Background optic flow modulates responses of multiple descending interneurons to object motion in locusts

optic flow modulates responses of multiple descending interneurons to object 1 motion in locusts 2 3 Short title: Optic flow modulates descending interneuron responses to object motion in locusts 4 5 Sinan Zhang 1*,# and John R. Gray 1 6 7 1 Department of Biology, College of Arts and Science, University of Saskatchewan, Saskatoon, 8 Saskatchewan, Canada 9 # Current address: Department of Medical Genetics, Cumming School of Medicine, University of 10 Calgary, Calgary, Alberta, Canada 11 12 * Corresponding author 13 E-mail: [email protected] (SZ) 14 15 16 17 18 19 20 21 22 23 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 2

24 Animals flying within natural environments are constantly challenged with complex visual 25 information. Therefore, it is necessary to understand the impact of the visual background on the 26 motion detection system. Locusts possess a well-identified looming detection pathway, 27 compromised of the lobula giant movement detector (LGMD) and the descending contralateral 28 movement detector (DCMD). The LGMD/DCMD pathway responds preferably to objects on a 29 collision course, and the response of this pathway is affected by the background complexity. 30 However, multiple other neurons are also responsive to looming stimuli. In this study, we 31 presented looming stimuli against different visual backgrounds to a rigidly-tethered locust, and 32 simultaneously recorded the neural activity with a multichannel electrode. We found that the 33 number of discriminated units that responded to looms was not affected by the visual 34 background. However, the peak times of these units were delayed, and the rise phase was 35 shortened in the presence of a flow field background. Dynamic factor analysis (DFA) revealed 36 that fewer types of common trends were present among the units responding to looming stimuli 37 against the flow field background, and the response begin time was delayed among the common 38 trends as well. These results suggest that background complexity affects the response of multiple 39 motion-sensitive neurons, yet the animal is still capable of responding to potentially hazardous 40 visual stimuli. 41 42

43 motion detection, optic flow, population coding, Locusta migratoria 44 45 46 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 3

47 Successful navigation in the natural environment requires animals to rapidly perceive and extract 48 important sensory cues, usually from a complex, “noisy” background. For example, self-motion 49 of animals generates optic flow, which is defined as the pattern of image shifts caused by the 50 motion of the observer [1]. Although animals can use optic flow to control locomotion speed and 51 correct for deviations in flight orientation [2–5], the presence of optic flow also makes it more 52 difficult to observe other moving objects, such as an approaching predator. Successful detection 53 and discrimination between different types of visual motion is critical to animal’s survival. 54 In mouse retina, four subtypes of direction-sensitive ganglion cell (DSGC) are involved in the 55 encoding of optic flow [6]. Each subtype of DSGC is selectively sensitive to optic flow along a 56 specific translatory orientation, including forward, backward, up, and down. Combination of 57 these subtypes can also encode rotatory degrees of freedom. Flying animals, however, 58 manoeuvre in additional degrees of freedom, including but not limited to vertical translation and 59 roll, neither of which is commonly executed by land animals. Therefore, flying animals need to 60 process more complicated optic flow parameters. With a relatively tractable nervous system and 61 capability to perform complex behaviours, insects are ideal subjects to study the neural coding of 62 complex visual backgrounds. In flies, similar to mice, four types of direction-sensitive neurons 63 are involved in the computation of the optic flow direction [7]. 64 The migratory locust, Locusta migratoria, is another classic subject for studying visual 65 processing. Migratory locusts have two morphological states: the solitary state and the 66 gregarious state. In the gregarious state, locusts can form large swarms, comprised of thousands 67 of individuals, and fly ~3 meters·s-1 [8]. With conspecifics moving around at various velocities 68 from different angles, along with the flow field generated from self-motion, locusts flying in a 69 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 4 swarm can still avoid collision with conspecifics [9] and predatory birds [10]. In locusts, an 70 established neural pathway, including the lobula giant movement detector (LGMD) and its 71 postsynaptic partner, the descending contralateral movement detector (DCMD), has been 72 identified to respond preferably to objects on a collision course [11–14]. 73 The LGMD/DCMD pathway is considered to play an important role in generating collision 74 avoidance behaviours. Each LGMD receives input from an entire compound eye [15]. When 75 presented with an approaching object, the LGMD responds with increased firing rate, which 76 peaks near the projected time of collision (TOC) and then decays [16–21]. Each LGMD spike 77 generates a spike in the DCMD [22], which synapses with the motor center related to avoidance 78 behaviours [23]. Different phases of responses recorded in the LGMD/DCMD pathway have 79 been associated with critical timing of avoidance behaviours, such as jumping [24] and flight 80 steering [25]. 81 Addition of visual background affects the response of the LGMD/DCMD pathway to looming 82 objects. In the LGMD, the number of spikes evoked by small looming objects is inhibited by the 83 presence of large-field optic flow [13,26,27]. In the DCMD, although the general peak response 84 is largely invariant between background types, flow field still caused a lower peak firing rate, 85 delayed peak time, shorter rise phase, and longer fall phase [28]. 86 Previously, we investigated putative population coding of motion sensitive neurons in response 87 to objects moving against a simple white background [29]. To further understand how motion 88 sensitive neural ensembles behave in a more natural environment, we presented a looming 89 stimulus approaching against different backgrounds to rigidly-tether locust and recorded the 90 simultaneous activity of multiple descending interneurons. Overall, most discriminated units 91 responded to looming stimuli regardless of background types, but the distribution of different 92 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 5 types of response was affected. For individual units that peaked near TOC, as well as common 93 trends that peaked near TOC, the peak time and/or the response begin time was delayed by the 94 flow field background. 95 96 97 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 6

