Research organism:D. melanogaster

Establishing a framework for large-scale analysis of descending neuron activation phenotypes

Mapping fly behavior using postural dynamics requires high temporal and spatial resolution video data from a large number of animals. Accordingly, we built a red light activation apparatus with an array of 12 USB cameras that allowed us to film 12 flies in separate chambers simultaneously at high resolution (Figure 1A). We crossed each split-GAL4 line to a UAS-CsChrimson line and filmed six experimental progeny that had been fed retinal, a co-factor necessary for neuronal activation via channelrhodopsin, and six genetically identical control flies whose food had not been supplemented with retinal. The flies were backlit using custom light tables, each consisting of an array of infrared and red LEDs covered by a diffuser. Each chamber was a 3 cm ‘fly bubble’ coated with silicone to encourage flies to remain on the flat floor of the chamber, which was in the focal plane of the camera (see Materials and methods). Each recording consisted of 30 trials of a 15 s pulse of red light followed by a 45 s recovery interval (Figure 1A).

If the descending neuron(s) labeled by a particular split-GAL4 line are involved in triggering, maintaining, or modulating a particular behavior, then activating these neurons withCsChrimson may be sufficient to activate that behavior. To identify behavioral phenotypes in an unbiased manner, we utilized the behavior mapping methods described inBerman et al. (2014). First, we generated a comprehensive ‘behavior space’ of stereotyped actions that single flies could produce in our assay. We collected a dataset of approximately 700 million images, which included behaviors recorded from activation of descending interneuron split-GAL4 lines, previously characterized sparse GAL4 drivers (fruitless-GAL4 andpIP10) that trigger courtship-related behaviors (Stockinger et al., 2005;von Philipsborn et al., 2011), and interneuron drivers targeting the flight neuropil. The additional lines that are not part of the descending neuron screen were included to sample fly behaviors as widely as possible, allowing for higher resolution mapping within the space of behaviors. We computed the behavior space by (1) aligning video images (Figure 1A), (2) decomposing the pixel value dynamics (which correspond to the fly’s posture changes) into a low-dimensional basis set using principal component analysis (Figure 1—figure supplement 1), (3) projecting the original pixel values onto this basis set and transforming those values using a spectral wavelet function to produce a time series that was (4) embedded into a two-dimensional ‘behavior space’ (Figure 1B) usingt-distributed Stochastic Neighbor Embedding (t-SNE) (van der Maaten and Hinton, 2008).

Each position in the behavior space corresponds to a unique set of postural dynamics. Nearby points represent similar motions, that is those involving related body parts executing similar temporal patterns. By observing the video data underlying sub-regions of the behavior space (Figure 1—figure supplement 2 andVideos 1–5), we generated a human-curated version of the behavior space to aid interpretation (Figure 1C). In this behavior space, anterior directed movements such as eye/antennal grooming and proboscis extension are located at the top (Video 1). Anterior-directed foreleg movements are on the upper left side (Video 2). Extremely slow or still postures are on the upper right side (Video 3), and complex wing and abdomen movements such as body and abdomen grooming, abdomen bending and wing extension are in the center (Video 4). Locomotion, ranging from slow (left) to fast (right) is at the bottom (Video 5).

Red peaks, or density maxima, represent the fly behaviors observed most frequently in our data set. These tend to be repetitive, stereotyped behaviors, such as walking or grooming, that our analysis methodology is most sensitive at detecting. By definition, we could not detect behaviors occuring over time-scales faster than 50 Hz, the Nyquist frequency of our system. Approximately 93% of all video image data points could be embedded in this space, including approximately equal fractions of frames when the red light was on or off (Figure 1—figure supplement 3), indicating that the majority of red light activated behaviors are well represented in the behavior space. Imaging errors, such as the fly wandering partially out of frame, are randomly distributed within the dataset.

Entropy of behavior space density provides a quantitative and sensitive measure of optogenetic activation phenotypes

Having established a behavior space representing the full repertoire of fly behaviors that could be captured with our apparatus, we next examined which parts of this space were occupied when individual or subsets of descending neurons were optogenetically activated by CsChrimson. We focused on 130 split-GAL4 lines that targeted descending neurons with little, or no, extraneous expression in other neurons. We first considered the timing and duration of red light triggered behaviors. If descending neuron activation triggered a particular behavior represented in the behavior space, then the density of that line in the behavior space should shift into the region that represents that behavior during periods of red light activation. For example, upon red light activation, retinal-fed flies expressingCsChrimson in a descending neuron line targeting DNg07 and DNg08 (SS02635) groomed their heads (Video 6). For these lines, we identified regions in the behavior space that displayed a statistically significant shift in density for experimental flies during the first three seconds of red light compared to a window at the end of the recovery period when the red light was off (Figure 2B,C, Figure 2—figure supplement 1). This same region in the behavior space did not undergo a significant shift in the control flies (Wilcoxon rank-sum test p<0.05 using the Dunn–Šidák correction for multiple hypotheses [Sidak, 1967]). Likewise, when considering densities over the whole behavior space in three second sliding windows, the experimental, but not the control, flies shifted into the head grooming region (arrowheads,Figure 2C).

