Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 74 of the 74 references it cites were found in a registry or an index; 38 of the 68 sources its claims cite were checked for support.

Multimodal cue integration and learning in a neural representation of head direction

Open access

Abstract

Navigation requires us to take account of multiple spatial cues with varying levels of informativeness and learn their spatial relationships. Here we investigate this process in the Drosophila head direction system, which functions as a ring attractor and a topographic map of head direction. Using population calcium imaging and multimodal virtual reality environments, we show that increasing cue informativeness improves encoding accuracy and produces a narrower and higher bump of activity. When cues conflict, the more informative cue exerts more weight. A familiar cue is weighted more heavily and used to guide the remapping of a less familiar cue. When a cue is less informative, it is remapped more readily in response to cue conflict. All these results can be explained by an attractor model with plastic sensory synapses. Our findings provide a mechanistic explanation for how the brain assembles spatial representations through inference and learning.

Overview

CoChat Summary

TL;DR

In the Drosophila head direction system, increasing cue intensity improves HD encoding accuracy and changes the bump profile, while more informative or familiar cues are weighted more heavily and guide remapping of less familiar cues. A ring attractor model with plastic sensory synapses reproduces these effects and highlights a stability–flexibility tradeoff.

Key findings

  • Increasing visual cue intensity progressively improved HD encoding accuracy, narrowed the EPG bump, and increased bump amplitude at the highest intensity (15 flies); the model explains this via inhibitory ER input plus a Hebbian LTD notch in ER→EPG weights.
  • With conflicting visual and wind cues, the cue that produced better HD encoding when presented alone was generally weighted more. Across 13 flies, mean single-cue accuracy was similar (P = 0.76), but accuracy differences correlated with bump width differences (R² = −0.68, P = 0.01) and bump amplitude differences (R² = 0.58, P = 0.04); bump preference index correlated with relative HD encoding accuracy (P = 0.03).
  • A stable two-cue configuration (18 flies) increased HD encoding accuracy and produced a narrower, higher-amplitude bump; the more recently familiar cue retained its offset and instructed remapping of the less familiar cue, with offset change predicted by cue conflict.
  • In an inverted-gain task, some flies learned to invert the visual cue mapping; remapping index was higher in the final versus initial inverted-gain block, and final remapping was predicted by wider bumps in the preceding normal-gain block (P = 0.02) but not by bump amplitude (P = 0.79) (14 flies; 12 for accuracy/orientation because two lacked a darkness trial).

Methods

  • Head-fixed Drosophila melanogaster (1-day-old virgin females) expressing jGCaMP7f in EPG neurons were imaged with two-photon calcium imaging while walking on a spherical treadmill; visual cues were presented on a 360° LED panorama and wind cues via a rotating air nozzle in closed loop.
  • Bump position, width, and amplitude were estimated with von Mises fits; HD encoding accuracy was defined as 1 − circular variance of the bump–HD offset. Analyses used linear mixed-effects models with Tukey/Bonferroni corrections, Wilcoxon tests, Pearson correlations, and Bayesian fitting for offset changes.
  • A ring attractor model with ER→EPG synaptic plasticity (associative LTD and nonassociative LTP) was simulated, with variations in ER input amplitude used to reproduce individual differences and remapping behavior.

Limitations

  • Stated in the paper: synaptic weights cannot be observed directly and are inferred from bump position and amplitude; the model cannot perform optimal Bayesian inference and does not store certainty about HD, with data implying no persistent bump-amplitude boost after an informative cue. Sample sizes were not predetermined, the experimenter was not blind for the randomized brightness-block order, normality was not formally tested for Tukey/Pearson tests, and starvation times varied by experiment.
  • Evident from the text: individual differences and model fits are correlational, with no direct causal manipulation of ER amplitude or synaptic weights; sample sizes are limited (12–18 flies in key experiments), and some effects are trends or non-significant.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41593-024-01823-z