98 Animals 99 Twenty-one adult male locusts (Locusta migratoria) were chosen from a colony maintained at 100 the University of Saskatchewan, Saskatoon, Canada, ensuring that they were at least 3 weeks 101 past the imaginal moult. These locusts were fed with a diet consisting of wheat grass and bran 102 and were subjected to a regular light-dark cycle of 12 hours of light and 12 hours of darkness at 103 an approximate temperature of 30 ° C. To maintain the gregarious state, the locust colony was 104 kept in a crowded condition [30]. 105 106 Preparation 107 During the light cycle, the locusts were carefully removed from the colony and transferred to a 108 wire mesh container. To provide both light and warmth, an incandescent lamp was placed on top 109 of the container. The experiments were conducted at a controlled room temperature of 110 approximately 25 ° C. 111 To prepare the locust for the experiments, the legs were first removed, and a rigid tether was then 112 securely attached to the ventral surface of the thorax using melted beeswax. A small portion of 113 the ventral cervical cuticle was removed to expose the paired connectives of the ventral nerve 114 anterior to the prothoracic ganglion. The prepared locust was then transferred to the stimulus 115 arena [31] (Fig 1.A). A silver wire hook electrode was used to lift and stabilize the left ventral 116 nerve cord. Subsequently, a sharp glass electrode was carefully inserted into the protective 117 sheath surrounding the left nerve cord, creating a small opening. A twisted wire tetrode, 118 fabricated following an established protocol [32], was then inserted into the left nerve cord 119 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 7 anterior to the hook electrode (Fig 1.B). Finally, a silver ground wire was inserted into the 120 abdomen. 121 122 Fig 1. Experimental setup and visual stimuli. 123 (A) The stimulus arena with the dome screen closed. (B) The locust preparation with tetrode 124 inserted (inset). (C) Diagram of the experimental setup. Visual stimuli were rendered and 125 projected onto the dome screen while neural activity was recorded simultaneously. A 5-V pulse 126 was used to align the neural recordings with stimulus timing. (D) Visual stimuli used in the 127 experiment. The first three visual stimuli included a 7-cm-diameter black disc approaching at 128 300 cm· s-1 from the bottom right at 45° angle, against three visual backgrounds: white (LW), 129 half-white and half-grey (LG), and flow field (LF). The last visual stimulus was composed of 130 flow field only (F). Each visual stimulus was presented 5 times and the order of all 20 visual 131 stimuli was randomized for each locust. 132 133 Neural responses were observed while waving a hand to verify a high signal-to-noise ratio. A 134 mixture of mineral oil and Vaseline was then applied around the electrodes and nerve cords to 135 insulate the recordings from the hemolymph and prevent desiccation. The entire setup was 136 rotated 180° , placing the locust in a dorsal-side-up position at the center of the rear projection 137 dome screen (Fig 1.C). The longitudinal axis of the locust was aligned parallel to the apex of the 138 dome, with a distance of 12 cm between the locust's eye and the screen. In this orientation, 0° 139 represented the front of the locust, 180° represented the direction directly behind, and +90° 140 represented a perpendicular angle to the right longitudinal axis of the locust's body. Negative 141 perpendicular angles were to the left of the locust. 142 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 8 143 Visual Stimulus 144 Visual stimuli were generated using a Python-based program EVSG [33]and incorporated 145 correctional factors to account for the curvature of the dome screen. The resolution of the stimuli 146 was set at 1024× 768 pixels. Each pixel on the screen corresponded to approximately 0.7 mm, 147 resulting in a subtense angle of approximately 0.4° , which is well below the angular resolution of 148 locusts' eyes (1° ) [34]. 149 Real-time rendering of the images was performed using a GeForce RTX 3080 video card 150 (NVIDIA Corporation, Santa Clara, United States), achieving a frame rate of over 150 frames 151 per second. The images were then projected onto the dome screen using an InFocus DepthQ 152 projector (InFocus, Portland, United States) with the colour wheel removed. The frame rate of 153 the projections exceeded the flicker fusion frequency of locusts' eyes, which is around 66 Hz 154 [35]. 155 A 5V pulse was generated at the projected time of collision (TOC), which was automatically 156 calculated based on the motion trajectory of the visual stimuli. This pulse was used to align 157 different trials. 158 In all visual stimuli, the backgrounds were divided vertically into two equal halves, simulating a 159 horizon line positioned at 0° elevation. The top half was rendered in solid white (R = 0, G = 0, B 160 = 0), while the bottom half was either solid white (LW), solid grey (LG, R = 100, G = 100, B = 161 100), or a flow field (LF), composed of grey concentric circles (R = 100, G = 100, B = 100) 162 expanding outward from the center of the dome below the horizon line. The Michelson contrast 163 ratio between the grey and the white backgrounds was 0.87. (Fig 1.D). Against each background, 164 a 7-cm diameter black disc (R=255, G = 255, B = 255) approached the locust at a velocity of 300 165 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 9 cm·s-1 (looming) from the bottom right at 45° azimuth and -45° elevation.The trajectory of the 166 looming disc was designed so that it remained within the lower half of the dome screen for the 167 majority of its motion. Additionally, a fourth visual stimulus (F) consisted of only the flow field 168 below the horizon line. 169 Each visual stimulus was presented 5 times, resulting in a total of 20 presentations, the order of 170 which was randomized for each locust. To prevent neural habituation, the visual stimuli were 171 presented with a 3-minute interval between each presentation. Additionally, a direct loom was 172 presented before and after the sequence of the 20 presentations to ensure that no attenuation of 173 the neural responses occurred during the length of the experiment. 174 175 Data Acquisition 176 The neural signals from all 5 channels, including 1 channel from the silver hook electrode and 4 177 channels from the twisted wire tetrode, were amplified using a differential amplifier (Model 178 1700, A-M Systems, Sequim, United States) with a high pass filter at 300 Hz, a low pass filter at 179 5 kHz, and a gain of 100× . These amplified neural signals, along with the stimulus pulse, were 180 then digitized using a USB data acquisition board (DT9818-OEM, TechmaTron Instrument Inc., 181 Laval, Canada) and recorded at a sampling rate of 25 kHz per channel using DataView version 182 11 (W.J. Heitler, University of St Andrews, St Andrews, Scotland). 183 184 Spike Sorting 185 Raw recordings were merged chronologically into a single file and then filtered using a Finite 186 Impulse Response (FIR) filter with a bandpass range of 500-2000 Hz and a Blackman window 187 type using DataView. The filtered data from the four channels of the tetrode was exported to 188 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 10 Offline Sorter version 4.6 (Plexon Inc., Dallas, United States) for waveform detection and spike 189 sorting. The waveform detection threshold was set to three times the standard deviation over the 190 mean of each channel, and the detected waveforms were aligned to the largest peak on either 191 side. Spike sorting was performed using a semi-automatic method based on the T-Dist E-M 192 algorithm [36]. The degree-of-freedom (DOF) multiplier was set to 3, and the initial sorting was 193 performed with 26 units in the 3D feature space. Subsequently, manual discrimination was based 194 on the shape and amplitude of the waveforms for final unit sorting. A Multivariate Analysis of 195 Variance (MANOVA) was used to assess the statistical distinctness of the sorted units in the 3D 196 cluster space (Fig 2.B). 197 198 Fig 2. Spike sorting and identification of responsive units. 