To detect activation-elicited behaviors performed by the experimental animals and the relative timing of these dynamics compared to the red light activation, we looked for a reduction in the entropy of the behavior space density (Figure 2A). Entropy measures the degree of uncertainty inherent in the distribution of the flies in the behavior space. When the red light was off, flies exhibited a range of different behaviors, and the probability that they performed any one behavior was low. This resulted in a low probability density distributed throughout the behavior space and correspondingly high entropy (Figure 2A). Upon red light activation, the experimental fly line engaged in red light triggered behaviors at the expense of other natural behaviors. This increased the probability that a small region within the behavior space showed a relatively high density, generating a drop in entropy whose timing and duration mirrored that of the red light triggered behaviors (Figure 2A,Video 7). We can therefore use entropy as a proxy for the duration and onset of red light triggered movements in the behavior space without needing to know, a priori, which behaviors are activated (i.e. which part of the behavior space to examine).

The region density and entropy are quantitative measurements sensitive to small changes in behavior map distribution. We therefore used these values to identify subtle phenotypes that could not be easily identified by manual inspection of the movies. For example, the activation of descending neuron DNg25 induced a short-lived rapid running phenotype (Figure 2E and F) that could be identified by a transient drop in entropy in the behavior space (Figure 2D) and a transient increase in density in the fast locomotion region of the space (Figure 2E and F,Video 8).

Comprehensive characterization of descending neuron split-Gal4 line activation phenotypes

We searched for optogenetically induced phenotypes across the entire collection of descending neuron lines by examining the entropy time course of each line (Figure 3). We found that most lines displayed the largest entropy drop immediately after red light activation (Figure 3A). For roughly a third of the lines, this entropy drop persisted throughout the entire red light activation window (Figure 3B). For most of the rest of the lines, however, the entropy drop was transient and diminished after several seconds (Figure 3B). For a minority of lines, the entropy minimum occurred near the middle or end of the activation window (Figure 3B). These dynamics were almost fully absent in the control animals (Figure 3C and D). We reviewed the raw video data for lines displaying entropy minima in the middle or end of the activation window and found that most of these flies performed some action upon red light activation, followed by a pause. This explained why the entropy was lower in the later part of the activation window, because consistent stillness generates a low entropy behavior space (seeFigure 2—source data 1 for a line by line description of phenotypes). We therefore performed our system-wide analysis using the first 3 s of the red light activation period, because this time period captured the majority of CsChrimson activated behaviors.

In our initial analysis, we looked for behaviors produced when our descending neuron lines were activated using a relatively low level of red light, (5 mW/cm²). Under these conditions, 91 of the 130 lines (69%) displayed a statistically significant increase in density of some area of the behavior space. We then re-tested most of the 41 lines that did not produce a significant density increase by driving CsChrimson at higher levels by growing the flies on food containing an increased retinal concentration and exposing flies to higher intensity red light (9 mW/cm²). Under these conditions, 80% of the lines that had previously displayed no phenotype produced a statistically significant increase in density in the behavior space (Figure 2—figure supplement 2).

Pooling the data from the low and high activation protocols, we detected statistically significant increases in the behavior space in 119 of the 130 (90%) descending neuron lines (Figure 4A). In 86 cases, we observed an increased density in only a single statistically significant region in the behavior space. However, some lines generated density increases in multiple non-contiguous regions of the behavior space (Figure 4A, examples shown inFigure 4B–D,Videos 9–11,Figure 2—source data 1). Many of these cases reflect multiple behaviors performed approximately simultaneously by the flies. For example, a line targeting DNp10 induced anterior reaching movements and wing flicking with similar timing (Figure 4D, red and blue regions, respectively).

In other cases, however, multiple activated regions reflect a stereotyped sequence of behaviors. For example, the DNp09 line shown inFigure 4B repeatedly ran and then paused throughout the entire 15 s activation period. The pausing observed in this case has been shown to be a defensive freezing behavior (Zacarias et al., 2018). The increased density in the run region of the behavior space (Figure 4B, red) appeared before the increased density in the paused region (Figure 4B, blue), reflecting the sequential timing of the two behaviors. Running followed by freezing upon DNp09 activation is also reported by Zacarias and colleagues (Zacarias et al., 2018). However, the flies rapidly became asynchronous as they repeated this series of behaviors, so this behavioral series was detected as simultaneous density increases in the running and still regions throughout the red light activation window. A line targeting descending neuron DNb01 displayed a simple behavior series; flies produced an anteriorly directed twitch of the front legs when the red light was turned on (Figure 4C, red region), then froze for the majority of the red light activation period (Figure 4C, black region), and then twitched when the light was turned off (Figure 4C, blue region). Thus, examining the timing of density shifts illuminates the more complicated behavior series produced by red light activation. This level of analysis is provided for all lines inFigure 2—source data 1. To facilitate study of these activation phenotypes by others, we also provide a compilation of videos slowed down 4X showing one second before and after activation for all animals and all trials (available via Dryadhttps://doi.org/10.5061/dryad.fr89c0c).

