Motion processing streams in Drosophila are behaviorally specialized

Summary

Motion vision is an ancient faculty, critical to many animals in a range of ethological contexts, the underlying algorithms of which provide central insights into neural computation. However, how motion cues guide behavior is poorly understood, as the neural circuits that implement these computations are largely unknown in any organism. We develop a systematic, forward genetic approach using high-throughput, quantitative behavioral analyses to identify the neural substrates of motion vision in Drosophila in unbiased fashion. We then delimit the behavioral contributions of both known and novel circuit elements. Contrary to expectation from previous studies, we find that orienting responses to motion are shaped by at least two neural pathways. These pathways are sensitive to different visual features, diverge immediately postsynaptic to photoreceptors, and are coupled to distinct behavioral outputs. Thus behavioral responses to complex stimuli can rely on surprising neural specialization from even the earliest sensory processing stages.

Introduction

A central challenge in neuroscience is to link sensory experience to motor output at the level of identified cells and circuits. In the context of vision, electrophysiological and psychophysical studies have provided considerable insight into the neural computations that underlie motion, color and form vision (Gegenfurtner and Kiper, 2003; Orban, 2008). However, the neural substrate that transforms retinal signals into changes in animal behavior remains poorly defined. Recent advances in the development of genetic tools for rapidly and reversibly manipulating neuronal activity in the fruit fly open the possibility of using these techniques to identify and define the neural circuits that underlie complex behaviors. To do so, however, requires the development of high-throughput, quantitative behavioral analyses that can adequately explore the stimulus-response relationship, and can be used to conduct forward genetic screens to identify functionally-critical neurons in unbiased fashion. Here we develop such an approach to dissect the neural circuits underlying motion vision. We find that the organization of the motion processing system in flies is strikingly different from that inferred from many previous electrophysiological and behavioral approaches (Gotz and Wenking, 1973; Gotz, 1975; Egelhaaf et al. 2002; Rister et al. 2007).

Motion vision has long served as a well-defined, robust paradigm in which to study neural processing. What functional principles guide how visual circuits in general, and motion circuits in particular, are organized with respect to behavior? On the one hand, it is well established that simple visual cues can guide particular behavioral responses via specialized neural pathways in cases where a broader visual context is unnecessary to interpret the signal. For example, luminance signals received by melanopsin-expressing retinal ganglion cells control the pupillary light response and circadian entrainment through targeted axonal projections to distinct brain regions (reviewed in Peirson and Foster, 2006). On the other hand, how features that require both analysis of local signals, as well as integration into a more global context, are processed with respect to behavior, is less well understood. It is generally thought that such features, for instance, edges, are first extracted by a generic processing step, and later assembled into more complex patterns, in this case shapes or objects, by specialized circuits (Orban, 2008). Only then are the outputs of these circuits used to guide behavioral decisions.

In the context of motion vision, distinct global patterns of motion, such as optic flow, carry information that guides different behavioral decisions (Gibson, 1950). These global patterns must be synthesized from local luminance changes. However, whether local motion signals already contain information to differentiate behavioral responses, and whether visual systems make use of such information, remains largely unknown. On the one hand, it is possible that a uniform processing step first extracts motion cues from each local region of a visual scene, irrespective of behavioral context. Then, multiple, specialized circuits would interpret this common motion signal to guide different aspects of behavior. On the other hand, motion-sensitive behaviors could be served by specialized local motion estimators, each tuned to different stimulus parameters, and coupled to a distinct behavioral response. Evidence in insects and vertebrates supports both possibilities. For example, both a common elementary motion detecting algorithm (Hildreth and Koch, 1987; Reichardt and Poggio, 1976; Buchner, 1984), and multiple processing pathways specific to particular stimulus conditions have been proposed (van Santen and Sperling, 1984; Heisenberg and Buchner, 1977; Franceschini et al., 1989). To understand how tradeoffs between these distinct organizing principles constrain circuit architecture, it is necessary to understand the behavioral contributions of neurons along the full extent of motion-processing pathways.

