Neural network features distinguish chemosensory stimuli in Caenorhabditis elegans

Nervous systems extract and process information from their environment to alter animal behavior and physiology. Despite progress in understanding how different stimuli are represented by changes in neuronal activity, less is known about how they affect broader neural network properties. We developed a framework to use graph-theoretic features of neural network activity and predict ecologically-relevant stimulus properties – namely, stimulus identity and valence. Specifically, we used the transparent nematode, Caenorhabditis elegans, with its small nervous system, to define neural network features associated with various chemosensory stimuli. We trapped animals using a microfluidic device and exposed their noses to chemical stimuli known to be attractive or repellent, while monitoring changes in neural activity in more than 40 neurons in their heads. We found that repellents trigger higher average neural activity across the network, and that the tastant salt increases neural variability. In contrast, graph-theoretic features, which capture patterns of interactions between neurons, are better suited to decode stimulus identity than measures of neural activity. Furthermore, we show that a simple machine learning classifier trained using graph-theoretic features alone or in combination with neural activity features can accurately predict stimulus identity. These results indicate that graph theory reveals network characteristics that are distinct from neural activity, confirming its utility in extracting stimulus properties from neural population data. Significance Statement Changes in the external environment (stimuli) alter patterns of neural activity in animal nervous systems. A central challenge in computational neuroscience is to identify how stimulus properties alter interactions between neurons. We recorded neural activity data from C. elegans head neurons while the animal experienced various chemosensory stimuli. We then used a combination of activity statistics (i.e., average, standard deviation, and several frequency-based measures) and graph-theoretic features of network structure (e.g., modularity – the extent to which a network can be divided into independent clusters) to define neural properties that can accurately predict stimulus identity. Our method is general and can be used across species.