Comparison of multiple video tracking-based behavioral summary approaches for compound discrimination
M. A. RitteriD, Serena DeianaiD, Alina Ritter, Carsten Tobias WotjakiD, Michael BrechtiD, Amarender R. BogadhiiD
Abstract
The rapidly increasing number of video tracking-based behavioral summary tools and methods raises the question as to the most suitable approaches for pharmacological fingerprinting in pre-clinical research. We have recently shown that social context has a strong effect on behavioral syntax in mice, further suggesting that treatment effects can be context dependent. Here, we aim to answer the question whether there is an optimal combination of context and behavioral summary method for effect detection and discrimination of different psychoactive substances in a controlled environment. To this end, we applied eight different treatment-dosage pairs (amphetamine 1.5,3,6 mg/kg; modafinil 5,10,50 mg/kg; seltorexant 3,10 mg/kg) and evaluated five different approaches to behavioral summary: parametric aggregation, unsupervised segmentation in Keypoint-MoSeq (KPMS) & Variational Animal Motion Encoding (VAME), and supervised segmentation in Simple Behavioral Analysis (SimBA) & A-SOiD, across two different contexts (Solitary & Social) in 314 recordings of freely moving mice in an open-field arena. Surprisingly, our results show no significant differences in performance across models and context. Across treatment effect detection to treatment-dose discrimination, all models showed performance significantly above chance that was insensitive to various data limitations and extensions. Overall, our study shows that under the tested conditions and treatments, the choice of a behavioral summary model does not meaningfully affect the description of treatment effects. Simple aggregate measures from tracking data and machine learning based behavioral summary approaches that are expensive, in terms of training data and computational resources, performed equally well. Our findings taken together with literature suggest that the fuller decomposition of complex behavior through unsupervised machine learning might be necessary for the description of large-scale datasets but does not necessarily align with the goals present in smaller-scale treatment discrimination tasks common in pre-clinical research.
02 /Specification & Classification
Click any tag to browse matching papersIn-depth summary
Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (28 Jul 2026) · CC BY 4.0
Scalar keypoint aggregates matched Keypoint-MoSeq, VAME, SimBA and A-SOiD at discriminating eight drug-dose treatments across 314 mouse open-field recordings; social context did not improve recall.
Many pose-tracking-based behavioral summary tools now exist, but they had not been benchmarked against each other across behavioral contexts and levels of drug detection and discrimination, nor tested for whether a social partner widens the range of detectable drug effects. This work compares parametric, unsupervised and supervised summaries under one classification scheme, in solitary and social contexts, at three levels of treatment discrimination.
Key results
ShownAll five summary approaches detected drug effects against pooled vehicle above chance in both contexts. Neither model choice nor solitary versus social context made a significant difference.
Two-sided binomial tests against chance 1/2 were significant for every model (Bonferroni, p≤0.05). Wilcoxon signed-rank tests of solitary vs social were all non-significant. A two-way type 2 ANOVA over model and context found no main effects or interactions. · 3.1 Treatment effect detection is similar across…; 2.8.4 Binomial test and confidence interval; 2.8.7 Two-way ANOVA
ShownEvery model assigned mice to the correct compound, ignoring dose, above chance, although recall was slightly lower than for effect detection. Model and context again had no significant effect.
Chance level was 1/4 (correct compound vs 2 other compounds + vehicle). Binomial tests were significant for all models (Bonferroni, p≤0.05). Wilcoxon signed-rank tests for context were non-significant, and a second two-way ANOVA showed no main effects or interactions. · 3.2 Treatment discrimination performance is…; 2.8.4 Binomial test and confidence interval
ShownClassifying the full treatment-dose pair was much harder for every model, except among the amphetamine dose groups, but all models stayed above chance. Parametric, supervised and unsupervised summaries were statistically indistinguishable.
Chance level was 1/9 (correct pair vs 7 other treatments + vehicle). Binomial tests showed above-chance recall and Wilcoxon signed-rank tests showed no context effect (Bonferroni, p≤0.05). ANOVA found no significant main effects or interactions. · 3.3 Treatment-dosage discrimination performance…; 2.8.4 Binomial test and confidence interval
ShownRecall held up under several feature changes. Removing F-test fingerprint features or enclosure-contact features, limiting unsupervised outputs to the frames the supervised labels covered, or adding motif transition probabilities did not significantly change performance.
Wilcoxon signed-rank tests against original-data recall were non-significant (Bonferroni; 60 tests in the enclosure-removal analysis). Dropping mouse-enclosure features in social recordings pushed some models, such as parametric aggregation, below significance. Logistic Regression and Support Vector classifiers gave matching results. · 3.4 Contribution of contextual features to…; 4 Discussion and conclusion; 2.8.8 Bonferroni correction
ShownEvery model could tell solitary from social context in vehicle-treated mice, yet context did not raise treatment recall. Single features were often enough on their own to classify above chance.
