Thirty-five wild-caught dragonflies (35 Hemicordulia tau, 33 male, two female) were immobilized with 1:1 beeswax and rosin mixture and fixed to an articulated magnetic stand with the head tilted forward to access the posterior surface. A hole was cut above the brain to gain access to the lobula and lateral midbrain, but the preparation was otherwise left with the perineural sheath and overlying haemolymph sacs intact. All dissections were performed on the left side of the animal corresponding to a right excitatory hemifield. We penetrated the sheath and recorded intracellularly using aluminosilicate micropipettes (OD = 1.00, ID = 0.58 mm), pulled on a Sutter Instruments P-97 puller and backfilled with either KCl (2 M, electrode tip resistance typically 50–150 MΩ) or 4% Lucifer Yellow solution in 0.1 M LiCl. Electrodes were placed in the medial portion of the lobula complex and stepped through the brain from posterior to anterior through the lobula complex, using a piezoelectric stepper (Marzhauser-Wetzlar PM-10, Wetzlar, Germany). Intracellular responses were digitized at 5 kHz with a 16-bit A/D converter (National Instruments) for offline analysis.

We presented stimuli on high-definition LCD monitors (Asus ROG Swift PG279Q 165 Hz). The animal was placed 20 cm away and centered on the visual midline with the head tilted forward such that back surface of the eye was in line with the top of the monitor (approximately 60° tilt). The display projection was distorted using OpenGL to ensure each 1° onscreen was 1° from the dragonfly’s perspective, extending 104° (−52 to 52° azimuth) by 58.5° (21 to 80° elevation from equator).

Neuronal classification was performed by presenting a sequence of stimuli: a gyrated, randomly generated texel pattern (1°), gray to black and gray to white full screen flicker (White 338 cd/m², Black 0.5 cd/m²), moving edges (up, down, left and right, 25°/s), moving bars (2° width, up, down, left and right, 25°/s) and a square-wave grating pattern (0.025 cycles/°, 6.25Hz, up, down, left and right) and a small target (1.5°) moving at 80°/s horizontally (both left and right) at 21 different elevations to map the receptive field (Figure 1A). CSTMD1 was identified based on characteristic responses to these visual stimuli including distinctive spike waveforms (Fabian and Wiederman, 2021) and unique receptive fields.

A 1 s prestimulus period of a mid-gray screen for experiments involving natural images. A post-stimulus period of 1 s was also recorded. For experiments involving repetitive stimulation of a single region, an inter-trial rest of between 15 s and 30 s was used to minimize influences from long-term adaptation. Between trials there is also some degree of variability in response strength. To assay across these factors (images and conditions), intracellular recordings are held for up to several hours, therefore we control for inter-trial variation (fatigue or adaptation) via randomization of the order of stimulus presentation.

Natural imagery used in experiments were 360° panoramas captured using a Nikon D-70 digital camera and panoramic tripod head (Brinkworth and O’Carroll, 2009). A subregion of each image was displayed to maximize vertical extent and maintain aspect ratio with the initial azimuth (referred henceforth as image phase) defined by the stimulus.

We developed custom-written MATLAB scripts for spike counting and analysis. All statistical tests were either paired t-tests (with Bonferroni correction where appropriate) or n-dimensional ANOVA using tukey-kramer for multiple comparisons. All p values are reported as raw numbers in text if significant differences exist (unmarked otherwise) or as <0.0001 if sufficiently small. Box and whisker plots represent the 75th, 50th and 25th quartiles (lines) with raw data overlaid.

When presenting a small moving target in visual clutter the effective contrast of the target is modulated, moving over dark or bright regions. To remove this source of ambiguity, we introduced a solid gray (0.5 Weber contrast) strip over the intended trajectory of the target (similar to Nicholas et al., 2018), allowing the target to move across the field of view with constant local contrast. The strip was 11° in height and the edges of the strip were partially transparent (using a linearly declining alpha channel over 2°) to prevent the display of hard edges.

