Fig 1.
Abbreviation of brain structures and neurons, and schematic of the CX model.
Indication of the different neuronal pathways used in the model, each pathway is represented by only 1 neuron example showing input and output(s) termination. The left part of the diagram shows the EB compass pathway, the right part shows the steering control pathway. Known pathways are indicated in plain lines while hypothetical pathways are shown with dotted lines.
Fig 2.
A. Map of the 1296 (48x27) elementary units composing the visual system. The visual span is fully panoramic (360°) on the azimuth and 81° upward on the elevation. The red framed visual unit is the one used as example in B. B. Elementary visual unit process. Each unit acts as a orientation-free edge detector. The comparison of the summed intensity of the left vs the right of the unit and of the top vs the bottom gives an index from 0 (no difference ↔ no edge) to 1 (corner edge) for each unit.
Fig 3.
A. Connectivity diagram between the main neurons of the EB-PB model, EPGs, PEGs and PENs. EPGs have inhibitory connections with each other as indicated, via Delta7 neurons (black lines; one example is shown from the EPG4 in the central part of the diagram). EPGs (light blue) form a recurrent circuit with PEGs (green) while forming connection with neighbouring EPGs through PENs (pink) on each side. B. Connection diagram from the visual circuit to the EPGs. Each angular segment connects in a retinotopic manner to the EPGs. Intrinsically this results in an activity bump forming in the circuit that corresponds to the direction of highest visual contrast. C. Proprioceptive inputs to the PENs. When the agent is engaged in a left turn, PEN1−8 are stimulated, while during a right turn PEN9−16 are stimulated. This will trigger a counter-motion of the bump (posterior view) so that it still indicates the relative direction of the external cue(s). D. Example of the function of the EB model in an arena surrounded by a single cue (here a cylinder represented by the black circle). The blue line shows the path of the agent without any influence of the CX model, i.e. only produced by the steering noise. E. Activity rate of the three neuron types constituting the EB model. On the first line, the activity of EPGs (blue for no activity to yellow for active) shows a perfect following of the cue orientation in the agent’s visual field (red line, scale on the right). Even during a darkness episode (from 1500th to 2500th steps), the model keeps track of the cue position with a relatively low error thanks to the PENs (second line) and PEGs (third line) combined actions.
Fig 4.
The EPG to PFL3 connectome at the interface between the compass and the steering control.
The full connectome is assessed from online Drosophila brain databases (https://v2.virtualflybrain.org, [35–38]). The full list of the synaptic connection between neurons annoted PFL3 and EPGs is given in S1 Table. A. (a) Division of the EB in wedges 15. (b) Quantity of synapses expressed as a function of the EB wedges and the PFL3 columns. B. (a) Relationship between EB wedges and the EPGs modeled here (Fig 3A). (b) The connectome is transformed into EPG-PFL3 synapse weights to create the steering model (see section 2.5). Weights are obtained by summing for each EPG the synapses on the right or on the left part of the PB. The gains are then normalized to fit between 0 and 1.
Fig 5.
EPG to PFL3 connectivity scheme.
A. Diagram of the EPGs to PFL3s connectivity that drives the steering. Each EPG makes inhibitory connection to two PFL3 neurons, one on each side. The strength of the synapse depends on the index of the EPGi as shown in B. The summed activity rates of the right and left PFL3 neurons are compared to obtain the CX steering signal. Finally, gaussian noise (ν = 0;σ = 10) is added to obtain the final steering command. B. (a) Synapse strength as a function of the index of EPGi, obtained from the connectomic approach (Fig 4B; [30]). Synapses are inhibitory and therefore multiplied by -1. Red lines represent the connection with the subset of PFL3s from the right part of the PB and green from the left part. Synapse weights are presented in a circular representation to show the correspondence with the EPG input geometry. Index 1 corresponds to the synapse between ellipsoid body tile EPG1, connected to the rear right FOV (-157.5°), and PFL31, and so on, through to index 8, the synapse between EPG8, connected to the rear left (157.5°), and PFL38. (b) Heatmap in the arena from 100 simulations in total. The data are normalized to the landmark direction (indicated on the right). (c) Final direction/position of each simulation (n = 100) with the cue placed in different positions around the arena, expressed in the cue reference frames. C. (a) Simulations (n = 100 for each shift condition) where EPG-PFL3 synapse weights were shifted for 1 cell to the right/clockwise (compared to B.). Results show the heatmap relative to the cue (top) of the 100 simulations. (b) Simulations (n = 100 for each shift condition) where EPG-PFL3 synapse weights were shifted for 1 cell to the left/anti-clockwise. Results show the heatmap relative to the cue (top) of the 100 simulations. (c)-(d) Equivalent to (a) and (b), respectively, with a 2 cells shift.
Fig 6.
Steering CX model with external control.