199 (A) Raw recording from the hook electrode, the stimulus event pulse (stim), and the tetrode are 200 shown on the top. Waveforms were detected using average plus three times the standard 201 deviation as the threshold, and sorted based on the cluster variance in the 3D feature space. Spike 202 sorting results from the same recording time window (tetrode sorted) is shown underneath the 203 raw recordings. Each colour represents a different unit. (B) Overlapped waveforms of three units 204 recorded from all four channels of the tetrode (top), and the distribution of the same units in the 205 3D feature space (bottom). (C) Identification of responsive units. For each discriminated unit, the 206 peristimulus time histogram (PSTH), representing the change of firing rate during the stimulus 207 presentation, is shown on the bottom of each block. The cumulative sum and 99% confidence 208 level (represented by the ellipses) are plotted on the top half of each block. If the cumulative sum 209 (represented by the line inside the ellipse) extended outside the ellipse (example shown in the 210 orange shade), this unit was determined to have responsed to the specific visual stimulus. If the 211 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 11 cumulative sum remained within the ellipse for the entire duration (example shown in the blue 212 shade), this unit was determined to have not nonresponded. Note that not all units generated 213 spikes during this stimulus presentation. Inactive and non-responsive units were removed from 214 subsequent analyses. 215 216 Spike Train Analysis 217 The spike times of discriminated units were exported to NeuroExplorer Version 5 (Plexon, 218 Dallas, United States) for further analysis. Peristimulus time histograms (PSTHs) were generated 219 using a 1-ms bin width and smoothed with a 50-ms Gaussian filter. All trials were aligned to the 220 projected time of collision (TOC). For trials involving a looming disc, a 2.2-second window was 221 used, starting 2 seconds before TOC, and ending 0.2 seconds after TOC. For trials with the flow 222 field only, a 10-second window was employed, spanning 5 seconds before the flow field onset 223 and ending 5 seconds after. 224 To evaluate the responsiveness of individual units, cumulative sum plots were created along with 225 an ellipse representing the 99% confidence level (Fig 2.C). The cumulative sum was calculated 226 in NeuroExplorer using the following algorithm: 227 For bin 1: cs(1) = bc(1) - A; 228 For bin 2: cs(2) = bc(1) + bc(2) - A× 2; 229 For bin 3: cs(3) = bc(1) + bc(2) + bc(3) - A× 3; 230 ... 231 In this algorithm, cs(n) represents the cumulative sum of the n-th bin, bc(n) represents the spike 232 bin count of the n-th bin, and A represents the average of all bin counts. If the cumulative sum 233 remained within the 99% confidence level ellipse, it indicated that the firing rate of the 234 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 12 corresponding unit did not exhibit a significant change during the stimulus presentation, 235 implying that the unit did not respond to the visual stimulus. Conversely, if the cumulative sum 236 touched or extended beyond the ellipse, it was considered responsive to the visual stimuli [37–237 39]. Units that did not exhibit a response were excluded from subsequent analyses. 238 239 Dimensionality Reduction 240 The responses of individual units were first examined to determine how many types of visual 241 stimuli that each unit responded to. Between different animals, the same neuron could have been 242 recorded multiple times. Therefore, we pooled the responsive units across all animals based on 243 the visual stimulus. For each stimulus, the response types were initially categorized by observing 244 the peristimulus time histogram (PSTHs) of individual units. The categorization was adopted 245 from previous studies [37,38]. Firing parameters, including peak firing rate, peak time, rise phase 246 (from the time when the unit firing rate exceeded the upper 95% confidence interval to the peak 247 time), and decay phase (from the peak time to when the firing rate decayed below 15% of the 248 peak firing rate) [20,40], were measured from the PSTHs of those units that displayed a clear 249 peak, and compared between different backgrounds. 250 Dynamic Factor Analysis (DFA) was then used to extract common trends from the responsive 251 units for each visual stimulus [37,41]. In this model, a set of n observed time series (y) (units) is 252 explained by a set of m hidden random walks (x) (common trends) through linear combination, 253 represented by factor loadings (Z) and offsets (a). The model equations can be written as 254 follows: 255 256 yt = Z × xt + a + vt, where vt ~ MVN (0, R) 257 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 13 xt = x(t-1) + wt, where wt ~ MVN (0, Q) 258 259 In the model, the error terms of hidden processes (x) (common trends) and observations (y) 260 (units) were represented by wt and vt, respectively. Both wt and vt follow a multivariate normal 261 distribution (MVN), with a mean of 0 and covariate metrices of R and Q, respectively. To 262 identify the model, certain constraints were applied: the first m elements of a were set to 0, zi,j 263 was set to 0 if j > i for the first m-1 rows of Z, and Q was set to the identity matrix Im. 264 DFA was performed using the MARSS package v3.11.4 [42] in R 4.1.3 (R Core Team, 2022), 265 using the Broyden–Fletcher–Goldfarb–Shanno (BFGS) method. The model quality was 266 evaluated using the Akaike Information Criterion with a correction for small sample size (AICc) 267 [43]. For each visual stimulus, DFA was iteratively performed, starting with one common trend. 268 The number of common trends gradually increased until AICc began to increase, and the model 269 with the lowest AICc was selected as the optimal model. 270 The extracted common trends were further analyzed. The factor loadings (Z) of each common 271 trend were used to rectify any potentially reversed trends and scale the common trends. 272 According to the DFA model, yt = Z × xt + a + vt, for the i-th common trend ((xT)i) and the 273 corresponding factor loadings (Zi), if (xT)i was multiplied by a scale factor, while Zi was divided 274 by the same scale factor, the overall model is not affected. The sign of the scale factor was 275 determined by the average of Zi, while the value was determined by the largest absolute value 276 among all elements within Zi. If the average of Zi was negative, it indicates that the common 277 trend contributes negatively to most unit responses (observations), and therefore reversing the 278 sign would make the common trend consistent with the majority of units. The value was chosen 279 to scale all factor loadings to (-1, 1), and therefore, to make the scale of all common trends 280 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 14 comparable to each other. For common trends showing clear peaks, various parameters, such as 281 peak time, rise phase, and decay phase, were measured from the PSTHs. These parameters were 282 then compared between different backgrounds. 283 284 Statistical Analysis 285 Statistical analyses were executed using SigmaPlot version 12.5 (Systat Software, San Jose, 286 USA) or R 4.1.3 (R Core Team, 2022), and plotted using SigmaPlot. Prior to applying specific 287 tests, datasets were preliminarily assessed for normal distribution (Shapiro-Wilk test) and 288 homoscedasticity. Parametric data was articulated through arithmetic mean and standard 289 deviation, and plotted with a column graph. Nonparametric data was articulated through median 290 and quartiles, and plotted with a box plot. Variables with no inherent relationship were compared 291 through either one-way Analysis of Variance (ANOVA) (for parametric data) or Kruskal-Wallis 292 One Way Analysis of Variance on Ranks (for nonparametric data). Variables with an inherent 293 relationship were compared using either one-way repeated measure (RM) ANOVA (for 294 parametric data) or Friedman Repeated Measures Analysis of Variance on Ranks (for 295 nonparametric data). Two-way comparisons were executed utilizing two-way RM ANOVA and 296 Holm-Sidak post-hoc test. All statistical examinations were two-tailed, and the significance level 297 (α) was determined to be 0.05. 298 299 300 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 15