Behavioral effects of descending neuron activation are often context dependent

Why does activation of some descending neurons result in multiple, distinct behavioral outputs? One possibility is that the behavioral output of some descending neurons depends on the behavioral context of the fly when the descending neuron is activated. To address this possibility, we calculated the mutual information between the density distribution of the experimental flies in the behavior space at 1.5 to 0.5 s before the red light was turned on versus the first second after red light activation. Mutual information is a non-linear measure of the degree of dependence between two variables and is typically measured in units of bits (Cover and Thomas, 2005). The higher the mutual information, the more the first variable, here the behavior of flies immediately prior to red light activation as measured by their distribution in the behavior space at t = −1.5 to −0.5 s, informs the value of the second variable, the region of the behavior space occupied in the first second of red light activation.

We found that all experimental animals displayed non-zero mutual information between the pre- and post-stimulation behaviors (Figure 5A). In addition, for most lines, more information was available in the experimental flies than in the controls (Figure 5B). This means that even in those cases where red light activation produced only one significant region in the behavior space, the fly’s activity prior to red light activation influenced whether or not it performed the behavior. However, lines with multiple red light activated regions in the behavior space were also those with a relatively high level of mutual information (Figure 5A–B). Thus, a given fly’s behavior before red light activation was highly informative of which behavior that fly would perform after red light activation, as indicated by the different significantly activated regions in the behavior space.Figure 5C displays this phenomenon for one of the lines with the highest mutual information, SS02542 (asterisk inFigure 5A and B, also shown inFigure 4C). Here, if the flies were performing an action in the wing/abdomen movement regions of the behavior space prior to stimulation, then they were likely to perform an anterior movement (region 1) immediately following stimulation. Similarly, flies performing anterior grooming were likely to transition to the small anterior twitch region (region 2), and flies that were initially still tended to remain still post-stimulation (region 3).

Individual descending neurons produce mainly stereotyped, modular behaviors

So far, we have analyzed split-GAL4 lines as if they were a proxy for individual descending neurons or anatomical classes of descending neurons. However, these lines vary in both their strength of expression and in the number and identity of additional cells labeled. To estimate phenotypes for individual descending neurons, we averaged the behavior space densities of multiple lines for those cases where we had multiple lines targeting the same descending neuron (Figure 6). Using this method, and combining it with those descending neurons for which we had only a single representative split-GAL4 line, we estimated phenotypes for 47 of the 58 descending neuron cell types. We have also included six lines and line averages that target two different types of descending neurons cleanly, but for which we have no lines that target each type individually. Twenty-six descending neurons drove locomotion phenotypes and ten drove anterior directed foreleg movements. We also identified six new descending neurons that triggered wing and abdomen movements (plus the previously published pIP10 [von Philipsborn et al., 2011]), two that drove anterior grooming, one that drove abdomen stroking, and four that drove still or slow behaviors.

In general, we found that activation of each type of descending neuron drove behaviors that mapped to a relatively small region of the behavior space. For example, some descending neurons drove slow locomotion, whereas others drove fast locomotion. Only a few, such as DNa01, DNa02 and DNp26, seemed to produce a global increase in locomotor activity. Likewise, we found descending neurons that produced different types of grooming, such as head grooming (DNg07 and DNg08, and DNg12) or abdomen grooming (DNp29), different types of anterior reaching movements (DNg10 versus DNg13) and different types of slow movements (e.g. DNd02 versus DNp02).

Fly stocks and fly handling

The descending neuron split-GAL4 driver collection is described inNamiki et al. (2018). Male flies were crossed to virgin females carrying20xUAS-CsChrimson-mVenus (Klapoetke et al., 2014) integrated into theattP18 landing site (Markstein et al., 2008) and transferred to Dickson lab power food (1L water, 10 g agar, 80 g Brewer’s yeast, 20 g yeast extract, 20 g peptone, 30 g sucrose, 60 g dextrose, 0.5 g MgSO_4*6H20, 0.5 g CaCl_2*2H20, 6 mL propionic acid and 7 mL 15% Nipagin). For the initial screen, experimental animals were raised on power food supplemented with 0.2 mM retinal. This concentration was increased to 0.4 mM for animals that were re-assayed at a higher light intensity. All flies (except parental stocks) were handled under 453 nm blue LEDs and reared in dark blue acrylic boxes (acrylic available from McMaster-Carr, # 8505K84) at 22°C on a 12 hr lights on:12 hr lights off day:night cycle. Individual male flies were collected upon eclosion and housed singly in 2 mL wells in a 96-well ‘condo,’ with power food (with or without retinal) deposited in the bottom of each well, which was sealed at the top with an airpore sheet (Qiagen #195761). Flies were imaged at age 7–12 days, within 4 hr of lights on.