Cites

74 of 74 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 The Drosophila HD system. a, HD cells (EPG neurons) form a ring attractor network in the EB. Their axons project to the PB, where they form two linearized topographic maps of HD. b, The position of the EPG activity bump is influenced by ER neurons that encode the positions of visual HD cues or the direction of the wind. ER neurons are inhibitory and the most active ER neurons push the bump to the location where their inhibitory output is minimal. ER → EPG connections are anatomically all-to-all but their weights are shaped by Hebbian plasticity at ER → EPG synapses, such that each ER neuron generally makes functional synapses onto only a subset of EPG neurons. c, Schematic ER → EPG weights. Given a single visual cue and a steady wind direction, associative LTD is predicted to produce a diagonal notch of weak connections in each weight matrix. EPG neurons are sorted by their preferred HD. ER neurons are sorted by their preferred cue position. If the two cues are aligned in the simulated environment, Hebbian plasticity should align the notches. d, We hypothesize that cue salience and stability affect bump attractor dynamics and learning. e, We image EPG neurons in head-fixed flies walking on a spherical treadmill. As the fly turns on the spherical treadmill, the virtual environment rotates around the fly in the expected direction. Here, the environment contains a bright vertical stripe that serves as an HD cue. f, The bump of EPG activity tracks the fly’s fictive HD in a virtual reality environment, with a relatively constant angular offset. Bump position rotates clockwise in the EB (imaged from the posterior side of the head) as HD rotates counterclockwise; therefore, to account for this directionality, we always plot (−HD) to make it easier to visualize the correspondence between bump position and HD.
Fig. 1 The Drosophila HD system. a, HD cells (EPG neurons) form a ring attractor network in the EB. Their axons project to the PB, where they form two linearized topographic maps of HD. b, The position of the EPG activity bump is influenced by ER neurons that encode the positions of visual HD cues or the direction of the wind. ER neurons are inhibitory and the most active ER neurons push the bump to the location where their inhibitory output is minimal. ER → EPG connections are anatomically all-to-all but their weights are shaped by Hebbian plasticity at ER → EPG synapses, such that each ER neuron generally makes functional synapses onto only a subset of EPG neurons. c, Schematic ER → EPG weights. Given a single visual cue and a steady wind direction, associative LTD is predicted to produce a diagonal notch of weak connections in each weight matrix. EPG neurons are sorted by their preferred HD. ER neurons are sorted by their preferred cue position. If the two cues are aligned in the simulated environment, Hebbian plasticity should align the notches. d, We hypothesize that cue salience and stability affect bump attractor dynamics and learning. e, We image EPG neurons in head-fixed flies walking on a spherical treadmill. As the fly turns on the spherical treadmill, the virtual environment rotates around the fly in the expected direction. Here, the environment contains a bright vertical stripe that serves as an HD cue. f, The bump of EPG activity tracks the fly’s fictive HD in a virtual reality environment, with a relatively constant angular offset. Bump position rotates clockwise in the EB (imaged from the posterior side of the head) as HD rotates counterclockwise; therefore, to account for this directionality, we always plot (−HD) to make it easier to visualize the correspondence between bump position and HD.CC BY · PubMed Central
Fig. 2 Increasing cue intensity changes the bump profile. This figure shows data for 15 flies. a, EPG ΔF/F in the PB is used to infer the position of the EPG bump in the EB (green). When a visual cue is present, the bump tracks HD (blue). The offset between HD and bump is fairly constant when a bright visual HD cue is available. b, Offset distributions for the fly in a. Black lines are vector averages, with line length denoting HD encoding accuracy. Polar histograms are shown on the same scale. c, HD encoding accuracy in each condition. In c,d,f,g, single flies are shown in gray, mean values are shown in black and P values are shown (linear mixed-effects models with Tukey comparisons and Bonferroni corrections). d, Consistency of behavioral orienting in each condition. e, ΔF/F and fit (Extended Data Fig. 1) at two time points for the same fly. f, Bump width in each condition. g, Bump amplitude in each condition. h, Model of the effect of increasing cue intensity on HD encoding accuracy, EPG bump width, EPG bump amplitude and the range of ER → EPG weights (max–min). Mean of 100 simulation runs ± s.e.m. i, Model of the spatial profile of ER and EPG population activity at one HD for each condition. Bottom, ER → EPG weights at the end of each block. The maximum weight represents maximum inhibition. The bright cue produces a deeper notch in the weight matrix.
Fig. 2 Increasing cue intensity changes the bump profile. This figure shows data for 15 flies. a, EPG ΔF/F in the PB is used to infer the position of the EPG bump in the EB (green). When a visual cue is present, the bump tracks HD (blue). The offset between HD and bump is fairly constant when a bright visual HD cue is available. b, Offset distributions for the fly in a. Black lines are vector averages, with line length denoting HD encoding accuracy. Polar histograms are shown on the same scale. c, HD encoding accuracy in each condition. In c,d,f,g, single flies are shown in gray, mean values are shown in black and P values are shown (linear mixed-effects models with Tukey comparisons and Bonferroni corrections). d, Consistency of behavioral orienting in each condition. e, ΔF/F and fit (Extended Data Fig. 1) at two time points for the same fly. f, Bump width in each condition. g, Bump amplitude in each condition. h, Model of the effect of increasing cue intensity on HD encoding accuracy, EPG bump width, EPG bump amplitude and the range of ER → EPG weights (max–min). Mean of 100 simulation runs ± s.e.m. i, Model of the spatial profile of ER and EPG population activity at one HD for each condition. Bottom, ER → EPG weights at the end of each block. The maximum weight represents maximum inhibition. The bright cue produces a deeper notch in the weight matrix.CC BY · PubMed Central
Fig. 3 The influence of a cue varies across individuals. This figure shows data for the 15 flies from Fig. 2. a, An example fly navigating in the virtual environment with the dim visual cue (HD encoding accuracy = 0.80). The polar histogram shows offset distribution. b, Another fly in the same environment (HD encoding accuracy = 0.14). c, Individual variations in HD encoding accuracy correlate with bump width (P = 0.01, Pearson correlation). Data are from the first block where the dim visual cue was presented (200 s in duration). The dim cue produced the most individual variation (Extended Data Fig. 2). d, Model showing that, with the dim cue, small variations in ER amplitude produce strongly correlated changes in bump width and HD encoding accuracy. e, Individual variations in HD encoding accuracy do not correlate with bump amplitude (P = 0.18, Pearson correlation). All data are from the environment with the dim cue. f, Model showing that, with a dim cue, small variations in ER amplitude produce weakly correlated changes in bump amplitude and HD encoding accuracy.