301 Unit discrimination 302 All twenty trials from each animal were chronologically merged for spike discrimination and 303 unit sorting. Utilizing a positive threshold of three times the standard deviation above the mean, 304 we detected 571,515 spikes across all twenty-one animals (median = 25,637 spikes per locust, 305 1st quartile = 23,692 spikes per locust, 3rd quartile = 29,090.5 spikes per locust). From these 306 spikes we discriminated 352 distinct units (mean = 16.8 units per locust, standard deviation = 3.4 307 units per locust). Multivariate Analysis of Variance (MANOVA) verified the statistical 308 distinctiveness of these units in three-dimensional cluster space (p < 0.001). A comprehensive 309 summary of the discriminated spikes and units for all animals is provided in S1 Table. 310 311 Unit responses 312 S2 Table summarizes the number of units that responded to each stimulus type. Out of the total 313 352 discriminated units, 248 units (70%) exhibited responses to looming stimuli against a white 314

(LW), 238 units (68%) responded to looming stimuli against a half-white-half-grey 315

(LG), 242 units (69%) responded to looming stimuli against a flow field background 316 (LF), 222 units (63%) responded to the flow field only (F), and 28 units (8%) did not respond to 317 any of the visual stimuli. One-way ANOVA revealed that there was no significant difference in 318 the number of units responding to looming stimuli against different backgrounds (F2 = 0.14, p = 319 0.87), nor was there a significant difference in the percentage of responsive units from each 320 animal (One-way ANOVA, F2 = 0.49, p = 0.61) (Fig 3). 321 322 Fig 3. Number of units that responded to each type of visual stimulus. 323 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 16 Within each box, the horizontal line represents the median, while the lower and upper 324 boundaries, the whiskers, and the dots represent the 25/75%, 10/90%, and 5/95% percentiles, 325 respectively. One-way repeated measure (RM) ANOVA revealed that there was no statistical 326 difference between the number of units that responded to different types of visual stimuli. 327 328 However, it is worth noting that the responses of these units were not exclusive to a single 329 background. A unit could respond to looming stimuli against one or more visual backgrounds. 330 Fig 4 illustrates the distribution of the percentage of different types of looming stimuli that the 331 units responded to. Among the total 352 units, 194 units (55%) responded to looming stimuli 332 against all three backgrounds. 42 units (12%) responded to two out of the three looming stimuli, 333 among which 19 units responded to LW & LG, 15 units responded to LW & LF, and 8 units 334 responded to LG & LF. 63 units (15%) responded to only one type of looming stimulus, among 335 which 20 units responded to LW, 18 units responded to LG, and 25 units responded to LF. 336 Additionally, 25 units responded exclusively to the flow field (F) and none of the looming 337 stimuli. 338 Overall, the number of units responding to looming stimuli was relatively similar across all three 339 backgrounds, suggesting that the visual background did not affect the number of units 340 responding within a putative population. 341 342 Fig 4. Percentage of responsive units based on stimulus properties. Each column represents 343 the number of units (n) that responded to 1, 2, 3 looming stimuli, flow field only, or none. Out of 344 all 352 discriminated units, 194 units responded to looming stimuli against all three backgrounds 345 (55%), 42 units responded to looming stimuli against two out of three backgrounds (12%), while 346 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 17 63 units responded to looming against one background (18%). Overall, the background type did 347 not affect how many units responded to the looming stimulus. 348 349 Categorization of unit response 350 The peristimulus time histograms (PSTHs) of individual units were examined visually and 351 categorized based on the trend of firing rate change over time for visual stimuli containing a 352 looming stimulus [38]. Five categories of responses were identified (Fig 5): 1) firing rate peaked 353 near the projected time of collision (TOC); 2) firing rate showed a valley near TOC; 3) firing rate 354 gradually increased over time; 4) firing rate remained relatively constant; 5) firing rate remained 355 constant during the looming and increased near the end of the recording. S3 Table provides a 356 summary of the distribution of response categories for looming against different visual 357 backgrounds. The distribution of response categories was similar between LW and LG (Chi-358 square test, χ2 (4, N = 486) = 1.615, p = 0.806). However, the distribution was significantly 359 affected by the flow field (LW vs LF: Chi-square test, χ2 (4, N = 490) = 12.541, p = 0.014; LG vs 360 LF: Chi-square test, χ2 (4, N = 480) = 19.672, p < 0.001). When presented with LF, only 5 units, 361 compared to 21 for LW and 27 for LG, displayed a gradual increase (category 3). Additionally, 362 more units showed a constant firing rate that increased near the end (category 5). It is likely that 363 some units in category 1 and 3 exhibited a delayed response to LF, leading to their placement in 364 category 5. Fig 5 illustrates the average PSTH and 95% confidence interval for each category in 365 response to LW. 366 367 Fig 5. Categorization of units responding a looming stimulus against white background 368 (LW). 369 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 18 All peristimulus time histograms (PSTHs), represented by the grey lines, were aligned to the 370 projected time of collision (TOC), represented by the red vertical line. The black solid lines 371 represent the average PSTH, while the black dotted lines represent the 95% confidence intervals. 372 Category 1 included units that showed a peak near TOC. Category 2 included units that showed a 373 valley near TOC. Category 3 included units that showed a gradual increase during the stimulus 374 presentation. Category 4 included units that showed a tonic firing during the stimulus 375 presentation. Category 5 included units that showed an increase towards the end of the displayed 376 window, after TOC. 377 378 For response category 1, which peaked near TOC, various parameters were calculated, including 379 the peak firing rate, peak time (relative to TOC), rise phase, and decay phase [20,38,40] (Fig 6). 380 The rise phase was defined as the time from when the unit's firing rate last exceeded the upper 381 95% confidence interval to the peak time, while the decay phase was measured from the peak 382 time to when the firing rate decayed to 15% of the peak firing rate. Among the 194 units that 383 responded to looming against all three backgrounds, 135 units (70%) exhibited a category 1 384 response to all three looming stimuli. Across different backgrounds, there was significant 385 differences in both peak time (Friedman Repeated Measures Analysis of Variance on Ranks, χ22 386 = 44.236, p < 0.001) and the length of the rise phase (Friedman Repeated Measures Analysis of 387 Variance on Ranks, χ22 = 70.673, p < 0.001). However, no significant differences were found in 388 peak firing rate (Friedman Repeated Measures Analysis of Variance on Ranks, χ22 = 4.326, p = 389 0.115) or the length of the decay phase (Friedman Repeated Measures Analysis of Variance on 390 Ranks, χ22 = 3.318, p = 0.190). Tukey post-hoc tests indicated that all pairwise comparisons were 391 significantly different for peak time (LW vs LG: q = 4.088, p < 0.05; LW vs LF: q = 9.338, p < 392 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 19 0.05; LG vs LF: q = 5.250, p < 0.05), although the difference in peak time between LW and LG 393 was relatively small. For the rise phase, there was no statistical difference between LW and LG 394 (Tukey post-hoc test, q = 2.711, p > 0.05), while both LW and LG showed significantly higher 395 values than LF (LW vs LF: q = 8.650, p < 0.05; LG vs LF: q = 11.361, p < 0.05). It is important 396 to note that, as peak times were delayed in LF, the selected trial length (-2 to 0.2 seconds relative 397 to TOC) did not always include a full decay phase, i.e. firing rate did not fall below 15% of the 398 peak firing rate by the end of the trial. Consequently, the sample size for the decay phase 399 comparison was smaller than that oof the other parameters, making it more difficult to find 400 significant difference. 401 402 Fig 6. Comparison of response parameters of units in Category 1 (peak near TOC) between 403

types. 404 (A) Sample PSTH of a Category 1 unit in response to a looming stimulus. The red vertical line 405 represents the time of collision (TOC), while the grey dashed line represents the upper 95% 406 confidence interval of the histogram. The last time the instantaneous firing rate exceeded the 407 upper 95% confidence interval (t95) was defined as the beginning of the response, while the time 408 when the firing rate decayed below 15% of the peak firing rate (t15) was defined as the end of the 409 response. The duration from t95 to the peak was defined as the rise phase, and the duration from 410 the peak to t15 was defined as the decay phase. (B) Within each box, the horizontal line 411 represents the median, while the lower and upper boundaries, the whiskers, and the dots 412 represent the 25/75%, 10/90%, and 5/95% percentiles, respectively. The peak time, peak firing 413 rate, rise phase, and decay phase of each PSTH was measured and compared using one-way 414 repeated measure (RM) ANOVA. Different letters above the boxes represent statistically 415 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 20 significant differences. The white and grey backgrounds evoked similar peak times (grey being 416 slightly later), while flow field evoked a significant peak delay. The flow field also resulted in 417 shortened rise phases. 418 419 In summary, among individual responsive units, LW and LG elicited similar responses, while the 420 addition of a flow field background (LF) resulted in delayed peak firing and a shorter rise phase 421 for units in category 1, indicating a delayed and more brief response. 422 423 Common trends 424 Dynamic factor analysis (DFA) was conducted using the MARSS package in R 4.1.3 to extract 425 common trends from all responsive units. Initial attempts to perform DFA on Gaussian-426 smoothed 1-ms-binned data resulted in consistent crashes of R due to the model's complexity. 427 Consequently, the firing rate was recalculated using a 50-ms bin width without smoothing. 428 Although temporal resolution was reduced, the 50-ms-binned data captured the dynamic 429 modulation of firing rate while significantly reducing computational complexity by a factor of 430 fifty. 431 For each of the 4 visual stimuli, DFA was iteratively performed with an increasing number of 432 common trends, starting from 1. The Akaike information criterion with a correction for small 433 sample sizes (AICc) was used to monitor the performance of each model. When the AICc began 434 to increase, the iteration was halted, and the model with the lowest AICc was deemed the optimal 435 model. S4 Table shows that the optimal models for LW and LG both included 7 common trends. 436 This consistency with the optimal model for looming at 0° elevation against a white background 437 [29] indicates that neither the looming trajectory (45° upwards) nor the half-white-half-grey 438 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 21

influenced the level of variability in the locust's response to a black looming 439 stimulus. In contrast, despite evoking a similar number of responsive units, the optimal model for 440 LF only contained 5 common trends, suggesting less variation among unit responses. 441 Fig 7 depicts activity of all the extracted common trends in response to each stimulus. In 442 response to a looming stimulus, irrespective of background types, most common trends exhibited 443 a positive peak near the time of collision (TOC). Against all three backgrounds, only one 444 common trend did not show a peak, and instead, continuously increased until the end of the 445 stimulus presentation (CT5 for LW, CT7 LG, and CT4 LF). When presented with the flow field 446

