Automatic classification of behavior in zebrafish larvae

Adrien Jouary, Germán SumbrebioRxiv 2016 · DOI

TL;DR

This paper introduces a novel, supervised soft-clustering method to automatically classify zebrafish larval tail movements into five distinct categories. The approach achieves 82% classification accuracy and, crucially, does not rely on trajectory analysis, making it applicable to both free-swimming and head-fixed experimental setups, including those used for functional imaging.

What Problem Does This Paper Address?

Understanding the fine-grained motor control of developing organisms is crucial for neuroscience. Zebrafish larvae, a popular model organism, exhibit complex swimming behaviors driven by tail undulations, which can be broken down into discrete "tail bouts." Previous methods for classifying these tail bouts often relied on analyzing the larva's trajectory in 3D space. However, this approach is incompatible with head-fixed experimental conditions, which are essential for correlating neural activity with behavior using techniques like functional imaging. This creates a gap in our ability to study the neural basis of larval locomotion in a controlled, high-throughput manner.

Methods

The authors developed a supervised soft-clustering algorithm to classify tail bouts. The core of their method involves:

  • Feature Extraction: They computed the curvature along the larva's tail and a "deflection index" over time for each tail bout.
  • Similarity Measurement: Dynamic Time Warping (DTW) was employed to measure the similarity between tail movement time series. DTW is robust to temporal distortions, making it suitable for comparing biological signals that might have slight variations in timing. They found DTW to be more effective than feature-based similarity measures for this task.
  • Supervised Classification: A Fuzzy K-Nearest Neighbor (FKNN) algorithm was used. This "soft-clustering" approach allows a single tail bout to belong to multiple categories with varying degrees of membership (interpreted as probabilities). This is important because the authors observed that tail movements form a continuum rather than discrete, isolated groups.
  • Visualization: t-Distributed Stochastic Neighbor Embedding (t-SNE) was used to visualize the high-dimensional tail movement data in a 2D space, illustrating the continuous nature of the kinematics.

Datasets

The study utilized data from 25 head-restrained zebrafish larvae recorded for 4 hours. This dataset comprised approximately 16,000 individual tail bouts. The data was collected under different conditions: initial visual stimulations (whole-field motion, dark flashes) and subsequent homogeneous illumination. A subset of these tail bouts was manually labeled to serve as the training and testing set for the supervised classifier.

Key Insights and Ideas

  • Continuum of Movement: Zebrafish larval tail movements, even in head-restrained conditions, form a continuum rather than distinct, easily separable categories. This challenges the notion of rigidly defined motor primitives and suggests a more fluid, modulated control system.
  • DTW's Efficacy: Dynamic Time Warping proved to be a powerful tool for capturing the subtle similarities between tail movements, outperforming traditional feature-based similarity measures. This highlights its utility for analyzing time-series biological data with temporal variability.
  • Soft-Clustering for Ambiguity: The use of fuzzy clustering (FKNN) is crucial for handling the inherent ambiguity and overlap between movement categories, reflecting the continuous nature of the data.
  • Trajectory-Independent Classification: The method's independence from trajectory analysis is a significant advancement, opening doors for studying neural control of locomotion in head-fixed preparations, which are essential for combining behavior with neurophysiological recordings.

Key Results

  • The developed supervised classifier achieved an 82% accuracy in categorizing tail bouts into five classes: Scoot, Asymmetrical Scoot, Routine Turn, C Bend, and Burst.
  • Misclassifications primarily occurred between similar categories (e.g., Scoot and Asymmetrical Scoot), indicating that the classifier captured meaningful distinctions while acknowledging the continuum.
  • The t-SNE visualization confirmed the continuous nature of tail kinematics and showed that the learned categories, while distinct, were embedded within this continuum.
  • The distribution of movement categories varied significantly depending on the sensory stimulus (e.g., dark flashes strongly induced C bends), demonstrating the method's ability to capture stimulus-evoked behavioral changes.

Limitations and Future Work

  • Manual Labeling Effort: The supervised approach requires a significant effort in manually labeling tail bouts for training.
  • Ambiguity in Classification: While the soft-clustering approach handles ambiguity, the paper notes that errors in classification were associated with more uniform membership levels, suggesting that some movements are genuinely difficult to assign to a single category.
  • Future Directions: The authors suggest that this method can be used to produce a fine description of the relationship between neuronal activity and locomotion, particularly in functional imaging experiments. Further exploration of the continuum and the underlying neural control mechanisms is implied.

Summary updated

Abstract

ABSTRACT Zebrafish larvae navigate the environment by discrete episode of propulsion called bouts. We introduce a novel method for automatically classifying tail bouts. A supervised soft-clustering algorithm to categorize tail bouts into 5 categories of movements: Scoot, Asymmetrical Scoot, Routine Turn, C Bend and Burst. Tail bouts were correctly classified with 82% chance while errors in the classification occurred mostly between similar categories. Although previous studies have performed categorization of behavior in free-swimming conditions, our method does not rely on the analysis of the larva’s trajectory and is thus compatible with both free-swimming and functional imaging in head-fixed condition.
BibTeX
@misc{jouary2016automatic,
  author = {Adrien Jouary and Germán Sumbre},
  title = {{Automatic classification of behavior in zebrafish larvae}},
  year = {2016},
  doi = {10.1101/052324},
  howpublished = {bioRxiv},
}
Figure