Sensorimotor experience remaps visual input to a heading direction network

In the Drosophila brain, “compass neurons” track the orientation of the body and head during navigation (the fly’s heading) 1 , 2 . In the absence of visual cues, the compass neuron network estimates heading by integrating self-movement signals over time 3 , 4 . When a visual cue is present, the network’s estimate is more accurate 1 , 3 . Visual inputs to compass neurons are thought to originate from inhibitory neurons called R neurons; R neuron receptive fields tile visual space 5 . The axon of each R neuron overlaps with the dendrites of every compass neuron 6 , raising the question of how visual cues are integrated into the compass. Here, using in vivo whole-cell recordings, we show that a visual cue can evoke synaptic inhibition in compass neurons, and R neurons mediate this inhibition. Each compass neuron is only inhibited by specific visual cue positions, implying that many potential connections from R neurons onto compass neurons are actually weak or silent. Notably, we show that the pattern of visually evoked inhibition can reorganize over minutes as the fly explores an altered virtual reality environment. Using ensemble calcium imaging, we demonstrate that this reorganization causes persistent changes in the compass coordinate frame. Our results suggest a model where correlated pre- and postsynaptic activity triggers associative long-term synaptic depression of visually evoked inhibition in compass neurons. Our findings provide evidence for the theoretical proposal that associative plasticity of sensory inputs, when combined with attractor dynamics, can reconcile self-movement information with changing external cues to generate a coherent sense of direction 7 – 12 .

Although the axon of each R neuron overlaps with every E-PG dendrite ( ), R→E-PG connections should be functionally selective; otherwise, information about the position of a visual cue would be discarded. We hypothesized that the all-to-all matrix of R→E-PG anatomical connections ( ) represents a set of “potential” functional connections which can be re-patterned during spatial learning. We therefore set out to test two hypotheses – first, that individual E-PG neurons respond selectively to specific visual cue positions, and second, that changes in visual-heading associations can trigger systematic, time-locked changes in the pattern of E-PG visual inputs.

Similar to mammalian head-direction cells, Drosophila compass neurons (called E-PG neurons) exhibit properties of a ring attractor 2 . Indeed, the dendrites of E-PG neurons are arranged in a ring in the brain ( ). At any point in time, there is one “bump” of activity in the E-PG ensemble which rotates as the fly turns 1 . This network receives continuous input from brain regions that track the fly’s rotational velocity via optic flow signals, proprioceptive signals, and/or motor efference signals 3 , 4 . These rotational velocity inputs push the bump around the circle. Visual cues make the bump’s position more accurate and stable 1 , 3 . However, we do not know whether visual inputs to E-PG neurons are plastic: The “offset” between the E-PG bump and the visual world is different in different individuals and it can occasionally change unpredictably within an individual 1 , 3 , but network instability alone is not evidence for synaptic plasticity.

The compass neurons in the Drosophila brain exhibit some resemblance to the head direction cells of the mammalian brain 13 – 16 . Visual cues stabilize the tuning preferences of mammalian head direction cells 15 , and when the experimenter rotates a visual cue to a new horizontal position, the preferences of all the head direction neurons rotate together 14 , 16 . It has been proposed that the mammalian head direction system represents a ring attractor – a network whose global dynamics exhibit multiple stable states which unfold in a repeated sequence in response to an input 7 , 17 , 18 . However, we do not know how visual cues anchor the mammalian head-direction system at a mechanistic level. It has been suggested that Hebbian synaptic plasticity of visual inputs could enforce the correct mapping between sensory cue and attractor network states 7 .

Results

Our first challenge was to isolate the synaptic input to E-PG neurons that is related to visual cue position, separate from the synaptic input related to the fly’s rotational velocity. We reasoned that this should be possible if we flashed visual cues transiently at randomized positions, preventing the fly from behaviorally fixating the stimulus. We therefore performed in vivo whole-cell recordings from E-PG neurons while flashing a bright vertical bar on a dark circular panorama at randomized horizontal positions ( ). In a typical neuron, we observed hyperpolarization that was time-locked to flashes at specific positions ( ). To verify that these neural responses are not related to the fly’s rotational velocity, we analyzed the movement of the air-cushioned ball that the fly was standing on ( ). Neural responses were unrelated to the fly’s rotational velocity around the time of the visual flash ( ), and there was no correlation between the fly’s rotational velocity and the flash ( ). Therefore, we can interpret visually-locked responses as synaptic inputs related to visual cue position. We call this the cell’s “visual receptive field”. The finding of visually-evoked hyperpolarization is consistent with the fact that R neurons release the inhibitory neurotransmitter γ-aminobutyric acid (GABA)19,20.

In almost every E-PG neuron, we found that some visual cue positions elicited hyperpolarization while other positions elicited no hyperpolarization ( , ). This implies that each E-PG neuron receives relatively strong input from some R neurons but weak or nonexistent input from other R neurons. In about half of E-PG neurons, we also found that some cue positions elicited depolarization ( ). Depolarization may represent disinhibition: because there is ongoing mutual inhibition between E-PG neurons2, a visual cue that inhibits one E-PG neuron will disinhibit other E-PG neurons.

We found that different E-PG neurons had distinct visual receptive fields ( ). When we sorted cells by the position eliciting minimal hyperpolarization; we found a uniform mapping of cue positions onto E-PG neurons. Notably, hyperpolarization was more prominent for lateral cue positions ( , ); this spatial bias is probably inherited from R neurons, because R neuron receptive fields are similarly biased towards lateral positions5.