only (F), without the looming stimulus, the primary type of response was a brief 447 peak right after time when the flow field started (Tff), followed by a gradual decrease, 448 suggesting adaptation to the flow field. Despite that the best DFA model contained 6 CTs, the 449 first three CTs all showed this type of response, while the other CTs maintained at a relatively 450 stable level with no clear change near Tff. 451 452 Fig 7. Common trends (CTs) responses to all four types of visual stimuli. 453 In each column, peristimulus time histograms (PSTHs) of the CTs from a stimulus type were 454 plotted, aligned to either the projected time of collision (TOC) or the time when flow field 455 started displaying (Tff). In DFA models, the order of CTs was randomized, i.e., CT1 from the 456 first and second column are not necessarily the same. The factor loadings of constituent 457 discriminated units in each CT are shown next to the PSTH in a heatmap. Within a heatmap, 458 each horizontal line represents one discriminated unit. The order of units in the heatmap is also 459 consistent within a column. Irrespective of background types, most CTs responded to a looming 460 stimulus with a peak near TOC, similar to a DCMD response. The remaining common trend 461 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 22 from each background (CT5 in the white background, CT7 in the half-white-half-grey 462 background, and CT4 in the flow field background) continuously increased until the end of 463 stimulus presentation. For the last visual stimulus, flow field only, CT 1-3 displayed a peak at 464 Tff, when the flow field started, while CT 4-6 maintained at a relatively constant level without 465 clear change near Tff. 466 467 Among the common trends displaying clear peaks, the peak times occurred before or after TOC 468 in response to LW and LG, while in response to LF, only one common trend (CT 3) had a peak 469 time before TOC, and all other common trends had peaks after TOC. The median peak time of 470 LF common trends was the highest (i.e., latest) among all three looming stimuli, and the rise 471 phase for LF was the shortest. These findings aligned with the peak times observed in category 1 472 unit responses, although the comparisons of peak time and rise phase were not statistically 473 significant (One-way ANOVA; peak time, F2 = 0.507, p = 0.618; rise phase, F2 = 0.798, p = 474 0.480) (Fig 8), likely due to small sample size and low statistical power. However, it is worth 475 noting that the response start time (t95), defined as when the unit firing rate exceeded the upper 476 95% confidence interval, was significantly influenced by the type of visual background 477 (Kruskal-Wallis One-Way Analysis of Variance on Ranks, H2 = 7.096, p = 0.014). The response 478 began later in response to LF, compared to LG (Dunn's post-hoc test, p < 0.05). Since the firing 479 rate did not decay below 15% of the peak firing rate before the end of stimulus presentation in 480 some common trends (e.g., CT 1, 3, and 4), the sample size of t15 and decay phase was too small 481 (≤ 3 for each background) to derive meaningful comparison. 482 483 Fig 8. Comparison of response parameters of common trends (CTs) that peaked near TOC. 484 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 23 (A) Sample PSTH of a common trend in response to a looming stimulus. The red vertical line 485 represents the time of collision (TOC), while the grey dashed line represents the upper 95% 486 confidence interval. The last time the instantaneous firing rate exceeded the upper 95% 487 confidence interval (t95) was defined as start of the response. We defined the duration from t95 to 488 the peak as the rise phase. (B) Box plots comparing CT firing parameters. Within each box, the 489 horizontal line represents the median, while the lower and upper boundaries and the whiskers 490 represent the 25/75% and 10/90% percentiles, respectively. Different letters above boxes 491 represent statistically significant differences. The peak time, rise phase, and response begin time 492 of each PSTH was measured and compared using one-way repeated measure (RM) ANOVA. 493 Similar to Fig 5, flow field background resulted in later peak and shortened rise phase in the 494 CTs, but the difference was not statistically significant. The response start, however, was 495 significantly later in the flow field background, compared to the white and grey background. 496 497 498 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 24

499 During flight, optic flow is a main source of spatial information an animal can acquire. 500 Components of the optic flow can be used to calculate the distance and trajectory of nearby 501 objects. This is the first study to investigate the effect of a flow field background on looming 502 responses of multiple interneurons in the locust visual system. We found that multiple units were 503 responsive to a looming disc approaching upwards from bottom right at 45° azimuth and -45° 504 elevation, yet displaying various types of responses. Compared to looming against the white 505

(LW) and half-white-half-grey background (LG), looming against the flow field 506

