Sensorimotor experience remaps visual input to a heading-direction network

Abstract

In the Drosophila brain, 'compass' neurons track the orientation of the body and head (the fly's heading) during navigation ^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 estimate of the network is more accurate^1,3. Visual inputs to compass neurons are thought to originate from inhibitory neurons called R neurons (also known as ring neurons); the receptive fields of R neurons tile visual space⁵. The axon of each R neuron overlaps with the dendrites of every compass neuron⁶, 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 that R neurons mediate this inhibition. Each compass neuron is inhibited only by specific visual cue positions, indicating that many potential connections from R neurons onto compass neurons are actually weak or silent. We also 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. Taken together, our data suggest a model in which 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.

Extended Data Fig. 1:. Measuring behavior and E-PG visual responses.

a. Side view of a fly walking on an air-cushioned ball during an electrophysiology experiment. b. Image of the ball and plastic holder. Air flows up through the holder and out the semi-spherical depression that cradles the ball. c. Schematic of the experimental set-up viewed from above. The fly is secured in an aperture in the center of a horizontal platform. The platform is surrounded by a circular panorama. The panorama is composed of square LED arrays (2 squares vertically × 12 squares horizontally). The ball is illuminated by an infrared (IR) LED which is visible as a red spot in (b). A camera captures an image of the ball to enable tracking using FicTrac. Inset shows FicTrac view. Camera and IR LED are not drawn to scale. d. Fly’s yaw velocity versus cue position. This is the data set that is the basis for Fig. 1f, but here broken down into individual-fly-averages, and with right (+) and left (−) cue positions kept separate. Positive velocities are right turns, and negative velocities are left turns. No tests showed a statistically significant yaw velocity (p<0.05, two-sided comparison to bootstrap distribution) for any individual fly at any cue position. For details of analysis, see Methods: yaw during open-loop epochs. e. Yaw velocity in response to the visual cue presentation. This analysis is the same as that shown in Fig. 1f, but here yaw velocity is plotted versus the distance of the cue jump between consecutive trials. As in Fig. 1f, we show mean (black) ± 1 standard deviation (gray) across experiments (73 experiments in 68 flies). Magenta lines show the bootstrapped 95% confidence interval of the mean across flies after randomizing cue positions, Bonferroni-corrected for multiple comparisons. Because the mean lies within these bounds, it is not significantly different from random. This analysis further supports the conclusion that there is no systematic yaw response to the random flashes of the vertical bar. For details of analysis, see Methods: yaw during open-loop epochs. f. The visual receptive field of an example cell measured multiple times over the course of a 40-min recording. Each row shows data from a separate visual mapping epoch. Data from this example cell are also shown in Fig 1e. Note the stability of the visual receptive field over this time period. Extended Data Fig. 1 genotype: UAS-mCD8::GFP / UAS-mCD8::GFP; R60D05-Gal4 / R60D05-Gal4

Extended Data Fig. 2:. Visually evoked hyperpolarization and depolarization, during and after cue presentation.

a. Example voltage responses of the same E-PG neuron to two cue positions. Dashed lines indicate the mean baseline voltage before the cue. This neuron is hyperpolarized by the cue at 90° and depolarized by the cue at −97°. Note that hyperpolarization decays more rapidly than depolarization. In (b), to quantify visual receptive fields, we measured the change in voltage during cue presentation and after cue removal in the 250-ms windows marked here with brackets, in both cases relative to baseline. b. Summary of E-PG visual receptive fields measured during cue presentation. Cells are sorted by the cue position that evokes maximal hyperpolarization. The histogram below shows the number of E-PG neurons with maximal hyperpolarization at each cue position (73 E-PG neurons in 68 flies). c. Summary of E-PG visual receptive fields measured after cue removal. Cell order is the same as in (b). Note that hyperpolarizing responses tend to decay, whereas depolarizing responses tend to persist; this is consistent with the hypothesis the hyperpolarization during cue presentation is due to direct synaptic inhibition from R neurons, whereas depolarization is polysynaptic and caused by withdrawal of tonic synaptic inhibition. The histogram below shows the number of E-PG neurons with maximal hyperpolarization after cue removal for each cue position. d. Same as (b), but sorted by the cue position that evoked maximal depolarization (minimal hyperpolarization), as in Fig. 1g. e. Same as (c), but with the cell order as in (d). f. Summed response across all neurons measured during (left) and after (right) the cue. The left curve has a pair of minima around ~±100°; this bias is likely inherited from R neuron receptive fields, which are biased toward positions offset from the visual midline. By contrast, the right curve is relatively flat. g. Visual cue position eliciting maximal depolarization (minimum hyperpolarization), plotted versus E-PG neuron location, for the 21 recorded E-PG neurons that were filled. No signification correlation was observed (circular correlation coefficient = −0.15, p = 0.49). Extended Data Fig. 2 genotype: UAS-mCD8::GFP / UAS-mCD8::GFP; R60D05-Gal4 / R60D05-Gal4

