A team at Emory University and Georgia Institute of Technology, led by Ahmad Abdal Qader and Ellen J. Hess, has now applied an unsupervised machine learning framework called MoSeq to break through this barrier. Their study, published in Dystonia on 15 August 2025, demonstrates that even when mutant mice look indistinguishable from controls at baseline, the anti-dystonia drug trihexyphenidyl (THP) evokes a robust and reproducible behavioral signature in both genotypes—one that also differs between males and females. The work suggests that the temporal structure of continuous free behavior, rather than the execution of any single task, may be the sensitive readout the field has been missing.

The rationale rests on the striatum, a brain region heavily implicated in dystonia and known to organize the probabilistic sequencing of ethologically relevant actions. Because the Tor1a mutation disrupts striatal function—producing aberrant cholinergic tone, reduced dopamine transmission, and defective corticostriatal synaptic plasticity—the behavioral defect may not lie in how well a mouse executes a movement, but in how often it expresses certain movements and in the stochastic order with which it transitions between them. Traditional assays such as the rotarod or hindlimb clasping test are designed to measure action execution, are sensitive to procedural variables, and suffer from observer subjectivity, making them poorly suited to capture such covert deviations.

MoSeq, developed originally by Wiltschko, Datta and colleagues, approaches behavior from an ethological premise: continuous free behavior is composed of organized probabilistic sequences of stereotyped sub-second motifs, or syllables. The pipeline begins with a top-mounted time-of-flight depth sensor recording freely moving mice at 30 Hz in a 30x30x30 cm arena. Custom Python scripts crop the arena, remove artifacts, and convert depth frames to lossless video. An expectation-maximization tracking model approximates the mouse as a 3D Gaussian in each frame, iteratively refining its parameters to extract the animal’s center, orientation, and kinematics such as body-averaged speed and height. An 80×80 pixel box centered on the mouse is then aligned along its spinal axis, flipped by convention so the mouse always faces right, and smoothed in space and time to suppress noise.

These aligned frames are projected onto the top 10 principal components—capturing 85 percent of pose variance—to form a compact time series of 3D posture dynamics. A hierarchical autoregressive hidden Markov model then segments the continuous stream into discrete syllables, each with its own autoregressive description of the stereotyped pose dynamics. In this dataset, MoSeq identified 44 coherent and distinct behavioral motifs, validated visually and through a cross-likelihood metric confirming effective separation between states. The result is a probabilistic map of how each mouse traverses behavioral space: which syllables it uses, how often, and with what likelihood it moves from one to another.

Seventeen mice—10 Tor1a+/ΔE mutants and 7 control littermates, balanced across sex—were tested in a crossover design with subcutaneous saline or 20 mg/kg THP hydrochloride, a dose previously shown to normalize extracellular dopamine in mutant striata. At baseline under saline, the two genotypes were remarkably similar. Total distance traveled over 30 minutes did not differ significantly (57.74 ± 4.10 m for controls versus 49.62 ± 3.71 m for mutants), average speeds were comparable, and syllable usage profiles overlapped. Quantitative comparison of bigram-normalized transition matrices using the Jensen-Shannon divergence, assessed with 10,000-label permutation tests, revealed no significant differences in transition dynamics between genotypes. The mutation’s behavioral footprint, if present, is subtle indeed.

THP changed everything. Both genotypes showed vastly different speed and height distributions after drug treatment, with significant increases in locomotor activity and a pronounced augmentation of fast behaviors, consistent with prior rodent studies of anticholinergics. More strikingly, MoSeq revealed a substantial reorganization of syllable usage profiles: in principal component space, saline-treated and THP-treated mice formed distinct clusters, while the genotypes overlapped within each cluster. Transition graphs confirmed that the drug, not the genotype, was the dominant force reshaping behavioral structure. Both control and mutant mice shared a common behavioral response to THP, indicating that the assay captures the drug’s signature with high sensitivity regardless of the underlying mutation.

A surprise emerged when the researchers examined sex. Females covered significantly more distance than males and exhibited faster speeds after both saline and THP treatments, and the THP-induced speed increase was more pronounced in females, revealed by a significant sex-by-treatment interaction in a three-way repeated measures ANOVA. In PC space, males and females separated cleanly—with the exception of one THP-treated male. This matters because biological sex is a well-established risk factor for most idiopathic focal dystonias, including blepharospasm, cervical dystonia, and spasmodic dysphonia. Notably, the female-predominant inherited dystonias DOPA-responsive dystonia and GNAL-related dystonia both disrupt dopamine neurotransmission, the very system THP modulates. Yet sex is rarely considered in dystonia therapeutic development because behavioral differences between the sexes are often too subtle for traditional assays to detect.

The mechanistic picture remains incomplete. THP is a nonselective muscarinic receptor antagonist binding with low nanomolar affinity to all five muscarinic subtypes, all expressed in the striatum. M1 and M4 receptors mediate corticostriatal plasticity and striatal dopamine release, and blocking either rescues both the dopamine deficit and the impaired plasticity in TOR1A models. However, because muscarinic receptors are expressed peripherally and throughout the brain, the authors caution that the behavioral effects of peripherally administered THP cannot be ascribed to a specific neuronal target. The similarity in temporal sequencing between mutant and control mice is itself remarkable given the 50 percent reduction in striatal dopamine release in mutants, suggesting that developmental compensation—perhaps homeostatic plasticity in spiny projection neurons or network-level cortical adaptation—stabilizes motor sequencing despite the underlying defect.

The implications for drug discovery are considerable. Few drugs effectively treat dystonia, and THP remains the only small molecule proven effective in a double-blind placebo-controlled study in people living with the disorder, yet dose-limiting side effects including dry mouth, constipation, and urinary retention curtail its chronic use. Preclinical testing has traditionally relied on labor-intensive ex vivo physiologic assays requiring specialized expertise. Here, a streamlined 30-minute open-field recording with automated segmentation captured a clear, reproducible drug signature. The authors draw an analogy to the Porsolt swim test, which accelerated antidepressant discovery by providing a simple behavioral predictor of therapeutic effect. Whether the THP-induced behavioral landscape predicts antidystonic efficacy in humans remains unproven given the limited pharmacopeia, but the framework—particularly its demonstrated sensitivity to sex-biased drug effects—offers a scalable path toward identifying the fundamental behavioral signatures of antidystonic therapeutics. Limitations exist, including potential confounds from size differences between male and female mice in depth-based segmentation, though size normalization or keypoint-MoSeq, which relies on tracked point trajectories instead of depth dynamics, could mitigate these concerns as the approach matures.

Subject of Research: Neuroscience

Article Title: Behavioral signature of trihexyphenidyl in the TOR1A (DYT1) knockin mouse model of dystonia

Article References: Abdal Qader, A., Donsante, Y., Markowitz, J. E., Jinnah, H. A., Pandarinath, C., & Hess, E. J. (2025). Behavioral signature of trihexyphenidyl in the TOR1A (DYT1) knockin mouse model of dystonia. Dystonia, 4, Article 15034. https://doi.org/10.3389/dyst.2025.15034

Image Credits: AI Generated

DOI: 10.3389/dyst.2025.15034

Keywords: dystonia, MoSeq, trihexyphenidyl, TOR1A, DYT1, striatum, behavioral segmentation, drug discovery, Behavioral, signature, knockin, mouse