Fig. 3 The influence of a cue varies across individuals. This figure shows data for the 15 flies from Fig. 2. a, An example fly navigating in the virtual environment with the dim visual cue (HD encoding accuracy = 0.80). The polar histogram shows offset distribution. b, Another fly in the same environment (HD encoding accuracy = 0.14). c, Individual variations in HD encoding accuracy correlate with bump width (P = 0.01, Pearson correlation). Data are from the first block where the dim visual cue was presented (200 s in duration). The dim cue produced the most individual variation (Extended Data Fig. 2). d, Model showing that, with the dim cue, small variations in ER amplitude produce strongly correlated changes in bump width and HD encoding accuracy. e, Individual variations in HD encoding accuracy do not correlate with bump amplitude (P = 0.18, Pearson correlation). All data are from the environment with the dim cue. f, Model showing that, with a dim cue, small variations in ER amplitude produce weakly correlated changes in bump amplitude and HD encoding accuracy.CC BY · PubMed Central
Fig. 4 More informative cues are accorded more weight. This figure shows data for 13 flies. a, A bright visual cue and wind cue were tested individually and then combined. One cue was shifted 120° in an alternating sequence every 5 min. In some experiments, the visual cue came first (as shown here) while, in others, the wind came first. b, In the initial single-cue epochs, mean HD encoding accuracy is similar for the two cues (P = 0.76, paired two-sided Wilcoxon test). Thick horizontal lines are mean values. Gray lines are individual flies. c, In the initial single-cue epochs, the difference in HD encoding accuracy for the two cues is correlated with the bump width difference (R2 = −0.68, P = 0.01, Pearson correlation) and the bump amplitude difference (R2 = 0.58, P = 0.04, Pearson correlation). d, Example responses to +120° shifts of the visual cue. In example 1, the bump shifts upward to follow the upward (positive) shift of the cue, whereas, in example 2, it does not. In both cases, the wind does not shift. Extended Data Figure 3a shows these examples in more detail. e, The bump preference index is significantly correlated with relative HD encoding accuracy in the two single-cue environments (P = 0.03, Pearson correlation). Each gray line at a single encoding accuracy connects the data for one fly (four wind shifts in orange and four visual shifts in blue). An index of −1 means that the bump follows the visual cue, whereas +1 means that the bump follows the wind. This index is close to −1 for example 1 and close to +1 for example 2 (blue numerals). f,g, Same as c (f) and e (g) but for model networks with two populations of ER neurons. In different simulation runs, we varied the amplitude of sensory activity of one ER population while holding the amplitude constant in the other population.
Fig. 4 More informative cues are accorded more weight. This figure shows data for 13 flies. a, A bright visual cue and wind cue were tested individually and then combined. One cue was shifted 120° in an alternating sequence every 5 min. In some experiments, the visual cue came first (as shown here) while, in others, the wind came first. b, In the initial single-cue epochs, mean HD encoding accuracy is similar for the two cues (P = 0.76, paired two-sided Wilcoxon test). Thick horizontal lines are mean values. Gray lines are individual flies. c, In the initial single-cue epochs, the difference in HD encoding accuracy for the two cues is correlated with the bump width difference (R2 = −0.68, P = 0.01, Pearson correlation) and the bump amplitude difference (R2 = 0.58, P = 0.04, Pearson correlation). d, Example responses to +120° shifts of the visual cue. In example 1, the bump shifts upward to follow the upward (positive) shift of the cue, whereas, in example 2, it does not. In both cases, the wind does not shift. Extended Data Figure 3a shows these examples in more detail. e, The bump preference index is significantly correlated with relative HD encoding accuracy in the two single-cue environments (P = 0.03, Pearson correlation). Each gray line at a single encoding accuracy connects the data for one fly (four wind shifts in orange and four visual shifts in blue). An index of −1 means that the bump follows the visual cue, whereas +1 means that the bump follows the wind. This index is close to −1 for example 1 and close to +1 for example 2 (blue numerals). f,g, Same as c (f) and e (g) but for model networks with two populations of ER neurons. In different simulation runs, we varied the amplitude of sensory activity of one ER population while holding the amplitude constant in the other population.CC BY · PubMed Central
Fig. 5 Cue combinations change the bump profile and drive learning. This figure shows data for 18 flies. a, Experimental protocol. Each cue was tested alone, then the two cues were combined and, finally, each cue was retested alone. The first cue was either the bright visual cue or wind. During the initial single-cue blocks, the two cues produce similar HD encoding accuracy and similar bump profiles, on average (Extended Data Fig. 4). b, HD encoding accuracy, bump width and bump amplitude for all three blocks; P values are for linear mixed-effects models with Tukey comparisons and Bonferroni corrections. c, Offset distributions for three example flies (Extended Data Fig. 4). d, Absolute difference between the offset with both cues and the initial offset. This difference is significantly smaller for cue 2, which is the more familiar cue (P value from paired two-sided Wilcoxon test). In most flies, cue 2 captures the bump during the epoch with both cues. e, Offset change (final − initial) versus offset conflict during the cue combination period (both − initial). There are two data points per fly, one for cue 1 (the less familiar cue) and the other for cue 2 (the more familiar cue). The line shown in the plot is a fit to all these points, based on the model that describes the offset change as a linear function of conflict with von Mises noise (Methods). f, Model showing the temporal evolution of max–min ER → EPG weight (that is, the range of weight values in the matrix), HD encoding accuracy, bump width and bump amplitude. g, Model showing ER → EPG weights at three time points (i, ii and iii). h, The same parameters tracked over time in the experimental data (mean ± s.e.m. across flies). Here, all parameters were measured in a rolling 60-s time window (Extended Data Fig. 5).