To test the efficacy of this gray-strip paradigm, we chose three natural image backgrounds (Figure 1B) to represent dense, urban and sparse environments. The panoramic background images were presented starting at one of two initial starting phases from the 360°. We tested the stimulus in five different conditions (Figure 1C): Target alone (TO) against a gray background, Background Alone (BO), Background with Strip (BS), Background with Target (BT) and Background, Strip and Target (BST). In all cases, the stationary background and strip (where present) were introduced together (Figure 1C, green bar) while the moving target (a 1.5 × 1.5° black square) was displayed 1 s after the background. The target appeared at the edge of the screen and moved horizontally at 25°/s (Figure 1C, black bar) in CSTMD1’s weakly preferred direction (left to right). Due to the excitatory and inhibitory subregions of CSTMD1’s receptive field, spiking activity elicited by the target is not apparent until halfway through the stimulus duration.

Figure 1C shows five individual spike traces from CSTMD1, one from each of the five conditions. The TO condition reveals a characteristically strong response beginning approximately half-way into the target’s appearance. In BO and BS (where no target is present) we observe limited activity over the same period. In the BT trials, we see a similar response to the TO trial, except that there is less activity part way through the target presentation. In the BST trial, we observe a similar response to the TO trial.

To examine aggregate responses, we determined spike rate (Figure 1D) over a 500 ms period starting 500 ms after the target cross the midline into excitatory receptive field (purple shaded region Figure 1C). When presented alone, the target (TO) generates strong responses in this window whereas in the BO and BS (no target present) trials the response matches the spontaneous spike rate (Figure 1C). When the background and target are presented without the strip (BT) the responses vary considerably with some low responses similar to the spontaneous spike rate. The increased variation in spike rate is expected as the local contrast of the different backgrounds varies, modulating the effective contrast of the target. With the introduction of the gray strip (BST), there was no significant difference between the TO and BST trials, whilst there was a significant difference between BT and BST trials (p = 0.0074, n = 3).

To examine these differences further, we generated raster plots for each of the individual trials separated by condition (Figure 1E). The BO and BS cases show no discernible change in response to the presentation of the background (1–7 s) except minor transient events at image onset and offset. The BT case shows increased variation with some trials exhibiting a muted response (lower overall spike-rate) due to poorer contrast and other exhibiting more variation due to the variable contrast of the image. The TO and BST cases are almost indistinguishable from one another. Thus, the gray strip paradigm is effective at eliminating the confounding effect of local, contrast variation between target and background.

Having demonstrated the effectiveness of the gray-strip paradigm, we examined how CSTMD1 responses are altered by moving backgrounds. We drifted the same 1.5 × 1.5° target across the gray strip at 25°/s whilst simultaneously having one of three background images moving horizontally at 15°/s (Figure 2A) presented at two different starting phases (0, 180°). To ensure differences in responses were not hidden within neuronal saturation, we chose test speeds away from the saturating, velocity optimum (60–90 deg/s). We presented five stimulus conditions (Figure 2B): Target only (TO), Leftward Background with Strip (BSL), Rightward Background with Strip (BSR), Leftward Background with Strip and Target (BSLT), Rightward Background with Strip and Target (BSRT).

Figure 2B shows individual traces measured in response to the targets and background. In the TO case, we observed strong inhibition (left hemisphere) followed by strong excitation (right hemisphere). In the BSL and BSR conditions we observed little sustained excitation or inhibition, but instead transient responses of varying strengths throughout the stimulus time course. This differs from fly TSDNs, which show consistent inhibitory responses to widefield motion (Nicholas et al., 2018). In this BSLT trial, we observed responses similar to the TO condition. In comparison, the BSRT example reveals some excitation even though the target was in CSTMD1’s inhibitory hemifield. Then, a weaker response corresponding to the period when the target was in the excitatory receptive field.

