along a single dimension is super-tightly correlated with the behavior of the animal. If there's a little bit of elevation, you get investigation; a little more, you get dominance mounting; a little more, you get outright fighting.

It’s like a volume knob for aggression. It tracks the level of aggressive motivation of the animal and if you remove the intruder mouse, it'll stay persistently active for a long time afterward, too.

It's the only dimension of population activity that acts like this. For the rest of them, you'll get activity that ramps up and decays back down quickly. So, we think of it as an approximate line attractor, because it's a single-dimensional scalable value. It’s a way for the VMHvl to encode a graded level of aggressive motivation.

The neighboring part of the VMHvl, the VMHdm, is involved in defensive behavior and has a similar persistent representation of fear-related stimuli. If you take a mouse, present a predator to it and then remove the predator, the cells in this region stay active for a long time.

While they're active, the animal shows persistent defensive behaviors; it will hug the walls of an open space and it'll freeze more. If you silence that persistent activity, the animal goes back to normal. So, the VMH seems to be involved in maintaining this persistent motivational state of the animals. It's not encoding actions, but it's keeping you in a motivated state for a longer time than the stimulus that elicited that state.

RJM: How does this ramping vary among animals?

AK: Across all of the animals that we've looked at – we're up to 14 now – the degree of persistence of this ramp is highly correlated with how aggressive the mouse is. We have some mice that never really fight, where you'll get a little ramping that dies back down very quickly. In others, activity will ramp up and stay persistently active. Those mice will get worked up and stay aggressive for a long time.

RJM: This all seems intuitive – if these neurons immediately switched off as soon as the threat was over, they'd be useless because the threat is more likely to come back if it has appeared once.

AK: Yes – as a mouse, you need to be able to remember that you ran into a hawk ten seconds after it has gone out of sight. So, this persistence fits very nicely with our intuitions of what you want to see in a motivational state. At least among the regions that we've looked at, this feature appears to be characteristic of the VMH. Other regions, like the mating-related MPOA, don’t show that kind of persistence in their neural activity. So, it seems like there are some parts of the hypothalamus that are maintaining these slow motivational signals, while others are encoding actions.

RJM: Looking ahead, are you examining things stepwise by cell population or are you looking at different behaviors?

AK: Within my own group, because we're computational, we're not very equipped to dissect out the cell types that are involved in this, although the Anderson lab is really digging into that now. I’m more interested in whether ramping dynamics exist in other brain areas, associated with other behaviors. I’ve heard from a couple of groups that have been seeing similar things in other populations of cells. So, I'm interested in working with those groups to see whether this is a general feature of computation in subcortical brain regions. And whether other regions show these scalable and persistent properties that we see associated with attack.

RJM: One tool that has propelled your research is your automated pose estimation software, the mouse activation estimation system (MARS). How can automation help us with understanding behaviors in mammals?

AK: There's been a big push in the past five years or so to take tools from computer vision and machine learning and automate the analysis of animal behavior. When I started in the Anderson lab, experimental postdocs would collect videos of mice while recording their neural activity, then either they or a bank of technicians in the lab had to sit there and then score what the mice were doing, frame by frame at 30 hertz. It's pretty miserable work. It’s also pretty subjective—we tried having different people in the lab annotate the same behaviors and found that they disagree quite often: different people have different internal rules for what they consider aggression or what they consider a frame of attack. MARS was our attempt to automate this process as a means of enabling high-throughput analysis of behavior.

So, for example, we used MARS to screen 40 hours’ worth of behavioral data from a set of mouse lines with autism-associated mutations, something that would be miserable to score by hand. With MARS, you can automatically detect how much the mice interact, how much they are fighting, and how they are fighting. You can very quickly obtain statistics about how the animals are interacting, and you can look for differences between lines that you wouldn't have just the bandwidth to look for in other ways.

RJM: How does your program vary from other pose estimation software like DeepLabCut, produced by the Mathis lab?

AK: We started working on MARS before DeepLabCut came out, and our approach to pose estimation was more data-intensive: we crowdsourced annotation of mouse pose to Amazon Mechanical Turk, and collected 15,000 labeled frames, each one labeled by five people. This gave us a good sense of how reliable these body part annotations are and how well our models are doing.  In contrast, the Mathis approach is to take a network that's already been trained on other pose estimation problems in humans, where we have big datasets, and then fine-tune them to animals. So, it works with smaller training sets.

What's also different about MARS is that after pose estimation, it performs a second analysis, namely behavior classification, that isn't present in tools like DeepLabCut. Once you estimate the poses of mice, you extract features, like the angles of the joints, the velocities of the mice and how they are facing each other, and use that as input to supervised classifiers to detect, when they are fighting or when they are investigating. So, MARS allows us to not only track the postures of mice, but it also learns to detect human-defined behaviors of interest.

Ann Kennedy was speaking to Ruairi J Mackenzie, Senior Science Writer for Technology Networks. The interview has been edited for length and clarity.