Fig. 5 Cue combinations change the bump profile and drive learning. This figure shows data for 18 flies. a, Experimental protocol. Each cue was tested alone, then the two cues were combined and, finally, each cue was retested alone. The first cue was either the bright visual cue or wind. During the initial single-cue blocks, the two cues produce similar HD encoding accuracy and similar bump profiles, on average (Extended Data Fig. 4). b, HD encoding accuracy, bump width and bump amplitude for all three blocks; P values are for linear mixed-effects models with Tukey comparisons and Bonferroni corrections. c, Offset distributions for three example flies (Extended Data Fig. 4). d, Absolute difference between the offset with both cues and the initial offset. This difference is significantly smaller for cue 2, which is the more familiar cue (P value from paired two-sided Wilcoxon test). In most flies, cue 2 captures the bump during the epoch with both cues. e, Offset change (final − initial) versus offset conflict during the cue combination period (both − initial). There are two data points per fly, one for cue 1 (the less familiar cue) and the other for cue 2 (the more familiar cue). The line shown in the plot is a fit to all these points, based on the model that describes the offset change as a linear function of conflict with von Mises noise (Methods). f, Model showing the temporal evolution of max–min ER → EPG weight (that is, the range of weight values in the matrix), HD encoding accuracy, bump width and bump amplitude. g, Model showing ER → EPG weights at three time points (i, ii and iii). h, The same parameters tracked over time in the experimental data (mean ± s.e.m. across flies). Here, all parameters were measured in a rolling 60-s time window (Extended Data Fig. 5).CC BY · PubMed Central
Fig. 6 Learning to invert the mapping of a visual cue. a, Schematic of the protocol. b, Data from an example experiment, shown for the last 300 s of the inverted gain block. The EPG bump begins to track the fly’s HD (arrowhead). c, Data from another example experiment. In this case, the HD system simply follows the visual cue for the entire inverted gain block and, thus, it does not accurately track HD.
Fig. 6 Learning to invert the mapping of a visual cue. a, Schematic of the protocol. b, Data from an example experiment, shown for the last 300 s of the inverted gain block. The EPG bump begins to track the fly’s HD (arrowhead). c, Data from another example experiment. In this case, the HD system simply follows the visual cue for the entire inverted gain block and, thus, it does not accurately track HD.CC BY · PubMed Central
Fig. 7 Bump width in normal gain predicts cue remapping in inverted gain. Panels b,e,f show data from 14 flies, while c,d show data from 12 flies because two flies did not receive a ‘darkness’ trial. a, Offset distributions for the two example flies from Fig. 6. The remapping index is positive for fly 1 because the offset with respect to HD is more stable. The remapping index is negative for fly 2 because the offset with respect to the visual cue is more stable. b, Remapping index in the initial and final part of the inverted gain block. Across all flies, the remapping index is larger in the final part (P value from two-sided Wilcoxon test). Black and gray lines denote individuals with final remapping indices above or below the mean (horizontal bar); this is preserved in c–f. c,d, HD encoding accuracy (c) and consistency of behavioral orientation (d) in darkness versus inverted gain (P values from linear mixed-effects models with Tukey comparisons and Bonferroni corrections). e,f, Final remapping index is predicted by bump width (e; P = 0.02, Pearson’s correlation) but not bump amplitude (f; P = 0.79, Pearson’s correlation) during the preceding normal gain block (n = 14 flies) (Extended Data Fig. 6). g, Model showing the effect of varying the amplitude of visual responses in ER neurons. Lower ER amplitude produces a faster recovery of HD encoding accuracy after the onset of inverted gain. Note that lower ER amplitude produces lower overall HD encoding accuracy; to facilitate a comparison between conditions, we show the normalized HD encoding accuracy (normalized to 1 at end of normal gain epoch) beginning halfway through the normal gain epoch (arrow). Results are shown as the mean ± s.e.m. across 100 simulation runs. Panel h shows weight matrices for the indicated time points (i, ii and iii). h, Model showing ER → EPG weights at the end of the normal gain block (i), soon after the onset of inverted gain (ii) and when remapping is nearly complete (iii). Results are shown for two ER amplitudes (1.327 and 2.123). At the intermediate time point (ii), remapping has progressed further with the lower ER amplitude.
Fig. 7 Bump width in normal gain predicts cue remapping in inverted gain. Panels b,e,f show data from 14 flies, while c,d show data from 12 flies because two flies did not receive a ‘darkness’ trial. a, Offset distributions for the two example flies from Fig. 6. The remapping index is positive for fly 1 because the offset with respect to HD is more stable. The remapping index is negative for fly 2 because the offset with respect to the visual cue is more stable. b, Remapping index in the initial and final part of the inverted gain block. Across all flies, the remapping index is larger in the final part (P value from two-sided Wilcoxon test). Black and gray lines denote individuals with final remapping indices above or below the mean (horizontal bar); this is preserved in c–f. c,d, HD encoding accuracy (c) and consistency of behavioral orientation (d) in darkness versus inverted gain (P values from linear mixed-effects models with Tukey comparisons and Bonferroni corrections). e,f, Final remapping index is predicted by bump width (e; P = 0.02, Pearson’s correlation) but not bump amplitude (f; P = 0.79, Pearson’s correlation) during the preceding normal gain block (n = 14 flies) (Extended Data Fig. 6). g, Model showing the effect of varying the amplitude of visual responses in ER neurons. Lower ER amplitude produces a faster recovery of HD encoding accuracy after the onset of inverted gain. Note that lower ER amplitude produces lower overall HD encoding accuracy; to facilitate a comparison between conditions, we show the normalized HD encoding accuracy (normalized to 1 at end of normal gain epoch) beginning halfway through the normal gain epoch (arrow). Results are shown as the mean ± s.e.m. across 100 simulation runs. Panel h shows weight matrices for the indicated time points (i, ii and iii). h, Model showing ER → EPG weights at the end of the normal gain block (i), soon after the onset of inverted gain (ii) and when remapping is nearly complete (iii). Results are shown for two ER amplitudes (1.327 and 2.123). At the intermediate time point (ii), remapping has progressed further with the lower ER amplitude.CC BY · PubMed Central