To examine these inter-trial variations more closely, we generated raster plots for each of the individual trials separated by condition (Figures 2C–E). We observed robust inhibitory and excitatory responses when the target was presented alone (Figure 2C). The variability observed in TO trials is largely due to variations in overall spike activity between animals. When the background was presented alone (BSL, BSR), responses were weaker, transient, and intermittent (Figure 2D) including less consistent inhibition in the inhibitory hemifield. These responses showed little consistency across trials, even with the same background image and background phase. In the BSLT and BSRT trials, the response to the target was still apparent in most trials but could be weaker, shorter and more variable (Figure 2E). Additionally, there are also intermittent responses when the target is not in the excitatory hemifield.

From the raster plots, we can identify where the changes are most (and least) prevalent. Most individual trials retained robust responses while the target was close to the midline of the animal (just after 4 s) with less spiking activity occurring when the target was located at a more peripheral location (5–6 s). The raster plots reveal differences in neuronal activity elicited by the different background images. Image B has the largest effect on target response (large gaps in both BSLT and BSRT). Image A has less pronounced changes, especially in the BSLT case where consistent target responses were most common. While variation of CSTMD1 activity across dragonflies is typical (Figure 1D), the strong sustained inhibition followed by sustained excitation is consistent (Figure 2C). Moreover, dragonfly Lobula Tangential Cells (LTCs), neurons sensitive to wide-field optic flow, give remarkably consistent responses to moving natural scenes, regardless of the image statistics (Evans et al., 2019). Thus, if CSTMD1 were inhibited by widefield motion in a manner similar to Eristalis TSDNs, we would expect to see a consistent effect of the background’s motion on CSTMD1 responses. This might appear as inhibitory or spontaneous activity in response to BSR trials and with facilitatory effects concomitant to increased relative motion between target and background in BSL trials. Instead, we see sporadic excitation of variable strength with differences between images.

We tested the differences in response between images by finding the mean spike rate in a 1 s window (4.5–5.5 s) for all BSLT and BSRT trials. We found a statistically significant difference of Image B in the BSLT case compared to both other images (A vs. B, p = 0.0026, B vs. C, p = 0.0145, 1-way ANOVA, Tukey-Kramer correction.) and no differences in the BSRT trials. Based on previous analysis of LTCs (Evans et al., 2019), we would expect Image C to be an optic flow “outlier” due to its lower image contrast. For a given speed, any inhibitory influences from an LTC would be consistent (across images). This is not observed, thus their role in CSTMD1 responses is unlikely.

We calculated mean spike rate as the Inverse Interspike Interval (1/ISI) across trials, averaging across different images and image phases (Figures 3A,B). The TO trials exhibit a characteristic period of inhibition (2–4 s) followed by strong excitation (4–6 s) as the target crosses the inhibitory and excitatory hemifields, respectively. When presented alone, the background (BSL and BSR) show responses above spontaneous (dotted blue, Figure 3A), stronger in the BSR case.

Background with rightwards target trials (BSRT) exhibited the least inhibition when the target was located in the inhibitory hemifield and the least excitation when in the excitatory hemifield (Figure 3B). We did not observe any facilitation compared to target alone, when target and background moved in opposing directions (BSLT, green) as observed in fly TSDNs (Nicholas and Nordström, 2021).

We further examined the variability observed in the raster plots (Figures 2C–D). Firstly, we calculated the instantaneous spike interval (ISI) by smoothing the spike times such that each spike event was replaced with the average spike time of its seven closest neighbors. This accounted for spike bursts, an additional source of variability when CSTMD1 is stimulated (Fabian and Wiederman, 2021). With bursts, the Inverse ISI oscillates around two means (burst and inter-burst) confounding variability measures. Additionally, we observe variation of the overall spiking activity across dragonflies. As a normalization factor, each Inverse ISI trial was divided by an overall excitation factor. This factor was the mean excitatory response (4.25–5.75 s) to TO trials for that animal. This normalized value was then multiplied by the mean of means across animals to return the result to spikes/s. We calculated the standard deviation in the instantaneous spike rate across trials for the five conditions (Figures 3C,D) and smoothed with a 50 ms moving average.

