Modeling neuromodulation at the single-neuron scale

Understanding the relationship between brain structure and behavior is a fundamental open challenge. Notably, connectivity information alone is insufficient to completely interpret neural processes: depending on internal states, two identically-wired neural circuits can yield different outcomes. For example, in the Drosophila melanogaster larva, hunger information affects the decision process between cowering and exploring through neuromodulation of a few neurons. However, the reasons why only certain neurons are modulated and not others and how circuit structure affects regulation remain poorly understood.

We here propose to study neuromodulation through simple machine-learning problems. We construct small artificial neural networks in which a few neurons have two possible states, altering their dynamics. The networks are trained to give opposite outputs depending on their internal state. We apply our model to synthetic feed-forward and recurrent networks, as well as real biological circuits found in the Drosophila larva brain. We find that the layout of modulated neurons, including their number and position, as well as the allocation of a restrictive budget for changes between states, impact performance and connectivity strength.

Our model could improve our understanding of connectivity rules under neuromodulation. Leveraging the growing number of whole-brain connectomes, our approach could help predict modulated neurons in specific circuits, for subsequent experimental testing.