Data collection

Single flies were loaded into individual trays made from 4.5 mm clear acrylic topped with a fly ‘bubble’ 3 cm in diameter and 4 mm at its tallest point, which was vacuum molded from clear 0.020’ PETG thermoform plastic (WidgetWorks Unlimited) (Berman et al., 2014;Klibaite et al., 2017). PETG was placed in a frame, heated in a Oster Convection Bake pizza oven set at 350 °F until the plastic started to deform (about 20 s), then placed on a vacuum former (WidgetWorks Unlimited). To further encourage the flies to remain on the two-dimensional acrylic surface, the bubbles were coated with Sigmacote siliconizing reagent 1 day prior to imaging and lightly wiped with ethanol to remove the excess silicone.

For each descending neuron split-GAL4 line, six retinal-fed experimental animals and six non-retinal-fed control animals were imaged simultaneously. For imaging, flies were placed in individual fly bubbles atop custom light tables (three identical light tables, each imaging four flies). These tables consisted of a custom light board topped with a 0.75’ 3D printed white plastic standoff that was lined with infrared reflective tape, and which was capped with a diffuser made from 0.125’ white plexiglass acrylic (available from eplastics.com, # ACRY24470.125PM24 × 48), which had 50% light transmittance. The light board itself consisted of an array of 256 IR LEDs (Osram Opto SFH 4050-Z, 850 nm wavelength) arrayed in a 16 × 16 pattern, spaced 7.14 mm apart, and 64 red LEDs (Philips Lumileds, LXM2-PD01-0050, 627 nm) arranged in an 8 × 8 pattern, spaced 14.28 mm apart. IR and red LED intensity was controlled separately by 0–2.5V control voltages, yielding 0-100mA for the IR LEDS and 0–400 mA for the red LEDs. We set the IR LEDs to 1V, which provided even illumination without overheating the flies. We used 0.2V (4.5 mW/cm²) red light in the initial screen and 1.0V (9 mW/cm²) red light when re-screening a subset of lines. All three light tables were connected to a 68-Pin unshielded I/O connector block (National Instruments, CB-68LP), then to an M Series multifunction DAQ board (National Instruments, NI USB-6281), so that all tables could be run simultaneously from a single computer. Each light table was topped with a 10” square frame constructed from off the shelf parts from Thorlabs, which supported four 1.3 MP grayscale USB cameras (Point Grey FL3-U3-13Y3M-C, one camera per fly) on optical rails, whose X/Y/Z coordinates could be adjusted relative to the fly bubble. Each camera was fitted with an HR F2/35 mm lens from Thorlabs and an 800 nm longpass filter (Thorlabs, FEL0800). Each set of four cameras was connected to a separate Dell Precision T3600 Tower Workstation containing two 100 GB internal solid state drives, so that two cameras wrote to each solid state drive.

Cameras were programmed using NI-MAX and custom software written in Labview (National Instruments). Data acquisition and the LED light tables were controlled by custom software written in Labview. In brief, a master computer ran a single program that (1) turned the red and infrared (Berman et al., 2014) LEDs on for all three light tables, the former running a program of 15 s, then off for 45 s, for 30 cycles; (2) started all 12 cameras recording; (3) recorded the position of the fly’s centroid for each frame for each camera; and (4) grabbed the frame number for each camera over the network every 2–3 frames, and wrote the frame number and red LED status from the light tables to a single text file. All movies were recorded as uncompressed avi files at 100 frames per second. Each camera was set to 1024 × 1024 pixel resolution that encompassed the entire 3 cm arena. However, flies were tracked using a blob detector and only a 150 × 150 pixel box centered on each fly was saved and used for analysis.

To facilitate direct examination of individual activated behaviors, we have provided a subset of the complete dataset, consisting of 1 s before and after stimulation for all flies and all trials, which is deposited at Dryad (doi:10.5061/dryad.fr89c0c).

Behavior space generation

Our approach for generating a behavior space largely follows the methodology originally described inBerman et al. (2014), which describes much of the procedure in additional detail. We first segmented flies using Canny’s method for edge detection (Canny, 1986) and morphological dilation to find the outline of the fly. All pixels within the corresponding closed curve were considered part of the fly. We assumed that all flies had identical morphology but variable sizes. We calculated a rescaling factor for each fly by segmenting 100 randomly selected images from a single fly and finding the pixels belonging to that fly’s body (head, thorax, and abdomen) in each of them, ignoring pixels associated with the wings and legs. Body pixels were assigned via a two-component Gaussian mixture model, and the average value of the number of pixels was chosen as the body area. All frames from a single movie were then uniformly re-scaled to make the number of body pixels in the average image equal to that in a reference image of a fly. We then rotationally aligned segmented, recalled images by finding the maximal angular cross-correlation of the magnitudes of the two-dimensional polar Fourier transforms between the image and a reference image. This reference image was common to all aligned images. Translational registration was then performed by maximizing the spatial cross-correlation.

