Using a forward genetic screen, we identified a GAL-4 enhancer trap expressed in a small cluster of neurons innervating the lobula complex, the Foma-1 neurons, as well as the γ lobe of the mushroom body [19] (Figure 1A, 1B and 1C). This enhancer trap was typically expressed in five cells in each optic lobe, and iontophoretic injections of fluorescent dye into these cells revealed three distinct morphological types (Figure 1D–1K): a wide-field translobula-plate neuron [4], with processes in both the lobula and lobula plate (Figure 1D and 1H, n = 6); a wide-field lobula projection neuron [4, 5], with processes in a proximal layer of the lobula and a projection into the protocerebrum (Figure 1E and 1I, n = 4); and a cluster of three lobula plate tangential cells, with sparse processes in the lobula plate and an axonal projection into the central brain (Figure 1F and 1J, n = 6).

We targeted Foma-1 neurons for loose patch recordings [20] to examine their visual responses. We accessed the neurons from the posterior of the fly’s head, while presenting visual stimuli generated on a high-speed CRT monitor to the eyes via two coherent fiber optic bundles (Figure S2). Due to the stereotyped locations of the neuronal cell bodies, we could reproducibly target individual cells expressing GFP under the control of Foma-1GAL-4. Of the five cells, we were able to record from the three largest cell bodies, verified by dye injection into the cell body following every recording, and obtained reproducible spiking responses to visual stimuli presented within the spatial extent of our display (n = 79). While these cells are morphologically distinct, their responses, as we will describe, were indistinguishable. We confirmed the identities of two recorded cells through dye labeling of the processes following data collection for the translobula-plate cell and the lobula projection neuron (data not shown).

We presented a variety of visual stimuli to the fly, examining the luminance and motion sensitivity of these cells (Figure 2). They responded to both increases and decreases in global light intensity (Figure 2A and 2D), and were sensitive to the onset and offset of local flashes of light within a dorsal region of the visual field (Figure 2B and 2E). Within this receptive field, these cells also responded in a direction selective fashion to the movement of a small dot, preferring downward motion (Figure 2C, 2F and Figure S1). However, these neurons did not exhibit direction selective responses to global motion stimuli, neither to drifting gratings nor to a dynamic dot stimulus (Figure S1).

Foma-1 neurons did, however, exhibit strong, selective responses to looming stimuli that mimicked an object on a direct collision course. Such stimuli expand as a function of the object’s size and speed (Figure 3A and 3B). When the fly was presented with such a stimulus, the Foma-1 neurons’ firing rate increased until it reached a peak close to the anticipated time of collision, whereupon there was a marked suppression of spiking activity (Figure 3C). This shape of this response was the same regardless of whether the looming object persisted or disappeared upon reaching its final angle (Figure S2). This response is broadly similar to those of loom detectors found in locust [21–23], pigeon [24, 25], and other animals [26–29]. Further, this response was similar to loom sensitive activity recorded in the ventral nerve cord of Drosophila, whose peak firing rate was correlated with the timing of the flies’ escape response [17].

Many loom detectors exhibit four characteristic features. First, the time of the peak firing rate relative to the expected time of collision is linearly related to the ratio of the object’s size to its velocity towards the animal, l/|v|. As previous studies have described, this linear relationship suggests that these neurons implement an arithmetic multiplication, and act as angular threshold detectors such that the peak response occurs at a fixed delay after the stimulus reaches a given angle [30, 31]. Second, loom detectors respond with the same timing to looming stimuli that originate in different locations of the visual field. Third, for a loom detector, the time of peak firing rate is not affected by the contrast polarity of the object. Finally, a loom detector responds to an approaching object even when its approach does not cause a global change in luminance.

Foma-1 neurons exhibited all four of these characteristics. First, the time of the peak firing rate was linearly related to the ratio of the object’s size to its velocity towards the fly, l/|v| (Figure 3D). Second, these neurons responded identically to looming stimuli that originated in different locations within the visual field (Figure 3E, 3F, and S2). We compared looming stimuli that initiated near the dorsal edge of the visual field with those that began frontally. In these experiments, the edge of the looming dot that began at the dorsal periphery passed through the receptive field in the preferred direction of local motion, while the edge of the looming dot that began frontally passed through the receptive field in the non-preferred direction. If the local motion preferences (Figure S1) were sufficient to account for the loom response, this difference in initial position would elicit opposite responses. However, this was not the case: Foma-1 neurons responded almost identically to these two stimuli, both in their overall firing rate and in the timing of the peak response (Figure 3E and 3F). Moreover, varying the azimuthal position of the frontal stimuli, which would also change the pattern of expansion relative to the local motion preferences, also did not change the response (Figure S2). Third, these neurons responded to looming stimuli that were either brighter than the background (contrast increment) or dimmer than the background (contrast decrement) (Figure 3G). While the firing rate evoked by these two stimuli differed, the timing of the peak response was unchanged (Figure 3H). Finally, these cells responded strongly to the looming approach of a “checkerboard” stimulus comprising squares of alternating light and dark contrast against a gray background (Figure 3I and 3J). As the response of cells to this equiluminant stimulus was the same as that to a looming white object on a gray background (a stimulus that brightens as the object approaches) we infer that Foma-1 neurons can respond to looming cues in the absence of global luminance changes. Based on these criteria, the Foma-1 enhancer trap labels a small group of loom detecting neurons, thus providing a genetic entry point to the dissection of the circuit that underlies this computation.