(LF) affected the distribution of different response categories among responsive 507 units. Common trends (CTs) were extracted and compared between background types. Both LW 508 and LG evoked 7 CTs, the same as loom from 90° azimuth at 0° elevation, while LF evoked 5 509 CTs, suggesting that flow field background reduced the variability among unit responses. In both 510 individual units and common trends, the flow field background caused the responses to be 511 delayed. 512 513 Flow field design 514 Previous studies used flow field backgrounds that consisted of vertical bars that expanded across 515 the entire visual field [28,44], or squares expanding from the center of the screen [19]. The type 516 of flow field used here was designed to emulate the natural visual environment when the locust is 517 flying. Since flying animals are closer to the ground, compared to the sky, the objects on the 518 ground would appear faster and more conspicuous. Therefore, the optic flow in the lower half of 519 the visual field is more prominent than that in the top half. By dividing the visual background 520 vertically into two halves, and showing flow field in the bottom half only, we presented a 521 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 25 simplified bbackground that simulates a horizon line. Preliminary testing indicated that there was 522 no effect of a flow field on looming responses when the object approached from 45° azimuth at 523 0° elevation. For this trajectory, the top half of the object approached within the white 524 background, which likely attenuated any putative effects of a flow field [28], regardless of flow 525 field contrast or velocity. Therefore, we choose a a looming stimulus that moved upwards from -526 45° elevation to ensure the disc was within the flow field except for final 20.2 ms of each 527 approach. When presented with looming stimuli approaching within different areas across the 528 receptive field, some parameters of the DCMD response are different, yet the shapes of the 529 PSTH remain largely unaffected [45,46]. We found a similar effect among the other motion 530 sensitive units and common trends extracted from these units. The number of responsive units 531 and the number of common trends, representing the types of responses among all responsive 532 units, were consistent between looming at 0° elevation and -45° elevation. 533 Locusts display motion dazzle. When presented with a two-tone looming object, in which the top 534 half is darker than the background while the bottom half is lighter, the DCMD responds less than 535 the top (dark) half alone [47]. However, the flow field we used here was composed of white and 536 grey circles, both lighter than the black object (Michelson contrast ratio between black and white 537 = 0.98, Michelson contrast ratio between black and grey = 0.80). Therefore, the motion dazzle 538 was likely not present in our stimuli 539 540 Effect of flow field on collision detection 541 When an animal moves through its environment, optic flow is generated as a translational or 542 rotational pattern that is perceived by the eye [48]. Animals use optic flow to access self-motion 543 and stabilize trajectory. For example, honeybees use optic flow to maintain flight height [49], 544 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 26 while bumblebees use optic flow in the frontal visual field to control position and flight speed 545 [50,51]. 546 In locusts, a flow field consisting of white and grey vertical bars can elicit excitatory 547 postsynaptic potentials (EPSPs) in the LGMD, quickly followed by inhibitory postsynaptic 548 potentials (IPSPs) [40] via activation of lateral and feed-forward inhibition [13,52]. Therefore, 549 LGMD response to small looming objects is inhibited by large-field optic flow, represented by 550 fewer spikes and shorter rise phases [13,19]. In the DCMD, it was found that compared to white 551 background, more complex visual background, such as the flow field, caused reduced peak firing 552 rate, delayed peaks, and shorter rise phases as well [28,44]. We found that flow-field-induced 553 peak time delay and rise phase shortening were common among multiple motion-sensitive 554 neurons, although the peak firing rate was not statistically affected. 555 The compass cells [53] found in the central complex (CX) of locusts can use polarized light to 556 determine the orientation of the animal [54] and also respond to approaching and translating 557 objects [55,56]. On average, the peak time of the compass cells was delayed in the presence of a 558 flow field. However, the response of most compass cells to looming objects was inhibited by the 559 presence of large-field motion, while some were enhanced [57]. This suggests that a group of 560 neurons are only responsive to looming when optic flow is present, i.e., when the animal is 561 moving. Here we discriminated units (n=25) that responded to LF only, but not to LW or LG 562 (Fig 4). It is possible that these units can enhance motion detection during flight, and play an 563 important role in proper flight manoeuvre within a large swarm. 564 565 Units that responded to flow field only 566 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 27 We found that a group of units (n=25) responded to flow field only, but not to any of the 567 looming stimuli. Despite not directly responding to looming objects, these units can potentially 568 use flow field to access the self-motion of the animal, and thus factor into the execution of 569 successful collision avoidance behaviours. 570 Neurons that are exclusively sensitive to large-field motion have been identified in multiple 571 species, such as flies [58], moths [59], praying mantis [60], and crabs [61,62]. In hawkmoths, 572 two types of large-field motion-sensitive neurons, the horizontal cells and the vertical cells, were 573 identified in the lobula plate [63]. In honeybees, horizontal regressive- and progressive motion-574 sensitive neurons were found to be selectively sensitive to regression or progression motion, 575 respectively [64]. 576 In locusts, although the DCMD does not respond to the initiation of large-field grating stimuli 577 [65], large-field motion-sensitive neurons have been identified in the optic lobe [66–68] and 578 ventral nerve cord [69]. The two lobula directionally selective motion-detecting neurons 579 (LDSMD) respond preferably to flow field moving forwards or backwards, respectively, and one 580 of them synapses with the protocerebral descending directionally selective motion-detecting 581 neuron (PDDSMD) [68,69]. Those neurons found in the medulla, however, did not display 582 directional selectivity [67]. It is worth noting that although the axon of the PDDSMD extends in 583 the ventral nerve cord towards the metathoracic ganglion, it is ipsilateral to the cell body, while 584 we recorded from the contralateral nerve cord. In my study, only one type of flow field was 585 presented, and thus we could not test for directional sensitivity. Based on previous findings, we 586 assume that some of these units, as well as neural ensembles, will display directional preference, 587 while others do not. This can be tested in future studies using various flow field backgrounds and 588 larger-scale multichannel recordings. 589 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 28 590 Population coding of motion-sensitive neurons 591 The response of individual neurons is often “noisy”, displaying large variability over repeated 592 stimuli. Compared to individual neurons, functional groups of neurons, i.e., neural ensembles, 593 can represent complicated information and convey robust guides to generate appropriate 594 behavioural responses. In the rat gustatory cortex, neural ensembles progress through reliable 595 and stimulus-specific sequences of states, although the timing of transitions varies between trials 596 [70]. In the locust antennal lobe, multiple projection neurons encode different odorants, while the 597 ensemble activity display complex patterns over repeated presentation [71]. We found that 598 despite variability among individual neuron responses, the extracted common trends, which 599 represent ensemble activity, remain largely stable. The dominant types of common trends match 600 the response of previously identified motion-sensitive interneurons, providing a reliable 601 interpretation of the given stimuli. 602 While repeated co-activation of the same neurons can strengthen their synaptic connections and 603 increase the likelihood of the same neural ensemble reoccurring, mechanisms like changes in 604 short-term synaptic dynamics enable flexible reconfiguration of ensemble composition to meet 605 varying computational needs [72]. In the rat anterior cingulate cortex, ensemble activity during a 606 reward-searching task is affected by previous behavioural history [73]. Physiological states can 607 also determine the level of synchronization among neuron populations [74]. In Drosophila, 608 previous experience and hunger state affect the ensemble coding during the determination of the 609 food value [75]. Here we showed that when presented with identical looming stimuli against 610 different backgrounds, individual unit responses and common trends were both affected. Since 611 the units were generally from the same pool, changes in common trends, both the total number 612 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 29 and the peristimulus time histogram, reflect the fluidity of the contribution from individual units. 613 Future studies could also examine the effect of physiological states, such as hunger or 614 sleeplessness, which may affect the animal’s ability to detect visual cues, likely via dynamic 615 ensemble coding. 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 30