Extended Data Fig. 1 Fitting the bump of activity in the HD system. a. Schematic illustrating procedure for dividing the PB into spatial bins. First, the maximum z-projection was used to draw a mask around the protocerebral bridge (gray). Then, the drawpolyline function in Matlab was used to trace a line (black) composed of many short segments (25-40 segments) running along the midline of the PB. Each segment defines a spatial bin (that is,a region of interest or ROI). Next, a custom algorithm was used to find the line bisecting each segment perpendicularly (a normal line), and then to find the shortest distance between each pixel in our PB mask and one of the normal lines. The signal from each pixel is then assigned to the ROI corresponding to that normal line, with the entire array of ROIs representing an angular distance of 675°, that is one complete HD representation (360°) plus 7/8 of a second representation (315°)41. Note that previous studies have divided the PB into 16 spatial bins37,38, but here we use many more spatial bins (up to 40). This is justified because the fluorescence emission from each EPG neuron will tend to “spill out” from its home glomerulus, due to (1) the scattering of fluorescence emission into adjacent glomeruli, and (2) the presence of stray EPG axonal branches in adjacent glomeruli. As a result of this spillover, fluorescence signals are not perfectly compartmentalized within glomeruli. Thus, dividing the protocerebral bridge into more spatial bins should improve our ability to resolve the amplitude and shape of the bump of activity. Indeed, we found that this approach produced a more reliable measure of bump width, as indicated by higher variance explained (adjusted R2) when fitting linear models to bump width data (as in Figs. 3, 4, 7 and Extended Data Figs. 2, 5, 7). b. ΔF/F and fitted function for two example time points. Shown are a case with very good goodness of fit (top), and a case with a borderline fit quality. In the latter case, adjusted R2 = 0.5, which is our threshold for discarding data points from the analysis. c. Fit from the top example in panel b, showing the two parameters that we extract from the fit.
Extended Data Fig. 1 Fitting the bump of activity in the HD system. a. Schematic illustrating procedure for dividing the PB into spatial bins. First, the maximum z-projection was used to draw a mask around the protocerebral bridge (gray). Then, the drawpolyline function in Matlab was used to trace a line (black) composed of many short segments (25-40 segments) running along the midline of the PB. Each segment defines a spatial bin (that is,a region of interest or ROI). Next, a custom algorithm was used to find the line bisecting each segment perpendicularly (a normal line), and then to find the shortest distance between each pixel in our PB mask and one of the normal lines. The signal from each pixel is then assigned to the ROI corresponding to that normal line, with the entire array of ROIs representing an angular distance of 675°, that is one complete HD representation (360°) plus 7/8 of a second representation (315°)41. Note that previous studies have divided the PB into 16 spatial bins37,38, but here we use many more spatial bins (up to 40). This is justified because the fluorescence emission from each EPG neuron will tend to “spill out” from its home glomerulus, due to (1) the scattering of fluorescence emission into adjacent glomeruli, and (2) the presence of stray EPG axonal branches in adjacent glomeruli. As a result of this spillover, fluorescence signals are not perfectly compartmentalized within glomeruli. Thus, dividing the protocerebral bridge into more spatial bins should improve our ability to resolve the amplitude and shape of the bump of activity. Indeed, we found that this approach produced a more reliable measure of bump width, as indicated by higher variance explained (adjusted R2) when fitting linear models to bump width data (as in Figs. 3, 4, 7 and Extended Data Figs. 2, 5, 7). b. ΔF/F and fitted function for two example time points. Shown are a case with very good goodness of fit (top), and a case with a borderline fit quality. In the latter case, adjusted R2 = 0.5, which is our threshold for discarding data points from the analysis. c. Fit from the top example in panel b, showing the two parameters that we extract from the fit.CC BY · PubMed Central
Extended Data Fig. 2 Individual variation in responses to cue brightness. This figure provides additional detail related to the 15 flies in Fig. 3. a. Variations in HD encoding accuracy correlate with bump width, pooling data across all three cue conditions and all 15 flies (p < 0.001, Pearson’s correlation). Gray points (dim cue) are reproduced from Fig. 3c. Black and blue points show data from the other two conditions (no cue and bright cue). b. There is a trend for variations in HD encoding accuracy to be correlated with bump amplitude, but the relationship falls just short of significance (p = 0.06, Pearson’s correlation). c. Mean head direction for each individual for all three cue conditions. The length of each line represents vector strength (that is, “consistency of behavioral orientation”, or 1-circular variance). Note that consistency of behavioral orientation increases with increasing cue brightness. Different flies choose different goal directions, which is typical of navigation driven by the HD system32,33,42.
Extended Data Fig. 2 Individual variation in responses to cue brightness. This figure provides additional detail related to the 15 flies in Fig. 3. a. Variations in HD encoding accuracy correlate with bump width, pooling data across all three cue conditions and all 15 flies (p < 0.001, Pearson’s correlation). Gray points (dim cue) are reproduced from Fig. 3c. Black and blue points show data from the other two conditions (no cue and bright cue). b. There is a trend for variations in HD encoding accuracy to be correlated with bump amplitude, but the relationship falls just short of significance (p = 0.06, Pearson’s correlation). c. Mean head direction for each individual for all three cue conditions. The length of each line represents vector strength (that is, “consistency of behavioral orientation”, or 1-circular variance). Note that consistency of behavioral orientation increases with increasing cue brightness. Different flies choose different goal directions, which is typical of navigation driven by the HD system32,33,42.CC BY · PubMed Central
Extended Data Fig. 3 Effects of cue shifts. This figure shows additional details regarding the 13 flies in Fig. 4. a. Responses to cue shifts. In both these examples, the visual cue shifted while the wind did not shift. In Example 1, the EPG bump follows the shift in the visual cue, so that the offset relative to that cue is restored. In Example 2, when the visual cue shifts, the bump does not shift, and so it retains its offset relative to the wind. b. Mean bump amplitude and mean bump width (± s.e.m.) around the time of cue shifts. Cue shifts produce no significant change in bump amplitude or width (p = 0.10 for bump amplitude, p = 0.26 for bump width, 2-sided Wilcoxon signed rank tests, comparing 1 sec before vs 1 sec after a shift). c. Bump preference index for each cue shift, for each fly, with flies sorted by mean bump preference index. Each thin vertical line connects the data for 4 wind shifts (orange) and 4 visual shifts (blue) for the same fly. Thick horizontal lines show mean values for each fly. This plot reproduces data from Fig. 4e. d. Same but for the “stickiness index”. This index is a metric for how much the bump tends to follow the cue that does not shift (that is, how sticky’ the bump is). A value of +1 means that the bump sticks with the non-shifted cue, whereas a value of -1 means that the bump follows the shifted cue. There were no flies where the bump tended to prefer either the shifted cue or the non-shifted cue. Flies are sorted as in (c). e. Same but for the “behavioral preference index”. Here, -1 means the visual cue dominates the fly’s behavioral orientation, whereas +1 means the wind dominates. For example, if the wind shifts by -120° and the fly then executes a -120° turn, the index would be +1 for that trial. The mean behavioral preference index was near zero for all individuals, meaning they reoriented with equal frequency in response to visual cue shifts and wind shifts. This was true even in the flies with a strongly biased HD system (for example, Fly 1 or Fly 13).