Classifying context in the vehicle group exceeded chance for every model (binomial, Bonferroni, p<=0.05). Of 536 model/context/motif combinations, 29 that never occurred in vehicles were dropped. Re-testing the rest across the three tasks gave 307 significant single-feature results (~20% of all possible tests). · 3 Results; 3.4 Contribution of contextual features to…; 4 Discussion and conclusion
Methods, data and reproducibility
SLEAP keypoints from 20-minute, two-phase open-field recordings (solitary first, then social, with an enclosed conspecific and an object) were summarized by scalar aggregation, Keypoint-MoSeq, VAME, SimBA or A-SOiD. Features were split by enclosure contact, normalized to pooled vehicle, and classified by leave-one-out nearest-centroid on Manhattan distance.
- Mice given amphetamine, modafinil, seltorexant or route-matched vehicle, then filmed in solitary and social open-field phases · mouse (C57BL/6J, male, ~5 weeks old) · n = 157 mice; 314 recordings; groups of 16 to 32 mice · 10.5281/zenodo.21460909
- Untreated, novel stimulus mice from a separate cohort, placed in one enclosure during the social phase · mouse (same age, sex and source as treated mice) · n = same number as treated mice
- Vehicle-treated recordings without dropped frames, used to train every supervised and unsupervised model · mouse · n = 49 videos (~8.2 hours) of a possible 64
- BORIS-annotated training videos with 6 behaviors (climbing, grooming, rearing, three sniffing types), used for SimBA and A-SOiD · mouse · n = 36 videos (30 training, 6 validation)
- Methods & tools
- SLEAP (12-keypoint pose tracking), Keypoint-MoSeq 0.5.2, VAME 0.10.0, SimBA 3.9.3 (random forest classifiers), A-SOiD 0.3.1.3 (active learning), BORIS 9.0.6, Stoelting Any-Maze 7.4, OpenCV2 4.11.0, scikit-learn 1.5.1 (NearestCentroid, LeaveOneOut), SciPy 1.14.0, statsmodels 0.14.2, Logistic Regression and Support Vector classifiers
- Data & accessions
- https://doi.org/10.5281/zenodo.21460909
- Software & environment
- Python 3.11.9; Keypoint-MoSeq 0.5.2; VAME 0.10.0; SimBA 3.9.3; A-SOiD 0.3.1.3; BORIS 9.0.6; Any-Maze 7.4; OpenCV2 4.11.0; scikit-learn 1.5.1; SciPy 1.14.0; statsmodels 0.14.2
- Reproducibility gaps
- Settings for SimBA, A-SOiD and VAME, and the keypoint renaming, are left to Supplementary/Extended Methods that are not in the text. SLEAP model details are deferred to Ritter et al. (2025).
Limitations (6)
- The authors say they used relatively few mice (groups of 16 to 32) and only three compounds. They add that these treatments may already saturate their effects in the solitary condition. · stated by the authors
- The authors mention the analog camera's frame rate (25.23Hz) and the smaller training set (49 vehicle videos, after dropped frames in 46 of 320 videos). They argue these limits affect all models equally. · stated by the authors
- The conspecific and object were confined in enclosures, not free. The authors suggest a more complex, semi-naturalistic setting may be needed to reveal interactions between social behavior and drug effects. · stated by the authors
- The authors note that unsupervised segmentation yields many motif parameters relative to the number of mice. They countered this with a nearest-centroid classifier and leave-one-out cross-validation that do not depend on parameter count, and report similar results with other classifiers. · stated by the authors
- The claim that models perform alike rests on non-significant, Bonferroni-corrected tests in modest groups. No equivalence testing or power analysis is reported, so small but real differences in performance cannot be ruled out. · noted in this summary
- Vehicle controls differed by route (SC, IP, PO) but were pooled into one reference group for normalization and classification. This could blur route-specific baseline effects in each drug's comparison. · noted in this summary
Who should read:Read this if you are choosing between scalar tracking metrics and machine-learning behavior segmentation (Keypoint-MoSeq, VAME, SimBA, A-SOiD) for preclinical drug discrimination in mice.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- Rapid, open-source, and automated quantification of the head twitch response in C57BL/6J mice using DeepLabCut and Simple Behavioral Analysis
Alexander D. Maitland, Nicholas R. Gonzalez, … Grant C. Glatfelter·ACS Pharmacology & Translational SciencePeer-reviewed·Jul 31, 2025
- A novel baseline-effect shift tracking model for more sensitive detection of differences in the effects of closely related dopamine transporter inhibitor/ sigma receptor antagonist drug combinations on psychostimulant use.
Reshma Paul, Roshni Gandhi, Martin O. Job·bioRxiv·Aug 27, 2024
- An optimized procedure for robust volitional drug intake in mice
Alberto J. López, Amy R. Johnson, … Erin S. Calipari·Experimental and Clinical PsychopharmacologyPeer-reviewed·Jul 13, 2020
- Pharmacobehavioral space of MoSeq syllables significantly overlaps with scalar locomotion features
M. A. Ritter, Amarender R. Bogadhi·bioRxiv·Aug 20, 2026
- Effects of dopamine receptor antagonism and amphetamine-induced psychomotor sensitization on sign- and goal-tracking after extended training
Shaun Yon‐Seng Khoo, Alexandra Uhrig, … Nadia Chaudhri·Behavioural Brain ResearchPeer-reviewed·Mar 18, 2021
- Deep learning based behavioral analysis enables high precision rodent tracking and is capable of outperforming commercial solutions
O. Sturman, L. M. von Ziegler, … J. Bohacek·bioRxiv·Jan 21, 2020