Extended Data Fig. 3:. E-PG neuron pairs recording sequentially from the same brain.

a. Two biocytin-filled dendrites (green) from sequentially recorded E-PG neurons that innervate adjacent wedges within the ellipsoid body. Neuropil reference marker is shown in gray (anti-nc82 antibody). Images are maximum intensity z-projections. Scale bar is 10 μm. The schematic shows the approximate position of ellipsoid body and E-PG dendrites from a coronal view of the fly brain. b-c. Heading tuning (red, measured in virtual reality) and visual receptive field (blue, measured with random flashes) from sequentially recorded E-PG pairs from two example flies. Dendritic locations of the recorded neurons are green in the ellipsoid body schematic above each set of plots. In both cases, by chance, the two dendrites were physically adjacent. In both cases, adjacent E-PG neurons from the same fly exhibited similar visual receptive fields and heading tuning curves, supporting the conclusion that adjacent E-PG cells typically receive inhibition from adjacent regions of visual space and represent adjacent heading directions. Comparing the visual receptive field and the heading tuning curve for each neuron yielded correlation coefficients (Pearson’s) of 0.76 (fly 1 neuron 1), 0.90 (fly 1 neuron 2), 0.95 (fly 2 neuron 1), and 0.65 (fly 2 neuron 2). Extended Data Fig. 3 genotype: UAS-mCD8::GFP / UAS-mCD8::GFP; R60D05-Gal4 / R60D05-Gal4

Extended Data Fig. 4:. E-PG neuron visual receptive fields and heading tuning.

Heading tuning (red, closed-loop mode) and visual receptive fields (blue, open-loop mode) for all 40 recorded E-PG neurons (from 39 flies). For each neuron, the correlation coefficient (Pearson’s) is reported for the comparison between the visual receptive field and the heading tuning curve. Asterisks denote data also shown in Fig. 2. Extended Data Fig. 4 genotype: UAS-mCD8::GFP / UAS-mCD8::GFP; R60D05-Gal4 / R60D05-Gal4

Extended Data Fig. 5:. R neurons types labeled by R20A02-Gal4 and R54E12-Gal4 described by Multi-Color Flip-Out.

a. Observed numbers of R neurons belonging to each type from a dataset of n = 78 single-neuron Multi-Color Flp-Out clones (MCFO) from the R20A02-Gal4 line. R neuron types were classified according to Omoto et al.. b. Same as (a) but for the R54E12-Gal4 line (n = 61 single-neuron MCFO clones). c-h. Examples of single R-neuron MCFO clones. Images are maximum intensity z-projections. Background labeling was manually removed to improve clarity of specific neuronal morphologies. i. Multiple R-neuron MCFO clones labeled in different colors using the R20A02-Gal4 line. Image is a maximum-intensity z-projection. Scale bars are 20 μm. Extended Data Fig. 5 genotypes: R57C10-FLPG5.PEST; UAS(FRT.stop)myr::smGdP-HA, UAS(FRT.stop)myr::smGdP-V5, UAS(FRT.stop)myr::smGdP-FLAG / R20A02-Gal4, R57C10-FLPG5.PEST; UAS(FRT.stop)myr::smGdP-HA, UAS(FRT.stop)myr::smGdP-V5, UAS(FRT.stop)myr::smGdP-FLAG / R54E12-Gal4