Figure 3C shows the standard deviation of the TO (black), BSL (green), and BSR (orange). In TO trials, we observed a small standard deviation during the prestimulus duration (0–1 s, gray screen) and the inhibitory period (2–4 s). Following this, we observe the largest variation at the start (4 s) and end (6 s) of the window when the target is within the excitatory region of CSTMD1’s receptive field. This represents variation in the onset of facilitation (Fabian et al., 2019) and interestingly also observed at the offset of the target’s presentation.

In the BSL and BSR trials we observed large variation for both background directions throughout the stimulus period, building over time. This is despite only a modest mean response (Figure 3A). We also observe that this variation continues after the offset of the stimulus (7–8 s) during which a gray screen is presented. This increased variation is likely due to post-excitatory rebounds common after stimulation of CSTMD1. In the TO cases, these post-excitatory inhibitory periods are highly consistent after stimulus offset (6 s), therefore eliciting a lower standard deviation (Figure 3A). Whereas the sporadic excitation seen in BSLT or BSRT trials will likely only induce sporadic post-excitatory inhibition.

In the BSLT and BSRT trials (Figure 3D), we observe high variation across the entire stimulus period. When the target is in the inhibitory hemifield (2–4 s), the variation is consistently higher than the TO trials. This is also true for when the target is in the excitatory hemifield (4–6 s). This is despite the mean spike rate for both conditions being lower than the TO case (Figure 3B).

We calculated a histogram of spike rates (1/ISI) for each stimulus condition over three periods of the stimulus; spontaneous period (0.25–0.75 s); target in inhibitory (3.25–3.75 s); and excitatory (4.25–4.75 s), with bins distributed logarithmically to better illustrate the differences at both high and low spike rates. In the analysis window corresponding to excitatory target responses (Figure 3E), the TO case (black) generates the highest spike rate overall. BSR (orange dashed) trials had higher spike rates over this period than BSL (green dashed) trials or spontaneous (blue dashed), reflective of CSTMD1’s weakly preferred direction (left-to-right). However, when the target is introduced, this relationship is reversed with higher spike rates for the BSLT trials (unbroken green) than the BSRT trials (unbroken orange). Thus, in background alone, rightward motion produces more excitation than leftwards (both low without a target). However, the inclusion of a target reverses this ordering, with the rightwards background (same direction as target) producing weaker responses than the opposing leftwards background.

In the analysis window corresponding to when a target may be within the inhibitory RF (Figure 3F), we see the TO cases (black) producing strong inhibition compared to the spontaneous activity. We also observe that both BSR (orange dashed) and BSRT (orange unbroken) produce the highest spike rates, stronger than spontaneous. If the background were merely inhibiting target responses, we would expect responses similar to spontaneous activity (blue dashed). Rather, CSTMD1 responses are elevated during this period.

In response to the inclusion of a background, what could cause weaker activity when the target is in the excitatory hemifield, increased responses when in the inhibitory hemifield and greater variability between trials? As previously mentioned, an inhibitory drive from a wide-field LTC neuron would not increase variability, nor induce sporadic responses such as those seen in BSL and BSR (Figure 2D). Instead, our results reveal variation in responses due to selective attention (Wiederman and O’Carroll, 2013). Natural images have been shown to have features which can excite CSTMD1 (Wiederman and O’Carroll, 2011) and the selection between the target and these background features would elicit sporadic responses observed during the BSL and BSR trials. The overall excitation observed in BSR trials (Figures 2D,E) would also be expected as this is CSTMD1’s preferred direction for the excitatory hemifield (left to right). CSTMD1 typically selected the high-contrast target and ignored features from within the background. However, when CSTMD1 selected background features rather than the target, we observe weaker responses, irrespective of whether the target is in CSTMD1’s inhibitory or excitatory hemifield.

