Generalizable Machine Learning in Neuroscience Using Graph Neural Networks.

Area of Science:

• Neuroscience
• Machine Learning
• Computational Biology

Background:

• Deep learning applications in neuroscience are expanding, but focus on microscopic neural systems is novel.
• Calcium imaging in C. elegans provides a unique opportunity to study neuron-level dynamics and emergent behaviors.

Purpose of the Study:

• To evaluate deep learning models for predicting microscopic neural dynamics and classifying behaviors in C. elegans.
• To compare structure-agnostic neural networks with graph neural networks (GNNs) for exploiting neural graph structures.

Main Methods:

• Utilized calcium imaging data from C. elegans.
• Developed and applied deep learning models, including structure-agnostic networks and a novel GNN that infers neuron relations from activity.
• Compared model performance on neuron-level dynamics prediction and behavioral state classification.

Main Results:

• Neural networks demonstrated high performance in predicting neural dynamics and classifying behaviors.
• The developed GNN, leveraging inferred graph structure, generally outperformed structure-agnostic models.
• GNNs showed enhanced generalization capabilities on unseen organisms.

Conclusions:

• Deep learning models are effective for analyzing microscopic neural dynamics and behaviors.
• Graph structure, when explicitly inferred and utilized, provides a beneficial inductive bias for machine learning in neuroscience.
• GNNs offer a promising approach for developing generalizable AI in neuroscience research.