by Tatsuya Tsukahara

Behavioral analysis is a kind of lens that can help us to understand the function of the nervous system and how that function is altered by neuro- or psychoactive drugs. Traditionally, researchers have relied on summary statistics that are thought to reflect psychological processes. For example, to quantify “anxiety” in rodent studies, researchers count the number of times a mouse placed in an open field goes to the center, instead of staying along the sidelines—with a higher number of ‘center entries’ signifying a lower anxiety state.

However, the limitations of these reduced metrics become apparent when considering large-scale behavioral datasets, in which distinguishing subtle differences between patterns of behavior is a challenge.

Can we develop methods that can capture the full repertoire of animal behaviors in a given context? And if we could, how effectively might such methods organize information about the similarities and differences within large-scale datasets?

Our lab has recently developed a data-driven method built on 3D animal postures and unsupervised machine learning called Motion Sequencing (MoSeq). MoSeq automatically identifies a set of reused and stereotyped behavioral motifs, which we called behavioral “syllables” (typically 40-100 including rears, turns, head-bobs), from any given behavioral dataset.

To ask if MoSeq can organize large and complex behavioral data in a meaningful way, we generated a dataset with hundreds of wild-type mice that had been treated with more than 30 distinct drug-dose pairs (a drug-dose pair being the name and amount of a drug, a familiar example for humans being Advil-200mg). MoSeq could accurately predict the drug identity and dose of a given mouse, and in addition could identify relationships among drugs within and between traditionally defined pharmacological classes.

We also employed MoSeq to study the CNTNAP2 mouse model of autism spectrum disorders (ASD)—characterizing the phenotype in these animals by identifying behavioral syllables distinct from those of wild type mice. Even though CNTNAP2 model is known to be hyperactive, MoSeq revealed more detailed phenotype (for example, only a subset of hi-speed syllables was altered). By taking advantage of this behavioral fingerprint, we could use MoSeq to identify new candidate therapeutics and to define their specific on- and off-target effects (in terms of reverted or altered behavioral syllables). As a proof of principle, we first tested a known drug for hyperactivity (Risperidone). Although Risperidone reverted some of CNTNAP2 model’s phenotype, it also induced a lot of off-target effects even in wild-type animals. We further tested two more drugs that have not been tested in CNTNAP2 model but induce similar effects in our wild-type dataset, and identified Sulpiride as a novel candidate drug (similar on-target and less off-target effects compared to Risperidone).

Given that MoSeq summarizes complex behavioral phenotypes induced by pharmacological and genetic perturbations as changes in subsets of behavioral syllables, MoSeq may serve as a useful discovery platform for characterizing both the disease-relevant effects and the side effects of candidate drugs in a wide range of disease models. The identification of novel candidate therapeutics in our current study demonstrates this potential of MoSeq for drug development.

Tatsuya Tsukahara is a research fellow in the lab of Sandeep Robert Datta, in the Department of Neurobiology at Harvard Medical School. Alexander B. Wiltschko, presently at Google Brain, is the first author on this study.

This story will also appear in the HMS Neurobiology newsletter, The Action Potential.

Learn more in the original research article:
Revealing the structure of pharmacobehavioral space through motion sequencing.
Wiltschko AB, Tsukahara T, Zeine A, Anyoha R, Gillis WF, Markowitz JE, Peterson RE, Katon J, Johnson MJ, Datta SR. Nat Neurosci. 2020 Nov;23(11):1433-1443.