The angular threshold for a loom detector can be computed from the parameters of the linear fit between peak firing rate and l/|v| [31]. Using this analysis, our data predict an angular threshold of 67.6 ± 2.5° (mean ± s.e.m., see Methods), an unusually large threshold, greater than those of loom sensitive activity in Drosophila and other species [17, 28, 31]. To directly test this angular threshold we presented looming stimuli that ceased to expand before reaching this size, and found that cells did indeed respond to this truncated looming stimulus (n = 3 out of 5 cells responded, Figure S2). Thus, this parameter appears to provide an incomplete description of the response threshold of these particular cells.

To more completely characterize the response to approaching stimuli, we created a random looming stimulus in which a virtual object jittered towards and away from the fly, its velocity being chosen from a uniform distribution every 25ms. This stimulus samples a wide range of approach trajectory statistics, and hence should provide an extensive exploration of a loom detector’s tuning properties. We used spike triggered analysis to compute a Linear-Nonlinear (LN) model of the stimulus-response function [13] (Figure 4A; see Methods). This model uses two functions to predict a cell’s response to the stimulus: a linear filter and a static nonlinearity. The linear filter captures the neuron’s sensitivity to the visual angle of the stimulus as a function of time, while the static nonlinearity captures nonlinearities in the neuron’s response, including its gain and threshold. The Foma-1 neurons responded robustly to the random looming stimulus, and the LN model captured this response well such that the LN prediction matched the firing rate as well as the trial-to-trial noise in the response (Figure 4B, See Methods). This model also accurately predicted the timing and magnitude of the cells’ responses to presentations of a looming stimulus comprising an isolated dot on a direct approach trajectory (as used in Figure 3; data not shown, Pearson’s correlation coefficient 0.53 ± 0.02 (mean ± s.e.m.), n = 14).

To test the ability of this model to capture the characteristic features of a loom detector, we calculated the LN model using a random looming stimulus that originated either at the dorsal periphery or anteriorally in the visual field. Both the linear filters (Figure 4C) and the static nonlinearities (Figure 4D) were indistinguishable between these two conditions, further confirming the position invariance of the neurons’ responses. We also calculated the LN model under both contrast increment and decrement conditions. For both conditions we observed a linear filter with positive polarity (Figure 4E), indicating that the Foma-1 neurons responded to increasing angle of the dot regardless of its contrast.

However, the filter calculated under the contrast decrement condition was slower than that for contrast increment, with a peak sensitivity 20 ± 4 ms after that for contrast increment (mean ± s.e.m., n = 6). Such temporal differences could result from luminance sensitivity differences in motion detecting pathways [32, 33]. The static nonlinearity for the contrast decrement condition reveals a lower gain than for the contrast increment (Figure 4F). This is consistent with the lower firing rate observed when a dark dot approaches with constant velocity (Figure 3G). Finally, the linear filter (Figure 4G) and nonlinearities (Figure 4H) we calculated using a white dot or an equiluminant checkerboard stimulus were very similar. Thus, the LN loom model captures the response selectivity of Foma-1 neurons and can robustly predict their responses to looming stimuli of varied velocities and trajectories.

While our experiments using looming checkerboard stimuli demonstrate that Foma-1 cells can respond to looming in the absence of global luminance change they are insufficient to exclude the possibility that luminance changes might nonetheless contribute to the loom response. To quantitatively examine this possibility, we computed a LN model to a full field flicker stimulus (Figure 5A). The luminance of the monitor was chosen at random from a Gaussian distribution every 15ms, and thus constitutes a pure luminance stimulus with no looming component. The resulting model, therefore, captures the neuron’s sensitivity to luminance (Figure 5A). We then used this luminance model to predict the neurons’ responses to the random looming stimuli, for both bright and dark dots. We calculated the luminance changes in the random looming stimuli, and used this as the input for the luminance LN model (Figure 5B – 5G). We found that the LN_loom model (see Figure 4) more accurately predicted the response of the Foma-1 neurons to the random loom stimuli than the LN_luminance model. In particular, the LN_loom model better captured the predicted firing rate of Foma-1 neurons to the contrast increment stimulus, and better captured both the timing and the firing rate of the neurons’ activity to contrast decrements (Figure 5C, 5D, 5F, 5G). Thus the responses of Foma-1 neurons are mainly induced by increasing angle and not by luminance changes.