Postural decomposition was performed as described inBerman et al. (2014). Images were Radon-transformed using a two-degree spacing, and the 9781 Radon-space pixels that contained the most variance were kept for further analysis (>95% of the total variance). We then performed principal components analysis (PCA) on these data, keeping the 50 modes capturing the most variance (>90% of the total variance). We projected the segmented and aligned images onto the found eigenvectors to create a set of time series that were representative of the postural movements of the fly. To obtain dynamic information about these time series, we applied a Morlet continuous wavelet transform to these time series. We transformed each mode separately, using 25 frequency channels that were dyadically spaced between 1and 50 Hz, retaining only the amplitudes of the resulting complex numbers.

Low-dimensional embedding of these wavelet time series using t-Distributed Stochastic Neighbor Embedding (t-SNE) (van der Maaten and Hinton, 2008) largely followed the approach inBerman et al. (2014) as well. A distance metric between points in time was calculated via the the Kullback-Leibler divergence (Cover and Thomas, 2005) between their associated normalized mode-frequency spectra. Because this data set contains several orders of magnitude more data than can be calculated through brute-force minimization of the t-SNE cost function, we used the sub-sampling technique described inBerman et al. (2014) to identify 600 representative data points from each of the recording sessions. From here, points were randomly assigned subsequent groupings such that each of these groups contained 36,000 data points. The same sub-sampling process was performed amongst these data points, but now keeping twice as many data points as in the previous iteration. This process was repeated until a data set of 36,000 points was obtained. We minimized t-SNE for this data set to create a low-dimensional embedding. We used the re-embedding procedure described inBerman et al. (2014) to include data from outside the 36,000-point training set into the embedding, resulting in the overall density seen inFigure 1.

Statistical analysis

Our main goal for the statistical analysis of the behavior space data was to isolate regions of the map that were significantly affected by optogenetic stimulation. Here, we assessed significance by (1) comparing the flies’ behaviors when the LED was on versus when the LED was off, and (2) requiring that the effect of the LED stimulation be larger in the experimental flies than in the control flies. Specifically, we compared the flies’ behaviors during the first three seconds of stimulation (t=0s to t=3s, where the LED turns on at t=0) to their behavior between stimulation (t=30s to t=45s). To statistically assess whether a particular region of the behavior space was significantly affected by the stimulus, we first defined${\rho }_{i,n}^{on}(x,y)$ to be the average behavior space density for fly$i$ during the$n$th cycle at location$(x,y)$ during the first 3 seconds of excitation and${\rho }_{i,n}^{off}(x,y)$ to be the same, but during the 15s window furthest from the stimulation. We then tested whether${\rho }_{i,n}^{on}\left(x,y\right)$ was significantly different from${\rho }_{i,n}^{off}\left(x,y\right)$ through a Wilcoxon rank sum test with Šidák corrections (p<0.05 after corrections) (Šidák 1967). To calculate the number of corrections, we conservatively assumed that the number of measurements was equal to$2^H$, where$H$ was the entropy of the mean density of the behavior space. This is likely an over-estimate of the number of comparisons, but it provides an upper-bound for the number of distinctions that could be made.

To compare the effect of the optogenetic stimulus on the experimental flies to that of the effect on the control flies, we computed the quantity${{\chi }_{i,n}(x,y)={\rho }_{i,n}^{on}\left(x,y\right)-\frac{1}{2}(\rho }_{i,n-1}^{off}\left(x,y\right)+{\rho }_{i,n}^{off}(x,y))$, which was the behavior space density during light stimulation compared to the average of the two preceding time periods with no light stimulation. We thus assessed statistical significance by using a Wilcoxon rank sum test with Šidák corrections (p<0.05 after corrections) to compare${\displaystyle {\left\{{\chi }_{i,n}\left(x,y\right)\right\}}_{i\in experimentalflies}}$ with${\displaystyle {\left\{{\chi }_{i,n}\left(x,y\right)\right\}}_{i\in controlflies}}$. For a point,$(x,y)$, in the behavior space to be considered significantly affected by the stimulus, we required that both of these tests—within experimental flies test and experimental versus control flies test—yielded a significant result.

Behavior activation maps for individual descending interneurons (Figure 6) were calculated by averaging together the maps of significant activations (${\displaystyle E{\left[{\chi }_{i,n}\left(x,y\right)\right]}_{i,n}>0}$) from each of the lines exciting that neuron.

Stimulation-response entropy curves (Figure 2A and D) were generated by first aligning each time point to its associated phase within the 60 second LED on-off cycle. For each phase within the cycle, we found all embedding points from all relevant trials that were detected within$\pm 200$ ms (using periodic boundary conditions). We then generated a histogram of these points, normalized and convolved the resulting values with a symmetric two-dimensional Gaussian of width$\sigma =2$, to generate a probability density function,$p_t(x,y)$. From this, the entropy curve value at phaset was given by$H\left(t\right)=\int dxdy{p}_t\left(x,y\right)\log p_t(x,y)$. We then pooled data from all individuals of a specific type together (i.e. all control flies from a given line or all experimental flies from a given line) to calculate these curves.