The central part of the diagram is similar to the model of EPG to PFL3 connection presented in Fig 5A except the synapse weights are initially set homogeneously (). On each side the model copies the [21] model architecture such that the connection between FBns [red (right FBns) and green (left FBns) cells] and steering neurons (PFL3s, yellow cells) are shifted by one column, in opposite directions on each side. FBns also receive rotatory self-motion information inhibiting inputs [light red (right turns) and green (left turns) lines] for the ipsilateral side, therefore the right FBns are activated during left turns and viceversa. The reward signal (pink) is provided by the sensory pathway indicating goodness-badness of the current sensory input. The modulation of the EPG-PLF3 synapses is under the control of this reward signal and the FBns as shown in the bottom-right inset. The steering command is obtained by comparing the right and left summed activity of the PFL3s after the addition of a gaussian noise (ν = 0;σ = 10) as previously (Fig 5A).
Fig 7.
Innate behaviour under the control of a visual reward signal.
Simulations of the FB steering model (Fig 6) using a reward signal provided by the visual processes to modify the EPG-PFL3 synapse weights. We created the visual reward signal to the CX using different masks. Results for each panel include (a) the final path directions (n = 50 simulations), (b) the probability density function of the final direction vector relative to the cue orientation, (c) the averaged EPG-PFL3s synapse weights and (d) examples of 5 simulation paths. See also Fig B in S1 Text for results with continuous masks. A. Visual reward to the FBs is equal to the sum of the visual units signal through a discrete mask equal to 0 outside the 30° frontal area or 1 inside. B. Visual reward to the FBs is equal to the sum of the visual units signal through a discrete mask equal to 0.5 inside the 30° frontal area, -0.5 inside the 30° rear area, or 0 otherwise. C. The mask is shifted by 45° to the left of the visual field. The reward area extends from 30° to 60°. D. The mask is shifted by 45° to the left of the visual field. The reward area extends from -30° to -60°.
Fig 8.
Persistence of direction in the absence of the landmark.
A. (a) Simulations are done in a static mode, i.e. the agent never leaves the center of the arena and only rotates around its vertical axis. (b) After the 2000th step the cue disappears. B. (a) We tested the model in the same configuration as before, where the landmark is used both as the main cue for the compass and to generate the directed behaviour (Fig 7C). Results are shown in C.a. (b) We tested another configuration where the compass was provided with cues from the absolute orientation in the environment (such as sun position or sky polarization could provide). The attraction behaviour is kept based on the landmark as before. Results are shown in C.b. C. EPG activity rate (blue to yellow) and relative landmark orientation (red line, right scale) during simulation of the cue disappearance paradigm showed in A. (a) The compass and the attraction are both based on the landmark (B.a). The direction can be maintained but slowly drifts as there is no external reference. (b) The compass is based on an absolute orientation perception (potentially from sky cues) while the attraction is based the landmark (B.b). Note this creates a potential offset in the position of the bump and the landmark, but the agent still moves towards the landmark. The heading can be accurately maintained when the landmark disappears.
Fig 9.
Orientation to a goal direction supported by the MB long-term visual memory.
A. Mushroom bodies model diagram. The visual Projection Neurons (vPN) correspond to the visual units (Fig 2) extracting edge from the panoramic view. Each KC has a random post-synaptic connection pattern with 2 to 5 vPNs and a pre-synaptic connection to the MB output neuron (MBON). During the learning procedure, the memory reward signal (DAN) is activated and induces the decrease of the synaptic strength between the activated KCs and the MBON. The MBON activity is then used as the reward signal to adapt the CX circuit. B. Schematic of the learning phase before the simulation. The agent faces the feeder during 100 steps from a position in between the feeder and the center of the arena (learning pose), with an orientation error from −15 to 15°. The MB model memory is updated during this phase (DAN active). C. Example paths from the retrieval experiments using the MB model as an input for the CX model. D. Final direction vectors for 50 simulations with the feeder at 0° and the cue oriented at 45°. The red arc shows the median (dot) 95% C.I. obtained by bootstrapping (rep = 10000). E. Averaged right (red) and left (green) EPG-PFL3 synapses weights (shaded area: ±s.d.) obtained from 50 simulations.
Table 1.
Mushroom bodies visual long-term memory learning rules.
KC to MBON synapses are updated depending on the KC activity level and a reward signal coming from dopaminergic neuron (DAN), considered on (1) during the learning phase and off (0) otherwise (during the test phase). Only the combination of an active KC with an active DAN induce the reduction of the corresponding KC-MBON synapse weight.
Fig 10.
Replication of unilateral mushroom body lesion experiments from [13].
A. Schematic of the model including both MB and innate visual attraction. Both elements are linearly summed to form a single reward signal, using individual weights (ωVin & ωMB) to modulate the influence of each component. Lesions are indicated by the black scissors and the results are reported in C and D (Left and right lesion respectively). B. Simulation with the combination of innate attraction (ωVin) and the mushroom bodies (ωMB) pathway. The learning procedure is similar as presented in Fig 9B. From top to bottom, final directions (blue arrows, nsim = 100) taken during the simulations [blue arc indicate the median (dot) 95% C.I. obtained by bootstrap (nrep = 10000)], examples of path obtained during simulations, and averaged right (red) and left (green) EPG-PFL3 synapse weights (shaded area: ±s.d.) obtained during 50 simulations. C. Unilateral lesion of the left mushroom body. Final direction vectors and example paths. D. Unilateral lesion of the right mushroom body. Final direction vectors and example paths.