Fly Larva Behavior

Hi @MWMathis,

The following is from a colleague here at MBL:

I’m a student in the MBL Neurobiology course and we are setting out to use DeepLabCut to analyze fly larvae behavior. We are planning to run DLC with 18 points placed around the larva to characterize behavior in response to stimulus, such as C-bending, S-bending, elongating, and rolling. I wanted to see if you have advice on how to perform this analysis? We were thinking about calculating all pair-wise distances between the points and to cluster them with tSNE + NNMF, then attempt to pull out individual frames to see and describe the behavior. Our main goal is to construct a raster plot of larva behavior over time. Any suggestions or advice would be really helpful!

Thanks,

Chris & Doug

Dear Chris and Doug,

“We were thinking about calculating all pair-wise distances between the points and to cluster them with tSNE + NNMF, then attempt to pull out individual frames to see and describe the behavior.”

That sounds like a great approach - perhaps also consider stacking multiple frames before clustering!

Also check out https://github.com/AlexEMG/DLCutils and the new skeleton analyzer: https://github.com/AlexEMG/DeepLabCut/blob/master/deeplabcut/post_processing/analyze_skeleton.py

Cheers,
Alexander

Hi @douglasmiller,

Our lab is interested in implementing DLC for Drosophila larval behavior. Did y’all have any luck with implementing this? Or any insights into the unique challenges of using DLC for larval behavior?

Thanks so much!
Elizabeth

Feel free to give it a shot — tracking of poses of Drosophila larvae has been done by quite a few labs and it works well. For behavioral clustering you could give this a shot: https://github.com/YttriLab/B-SOID

Thanks, @AlexanderMathis! I am confident that DLC could successfully track poses of Drosophila larvae - the challenge is figuring out how to extract out information on rolling behavior (in my case) from the positional information. B-SOID definitely looks like it could help!

1 Like

I see! Well, if there are any visual features along the roll axis that you could label, then this would greatly simplify the readout…

1 Like