Mutual information between pre-stimulus behavior space densities and post-stimulus regions was computed by numerically integrating the integral:

where${\displaystyle p_{pre}\left(\overrightarrow{x}|R_k\right)}$ is the conditional probability of observing the fly’s behavior to be at location$\overrightarrow{x}$ between 1.5 and 0.5 seconds before the stimulus onset and$p\left(R_k\right)$ is the probability that the fly transitions to region$R_k$ following the stimulus onset. Finite data-size corrections were performed by drawing subsets of the data with replacement and extrapolating to an infinite number of trials, and error bars were generated by extrapolating the calculated variance in a similar manner (Bialek, 2012). The region of transition for each trial was assigned by finding the mode of the behavior space distribution during the first second subsequent to the onset of the stimulus. If the location of the mode of the distribution for that trial was within or closer than 5-pixels to the edge of a region, it was assigned to that region, unless another region was closer. Trials not assigned to any of the regions were given a ‘zero’ label, as reflected in the previous equation.

To provide a sense of scale, if there are$N$ significantly activated regions, the maximum possible mutual information one could potentially measure between the prior distribution and the activated region would be${\log }_2\left(N\right)$ bits. Note, however, that we assigned an additional state corresponding to the fly performing a behavior outside of the significantly activated regions subsequent to the light turning on, thus making the maximal possible mutual information${\log }_2(N+1)$. This additional ‘zero’ state is necessary to account for the possibility that the significant regions might be exhibited only in a context-dependent manner, leading to no significant phenotype when the fly is performing some behaviors at the onset of red light stimulation and leading to a phenotype if other actions are being exhibited.

Competing interests

Author contributions

Data availability

Videos including one second before until one second after activation for all flies during all treatments have been uploaded to Dryad (doi:10.5061/dryad.fr89c0c). Videos were slowed down 4X to allow easier examination.

The following dataset was generated:

Cande J, author; Namiki S, author; Qiu J, author; Korff W, author; Card G, author; Shaevitz JW, author; Stern DL, author; Berman GJ, author. Data from: Optogenetic dissection of descending behavioral control in Drosophila. 2018http://dx.doi.org/10.5061/dryad.fr89c0c Available at Dryad Digital Repository under a CC0 Public Domain Dedication.