Extended Data Fig. 6:. Suppressing R neuron activity with two independent driver lines reduces E-PG visually evoked hyperpolarization.

a. Same as Fig. 3c (right), except instead of measuring peak visually evoked hyperpolarization, we measured mean visually evoked hyperpolarization (by zeroing all non-negative visual responses and then averaging visual responses across all cue positions) n = 8, 10, 12, 10, 9). Both Kir means are significantly different from corresponding genetic controls using two-sided Wilcoxon rank-sum tests. R20A02 Kir vs. R20A02 / + and UAS / + (p= 0.0013 and 0.0003), R54E12 Kir vs R54E12 / + and UAS / + (p = 0.005 and 0.0025). b. R neuron population labeled by Kir2.1::EGFP. Images are maximum intensity z-projections. c. Numbers of R neurons per hemisphere expressing Kir2.1::EGFP in each experimental genotype n = 9 (R20A02), n = 11 (R54E12) (horizontal lines are means). Based on the total number of R neurons of each type reported by Omoto et al., and our MCFO quantification of the R neuron types labeled by R20A02-Gal4 and R54E12-Gal4 (Extended Data Fig. 5), these cell counts suggest that R20A02-Gal4 targets approximately 20% of R2, 30% of R4m, and all R4d neurons. These counts suggest that R20A02-Gal4 targets approximately 40% of R2 neurons and all R4m and R4d neurons. This incomplete targeting of outer R neurons may provide one explanation for the remaining visually evoked inhibition observed in some recordings (Fig. 3). Note that while both driver lines label other neurons in the central brain and visual system, R neurons appear to be the only cell type that is labeled by both lines. In the visual system, driver line R20A02-Gal4 targets one medulla intrinsic neuron, likely Mi12 and one cell type that arborizes in ~layer 4–6 of the lobula, while driver line R54E12-Gal4 appears to target the medulla neuron Tm3. Extended Data Fig. 6 genotypes: + / w; R60D05-LexA / LexAop-mCD8::GFP; + / UAS-Kir2., (UAS-only control), + / w; R60D05-LexA / LexAop-mCD8::GFP; R20A02-Gal4 /+ (R20A02 Gal4-only control), + / w; R60D05-LexA/LexAop-mCD8::GFP; R54E12-Gal4/+ (R54E12 Gal4-only control),+ / w; R60D05-LexA / LexAop-mCD8::GFP; R20A02-Gal4/UAS-Kir2.1 (R20A02 Kir), + / w; R60D05-LexA/LexAop-mCD8::GFP; R54E12-Gal4 / UAS-Kir2.1 (R54E12 Kir)

Extended Data Fig. 7:. Offset probability histograms in training experiments.

Offset probability histograms during each segment of the training experiments shown in Fig. 4, for all 19 GCaMP imaging experiments (in 19 flies). As in Fig. 4, the circular mean during the pre-training period is defined as offset₀ (here marked with an arrowhead), and for display purposes we horizontally aligned all the offset₀ values in different flies. Asterisks mark data shown in Fig. 4. Extended Data Fig. 7 genotype: + / w; UAS-GCaMP6f / +; R60D05-Gal4 / +

Extended Data Fig. 8:. Heading tuning and visual receptive field measurements in training experiments.

Heading tuning curves and visual receptive fields for all additional 17 E-PG neurons (from 17 flies) from the training experiments in Fig. 5. As in Fig. 5, red solid curves are heading tuning. Red dashed curves are the change in heading tuning (training minus pre-training). Blue curves are visual receptive fields. Blue dashed curve is the change in the visual receptive field (2^nd probe minus 1^st probe). Seven neurons from this data set also appear in Figs. 1–2.