Since background features can out-compete the high-salience target for selection, what happens when the target’s salience is lowered? To test this, we repeated previous experiment with varied target contrast. We measured spike activity while the target was within the excitatory receptive field (4.5–5.0 s) for each condition and calculated averages for each image and direction pair (Figure 4A—BSLT, Figure 4B—BSRT, n = 6).

The presence of background clutter reduced neuronal activity over the range of contrasts measured (Figures 4A,B). The varying degrees of shift in the BSLT and BSRT contrast sensitivity functions (across images) show that interaction between features, rather than optic flow, underlie neuronal responses (i.e., shifts would be the same to a single background speed).

As exemplars, Figure 4C shows stochastic activity from three individual neurons. Each plot shows the response to the TO (black) and two different backgrounds with target (BSLT, BSRT), where the only difference is the starting phase of the background image (red or blue line). The change in starting image phase alters when background features are within CSTMD1’s receptive field, as well as their relative position to the target. However, optic flow remains relatively constant across image phases, with minimal pattern noise expressed in LTC neurons. Here we see in some trials, the presence of background had little effect on the overall spiking activity (Figure 4C top left). In other trials (Figure 4C top right, bottom left) the starting phase of the background caused a large reduction in response at one phase (red) but not the other (blue). We also saw the introduction of sporadic responses to background features (Figure 4C bottom right, black arrow) when the target would usually generate inhibitory responses. This variability was also observed with low contrast BSLT/BSRT trials (Figure 4C bottom right, red). In this example, despite the lower contrast, we observed responses similar to the TO control.

We have shown that weaker responses can be elicited in CSTMD1, when a slowly moving background (15°/s) is presented with the target. How might a faster-moving background, as experienced during rapid pursuit flights, affect neuronal responses to a fast or slow-moving target? To investigate this, we tested a single image (Figure 5A) and recorded responses to a 1.5 × 1.5° black target moving at one of three velocities (15, 35, 90°/s) against a background. The background moved at one of three different velocities (15, 35, 90°/s) in the neuron’s preferred direction (rightwards) or anti-preferred direction (leftwards), starting at one of four initial phases (0, 90, 180, 270°). We measured spiking activity in an analysis window corresponding to the moving target. As faster target velocities evoke responses over a shorter duration of time, we changed our analysis window (Figure 5B) to correspond to a period when the target moved over a constant region of space (50°) within CSTMD1’s receptive field. When the target was not present, we analyzed the corresponding period, ensuring an equivalent epoch for analysis.

Figure 5C shows the spike rate from these windows for each of fifteen conditions when the target and background direction are the same: three target-only controls for the three velocities (red); three corresponding background-only controls (green); and nine background and target speed combinations (black). The target-only (TO) trials exhibited responses matched to the underlying velocity tuning (optimum ~90°/s), as previously described (Dunbier et al., 2012). Although higher speeds produced slightly stronger responses in background-only (green) trials (one way ANOVA, n = 8, p < 0.0001), these were still very weak at all speeds compared with even the slowest target-only condition.

In BSRT trials, CSTMD1 responses strengthened as target speed increased, consistent with the inherent velocity tuning. However, at each target speed, CSTMD1 responses decreased with higher background speeds (two-way ANOVA, n = 8; background speed, p = 0.0005; target speed, p < 0.0001). To examine effects independent of the underlying velocity tuning, we divided all the BSRT responses by the mean of the TO trials matching their target speed (i.e., a trial with background speed of 15°/s and target speed of 35°/s was divided by the TO trial with speed 35°/s) to provide a normalized response compared to the TO alone case. After this normalization, differences we observe due to the target speed are not due to the normal velocity tuning of CSTMD1, but rather other indirect effects such as increased or decreased likelihood of selection. After normalization (Figure 5E), increases in target speed increased average responses and increases in background speed reduced average responses (two-way ANOVA, n = 8; background speed, p = 0.0026; target speed p = 0.0015).

