Johannes Bohacek🚨In our latest preprint🚨 we propose a radical change to the analysis of rodent #behavior in preclinical research.
Featuring
@deeplabcut
♥️
Check out the manuscript and the 🧵👇
First, we address the elephant in the room: even the fanciest data-driven analyses of behavior run up against punishing multiple testing correction, making it hard to detect treatment effects. We solve this using Behavior Flow Analysis (BFA).
BFA identifies behavioral clusters in a given behavior test, and maps out all the dynamic transitions between them.
From this high-dimensional data, we compute a single metric to detect phenotypes. This massively increases statistical power.
More power = fewer animals.
Applied to open field data, BFA repeatedly identifies treatment effects when the most advanced clustering approaches fail.
BFA is a simple add-on analysis (all code open source) that works with any clustering algorithm!
We show it with K-means (our first choice), B-SOiD & VAME. But try it with your favorite clustering method and tell us how it worked for you!
Next, we train a machine learning classifier to faithfully reproduce our set of clusters, so we can start comparing clustering results across experiments. That was a very important improvement for us, to stabilize all clusters so we know e.g. cluster1=onset of a supported rear etc.
We apply this across 6 different experiments using #stress exposures, pharmacology, and noradrenaline circuit manipulation.
Across these datasets (>200 mice tested), we find that BFA increases statistical power and reveals phenotypes that would otherwise be missed.
The snapshot below shows an example where we were underpowered (n=8) to detect treatment effects (DREADD activation of the #LocusCoeruleus), but BFA unveils the latent phenotype! 🔵
Next: To profile individual mice, we introduce Behavioral Flow Fingerprinting (BFF). BFF projects the entire transition matrix of every animal onto a single datapoint in 2D.
Check out how well we can separate mice that received different doses of yohimbine, a noradrenaline drug.
With stabilized clusters and a single, high-dimensional datapoint for each mouse, we now compare all🐭across all 6 experiments.
Very cool to see: Chronic stressors look very different than acute stressors. Acute stress aligns with noradrenaline release, as we would expect! 🥳
This is my favorite plot, it shows how lots of experiments can be compared in one graph 🤩
What's maybe even more useful: BFF enables predictions about future behavior! 🔥
We introduce a novel approach to segregate responders from non-responders. Rather than using one metric, we use the entire, high-dimensional behavioral flow data per🐭. It predicts who can extinguish fear after stress!
Finally, we apply the BFA & BFF pipelines to a dataset from our collaborators at Roche Pharma.
.
They use an entirely different open field setup (size, bedding, lighting, camera, frame rate).
Again, we increase power and resolve subtle effects, finding yohimbine effects at low doses.
.
They use an entirely different open field setup (size, bedding, lighting, camera, frame rate).
Again, we increase power and resolve subtle effects, finding yohimbine effects at low doses.
Although applicable across behavior tests, we focus on the widely-used open field test.
We share a unique resource: 443 open field videos containing 7 experimental designs, all fully analyzed with BFA and BFF and with the corresponding @deeplabcut tracking data!
All code is available in our github repository: https://github.com/ETHZ-INS/BehaviorFlow
The major ideas behind this work came from superstar Lukas von Ziegler, with lots of creative work from Fabienne Rössler and @osturmscience. Huge team effort as always 💪
Funded by the Swiss 3R Competence Center, helping to reduce/refine the use of animals in research 🙏💙