636 1. Gibson JJ. The perception of the visual world. Oxford, England: Houghton Mifflin; 1950. 637 pp. xii, 242. 638 2. Baird E, Srinivasan MV, Zhang S, Cowling A. Visual control of flight speed in honeybees. 639 J Exp Biol. 2005;208: 3895–3905. doi:10.1242/jeb.01818 640 3. Barron A, Srinivasan MV. Visual regulation of ground speed and headwind compensation 641 in freely flying honey bees (Apis mellifera L.). J Exp Biol. 2006;209: 978–984. 642 doi:10.1242/jeb.02085 643 4. Borst A, Haag J. Neural networks in the cockpit of the fly. J Comp Physiol A. 2002;188: 644 419–437. doi:10.1007/s00359-002-0316-8 645 5. Stöckl AL, O’Carroll D, Warrant EJ. Higher-order neural processing tunes motion neurons 646 to visual ecology in three species of hawkmoths. Proc R Soc B Biol Sci. 2017;284: 647 20170880. doi:10.1098/rspb.2017.0880 648 6. Sabbah S, Gemmer JA, Bhatia-Lin A, Manoff G, Castro G, Siegel JK, et al. A retinal code 649 for motion along the gravitational and body axes. Nature. 2017;546: 492–497. 650 doi:10.1038/nature22818 651 7. Borst A, Haag J, Mauss AS. How fly neurons compute the direction of visual motion. J 652 Comp Physiol A. 2020;206: 109–124. doi:10.1007/s00359-019-01375-9 653 8. Uvarov BP. Grasshoppers and locusts: A handbook of general acridology. Vol. 2. 654 Behaviour, ecology, biogeography, population dynamics. Centre for Overseas Pest 655 Research; 1977. 656 9. Waloff Z. Orientation of flying locusts, Schistocerca gregaria (Forsk.), in migrating 657 swarms. Bull Entomol Res. 1972;62: 1. doi:10.1017/S0007485300003771 658 10. Santer RD, Rind FC, Simmons PJ. Predator versus Prey: Locust Looming-Detector Neuron 659 and Behavioural Responses to Stimuli Representing Attacking Bird Predators. PLOS ONE. 660 2012;7: e50146. doi:10.1371/journal.pone.0050146 661 11. Hatsopoulos N, Gabbiani F, Laurent G. Elementary computation of object approach by a 662 wide-field visual neuron. Science. 1995;270: 1000–1003. 663 doi:10.1126/science.270.5238.1000 664 12. Rind FC, Simmons PJ. Orthopteran DCMD neuron: a reevaluation of responses to moving 665 objects. I. Selective responses to approaching objects. J Neurophysiol. 1992;68: 1654–66. 666 doi:10.1152/jn.1992.68.5.1654 667 13. Rowell CHF, O’Shea M, Williams JL. The neuronal basis of a sensory analyser, the acridid 668 movement detector system. IV. The preference for small field stimuli. J Exp Biol. 1977;68: 669 157–85. 670 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 31 14. Schlotterer GR. Response of the locust descending movement detector neuron to rapidly 671 approaching and withdrawing visual stimuli. Can J Zool. 1977;55: 1372–1376. 672 doi:10.1139/z77-179 673 15. O’Shea M, Williams JLD. The anatomy and output connection of a locust visual 674 interneurone; the lobular giant movement detector (LGMD) neurone. J Comp Physiol. 675 1974;91: 257–266. doi:10.1007/BF00698057 676 16. Dick PC, Gray JR. Spatiotemporal stimulus properties modulate responses to trajectory 677 changes in a locust looming-sensitive pathway. J Neurophysiol. 2014;111: 1736–1745. 678 doi:10.1152/jn.00499.2013 679 17. Gabbiani F, Krapp HG, Laurent G. Computation of object approach by a wide-field, 680 motion-sensitive neuron. J Neurosci. 1999;19: 1122–41. 681 18. Gabbiani F, Mo C, Laurent G. Invariance of angular threshold computation in a wide-field 682 looming-sensitive neuron. J Neurosci. 2001;21: 314–329. 683 19. Gabbiani F, Krapp HG, Koch C, Laurent G. Multiplicative computation in a visual neuron 684 sensitive to looming. Nature. 2002;420: 320–324. doi:10.1038/nature01190 685 20. Guest BB, Gray JR. Responses of a looming-sensitive neuron to compound and paired 686 object approaches. J Neurophysiol. 2006;95: 1428–1441. doi:10.1152/jn.01037.2005 687 21. McMillan GA, Gray JR. A looming-sensitive pathway responds to changes in the trajectory 688 of object motion. J Neurophysiol. 2012;108: 1052–1068. doi:10.1152/jn.00847.2011 689 22. O’Shea M, Rowell CHF, Williams JLD. The anatomy of a locust visual interneurone; the 690 descending contralateral movement detector. J Exp Biol. 1974;60: 1–12. 691 23. Simmons PJ. Connexions between a movement-detecting visual interneurone and flight 692 motoneurones of a locust. J Exp Biol. 1980;86: 87–97. 693 24. Fotowat H, Gabbiani F. Relationship between the Phases of Sensory and Motor Activity 694 during a Looming-Evoked Multistage Escape Behavior. J Neurosci. 2007;27: 10047–695 10059. doi:10.1523/JNEUROSCI.1515-07.2007 696 25. Santer RD, Rind FC, Stafford R, Simmons PJ. Role of an Identified Looming-Sensitive 697 Neuron in Triggering a Flying Locust’s Escape. J Neurophysiol. 2006;95: 3391–3400. 698 doi:10.1152/jn.00024.2006 699 26. Horn G, Rowell CHF. Medium and Long-term Changes in the Behaviour of Visual 700 Neurones in the Tritocerebrum of Locusts. J Exp Biol. 1968;49: 143–169. 701 27. Palka J. An inhibitory process influencing visual responses in a fibre of the ventral nerve 702 cord of locusts. J Insect Physiol. 1967;13: 235–248. doi:10.1016/0022-1910(67)90151-5 703 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 32 28. Silva AC, McMillan GA, Santos CP, Gray JR. Background complexity affects the response 704 of a looming-sensitive neuron to object motion. J Neurophysiol. 2015; 218–231. 705 doi:10.1152/jn.00478.2014 706 29. Zhang S, Gray JR. Complex motion trajectories are represented by a population code from 707 the ensemble activity of multiple motion-sensitive descending interneurons in locusts. 708 bioRxiv; 2024. p. 2024.09.29.615702. doi:10.1101/2024.09.29.615702 709 30. Gaten E, Huston SJ, Dowse HB, Matheson T. Solitary and gregarious locusts differ in 710 circadian rhythmicity of a visual output neuron. J Biol Rhythms. 2012;27: 196–205. 711 doi:10.1177/0748730412440860 712 31. Gray JR, Pawlowski V, Willis M a. A method for recording behavior and multineuronal 713 CNS activity from tethered insects flying in virtual space. J Neurosci Methods. 2002;120: 714 211–223. doi:10.1016/S0165-0270(02)00223-6 715 32. Guo P, Pollack AJ, Varga AG, Martin JP, Ritzmann RE. Extracellular wire tetrode 716 recording in brain of freely walking insects. J Vis Exp. 2014 [cited 30 Nov 2018]. 717 doi:10.3791/51337 718 33. Stott TP, Olson EGN, Parkinson RH, Gray JR. Three-dimensional shape and velocity 719 changes affect responses of a locust visual interneuron to approaching objects. J Exp Biol. 720 2018; jeb.191320. doi:10.1242/jeb.191320 721 34. Horridge GA. The separation of visual axes in apposition compound eyes. Philos Trans R 722 Soc B Biol Sci. 1978;285: 1–59. doi:10.1098/rstb.1978.0093 723 35. Miall RC. The flicker fusion frequencies of six laboratory insects, and the response of the 724 compound eye to mains fluorescent ‘ripple’. Physiol Entomol. 1978;3: 99–106. 725 doi:10.1111/j.1365-3032.1978.tb00139.x 726 36. Shoham S, Fellows MR, Normann RA. Robust, automatic spike sorting using mixtures of 727 multivariate t-distributions. J Neurosci Methods. 2003;127: 111–122. doi:10.1016/S0165-728 0270(03)00120-1 729 37. Dick PC, Michel NL, Gray JR. Complex object motion represented by context‐dependent 730 correlated activity of visual interneurones. Physiol Rep. 2017;5: e13355. 731 doi:10.14814/phy2.13355 732 38. Parkinson RH, Zhang S, Gray JR. Neonicotinoid and sulfoximine pesticides differentially 733 impair insect escape behavior and motion detection. Proc Natl Acad Sci. 2020;117: 5510–734 5515. doi:10.1073/pnas.1916432117 735 39. Ritzmann RE, Ridgel AL, Pollack AJ. Multi-unit recording of antennal mechano-sensitive 736 units in the central complex of the cockroach, Blaberus discoidalis. J Comp Physiol A. 737 2008;194: 341–360. doi:10.1007/s00359-007-0310-2 738 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 33 40. Gabbiani F, Cohen I, Laurent G. Time-Dependent Activation of Feed-Forward Inhibition in 739 a Looming-Sensitive Neuron. J Neurophysiol. 2005;94: 2150–2161. 740 doi:10.1152/jn.00411.2005 741 41. Zuur AF, Fryer RJ, Jolliffe IT, Dekker R, Beukema JJ. Estimating common trends in 742 multivariate time series using dynamic factor analysis. Environmetrics. 2003;14: 665–685. 743 doi:10.1002/env.611 744 42. Holmes E E, Ward E J, Wills K. MARSS: Multivariate Autoregressive State-space Models 745 for Analyzing Time-series Data. R J. 2012;4: 11. doi:10.32614/RJ-2012-002 746 43. Burnham KP, Anderson DR, editors. Model Selection and Multimodel Inference. New 747 York, NY: Springer; 2002. doi:10.1007/b97636 748 44. Yakubowski JM, McMillan GA, Gray JR. Background visual motion affects responses of 749 an insect motion-sensitive neuron to objects deviating from a collision course. Physiol Rep. 750 2018;4: e12801. doi:10.14814/phy2.12801 751 45. Krapp HG, Gabbiani F. Spatial distribution of inputs and local receptive field properties of 752 a wide-field, looming sensitive neuron. J Neurophysiol. 