Extended Data Fig. 9:. Controls for remapping experiments.

a. Data reproduced from Fig. 5e. Absolute change in visual receptive fields. Control flies navigated in a 1-cue world (rather than a 2-cue world) during the waiting period between the open-loop epochs used to compute the change in visual responses. In some cases (“matched control”), flies received exactly the same protocol as the experimental condition except with 1-cue closed-loop during the training period; in other words, these matched controls received 12 consecutive minutes of 1-cue (rather than 2-cue) closed-loop during the “training” period. In all other cases (“control”), flies received 4-minute blocks of 1-cue closed-loop interleaved with 150-second open-loop epochs during the “training” period, which lasted 12 minutes or more. b. Visual receptive fields from control cells. Blue dashed curve is the change in visual receptive field (2^nd probe – 1^st probe) over the control period. Typically, visual receptive fields were stable over time under control conditions (control neurons 2 and 3). On occasion, we observed spontaneous changes in an E-PG neuron’s visual receptive field during the control period (e.g., control neuron 1), although these changes were not as large as the changes we observed in many neurons in trained flies (see panel a). c. Heading tuning in the same three control cells. Note how the spontaneous changes in visual receptive fields seen in neuron 1 are accompanied by changes in heading tuning.

Figure 1:. E-PG neurons are inhibited by visual cues at specific positions.

a. Schematic: E-PG neuron dendrites form a circular array, with adjacent cells representing adjacent headings. b. Schematic: each R neuron axon forms a ring (left) which overlaps all E-PG dendrites (right). c. Schematic: an unwrapped R→E-PG matrix. d. Schematic: an E-PG neuron is recorded in whole-cell mode while the fly walks on a ball, surrounded by a panorama where a cue flashes at random horizontal positions. e. Top: three example E-PG responses per cue position, for three different positions, all from the same recording. Bottom: fly’s yaw velocity (+ right, − left). Note that the fly behaves differently on different trials, but the neural response is essentially the same regardless of the fly’s behavior. Cue flash is 500 ms. f. Fly’s rotational velocity around cue presentation, mean (black) ± 1 standard deviation (gray) across flies. Magenta lines show bootstrapped 95% confidence interval of the mean across flies after randomizing cue positions, Bonferroni-corrected; because the mean lies within these bounds, it is not significantly different from random. g. Summary of E-PG visual receptive fields (73 neurons in 68 flies). Cells are sorted by the cue position that evoked the most positive (least negative) response. Histogram counts cells preferring each cue position. Some cells were filled to determine their location (see h). h. Cue position eliciting peak inhibition, versus neuron location (no significant correlation: circular correlation coefficient = 0.097, p = 0.66, n=21; see Extended Data Fig. 2g and ref. ).

Figure 2:. E-PG neuron visual receptive fields align with heading tuning.

a. Schematic: interleaved blocks measuring the visual receptive fields and heading tuning. b. Top: E-PG voltage during a virtual reality (VR) epoch. Bottom: VR heading. A heading of 0° means the cue is in front of the fly. c. Comparison of visual receptive field and heading tuning from three example E-PG neurons (from 3 flies, with Pearson’s correlation coefficients). d. Pearson’s correlation coefficients from 40 cells in 39 flies (all cells from Fig. 1). The mean and 95% confidence internal (CI) are shown as horizontal/vertical lines. The mean of the data is outside the 95% CI of a bootstrap distribution (gray violin plot) computed on randomized visual-heading pairings.

Figure 3:. R neurons drive visually evoked inhibition in E-PG neurons.