When we managed to record sequentially from two adjacent E-PG neurons in the same brain, we found they had adjacent receptive fields adjacent receptive fields, as we would expect ( ). However, when we pooled data across brains, we found no systematic relationship between the location of the E-PG neuron’s dendrites and its receptive field ( ). Therefore, the mapping from visual space to compass coordinates is different across individuals.

Next, we asked how a neuron’s visual receptive field compares with its heading tuning. To measure heading tuning, we allowed the fly to walk in closed-loop virtual reality (VR) where the horizontal position of the cue was locked to the fly’s virtual heading ( , ). We periodically paused VR to map the same neuron’s visual receptive field using brief random flashes. In most neurons, we found that the visual receptive field was correlated with heading tuning ( – , & ). This result is notable because heading tuning reflects not only synaptic inputs related to visual cue position, but also synaptic inputs related to the fly’s rotational velocity. Imperfect alignment between these inputs may explain why some neurons showed poor correlations ( ).

To confirm that R neurons are actually the source of visual responses in E-PG cells, we focused on two R neuron types (R2 and R4d) that respond to sparse visual cues5. First, we used whole-cell recordings to confirm that these R neuron types can be excited by the visual cue ( ). Second, we verified that optogenetically activating either R2 or R4d neurons inhibits E-PG neurons ( ). Third, we established that R neurons are required for normal visually evoked hyperpolarization in E-PG neurons. We used two independent driver lines to hyperpolarize R2 or 4d neurons by overexpressing the potassium channel Kir2.1 ( ), and we confirmed that visually evoked hyperpolarization was attenuated ( , ). In both genotypes, a few E-PG neurons still showed some visual responses, likely because neither driver line achieves complete coverage of R2/4d neurons ( , ).

Next, we turned to our second hypothesis – that changes in visual-heading associations can trigger systematic, time-locked changes in E-PG visual receptive fields. After allowing the fly to navigate in VR with one visual cue (the pre-training block), we switched to VR with two cues positioned 180° degrees apart (the training block). In the training block, a full turn and a half-turn will arrive at an identical view of the world, meaning the correlation between rotational velocity signals and visual cue position signals will be altered.

To assess the effect of training on network dynamics, we imaged calcium signals from the entire E-PG ensemble ( ). During pre-training, there was a stable offset between the visual environment and the E-PG bump ( – ). During training, the offset toggled between two values ~180° apart. This result is expected, because there are two equally-valid interpretations of the visual scene, yet only one bump can exist in the E-PG ensemble2. When the fly made a 360° turn, we often saw the bump flow twice around 180° of the E-PG ensemble, skipping over the other 180° ( – ). Rotational velocity inputs to the E-PG network should drive the bump to traverse the full circle during a full turn3,4; the “skipping over” phenomenon thus implies the dominance of visual position inputs over angular velocity inputs. The E-PG neurons that were traversed twice essentially displayed two preferred heading directions; this is reminiscent of the finding that some rat head direction cells show two preferred directions in an environment with two-fold rotational symmetry21.

Upon returning to a one-cue environment (post-training), the offset sometimes immediately settled into its original value. Often, however, this was not the case. Rather, the offset continued to toggle for several minutes, or else it immediately settled in a new value rather than the original one ( – ). Both of the latter two outcomes suggest a persistent, systematic change in the way that visual cues are mapped onto E-PG neurons. We observed one of the latter outcomes in half of our experiments ( , ).

Finally, to investigate whether training changes visual receptive fields, we returned to E-PG whole-cell recordings ( ). We began each experiment with one visual cue in VR (pre-training). We then switched to two visual cues in VR (training). Between each block of VR, we periodically paused to map the neuron’s receptive field with brief random flashes. Whereas we used a 360° panorama during calcium imaging, the spatial constraints of electrophysiology required us to map the 360° environment onto a 270° panorama1,10.

During the training block, we found that some E-PG neurons were strongly modulated by the fly’s heading. In these neurons, training produced striking changes in the visual receptive field. These changes were bidirectional ( ), suggesting that visually evoked inhibition was depressed for some cue locations and potentiated for others. We quantified these changes by summing the absolute value of the change in the receptive field across all cue positions (“absolute change”, ). We also measured the change in the shape of the receptive field ( ). These metrics were correlated across experiments ( ); we never saw a large absolute change in the receptive field without a change in receptive field shape. We also never observed large receptive field changes under control conditions where flies only experienced one cue in VR (not two cues) during the period between the receptive field mapping epochs ( , & ).

By contrast, other E-PG neurons were essentially unmodulated by the fly’s heading during training ( , neuron 5). These neurons may reside in sectors of the ensemble that were “skipped over” by the bump during training. Interestingly, in these neurons, training had almost no effect on visual receptive fields ( , ). Overall, the magnitude of heading modulation during training was significantly correlated with the subsequent visual receptive field change ( ). This correlation implies that remapping depends on E-PG neuron activity. Simply exposing the fly to the altered visual environment is not sufficient; rather, visual cues must intersect with heading representations in E-PG neurons. Because R→E-PG synapses are the site of intersection between visual responses and heading representations, they are the most likely locus of plasticity. In a companion study, Kim et al.22 used optogenetic manipulations to reach the same conclusion. Because R neuron dendrites form a retinotopic map which is fairly consistent across flies5, it seems unlikely that the visual map in R neuron dendrites is experience-dependent, further supporting the notion that R→E-PG synapses are the locus of plasticity.