1. Alivisatos AP, Chun M, Church GM, Greenspan RJ, Roukes ML, Yuste R. The brain activity map project and the challenge of functional connectomics. Neuron. 2012;74:970–974. doi: 10.1016/j.neuron.2012.06.006. [DOI] [PMC free article] [PubMed] [Google Scholar]
2. Bentley D. Control of cricket song patterns by descending interneurons. Journal of Comparative Physiology. 1977;116:19–38. doi: 10.1007/BF00605514. [DOI] [Google Scholar]
3. Berman GJ, Choi DM, Bialek W, Shaevitz JW. Mapping the stereotyped behaviour of freely moving fruit flies. Journal of the Royal Society Interface. 2014;11:20140672. doi: 10.1098/rsif.2014.0672. [DOI] [PMC free article] [PubMed] [Google Scholar]
4. Bialek W. Biophysics: Searching for Principles. Princeton, NJ: Princeton University Press; 2012. [Google Scholar]
5. Bidaye SS, Machacek C, Wu Y, Dickson BJ. Neuronal control of Drosophila walking direction. Science. 2014;344:97–101. doi: 10.1126/science.1249964. [DOI] [PubMed] [Google Scholar]
6. Böhm H, Schildberger K. Brain neurones involved in the control of walking in the cricket Gryllus bimaculatus. The Journal of Experimental Biology. 1992;166:113–130. [Google Scholar]
7. Canny J. A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence. 1986;8:679–698. doi: 10.1109/TPAMI.1986.4767851. [DOI] [PubMed] [Google Scholar]
8. Cover TM. Thomas JA. Elements of Information Theory. 2nd ed. Hoboken, NJ: John Wiley and Sons, Inc; 2005. [DOI] [Google Scholar]
9. Gras H, Kohstall D. Current injection into interneurones of the terminal ganglion modifies turning behaviour of walking crickets. Journal of Comparative Physiology A: Sensory, Neural, and Behavioral Physiology. 1998;182:351–361. doi: 10.1007/s003590050185. [DOI] [Google Scholar]
10. Griss C, Rowell CH. Three descending interneurons reporting deviation from course in the Locust. I. anatomy. Journal of Comparative Physiology A: Sensory, Neural, and Behavioral Physiology. 1986;158:765–774. doi: 10.1007/BF01324820. [DOI] [PubMed] [Google Scholar]
11. Gronenberg W, Strausfeld NJ. Descending neurons supplying the neck and flight motor of Diptera: physiological and anatomical characteristics. The Journal of Comparative Neurology. 1990;302:973–991. doi: 10.1002/cne.903020420. [DOI] [PubMed] [Google Scholar]
12. Hedwig B. On the control of stridulation in the acridid grasshopper Omocestus viridulus L. I. interneurons involved in rhythm generation and bilateral coordination. Journal of Comparative Physiology. A, Sensory, Neural, and Behavioral Physiology. 1992;171:117–128. doi: 10.1007/BF00195967. [DOI] [Google Scholar]
13. Hedwig B. Control of cricket stridulation by a command neuron: efficacy depends on the behavioral state. Journal of Neurophysiology. 2000;83:712–722. doi: 10.1152/jn.2000.83.2.712. [DOI] [PubMed] [Google Scholar]
14. Heinrich R. Impact of descending brain neurons on the control of stridulation, walking, and flight in orthoptera. Microscopy Research and Technique. 2002;56:292–301. doi: 10.1002/jemt.10033. [DOI] [PubMed] [Google Scholar]
15. Hensler K. Neuronal co-processing of course deviation and head movements in locusts. Journal of Comparative Physiology. A, Neuroethology, Sensory, Neural, and Behavioral Physiology. 1992;171:257–271. [Google Scholar]
16. Hsu CT, Bhandawat V. Organization of descending neurons in Drosophila melanogaster. Scientific Reports. 2016;6:20259. doi: 10.1038/srep20259. [DOI] [PMC free article] [PubMed] [Google Scholar]
17. Kanzaki R, Ikeda A, Shibuya T. Morphological and physiological properties of pheromone-triggered flipflopping descending interneurons of the male silkworm moth, Bombyx mori. Journal of Comparative Physiology A. 1994;175:1–14. doi: 10.1007/BF00217431. [DOI] [Google Scholar]
18. Kien J. The initiation and maintenance of walking in the locust: an alternative to the command concept. Proceedings of the Royal Society B: Biological Sciences. 1983;219:137–174. doi: 10.1098/rspb.1983.0065. [DOI] [Google Scholar]
19. Kien J. Neuronal activity during spontaneous walking--I. Starting and stopping. Comparative Biochemistry and Physiology Part A: Physiology. 1990;95:607–621. doi: 10.1016/0300-9629(90)90747-G. [DOI] [PubMed] [Google Scholar]
20. King DG, Wyman RJ. Anatomy of the giant fibre pathway in Drosophila. I. Three thoracic components of the pathway. Journal of Neurocytology. 1980;9:753–770. doi: 10.1007/BF01205017. [DOI] [PubMed] [Google Scholar]
21. Klapoetke NC, Murata Y, Kim SS, Pulver SR, Birdsey-Benson A, Cho YK, Morimoto TK, Chuong AS, Carpenter EJ, Tian Z, Wang J, Xie Y, Yan Z, Zhang Y, Chow BY, Surek B, Melkonian M, Jayaraman V, Constantine-Paton M, Wong GK, Boyden ES. Independent optical excitation of distinct neural populations. Nature Methods. 2014;11:338–346. doi: 10.1038/nmeth.2836. [DOI] [PMC free article] [PubMed] [Google Scholar]
22. Klibaite U, Berman GJ, Cande J, Stern DL, Shaevitz JW. An unsupervised method for quantifying the behavior of paired animals. Physical Biology. 2017;14:15006. doi: 10.1088/1478-3975/aa5c50. [DOI] [PMC free article] [PubMed] [Google Scholar]
23. Larimer JL. The command hypothesis: a new view using an old example. Trends in Neurosciences. 1988;11:506–510. doi: 10.1016/0166-2236(88)90013-6. [DOI] [PubMed] [Google Scholar]
24. Lima SQ, Miesenböck G. Remote control of behavior through genetically targeted photostimulation of neurons. Cell. 2005;121:141–152. doi: 10.1016/j.cell.2005.02.004. [DOI] [PubMed] [Google Scholar]