Extended Data Fig. 3 Effects of cue shifts. This figure shows additional details regarding the 13 flies in Fig. 4. a. Responses to cue shifts. In both these examples, the visual cue shifted while the wind did not shift. In Example 1, the EPG bump follows the shift in the visual cue, so that the offset relative to that cue is restored. In Example 2, when the visual cue shifts, the bump does not shift, and so it retains its offset relative to the wind. b. Mean bump amplitude and mean bump width (± s.e.m.) around the time of cue shifts. Cue shifts produce no significant change in bump amplitude or width (p = 0.10 for bump amplitude, p = 0.26 for bump width, 2-sided Wilcoxon signed rank tests, comparing 1 sec before vs 1 sec after a shift). c. Bump preference index for each cue shift, for each fly, with flies sorted by mean bump preference index. Each thin vertical line connects the data for 4 wind shifts (orange) and 4 visual shifts (blue) for the same fly. Thick horizontal lines show mean values for each fly. This plot reproduces data from Fig. 4e. d. Same but for the “stickiness index”. This index is a metric for how much the bump tends to follow the cue that does not shift (that is, how sticky’ the bump is). A value of +1 means that the bump sticks with the non-shifted cue, whereas a value of -1 means that the bump follows the shifted cue. There were no flies where the bump tended to prefer either the shifted cue or the non-shifted cue. Flies are sorted as in (c). e. Same but for the “behavioral preference index”. Here, -1 means the visual cue dominates the fly’s behavioral orientation, whereas +1 means the wind dominates. For example, if the wind shifts by -120° and the fly then executes a -120° turn, the index would be +1 for that trial. The mean behavioral preference index was near zero for all individuals, meaning they reoriented with equal frequency in response to visual cue shifts and wind shifts. This was true even in the flies with a strongly biased HD system (for example, Fly 1 or Fly 13).CC BY · PubMed Central
Extended Data Fig. 4 Comparing one-cue environments and a two-cue environment. This figure shows additional detail for the 18 flies in Fig. 5. a–c. There is no difference between the visual cue and the wind in terms of HD encoding accuracy (p = 0.28; a), bump width (p = 0.27; b), or bump amplitude (p = 0.20, paired 2-sided Wilcoxon tests; c). These plots show data from the two initial single-cue environments, prior to the onset of the two-cue environment. Thick horizontal lines are mean values. d. The offset during the two-cue block is equally similar to the initial offset with respect to the visual cue and the initial offset with respect to the wind (p = 0.52, paired Wilcoxon test). e. Offset data for both initial single-cue environments and the two-cue environment, for all 18 flies. The angle of each line indicates the mean offset during that block; the length of each line indicates the vector strength of the offset (that is, HD encoding accuracy). Note that there is no relationship between the initial visual offset and the initial wind offset. Note also that there is no tendency for the visual cue or the wind to dominate the offset in the two-cue period. f. Same but now color-coded by cue order. Note that the offset in the two-cue environment is generally similar to the offset in the environment with cue 2. The exceptions to this rule are generally flies where the offset was not stable (for example, the fourth fly in the top row, or the fifth fly in the bottom row).
Extended Data Fig. 4 Comparing one-cue environments and a two-cue environment. This figure shows additional detail for the 18 flies in Fig. 5. a–c. There is no difference between the visual cue and the wind in terms of HD encoding accuracy (p = 0.28; a), bump width (p = 0.27; b), or bump amplitude (p = 0.20, paired 2-sided Wilcoxon tests; c). These plots show data from the two initial single-cue environments, prior to the onset of the two-cue environment. Thick horizontal lines are mean values. d. The offset during the two-cue block is equally similar to the initial offset with respect to the visual cue and the initial offset with respect to the wind (p = 0.52, paired Wilcoxon test). e. Offset data for both initial single-cue environments and the two-cue environment, for all 18 flies. The angle of each line indicates the mean offset during that block; the length of each line indicates the vector strength of the offset (that is, HD encoding accuracy). Note that there is no relationship between the initial visual offset and the initial wind offset. Note also that there is no tendency for the visual cue or the wind to dominate the offset in the two-cue period. f. Same but now color-coded by cue order. Note that the offset in the two-cue environment is generally similar to the offset in the environment with cue 2. The exceptions to this rule are generally flies where the offset was not stable (for example, the fourth fly in the top row, or the fifth fly in the bottom row).CC BY · PubMed Central
Extended Data Fig. 5 Acquiring familiarity with a virtual environment. This figure shows data from 28 flies walking in the environment with the bright visual cue, beginning when they are first exposed to this environment. In a-c, The first two plots show data from two example flies and the third plot shows data for all flies (mean ± s.e.m). Each variable was smoothed with a 60-s moving window, after first excluding time points where the fly was immobile or goodness of fit was below threshold. a. Mean HD encoding accuracy increases over time. Over all flies, there is a significant effect of time in a linear mixed-effects model (p < 0.001). b. Mean consistency of behavioral orientation increases over time. Over all flies, there is a significant effect of time in a linear mixed-effects model (p < 0.001). c. Mean bump width decreases over time. Over all flies, there is a significant effect of time in a linear mixed-effects model (p < 0.001). d. Mean bump amplitude increases during the first 10 min. in the environment, and then decreases. Over all flies, there is a significant effect of time in a linear mixed-effects model (p = 0.03). e. Mean rotational speed and forward velocity for the same flies over this time period (± s.e.m).
Extended Data Fig. 5 Acquiring familiarity with a virtual environment. This figure shows data from 28 flies walking in the environment with the bright visual cue, beginning when they are first exposed to this environment. In a-c, The first two plots show data from two example flies and the third plot shows data for all flies (mean ± s.e.m). Each variable was smoothed with a 60-s moving window, after first excluding time points where the fly was immobile or goodness of fit was below threshold. a. Mean HD encoding accuracy increases over time. Over all flies, there is a significant effect of time in a linear mixed-effects model (p < 0.001). b. Mean consistency of behavioral orientation increases over time. Over all flies, there is a significant effect of time in a linear mixed-effects model (p < 0.001). c. Mean bump width decreases over time. Over all flies, there is a significant effect of time in a linear mixed-effects model (p < 0.001). d. Mean bump amplitude increases during the first 10 min. in the environment, and then decreases. Over all flies, there is a significant effect of time in a linear mixed-effects model (p = 0.03). e. Mean rotational speed and forward velocity for the same flies over this time period (± s.e.m).CC BY · PubMed Central
