Behavioral cluster assignments per frame
Each frame assigned to a behavioral cluster without predefined labels. Cluster identities are consistent within and across sessions for direct comparison.
CSV
← Capabilities
Unsupervised Discovery
Identify behavioral patterns your scoring rubric doesn't include
Unsupervised clustering reveals behavioral structure you didn't define in advance — find phenotypes, drug effects, and strain differences in unbiased fashion.
UMAP
Embedding method
0
Predefined labels required
p<0.001
Density permutation test
Markov
Transition analysis
The problem
Predefined ethograms restrict what you can find
Supervised behavioral scoring only detects behaviors you define in advance. Subtle phenotypes, unexpected drug effects, and novel behavioral states are invisible when observers score only what their rubric includes. This confirmation bias limits discovery and can cause meaningful behavioral differences to go unreported.
- Observers only score behaviors listed in the ethogram — unlisted patterns go unrecorded
- Subtle behavioral variants (e.g., partial rearing, interrupted grooming sequences) are collapsed into coarse categories
- Cross-condition differences in behavioral microstructure are invisible to standard scoring rubrics
The solution
Unsupervised clustering reveals behavioral structure without categories
ConductVision embeds behavioral features into a low-dimensional space using UMAP and identifies clusters without any predefined labels. Novel behavioral states are flagged automatically. Cross-condition comparisons reveal phenotypic differences that standard ethograms would miss.
- Behavioral features embedded into UMAP space without predefined categories or training labels
- Clusters identified automatically — each represents a distinct behavioral motif or state
- Novel behavior flagging detects patterns that fall outside the distribution of control animals
Endpoints
Unbiased behavioral structure variables
UMAP embeddings and visualization
Low-dimensional embedding coordinates for each frame, color-coded by cluster, condition, or time. Publication-ready UMAP visualizations exported as images.
CSVPNG
Transition probabilities between clusters
Markov transition matrix between behavioral clusters. Reveals sequential structure — which behavioral states follow which — and how transitions differ across conditions.
CSVJSON
Density-difference mapping
Group-level kernel density estimates over the UMAP manifold with permutation testing. Reveals localized regions where one condition occupies behavioral microstates more frequently than another.
CSVPNG
Applications
Discovery-oriented experimental designs
Phenotype-first drug screening
Screen compounds for behavioral effects without specifying which behaviors to look for. Unsupervised clustering detects any behavioral shift relative to vehicle control.
Measures
- Cluster occupancy distribution
- Behavioral state entropy
- Effect size vs vehicle
Novel behavioral biomarker identification
Discover behavioral signatures that discriminate disease model from control without prior hypotheses. Identified clusters can be characterized post-hoc and validated in independent cohorts.
Measures
- Discriminative cluster identity
- Classification accuracy
- Biomarker sensitivity/specificity
Strain comparison without prior hypotheses
Compare behavioral repertoires across strains or genotypes using unbiased embedding. Detect strain-specific behavioral motifs that standard ethograms do not include.
Measures
- Strain-specific cluster occupancy
- Behavioral repertoire overlap
- Novel motif frequency
Drug effect discovery
Characterize the full behavioral profile of a novel compound. Identify which behavioral clusters are expanded, compressed, or newly introduced relative to control.
Measures
- Cluster shift magnitude
- New cluster emergence
- Transition structure alteration
Post-injury behavioral assessment
Detect subtle behavioral shifts following traumatic brain injury without predefined endpoints. Unsupervised embedding reveals injury-associated behavioral microstates that conventional motor scoring misses.
Measures
- Microstate occupancy shift
- Density-difference significance
- Kinematic effect sizes
Compared to typical systems
How ConductVision differs
| Feature | ConductVision | Typical systems |
|---|---|---|
| Discovery mode | Unsupervised — no predefined labels | Supervised — scores only predefined behaviors |
| Observer bias | Eliminated — data-driven clustering | Present — scorer selects what to score |
| Novel behavior detection | Automatic flagging | Not possible with fixed ethogram |
| Ethogram required | No — categories emerge from data | Yes — must define behaviors in advance |
| Cross-condition comparison | UMAP overlay with statistical testing | Limited to predefined measures |
| Density mapping | Kernel density estimation with permutation testing | Not available |
Related capabilities
Behavior Classification
ML-powered frame-level behavioral classification with temporal smoothing and custom classifier training.
Pose Estimation
Multi-keypoint body part tracking — head, paws, tail, body center — with pretrained and custom model support.
Social Interaction
Continuous automated scoring of all pairwise social interactions — approach, contact, following, avoidance — for up to 4 animals.
Discover behaviors your rubric doesn't include
Upload recordings from two conditions and see unsupervised clustering reveal behavioral differences you did not define.
