Statistical signature of subtle behavioural changes in large-scale behavioural assays

Alexandre Blanc, François Laurent, Alex Barbier–Chebbah, Ben Cocanougher, Benjamin M.w. Jones, Peter Hague, Marta Zlatic, Rayan Chikhi, Christian L. Vestergaard, Tihana Jovanic, Jean‐Baptiste Masson, Chloé BarrébioRxiv 2024 · DOI

TL;DR

This paper introduces a suite of novel statistical methods to detect subtle behavioral changes in large-scale Drosophila larval assays, overcoming challenges posed by noisy data and complex dynamics. By combining physics-informed shape regularization, unsupervised learning in a latent space, and a generative model for action sequences, the authors can identify nuanced behavioral phenotypes and link them to specific genetic manipulations.

What Problem Does This Paper Address?

Understanding how neural circuits generate behavior is a major challenge in neuroscience. Drosophila larvae are an excellent model system for this, offering complex behaviors and powerful genetic tools to manipulate neural activity. However, analyzing the massive datasets generated by high-throughput screens (millions of larvae) to find subtle behavioral changes—especially delayed responses or minor deviations—is difficult. Existing methods often struggle with noisy data, the inherent deformability of larvae, and the imbalanced nature of behavioral categories. This paper aims to bridge this gap by providing statistically robust tools to extract meaningful behavioral signatures from such large-scale, complex datasets.

Methods

The authors developed four key methodological components:

  1. Physics-informed regularization of larva shape: To address issues with low-resolution images and larval deformability, they use a physics-informed Bayesian inference model to regularize larval contours, ensuring accurate shape representation.
  2. Continuous self-supervised representation of behaviour: They employ a self-supervised learning approach using a compact neural network (an autoencoder) to create a continuous, low-dimensional "latent space" that captures larval behavior. This space is trained to predict future and past larva states, encoding temporal continuity without relying on a predefined behavioral dictionary.
  3. Kernel-based statistical testing: Within this learned latent space, they use the Maximum Mean Discrepancy (MMD) test to detect subtle differences in behavioral distributions between genetic lines. This non-parametric test is effective at identifying nuanced deviations.
  4. Probabilistic generative model of action sequences: A time-varying Markov chain model is used to generate synthetic behavioral sequences. This allows for a more structured comparison of action sequences and the identification of lines with altered temporal dynamics or specific sequence patterns, going beyond simple action frequencies.
  5. Suffix Tree Representation: For a broader analysis, they construct a suffix tree of all behavioral action sequences to cluster genetic lines based on shared patterns, providing a hierarchical view of behavioral phenotypes.

Datasets

The study utilized a large dataset comprising behavioral recordings from 280,716 Drosophila larvae across 568 genetic lines. This data was collected in response to an air puff stimulus. The authors also mention using data from previous studies ([3], [10], [34], [54]) for comparison and validation. No new benchmarks were explicitly introduced, but the scale and nature of the dataset allowed for the development and testing of their novel analytical techniques.

Key Insights and Ideas

  • Subtle behaviors are informative: The core idea is that even minor, hard-to-detect behavioral shifts can reveal significant information about neural circuit function.
  • Physics meets machine learning: Integrating physical principles (larval shape dynamics) into ML models improves robustness and accuracy, especially with noisy biological data.
  • Latent space as a behavioral fingerprint: A continuous, low-dimensional latent space derived from self-supervised learning can effectively represent complex, multi-faceted larval behaviors, allowing for sensitive statistical comparisons.
  • Generative models for sequence analysis: Modeling action sequences probabilistically provides a powerful way to identify deviations in temporal dynamics that might be missed by analyzing individual actions or simple frequencies.
  • Uncovering "hit" neurons: The methods are capable of identifying genetic lines (and by extension, the neurons they target) that induce subtle but significant behavioral changes, expanding the scope of discovery beyond previously known "strong" effects.

Key Results

The authors successfully applied their suite of methods to analyze a large-scale behavioral screen of Drosophila larvae responding to an air puff. They demonstrated the ability of their techniques to:

  • Regularize larval shapes: The physics-informed model improved the accuracy of shape inference.
  • Detect subtle behavioral deviations: The kernel-based statistical testing in the latent space identified genetic lines with nuanced behavioral changes that were not apparent with simpler methods.
  • Characterize action sequences: The generative model allowed for the identification of lines with altered temporal dynamics and specific action sequence patterns.
  • Cluster genetic lines: Suffix tree analysis and hierarchical clustering grouped lines based on shared behavioral sequences, revealing distinct phenotypic clusters.
  • Identify novel "hit" lines: Their methods identified new genetic lines exhibiting subtle behavioral modifications, complementing and extending previous findings. For instance, they identified lines showing distinct responses to varying stimulus intensities, suggesting roles in encoding stimulus intensity or regulating behavioral responses.

