DeepLabCut meets the brainstem: how deep learning for behavior yields insights into the neural basis of motion
How researchers are using DeepLabCut to study the neural basis of movement.
Brainstem neurons that command mammalian locomotor asymmetries
Researchers have uncovered neurons in the brainstem that can control turning motion in mammals when activated. This is part of a greater body of work led by Prof. Ole Kiehn, who has recently been awarded the Brain Prize 2022, alongside Prof. Silvia Arber and Prof. Martyn Goulding, for work that focuses on mapping of the neural networks in the brain and spinal cord that activate and control our movements.
A recent paper from the Kiehn Lab was published in Nature Neuroscience titled “Brainstem neurons that command mammalian locomotor asymmetries”. Cregg et. al. leverage DeepLabCut to measure movement after chemogenetic modulation of Chx10+ neurons in the brainstem. They found that excitation of Chx10 Gi neurons causes ipsilateral movements, whereas inhibition causes contralateral movement.
The details:
They also compared the natural and evoked gait properties, and found during spontaneous turns, for both the forelimbs and the hindlimbs the stride length was diminished on the side of the turn. For both turns in wild type (WT) mice and turns caused by chemogenetic modulation of Chx10 neurons, the hindlimbs and the forelimbs on the side of the turn traveled a shorter distance than the leg opposite to the turn.
How they used DeepLabCut
The authors used DeepLabCut 2.x toolbox (see our Nature Protocols paper on 2.x DLC and how to use it!). Specifically, here are their video, network and hardware stats:
“Q & A” with first author, Dr. Jared Cregg:
Alexander and Aristotelis caught up with Dr. Cregg to discuss his experience with DeepLabCut. He started using the tool in 2019, nearly exclusively through the GUI interface. He recalls being at a meeting in Japan where there was a “buzz around the paper,” and he recalls thinking “this might actually be useful!” We sure glad it was! Below are some Q&A that we discussed.
“DeepLabCut become essential to the work that I do.”- Dr. Jared Cregg
What are some key design decisions when you are setting up an experiment? Does the use of DeepLabCut affect these decisions and how?
- From the experiment’s perspective I’m always thinking “this needs to be able to run on DLC”. So I make sure that my new videos are similar enough to videos that I have already trained networks with. This way when the experiment is done, I just need to run inference on the new videos, and I have my keypoints ready for further analysis.
What was the greatest challenge you faced when using DeepLabCut?
- The only issue we face is installing the GPU drivers. One thing we noticed, and we appreciate a lot, is that every time there’s a windows or linux update with OS issues, or changes in python versions etc, the next release of DLC takes care of issues that arise. For us it seems like everything just keeps working magically, but we know that there must be quite some effort going on from your side to keep the software running smoothly.
What advice would you give to researchers who want to perform kinematic analysis on data?
- There’s one thing that many people have a misconception. They think that DLC will give you graphs for your paper. DLC gives you the (X,Y) coordinates [and confidence readout] of the points you are interested in. Analyzing these data to extract behavior, is up to the user to set up the way it fits their experiment. For the paper, we wrote the whole kinematic analysis ourselves. This way we can easily design it for specific things that we think are relative to the specific experiment we are working with.
If you could have one wish for DLC, what would it be?
- I’m actually not the best person to be asking this, since I use DLC for mostly straightforward scenarios. I’m a happy user.
Thanks Jared! 💜
- Blog post by Aristotelis, Mackenzie & Alexander
