Analysis of behavioral flow resolves latent phenotypes

This article has been Reviewed by the following groups

Read the full article See related articles

Listed in

Log in to save this article


The nuanced detection of rodent behavior in preclinical biomedical research is essential for understanding disease conditions, genetic phenotypes, and internal states. Recent advances in machine vision and artificial intelligence have popularized data-driven methods that segment complex animal behavior into clusters of behavioral motifs. However, despite the rapid progress, several challenges remain: Statistical power typically decreases due to multiple testing correction, poor transferability of clustering approaches across experiments limits practical applications, and individual differences in behavior are not considered. Here, we introduce “behavioral flow analysis” (BFA), which creates a single metric for all observed transitions between behavioral motifs. Then, we establish a “classifier-in-the-middle” approach to stabilize clusters and enable transferability of our analyses across datasets. Finally, we combine these approaches with dimensionality reduction techniques, enabling “behavioral flow fingerprinting” (BFF) for individual animal assessment. We validate our approaches across large behavioral datasets with a total of 443 open field recordings that we make publicly available, comparing various stress protocols with pharmacologic and brain-circuit interventions. Our analysis pipeline is compatible with a range of established clustering approaches, it increases statistical power compared to conventional techniques, and has strong reproducibility across experiments within and across laboratories. The efficient individual phenotyping allows us to classify stress-responsiveness and predict future behavior. This approach aligns with animal welfare regulations by reducing animal numbers, and enhancing information extracted from experimental animals

Article activity feed

  1. To determine the best number of clusters, we chose an approach used previously for segmenting behavior by VAME21 and MoSeq17. To this end, we first partitioned the recorded behavior into 100 clusters, and then chose the number of clusters which represented 95% of the imaging frames. This mark indicated about 70 clusters (Suppl. Figure 1A), so we subsequently re-ran the clustering approach using 70 clusters. Although nominal p-values revealed that CSI and control animals behaved differently on many of these clusters, only 4 out of 70 clusters survived the appropriate multiple testing correction (Benjamini-Yekutieli, Figure 1D). Visual inspection of these significant clusters reveals - in agreement with the classical analysis - that they capture the time the animals spent in the center of the open field. Specifically, these four clusters represent movement of the mouse from the periphery into the center (cluster 41, Suppl. Video 1), movement in or through the center (cluster 23, Suppl. Video 2), orienting and turning in the center (cluster 69, Suppl. Video 3) and movement from the center back to the periphery (cluster 45, Suppl. Video 4).

    I genuinely appreciate your work on this new algorithm for behavioral analysis! I really liked how you compared some classical behavior analysis versus the unsupervised clustering approach. Also, I appreciate how you tested the BFA with different clustering approaches. I'm curious to find out whether it was possible to assign distinct behavior to each of the 70 clusters in Figure 1.