Merci Claire!!

Thanks, Oded! Great question! For laminar flow, you can estimate the fish's relative position and adjust the trajectory accordingly. For strong turbulent flow, you might mask the trajectory and run the pipeline using only the tail tracking data. Let me know if I can help with that!

A huge thank you to all my amazing co-authors: Pedro Silva, Alexandre Laborde, Miguel Mata, João Marques, Elena Collins, Randall Peterson, Christian Machens, and Michael Orger (@daniobrain.bsky.social), who made this possible! 🙌
9/n

I'd love for fellow zebrafish researchers to try out Megabouts. We've put up extensive documentation and tutorials at megabouts.ai 🛠️

Feedback is very welcome!
8/n

Our method also improves sensitivity in detecting drug-induced phenotypes! 🔬
7/n

Megabouts builds on the repertoire of zebrafish behavior uncovered by João C. Marques. This clustering of swim movements is fantastic for quantifying how stimuli translate into actions.
6/n

What's really cool? Our model maintains high accuracy even at lower frame rates and without tail tracking! 📊
5/n

Megabouts uses a Transformer neural network at its core, with each input token representing a pose measurement. This helps us overcome issues like changes in frame rate, and missing data, offering robust classification of movements across various setups.
4/n

There are fantastic tools for tracking zebrafish behavior: DeepLabCut (@deeplabcut.bsky.social), SLEAP (@talmo.bsky.social), BonZeb, ZebraZoom (@clairewyart.bsky.social), and Stytra.
But we lack methods to process these tracking data into standardized ethograms.
3/n

Quantifying behavior is crucial in neuroscience, genetics, and pharmacology. With our growing ability to track movements comes the need to standardize how we describe them—despite the diversity in recording configurations.
2/n