Introduction
Drosophila larvae have become an excellent model for concurrently investigating behavior and the underlying neuronal processes. Advanced genetic tools enable the efficient activation or silencing of individual neurons or small neuronal groups. By combining these techniques with standardized stimuli across thousands of individuals, it is possible to establish causal relationships between neurons and behavior. However, extracting meaningful insights from extensive and noisy recordings necessitates the development of new, statistically robust methodologies.
Recent experimental studies focusing on defensive actions or neuromodulation in small neural networks have revealed significant deviations in typical behavioral patterns among larva populations. These deviations either necessitate a redefinition of behavioral features or pose challenges in their detection.
The projects
To address these challenges, we propose developing a Bayesian Program Synthesis (BPS) methodology for generating synthetic data that mimics key characteristics observed in experimental recordings. This approach will involve an inference phase to identify large-scale parameters characteristic of larva populations. Crucially, it will leverage conditional generative models to replicate behavioral traits associated with specific genotypes or to create novel behavioral features. We will use our database consisting of millions of larvae recordings to learn a parameterizable generative model of larva actions with varying durations using denoising diffusion approaches on graphs.
The effectiveness of this method will be validated by applying our behaviour analysis pipeline to the synthetic data and evaluating its potential to enhance transfer learning with annotated behavioural datasets.
The approach will then be extended to incorporate the muscle architecture of the larva within the generative process. We recently extracted the muscles of the Drosophila larva body from a CT-scan recording. Furthermore, we developed a finite element simulation of the drosophila larva body, allowing motion to be simulated directly from a Hill model of muscle activity. We will first use simulation-based inference to map muscle activity onto larva behaviour as recorded within our setups. Then we will map the behaviour generative program onto the muscle programs.
References
- Masson, J.-B. et al. Identifying neural substrates of competitive interactions and sequence transitions during mechanosensory responses in Drosophila. PLOS Genet. 16, e1008589 (2020).
- Jovanic, T. et al. Competitive Disinhibition Mediates Behavioral Choice and Sequences in Drosophila. Cell 167, 858-870.e19 (2016).
- Croteau-Chonka, E. C. et al. High-throughput automated methods for classical and operant conditioning of Drosophila larvae. eLife 11, e70015 (2022).
- Vogelstein, J. T. et al. Discovery of Brainwide Neural-Behavioral Maps via Multiscale Unsupervised Structure Learning. Science 344, 386–392 (2014).
- Winding, M. et al. The connectome of an insect brain. Science 379, eadd9330 (2023).
- Lehman, M. et al. Neural circuits underlying context-dependent competition between defensive actions in Drosophila larva. 2023.12.24.573276 Preprint at https://doi.org/10.1101/2023.12.24.573276 (2023).
- Tredern, E. de et al. Feeding-state dependent modulation of reciprocally interconnected inhibitory neurons biases sensorimotor decisions in Drosophila. 2023.12.26.573306 Preprint at https://doi.org/10.1101/2023.12.26.573306 (2023).
- Lake, B. M., Salakhutdinov, R. & Tenenbaum, J. B. Human-level concept learning through probabilistic program induction. Science 350, 1332–1338 (2015).
The successful intern should have a background in one of the following areas:
- Statistical Physics
- Applied Mathematics
- Statistics & Bayesian Inference
Proficiency in Python is also expected.
Contacts
dbc-epi-recrutement at pasteur dot fr