Extended Data Fig. 6 Learning in inverted gain. This figure shows additional details regarding the 14 flies in Fig. 7. a. In the initial normal-gain segment, bump width is significantly negatively correlated with HD encoding accuracy (p < 0.001, Pearson’s correlation). There is a trend for bump amplitude to be positively correlated with HD encoding accuracy, but this relationship falls short of statistical significance in this group of flies (p = 0.08, Pearson’s correlation). b. There is a trend for the remapping index (in the final part of the inverted gain block) to be negatively correlated with HD encoding accuracy (in the preceding normal gain block), but this relationship falls short of statistical significance (p = 0.06). c. We finished each experiment by returning to normal gain for 200 s. This panel shows the entire experiment (including this last normal gain block) for one example fly. This is the same example fly shown in Fig. 7b. Note that the EPG bump tracks HD accurately during the final normal gain block. d. On average, HD encoding accuracy was no different in the first and last normal gain block (p = 0.67, paired 2-sided Wilcoxon test).
Extended Data Fig. 6 Learning in inverted gain. This figure shows additional details regarding the 14 flies in Fig. 7. a. In the initial normal-gain segment, bump width is significantly negatively correlated with HD encoding accuracy (p < 0.001, Pearson’s correlation). There is a trend for bump amplitude to be positively correlated with HD encoding accuracy, but this relationship falls short of statistical significance in this group of flies (p = 0.08, Pearson’s correlation). b. There is a trend for the remapping index (in the final part of the inverted gain block) to be negatively correlated with HD encoding accuracy (in the preceding normal gain block), but this relationship falls short of statistical significance (p = 0.06). c. We finished each experiment by returning to normal gain for 200 s. This panel shows the entire experiment (including this last normal gain block) for one example fly. This is the same example fly shown in Fig. 7b. Note that the EPG bump tracks HD accurately during the final normal gain block. d. On average, HD encoding accuracy was no different in the first and last normal gain block (p = 0.67, paired 2-sided Wilcoxon test).CC BY · PubMed Central
Extended Data Fig. 7 Correlations between bump profile and rotational speed. Panels a-d show data from 28 flies walking in a virtual reality environment with a bright visual cue. Each variable was smoothed with a 10-s moving window, after first excluding time points where the fly was immobile or goodness of fit was below threshold. a. Example fly. Here and elsewhere in this figure, we used a 10-s rolling window to compute the rotational speed, bump width and bump amplitude values. b. Mean HD encoding accuracy (± s.e.m.) versus the fly’s rotational speed. As the fly’s rotational speed increases, HD encoding accuracy decreases (linear mixed-effects model, p < 0.001). c. Mean z-scored bump width (± s.e.m.) versus the fly’s rotational speed. As the fly’s rotational speed increases, so does bump width (linear mixed-effects model, p < 0.001). d. Mean z-scored bump amplitude (± s.e.m.) versus the fly’s rotational speed. As the fly’s rotational speed increases, so does the bump amplitude (linear mixed-effects model, p < 0.001). Panels e-g show data from 7 flies tested in separate experiments where we displayed a bright visual cue rotating around the fly. The cue rotated at a different speed for each block of the experiment. e. Mean cue position encoding accuracy (± s.e.m.) versus cue rotational speed. Cue position encoding accuracy is computed in the same way that we compute HD encoding accuracy, except that it measures encoding of cue position rather than HD. For example, a value of 1 would mean that the EPG bump perfectly tracks the cue, whereas a value of 0 would mean that the bump moves independently from the cue. Cue position encoding accuracy decreases as cue speed increases (linear mixed-effects model, p < 0.001). f. Mean bump width (± s.e.m.) versus cue rotational speed. Bump width increases as cue speed increases (linear mixed-effects model, p = 0.007). g. Mean bump amplitude (± s.e.m.) versus cue rotational speed. Bump amplitude decreases as cue speed increases (linear mixed-effects model, p < 0.001).
Extended Data Fig. 7 Correlations between bump profile and rotational speed. Panels a-d show data from 28 flies walking in a virtual reality environment with a bright visual cue. Each variable was smoothed with a 10-s moving window, after first excluding time points where the fly was immobile or goodness of fit was below threshold. a. Example fly. Here and elsewhere in this figure, we used a 10-s rolling window to compute the rotational speed, bump width and bump amplitude values. b. Mean HD encoding accuracy (± s.e.m.) versus the fly’s rotational speed. As the fly’s rotational speed increases, HD encoding accuracy decreases (linear mixed-effects model, p < 0.001). c. Mean z-scored bump width (± s.e.m.) versus the fly’s rotational speed. As the fly’s rotational speed increases, so does bump width (linear mixed-effects model, p < 0.001). d. Mean z-scored bump amplitude (± s.e.m.) versus the fly’s rotational speed. As the fly’s rotational speed increases, so does the bump amplitude (linear mixed-effects model, p < 0.001). Panels e-g show data from 7 flies tested in separate experiments where we displayed a bright visual cue rotating around the fly. The cue rotated at a different speed for each block of the experiment. e. Mean cue position encoding accuracy (± s.e.m.) versus cue rotational speed. Cue position encoding accuracy is computed in the same way that we compute HD encoding accuracy, except that it measures encoding of cue position rather than HD. For example, a value of 1 would mean that the EPG bump perfectly tracks the cue, whereas a value of 0 would mean that the bump moves independently from the cue. Cue position encoding accuracy decreases as cue speed increases (linear mixed-effects model, p < 0.001). f. Mean bump width (± s.e.m.) versus cue rotational speed. Bump width increases as cue speed increases (linear mixed-effects model, p = 0.007). g. Mean bump amplitude (± s.e.m.) versus cue rotational speed. Bump amplitude decreases as cue speed increases (linear mixed-effects model, p < 0.001).CC BY · PubMed Central
Extended Data Fig. 8 Bump profile around the time of a cue change. a. Bump width over time (mean ± s.e.m.), around transitions between darkness and the bright visual cue. This plot shows data from the 15 flies in Fig. 2. Note that bump width decreases abruptly at the time of cue onset, and it increases abruptly at the time of cue offset. b. Same but for bump amplitude.
Extended Data Fig. 8 Bump profile around the time of a cue change. a. Bump width over time (mean ± s.e.m.), around transitions between darkness and the bright visual cue. This plot shows data from the 15 flies in Fig. 2. Note that bump width decreases abruptly at the time of cue onset, and it increases abruptly at the time of cue offset. b. Same but for bump amplitude.CC BY · PubMed Central