2005;93: 2240–2253. 753 doi:10.1152/jn.00965.2004 754 46. Judge S, Rind FC. The locust DCMD, a movement-detecting neurone tightly tuned to 755 collision trajectories. J Exp Biol. 1997;200: 2209–16. 756 47. Santer RD. Motion dazzle: a locust’s eye view. Biol Lett. 2013;9: 20130811. 757 doi:10.1098/rsbl.2013.0811 758 48. Warren WH. 2.12 - Optic Flow. In: Masland RH, Albright TD, Albright TD, Masland RH, 759 Dallos P, Oertel D, et al., editors. The Senses: A Comprehensive Reference. New York: 760 Academic Press; 2008. pp. 219–230. doi:10.1016/B978-012370880-9.00311-X 761 49. Portelli G, Ruffier F, Franceschini N. Honeybees change their height to restore their optic 762 flow. J Comp Physiol A. 2010;196: 307–313. doi:10.1007/s00359-010-0510-z 763 50. Linander N, Dacke M, Baird E. Bumblebees measure optic flow for position and speed 764 control flexibly within the frontal visual field. J Exp Biol. 2015;218: 1051–1059. 765 doi:10.1242/jeb.107409 766 51. Linander N, Baird E, Dacke M. Bumblebee flight performance in environments of different 767 proximity. J Comp Physiol A. 2016;202: 97–103. doi:10.1007/s00359-015-1055-y 768 52. O’Shea M, Rowell CHF. Protection from habituation by lateral inhibition. Nature. 769 1975;254: 53–55. doi:10.1038/254053a0 770 53. Stone T, Webb B, Adden A, Weddig NB, Honkanen A, Templin R, et al. An Anatomically 771 Constrained Model for Path Integration in the Bee Brain. Curr Biol. 2017;27: 3069-772 3085.e11. doi:10.1016/j.cub.2017.08.052 773 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 34 54. Heinze S, Homberg U. Maplike Representation of Celestial E-Vector Orientations in the 774 Brain of an Insect. Science. 2007;315: 995–997. doi:10.1126/science.1135531 775 55. Bockhorst T, Homberg U. Compass Cells in the Brain of an Insect Are Sensitive to Novel 776 Events in the Visual World. PLOS ONE. 2015;10: e0144501. 777 doi:10.1371/journal.pone.0144501 778 56. Rosner R, Homberg U. Widespread Sensitivity to Looming Stimuli and Small Moving 779 Objects in the Central Complex of an Insect Brain. J Neurosci. 2013;33: 8122–8133. 780 doi:10.1523/JNEUROSCI.5390-12.2013 781 57. Rosner R, Pegel U, Homberg U. Responses of compass neurons in the locust brain to visual 782 motion and leg motor activity. J Exp Biol. 2019;222: jeb196261. doi:10.1242/jeb.196261 783 58. Hausen K, Egelhaaf M. Neural Mechanisms of Visual Course Control in Insects. In: 784 Stavenga DG, Hardie RC, editors. Facets of Vision. Berlin, Heidelberg: Springer; 1989. pp. 785 391–424. doi:10.1007/978-3-642-74082-4_18 786 59. Rind FC. A Directionally Sensitive Motion Detecting Neurone in the Brain of a Moth. J 787 Exp Biol. 1983;102: 253–271. 788 60. Yamawaki Y. Unraveling the functional organization of lobula complex in the mantis brain 789 by identification of visual interneurons. J Comp Neurol. 2019;527: 1161–1178. 790 doi:10.1002/cne.24603 791 61. Sandeman DC, Kien J, Erber J. Optokinetic eye movements in the crab, Carcinus maenas. J 792 Comp Physiol. 1975;101: 259–274. doi:10.1007/BF00657186 793 62. Scarano F, Tomsic D, Sztarker J. Direction Selective Neurons Responsive to Horizontal 794 Motion in a Crab Reflect an Adaptation to Prevailing Movements in Flat Environments. J 795 Neurosci. 2020;40: 5561–5571. doi:10.1523/JNEUROSCI.0372-20.2020 796 63. Wicklein M, Varjú D. Visual system of the European hummingbird hawkmoth 797 Macroglossum stellatarum (sphingidae, lepidoptera): Motion-sensitive interneurons of the 798 lobula plate. J Comp Neurol. 1999;408: 272–282. doi:10.1002/(SICI)1096-799 9861(19990531)408:23.0.CO;2-9 800 64. DeVoe RD, Kaiser W, Ohm J, Stone LS. Horizontal movement detectors of honeybees: 801 Directionally-selective visual neurons in the lobula and brain. J Comp Physiol. 1982;147: 802 155–170. doi:10.1007/BF00609840 803 65. Rind FC, Santer RD. Collision avoidance and a looming sensitive neuron: size matters but 804 biggest is not necessarily best. Proc R Soc Lond B Biol Sci. 2004;271: S27–S29. 805 doi:10.1098/rsbl.2003.0096 806 66. Gewecke M, Hou T. Visual brain neurons in Locusta migratoria. Sensory Systems of 807 Arthropods (Wiese K, Gribakin FG, Popov AV, Renninger G, eds). 1993. pp. 119–144. 808 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 35 67. Homberg U, Wü rden S. Movement-sensitive, polarization-sensitive, and light-sensitive 809 neurons of the medulla and accessory medulla of the locust, Schistocerca gregaria. J Comp 810 Neurol. 1997;386: 329–346. doi:10.1002/(SICI)1096-9861(19970929)386:33.0.CO;2-3 812 68. Rind FC. Identification of directionally selective motion-detecting neurones in the locust 813 lobula and their synaptic connections with an identified descending neurone. J Exp Biol. 814 1990;149: 21–43. 815 69. Rind FC. A directionally selective motion-detecting neurone in the brain of the locust: 816 physiological and morphological characterization. J Exp Biol. 1990;19: 1–19. 817 70. Jones LM, Fontanini A, Sadacca BF, Miller P, Katz DB. Natural stimuli evoke dynamic 818 sequences of states in sensory cortical ensembles. Proc Natl Acad Sci. 2007;104: 18772–819 18777. doi:10.1073/pnas.0705546104 820 71. Nizampatnam S, Zhang L, Chandak R, Li J, Raman B. Invariant odor recognition with ON–821 OFF neural ensembles. Proc Natl Acad Sci. 2022;119: e2023340118. 822 doi:10.1073/pnas.2023340118 823 72. Carrillo-Reid L. What Is a Neuronal Ensemble? 2020. 824 73. Proskurin M, Manakov M, Karpova A. ACC neural ensemble dynamics are structured by 825 strategy prevalence. eLife. 12: e84897. doi:10.7554/eLife.84897 826 74. Renart A, de la Rocha J, Bartho P, Hollender L, Parga N, Reyes A, et al. The Asynchronous 827 State in Cortical Circuits. Science. 2010;327: 587–590. doi:10.1126/science.1179850 828 75. Sareen PF, McCurdy LY, Nitabach MN. A neuronal ensemble encoding adaptive choice 829 during sensory conflict in Drosophila. Nat Commun. 2021;12: 4131. doi:10.1038/s41467-830 021-24423-y 831 832 833 834 835 836 837 838 839 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 36 Supporting information 840 S1 Table. Summary of spike detection and unit discrimination. On average, 27215 spikes 841 (waveforms) were detected, and 16.8 units were discriminated. The MANOVA results show that 842 the discriminated units were statistically distinct from each other. 843 S2 Table. Summary of the number of units that responded to each type of visual stimulus. 844 LW represents loom against the white background, LG represents loom against the grey 845 background, LF represents loom against the flow field background. 846 S3 Table. Distribution of units in different categories when presented with looming against 847 different visual backgrounds (LW, white; LG, half-white-half-grey; LF, flow field). The 848 distribution was similar betweeno LW and LG. However, in response to LF, the proportion of 849 Category 3 decreased, while the proportion of Category 5 increased. 850 S4 Table. Summary of dynamic factor analysis (DFA) models. For each stimulus type, the 851 DFA model was performed iteratively, starting with 1 common trend (CT). The Akaike 852 information criterion (AIC) and AIC corrected for small sample size (AICc) of each model are 853 shown above. Since AICc can prevent over-fitting, it was used to determine the best-fit 854 approximating model. For looming against both white and white/grey backgrounds, the best 855 models contained 7 common trends. For looming against the flow field background, the best 856 model contained 5 common trends. For flow field only, the best model contained 6 common 857 trends. 858 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 37 859 Figure 1 860 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 38 861 Figure 2 862 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 39 863 Figure 3 864 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 40 865 Figure 4 866 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 41 867 Figure 5 868 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 42 869 Figure 6 870 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 43 871 Figure 7 872 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint 44 873 Figure 8 874 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted October 17, 2024. ; https://doi.org/10.1101/2024.10.15.617990doi: bioRxiv preprint