The Drosophila larva is extensively used as model species in experiments where behavior is recorded via tracking equipment and evaluated via population-level metrics. Although larva locomotion neuromechanics have been studied in detail, no comprehensive model has been proposed for realistic simulations of foraging experiments directly comparable to tracked recordings. Here we present a virtual larva for simulating autonomous behavior, fitting empirical observations of spatial and temporal kinematics. We propose a trilayer behavior-based control architecture for larva foraging, allowing to accommodate increasingly complex behaviors. At the basic level, forward crawling and lateral bending are generated via coupled, interfering oscillatory processes under the control of an intermittency module, alternating between crawling bouts and pauses. Next, navigation in olfactory environments is achieved via active sensing and topdown modulation of bending dynamics by concentration changes. Finally, adaptation at the highest level entails associative learning. We could accurately reproduce behavioral experiments on autonomous free exploration, chemotaxis, and odor preference testing. Interindividual variability is preserved across virtual larva populations allowing for single animal and population studies. Our model is ideally suited to interface with neural circuit models of sensation, memory formation and retrieval, and spatial navigation.

The behavior of many living organisms is not continuous. Rather, activity emerges in bouts that are separated by epochs of rest, a phenomenon known as intermittent behavior. Although intermittency is ubiquitous across phyla, empirical studies are scarce and the underlying neural mechanisms remain unknown. Here we reproduce empirical evidence of intermittency during Drosophila larva free exploration. Our findings are in line with previously reported power-law distributed rest-bout durations while we report log-normal distributed activity-bout durations. We show that a stochastic network model can transition between power-law and non-power-law distributed states and we suggest a plausible neural mechanism for the alternating rest and activity in the larva. Finally, we discuss possible implementations in behavioral simulations extending spatial Levy-walk or coupled-oscillator models with temporal intermittency.

No abstract

Predicting positive or negative reinforcement from environmental clues is essential to guide decision making and goal-directed behavior. In insect brains the mushroom body is a central structure for learning such valuable associations between sensory signals and reinforcement. We propose a biologically realistic spiking network model of the Drosophila larval olfactory pathway for the association of odors and reinforcement to bias behavior towards either approach or avoidance. We demonstrate that prediction error coding through integration of present and expected reinforcement in dopaminergic neurons can serve as a driving force in learning that can, combined with synaptic homeostasis, account for the experimentally observed features of acquisition and extinction of associations that depend on the intensity of odor and reward, as well as temporal features of the odor/reward pairing. To allow for a direct comparison of our simulation results with behavioral data we model learning-induced plasticity over the full time course of behavioral experiments and simulate locomotion of individual larvae towards or away from odor sources in a virtual environment.