25. Luan H, Peabody NC, Vinson CR, White BH. Refined spatial manipulation of neuronal function by combinatorial restriction of transgene expression. Neuron. 2006;52:425–436. doi: 10.1016/j.neuron.2006.08.028. [DOI] [PMC free article] [PubMed] [Google Scholar]
26. Marder E. Variability, compensation, and modulation in neurons and circuits. PNAS. 2011;108:15542–15548. doi: 10.1073/pnas.1010674108. [DOI] [PMC free article] [PubMed] [Google Scholar]
27. Markstein M, Pitsouli C, Villalta C, Celniker SE, Perrimon N. Exploiting position effects and the gypsy retrovirus insulator to engineer precisely expressed transgenes. Nature Genetics. 2008;40:476–483. doi: 10.1038/ng.101. [DOI] [PMC free article] [PubMed] [Google Scholar]
28. Milde JJ, Strausfeld NJ. Cluster organization and response characteristics of the giant fiber pathway of the blowfly Calliphora erythrocephala. The Journal of Comparative Neurology. 1990;294:59–75. doi: 10.1002/cne.902940106. [DOI] [PubMed] [Google Scholar]
29. Namiki S, Dickinson MH, Wong AM, Korff W, Card GM. The functional organization of descending sensory-motor pathways in Drosophila. eLife. 2018;7:e34272. doi: 10.7554/eLife.34272. [DOI] [PMC free article] [PubMed] [Google Scholar]
30. Olberg R, Willis M. Pheromone-modulated optomotor response in male gypsy moths, Lymantria dispar L.: directionally selective visual interneurons in the ventral nerve cord. Journal of Comparative Physiology A. 1990;167:707–714. doi: 10.1007/BF00192665. [DOI] [Google Scholar]
31. Olberg RM. Pheromone-triggered flip-flopping interneurons in the ventral nerve cord of the silkworm moth,Bombyx mori. Journal of Comparative Physiology ? A. 1983;152:297–307. doi: 10.1007/BF00606236. [DOI] [Google Scholar]
32. Pfeiffer BD, Ngo TT, Hibbard KL, Murphy C, Jenett A, Truman JW, Rubin GM. Refinement of tools for targeted gene expression in Drosophila. Genetics. 2010;186:735–755. doi: 10.1534/genetics.110.119917. [DOI] [PMC free article] [PubMed] [Google Scholar]
33. Ritzmann RE, Büschges A. Adaptive motor behavior in insects. Current Opinion in Neurobiology. 2007;17:629–636. doi: 10.1016/j.conb.2008.01.001. [DOI] [PubMed] [Google Scholar]
34. Rowell CH, Reichert H. Three descending interneurons reporting deviation from course in the locust. II. Physiology. Journal of Comparative Physiology A Sensory, Neural, and Behavioral Physiology. 1986;158:775–794. doi: 10.1007/BF01324821. [DOI] [PubMed] [Google Scholar]
35. Sidak Z. Rectangular confidence regions for the means of multivariate normal distributions. Journal of the American Statistical Association. 1967;62:626–633. doi: 10.2307/2283989. [DOI] [Google Scholar]
36. Staudacher EM. Sensory responses of descending brain neurons in the walking cricket, Gryllus bimaculatus. Journal of Comparative Physiology A: Sensory, Neural, and Behavioral Physiology. 2001;187:1–17. doi: 10.1007/s003590000171. [DOI] [PubMed] [Google Scholar]
37. Staudacher E, Schildberger K. Gating of sensory responses of descending brain neurones during walking in crickets. The Journal of Experimental Biology. 1998;201:559–572. doi: 10.1242/jeb.201.4.559. [DOI] [PubMed] [Google Scholar]
38. Stockinger P, Kvitsiani D, Rotkopf S, Tirián L, Dickson BJ. Neural circuitry that governs Drosophila male courtship behavior. Cell. 2005;121:795–807. doi: 10.1016/j.cell.2005.04.026. [DOI] [PubMed] [Google Scholar]
39. Strausfeld NJ, Bassemir U, Singh RN, Bacon JP. Organizational principles of outputs from dipteran brains. Journal of Insect Physiology. 1984;30:73–93. doi: 10.1016/0022-1910(84)90109-4. [DOI] [Google Scholar]
40. Strausfeld NJ, Bassemir UK. Lobula plate and ocellar interneurons converge onto a cluster of descending neurons leading to neck and leg motor neuropil in Calliphora erythrocephala. Cell and Tissue Research. 1985;240:617–640. doi: 10.1007/BF00216351. [DOI] [Google Scholar]
41. Strausfeld NJ, Gronenberg W. Descending neurons supplying the neck and flight motor of Diptera: organization and neuroanatomical relationships with visual pathways. The Journal of Comparative Neurology. 1990;302:954–972. doi: 10.1002/cne.903020419. [DOI] [PubMed] [Google Scholar]
42. van der Maaten L, Hinton G. Visualizing data using t-SNE. Journal of Machine Learning Research : JMLR. 2008;9:2579–2605. [Google Scholar]
43. von Philipsborn AC, Liu T, Yu JY, Masser C, Bidaye SS, Dickson BJ. Neuronal control of Drosophila courtship song. Neuron. 2011;69:509–522. doi: 10.1016/j.neuron.2011.01.011. [DOI] [PubMed] [Google Scholar]
44. von Reyn CR, Breads P, Peek MY, Zheng GZ, Williamson WR, Yee AL, Leonardo A, Card GM. A spike-timing mechanism for action selection. Nature Neuroscience. 2014;17:962–970. doi: 10.1038/nn.3741. [DOI] [PubMed] [Google Scholar]
45. Zacarias R, Namiki S, Card G, Vasconcelos ML, Moita MA. Speed dependent descending control of innate freezing behavior in Drosophila Melanogaster. bioRxiv. 2018:234443. doi: 10.1101/234443. [DOI] [PMC free article] [PubMed] [Google Scholar]
46. Zorović M, Hedwig B. Processing of species-specific auditory patterns in the cricket brain by ascending, local, and descending neurons during standing and walking. Journal of Neurophysiology. 2011;105:2181–2194. doi: 10.1152/jn.00416.2010. [DOI] [PMC free article] [PubMed] [Google Scholar]