a. Left: visually evoked spike rates in an R neuron (mean ± SEM across trials, n = 5–6 trials, R2 neuron). Right: four responses to repeated presentation of the best cue position for this neuron. We observed spatially tuned responses in 3 of 7 R2 cells and 1 of 3 R4d cells; an additional 3 R2 cells and 1 R4d cell responded to full field illumination but were unresponsive to the cue or not spatially tuned. b. Left: responses of an E-PG neuron to optogenetic activation of R2 neurons via Chrimson (ChR), with four single trials in gray, mean in black. Middle: same but with no ChR in R neurons, n= 4 trials. Right: summary of mean evoked hyperpolarization with ChR in R neurons (■ R2 neurons, n=7; ▲ R4d neurons, n=4) and controls (n=5). c. Left: E-PG visual receptive fields in flies where R neurons were hyperpolarized using Kir2.1 expression driven by R54E12-Gal4 or R20A02-Gal4 (green shades) versus controls (R54E12-Gal4-only, R20A02-Gal4-only, UAS-Kir-only, gray shades). Right: summary of peak visually evoked hyperpolarization, color coded as before (horizontal lines are means; n = 8, 10, 12, 10, 9 cells; R54E12 Kir versus R54E12 / + and UAS / +, p = 0.021 and 0.0016; R20A02 Kir versus R20A02 / + and UAS / +, p= 0.0046 and 0.012; two-sided Wilcoxon rank-sum tests).

Figure 4:. Visuomotor experience can persistently change ensemble heading direction representations.

a. E-PG ensemble GCaMP6f signals. Here the circular E-PG ensemble has been linearized, with each row showing 8 sectors of the ensemble. The fly walked in a 1-cue environment (pre-training, ≥10 min), then a 2-cue environment (training, 20 min), and finally a 1-cue environment (post-training, ≥4 min). Three snippets of one experiment are shown. Brackets mark 360° turns when the bump skipped over half the ensemble. b. In the same experiment, VR heading (red) overlaid with the decoded neural representation of heading (blue). We double-plotted both traces and shifted the entire red trace horizontally so it overlapped with the blue trace during pre-training. c. The offset of the neural representation of heading relative to the fly heading, double-plotted. The circular mean during pre-training is marked with a vertical line (defined as offset₀). d. Offset probability histograms during each block, for seven example experiments. We found diverse values of offset₀ in different flies, as reported previously, but for display we horizontally aligned all offset₀ values in different flies. The “opposing range” is the range from (offset₀ + 90°) to (offset₀ − 90°). Examples 1–5 show a phase change post-training. e. Total offset probability in the opposing range. Each set of connected points is one experiment (n=19 flies). Training and post-training are both significantly different from pre-training (p=3.8×10⁻⁶ training vs pre-training and p=5.31⁻⁵ post-training vs pre-training, two-sided exact paired Wilcoxon signed rank tests).

Figure 5:. Visuomotor experience can remap visual input to E-PG neurons contingent on postsynaptic activity.

a. After the fly navigated in VR with one cue (pre-training), we measured the E-PG visual receptive field (1^st probe). Then the fly navigated in VR with two cues for 12 min (training) and we again measured the visual receptive field (2^nd probe). b. Five example neurons. Red solid curves are heading tuning (red tick is −50 mV). Red dashed curve is the change in heading tuning (training minus pre-training). Blue solid curves are visual receptive fields. Blue dashed curve is the change in visual receptive fields (2^nd probe minus 1^st probe). Arrowheads mark large changes. Neuron 5 is an example with little modulation by heading during training and little change in visual receptive field. c. Explanation of metrics in (d-f). d. Absolute change in visual receptive fields, versus change in receptive field shape (R² = 0.44, p = 0.00078 testing t-statistic slope ≠ 0, 22 E-PG neurons in 22 flies). e. Absolute change in visual receptive fields post-training (22 E-PG neurons in 22 flies) versus controls (17 E-PG neurons in 17 flies). Training is significantly different from control (p= 0.043, two-sided Wilcoxon rank-sum test). Controls walked in a one-cue VR (not two-cue VR) between the 1^st and 2^nd probe. Four training experiments had changes significantly larger than any controls (>2 SDs above control mean, vertical bar); these are neurons 1–4. f. Absolute change in visual receptive fields, versus modulation by heading during training (R² = 0.52, p = 0.00016 testing t-statistic slope ≠ 0, 22 E-PG neurons in 22 flies). g. Schematic of model. When a visual cue appears, it activates specific R neurons (highlighted magenta cell), and this pushes the bump toward the E-PG neuron with minimal inhibition (highlighted gray cell). Training changes R→E-PG weights so the bump toggles between two offsets during post-training. R neurons are ordered by receptive field position.