Limitations and Future Work

  • Curse of dimensionality: While the latent space is low-dimensional, the authors acknowledge the general challenge of high dimensionality in statistical testing. They suggest that a lower-dimensional latent space would enhance statistical significance.
  • Temporal resolution: For long-time window encoding, the approach might lose resolution in temporal sequences of action.
  • Action sequence frequency: The generative model, in its current form, does not fully account for the frequencies of sequences longer than three actions.
  • Future Work: The authors suggest extending their approach to larger timescales without retraining by encoding more significant epochs. They also highlight the potential for integrating their methods with connectome data and neural recordings to accelerate the discovery of neural circuits responsible for decision-making and complex behaviors. The development of a software tool for interactive navigation of the latent space and cluster hierarchy is also presented as a step towards facilitating further exploration.

Summary updated

Abstract

Abstract The central nervous system can generate various behaviours, including motor responses, which we can observe through video recordings. Recent advancements in genetics, automated behavioural acquisition at scale, and machine learning enable us to link behaviours to their underlying neural mechanisms causally. Moreover, in some animals, such as the Drosophila larva, this mapping is possible at unprecedented scales of millions of animals and single neurons, allowing us to identify the neural circuits generating particular behaviours. These high-throughput screening efforts are invaluable, linking the activation or suppression of specific neurons to behavioural patterns in millions of animals. This provides a rich dataset to explore how diverse nervous system responses can be to the same stimuli. However, challenges remain in identifying subtle behaviours from these large datasets, including immediate and delayed responses to neural activation or suppression, and understanding these behaviours on a large scale. We introduce several statistically robust methods for analyzing behavioural data in response to these challenges: 1) A generative physical model that regularizes the inference of larval shapes across the entire dataset. 2) An unsupervised kernel-based method for statistical testing in learned behavioural spaces aimed at detecting subtle deviations in behaviour. 3) A generative model for larval behavioural sequences, providing a benchmark for identifying complex behavioural changes. 4) A comprehensive analysis technique using suffix trees to categorize genetic lines into clusters based on common action sequences. We showcase these methodologies through a behavioural screen focused on responses to an air puff, analyzing data from 280,716 larvae across 568 genetic lines. Author Summary There is a significant gap in understanding between the architecture of neural circuits and the mechanisms of action selection and behaviour generation. Drosophila larvae have emerged as an ideal platform for simultaneously probing behaviour and the underlying neuronal computation [1]. Modern genetic tools allow efficient activation or silencing of individual and small groups of neurons. Combining these techniques with standardized stimuli over thousands of individuals makes it possible to relate neurons to behaviour causally. However, extracting these relationships from massive and noisy recordings requires the development of new statistically robust approaches. We introduce a suite of statistical methods that utilize individual behavioural data and the overarching structure of the behavioural screen to deduce subtle behavioural changes from raw data. Given our study’s extensive number of larvae, addressing and preempting potential challenges in body shape recognition is critical for enhancing behaviour detection. To this end, we have adopted a physics-informed inference model. Our first group of techniques enables robust statistical analysis within a learned continuous behaviour latent space, facilitating the detection of subtle behavioural shifts relative to reference genetic lines. A second array of methods probes for subtle variations in action sequences by comparing them to a bespoke generative model. Together, these strategies have enabled us to construct representations of behavioural patterns specific to a lineage and identify a roster of ”hit” neurons with the potential to influence behaviour subtly.
BibTeX
@misc{blanc2024statistical,
  author = {Alexandre Blanc and François Laurent and Alex Barbier–Chebbah and Ben Cocanougher and Benjamin M.W. Jones and Peter Hague and Marta Zlatic and Rayan Chikhi and Christian L. Vestergaard and Tihana Jovanic and Jean‐Baptiste Masson and Chloé Barré},
  title = {{Statistical signature of subtle behavioural changes in large-scale behavioural assays}},
  year = {2024},
  doi = {10.1101/2024.05.03.591825},
  howpublished = {bioRxiv},
}
Figure