How can an increased target speed cause an increase in average responses when normalized for the inherent velocity tuning? Similarly, why does increasing the background speed reduce the mean response? If features (whether target or background) are more likely to be selected when they better match the velocity tuning of CSTMD1, then we might expect that a fast background would generate more selections of “weak” background features (resulting in weaker CSTMD1 activity) and a fast target would generate more selections (stronger activity). Likewise, a slower background would generate less selections to background features (stronger activity) and a slower target is less selected (weaker activity). As there is a response difference between the well-matched target and poorly matched background features, this change in selection distribution explains the differences observed in BSRT responses.

When the background direction was reversed (BSL, BSLT, Figure 5D), we observed weaker responses from the BSL trials when compared to the BSR trials (unpaired T test, p < 0.0001). However, while we did observe a significant effect of target velocity on response prior to normalization (two-way ANOVA, n = 7, p < 0.0001), there was no significant effect of background speed on response (p > 0.5). After normalization, the effect of target speed also disappeared (two-way ANOVA, n = 7; background speed, p = 0.49; target speed p = 0.24, Figure 5F). Thus, with opposed target and background motion, the relative speeds are less influential. As the inhibitory hemisphere is more sensitive to leftwards motion (Bolzon et al., 2009), it is likely that the stronger competitive features are more likely to excite the inhibitory hemisphere. We have previously shown that selective attention can operate across hemispheres but is subject to different dynamics than intra-hemispheric selective attention (Lancer et al., 2022). This may explain the smaller dependence on the speed of a leftward moving background.

To determine whether our selective attention hypothesis matches the observed data we developed a series of six models based on previous models used to test for selective attention (Wiederman and O’Carroll, 2013). The models were designed to predict the background with target responses (BSLT, BSRT) from different combinations of the target-alone (TO) and background-alone (BSL, BSR) physiological data (Figure 6A). We used data from our Moving Background experiments as input.

For each animal (n = 12), we first found the mean response of the TO trials to represent the Target-Only input. For the Background-Only input, we separated trials based on the image used, the background direction and the background starting location (phase) and then found the corresponding mean image response. This resulted in a total of twelve different Background-Only inputs (three images, two phases, two directions). Both the mean TO and set of mean BSL/BSR responses were then segmented into 5 ms bins and combined according to the rules of each model variant (six variants, details below). We then calculated an error between model output and the physiological BSLT/BSRT data (matched to the corresponding image, phase and direction).

The six variant models tested are as follows: (1) Mean: the average of the Target-Only or Background-Only mean responses (each in 5 ms bins); (2) Max: the maximum of the Target-Only or Background-Only mean responses; (3) Min: the minimum of the Target-Only or Background-Only mean responses; (4) Sum: the addition of the minimum of the Target-Only or Background-Only mean responses; (5) Inhibition: each Target-Only input was multiplied by a constant factor depending on the image/direction/speed group. The factor was calculated as the mean response across all BSLT/BSRT trials divided by the mean response across all TO trials; and (6) Switch: the smaller of the errors between either the Target-Only or Background-Only mean responses (Figure 6B).

A distribution for the errors across all model variants is shown in Figure 6C. The Switch model resulted in the smallest errors indicated by the high peaks at zero and narrow flanks. The Sum and Max models generated large over-estimates of the response while the Min model produced underestimates. We further calculated the mean absolute error for each trial (averaged across 5 ms bins) and treated each trial as an independent error (Figure 6D). The Switch model produced the smallest mean error of the six variants tested. The Switch model was significantly better than all other variants (p < 0.0001 T-Test with Boneferroni correction, in all five comparisons) indicative of a selective attention mechanism underpinning background interactions.