As our LN_loom model contains a simple thresholded nonlinearity, we suggest that increasing the firing rate of the loom detecting neurons above this threshold serves to increase the probability of evoking a behavioral response. To test the functional requirements of Foma-1 neurons for behavior, we took advantage of the fact that Drosophila exhibit a robust escape response to looming stimuli, an ethologically relevant cue under many circumstances, including predator evasion. When presented with a looming stimulus, flies take off into flight [15–18]. Based on these previous behavioral studies, we constructed an apparatus that presented a computer-generated looming stimulus to individual flies (Figure 6A and Figure S3). High speed imaging revealed details specific to escape behavior, including raised wings prior to takeoff and a subsequent unstable flight trajectory [15, 16, 18] (Figure 6B, Movie S1). Specifically, 92% (n = 50 filmed takeoffs) of flies exhibited raised wings prior to takeoff. Under these conditions, a loom escape response was elicited from approximately 65–76% of control flies, depending on genotype (Figure 6C). In addition these escapes displayed a linear l/|v| relationship (Figure S3).

To directly demonstrate that these responses corresponded to visually-evoked escapes, we first silenced L2 neurons by expressing shibire^ts, a temperature sensitive synaptic silencer [34], in these cells. L2 neurons are lamina monopolar cells that are immediate post-synaptic targets of photoreceptors, and are critical for detecting the movement of dark edges [19, 32, 33, 35]. Under these conditions, escape behavior was strongly suppressed such that only 11% of flies jumped when presented with the looming stimulus (Figure 6C). Next, we tested whether the activity of Foma-1 neurons influenced this escape behavior by silencing these cells in the same manner. Under these conditions, only 30% of the flies jumped in response to the looming stimulus, a significant decrease relative to control flies (Figure 6C). By comparison with control flies, which typically escaped within the first or second presentation of the loom stimulus, flies with silenced Foma-1 neurons that did takeoff were as likely to escape to any of the six stimulus presentations (Figure S3). Thus, Foma-1 neurons are important for the normal frequency of initiation of the wild-type loom escape behavior.

If increasing the activity of Foma-1 neurons signified the presence of a looming stimulus, targeted stimulation of these cells should induce escape behavior in the absence of visual input. We therefore expressed channelrhodopsin in Foma-1 neurons and used light to selectively activate these cells [36]. To eliminate any photoreceptor mediated visual responses, we rendered flies blind using a null allele of a critical component of the phototransduction cascade, Phospholipase C-β, encoded by the norpA gene [37]. We then illuminated individual flies with blue light to evoke escape responses (Figure 6D and Figure S3). Under these conditions, blue light illumination induced escape responses that were behaviorally similar to those we observed with looming stimuli, with 90% of flies (n = 30) raising their wings prior to takeoff (Figure 6E, Movie S2). Quantitatively, while only 16–21% of control blind flies, depending on genotype, took off within a five second interval, 75% of flies that expressed channelrhodopsin in Foma-1 neurons did so (Figure 6F). Thus, stimulation of Foma-1 neurons was typically sufficient to evoke escape behavior, even when no visual input was present. To test whether such a strong behavioral response was specific to the activation of the Foma-1 neurons, we took advantage of cell-specific Gal4 elements that could target expression of channelrhodopsin to other neuron types (thereby providing specific access to the activation of these cells). Importantly, channelrhodopsin mediated stimulation of the -lobe of the mushroom body, the only other brain region in which Foma-1GAL-4 is expressed, did not significantly increase the frequency of escape response (Figure 6F). Finally, selective activation of L2 neurons was also insufficient to evoke escape responses (Figure 6F). Thus, while these cells are necessary for loom escape behavior, their outputs are not sufficiently specialized to evoke this particular response. These data argue strongly that escapes evoked by stimulation of Foma-1 neurons require the integration of motion signals that are specific to loom, information that is unavailable earlier in the visual pathway.

We recorded the spiking activity of Foma-1 neurons using targeted loose-patch recordings, while visual stimuli, generated on a high-speed monitor, were presented to the fly using two coherent fiber optic arrays. Looming stimuli comprised of either an expanding white or a black dot on a grey background. The visual angle of the dot was given by the equation θ(t)=2tan−1lvl, where l is half of the length of the object and v the velocity of the object towards the fly. For the single approach looming stimulus a single velocity was used for each approach, while for the random looming stimulus, the velocity was chosen from a uniform distribution every 25 ms, such that the object jittered both towards and away from the fly. A Linear-Nonlinear model was calculated through reverse correlation of the neurons’ activity to the visual angle of the looming stimulus. Individual cells were filled by electroporation of Texas Red dextran, and serial optical sections were taken in the intact preparation using a confocal microscope with a water immersion objective.

To evoke loom escape behavior, individual flies were placed on a platform, presented with a single approach looming stimulus, and scored as to whether they took off in response. To block neural activity, Shibire^tswas expressed in either Foma-1 or L2 neurons, and both experimental and control genotypes were warmed to the non-permissive temperature immediately prior to experiments. Optogenetic stimulation was produced by expressing channelrhodopsin in either Foma-1, L2 or mushroom body neurons. Individual flies were placed on a platform, illuminated by 470 nm light with an irradiance of 713 W/m², and scored by the time of their takeoff.

See Supplemental Experimental Procedures for complete methods.