Workflow for generating and analyzing neural network models.

(A) Basal locomotor patterns (filtered as in Fig 1D) are shown for two of ten Canton-S flies recorded for 30 min each. A stationary interval is highlighted for fly number ten. From these data, walking and stationary interval durations are aggregated and represented using weighted variable-width histograms for walking (top) and stationary (bottom) intervals. (B) A Continuous-Time Recurrent Neural Network (CTRNN) modeling framework used to investigate AS network dynamics. Models included up to five neurons, Gaussian noise inputs representing ongoing activity fluctuations, reciprocal and recurrent connections, odor inputs, and an output threshold to define the locomotor state of the virtual fly. (C) In the first step, a stochastic optimization approach is used to generate models that best reproduce Drosophila basal locomotor statistics. Locomotor statistics generated by a given model for 100 trials (virtual flies) are aggregated and compared to Canton-S strain data. The parameters for this model are then adjusted to reduce the difference from Canton-S data. This process is performed iteratively. (D) In the second step, odor inputs are added to the best previously generated models. For each model the strength of these inputs and the output threshold are adjusted iteratively to reduce the Root-Mean-Square Error (RMSE) between the model’s odor impulse locomotor pattern averaged across 200 trials (virtual flies) and a Drosophila strain’s odor impulse locomotor pattern averaged across ~200 flies. (E) In the third step, models that could match both Canton-S basal walking statistics and DGRP strain odor impulse locomotor patterns are characterized by the number of times specific neural activity levels are observed (‘Trajectory density’) color-coded from very frequent (dark red) to very rare (dark blue), and the tendency of neural activity, in the absence of fluctuations, to move toward (‘Stable’, cyan) or away from (‘Unstable’, orange) equilibrium activity levels (‘Phase portrait’). In each Trajectory density plot and Phase portrait, the bottom-left corner represents the lowest neural activity values.