Verification

Claims the paper makes

13 Valid · 2 Not valid · 6 Not checked · 104 Own results and interpretations

68 more claims

Create a free account to see all 125 claims and to verify the 2 references still open.

See all claims — free account
  • It may state two separate claims as one. 38
  • Probably not a substantial claim. 20

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 70 references

Research works printed but not linked: 1

  1. Touretzky, D. S. in Head Direction Cells and the Neural Mechanisms of Spatial Orientation (eds Wiener, S. I. & Taube, J. S.) Ch. 18 (MIT Press, 2005).

Cited by

See all 13

Metadata

Authors
Basnak, M. A., Kutschireiter, A., Okubo, T. S., Chen, A., Gorelik, P., Drugowitsch, J., & Wilson, R. I.
Affiliations
Harvard University (US)
Corresponding author
Wilson, Rachel I.
Venue
Nature Neuroscience
Published
2025-07-23
Volume
28
Issue
8
Pages
1729-1740
ISSN
  • 1097-6256 (ISSN-L)
  • 1546-1726
OpenAlex
W4412589641
Cited by
13
References
70
Citations by year
2026: 11 · 2025: 2
FWCI
7.36
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Neuroscience (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
Funders
  • Foundation for the National Institutes of Health (U19NS104655, P40OD018537, U19NS118246)
  • Howard Hughes Medical Institute (n/a)
  • Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung (P2ZHP2_184213, P400PB_199242)
  • Warren Alpert Foundation (n/a)
  • National Science Foundation (DGE2140743)
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Memory and Neural Mechanisms
  • Life Sciences › Neuroscience › Cellular and Molecular Neuroscience › Neurobiology and Insect Physiology Research
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Spatial Cognition and Navigation
Keywords
head direction system, ring attractor, topographic map, multimodal cue integration, cue conflict, attractor model, spatial navigation, Drosophila, virtual reality navigation, spatial representation learning
MeSH terms
Virtual Reality, Animals, Brain, Cues, Drosophila, Head, Learning, Neurons, Space Perception, Spatial Navigation

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Multimodal cue integration and learning in a neural representation of head direction. https://app.cochat.ai/papers/10.1038/s41593-024-01823-z
The paper
Basnak, M. A., Kutschireiter, A., Okubo, T. S., Chen, A., Gorelik, P., Drugowitsch, J., & Wilson, R. I. (2025). Multimodal cue integration and learning in a neural representation of head direction. Nature Neuroscience. https://doi.org/10.1038/s41593-024-01823-z

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.