Orienting responses to visual motion are found in every animal group with non-primitive eyes, providing critical signals for course control and navigation (Srinivasan et al., 1999). Indeed, arthropods, molluscs and vertebrates spontaneously follow, or turn against, the direction of motion cues, depending on their mode of locomotion and on how the stimulus is presented (Hecht and Wald, 1934; Kalmus, 1949; Wells, 1962; Gotz, 1970; David, 1979; Miles and Wallman, 1993). The robustness, ubiquity and apparent simplicity of these optomotor responses makes them powerful paradigms for examining the neural mechanisms of motion detection. In both walking and flying fruit flies, similar relationships between turning reactions and motion parameters across a range of stimulus configurations suggest that a single motion estimate guides DS behaviors (Buchner, 1984; Heisenberg and Buchner, 1977; Gotz, 1964; Gotz and Wenking, 1973; Buchner, 1976). On the other hand, fruit flies appear to process motion signals differently under different light adaptation states (Heisenberg and Buchner, 1977; Pick and Buchner, 1979) and anatomical considerations suggest that multiple motion-sensitive retinotopic pathways segregate by the second or third stage of visual processing (Sinakevitch et al., 2003; Bausenwein et al., 1992; Bausenwein and Fischbach, 1992). As the electrophysiological properties and anatomical connections of many interneurons linking the retinal input to deeper brain layers are unknown, the neural substrate of motion estimation remains to be defined. Thus, the relative contributions of a single motion estimator toward different behaviors, or the roles of potentially parallel pathways are unclear.

For an animal to respond to motion in a particular direction, direction-selective (DS) neurons must guide orienting behavior. Such neurons have been described in vertebrates and invertebrates (Clifford and Ibbotson, 2002; Egelhaaf et al., 2002). In primates, signals from DS neurons can drive perceptual decisions about motion direction (Salzman et al., 1992; Celebrini and Newsome, 1995). In insects, DS neurons tuned to global patterns of motion resemble vertebrate neurons in their tuning properties and can alter DS behavior when microstimulated (Egelhaaf et al., 2002; Blondeau, 1981). In both insects and primates, ablations of brain regions where DS neurons are found disrupt at least some responses to motion (Newsome et al., 1985; Heisenberg et al., 1978; Geiger and Nassel, 1981; Hausen and Wehrhahn, 1990). However, in no context do we understand how direction selectivity emerges in the nervous system to guide behavior.

To begin a systematic dissection of these pathways, we combined the forward genetic techniques available in the fruit fly with a quantitative analysis of its robust DS behaviors to identify motion-sensitive neurons required for specific behaviors in an unbiased fashion. This required development of a sensitive, high-throughput, quantitative assay of motion-evoked behavior. Using this novel approach, we define two distinct behavioral responses to motion, dominated by DS changes either in the rotation of the fly’s body, or in its translatory movement. Most moving stimuli modulate both responses, but do so to differing extents. That is, changes in rotation and translation are sensitive to different combinations of visual stimulus parameters. We then demonstrate, using a forward genetic screen, combined with a candidate neuron approach, that these behavioral responses are at least partially genetically separable. We find that specific neural populations are differentially coupled to each behavioral response. Taken together, these results demonstrate that most motion stimuli simultaneously activate multiple motion processing streams, corresponding to at least partially distinct neural circuits. Remarkably, these circuits separate early in visual processing, in the neurons immediately post-synaptic to photoreceptors. In this way, an initial genetic dissection has shed light on how the pathways of motion processing are organized to inform behavior, and provides a quantitative paradigm for applying similar approaches to other sensorimotor integrations.

Results

A high-throughput assay for motion-evoked behavior

Visual behaviors observed in both flying and walking Drosophila are likely to be ethologically relevant (Carey et al., 2006). Therefore, to apply a forward genetic screen to the dissection of motion processing, we constructed an apparatus that allowed presentation of arbitrary visual stimuli to large populations of walking flies, while capturing the behavioral responses of each individual (Figure 1A–C; Suppl. Figure 1). This system enabled us to monitor the immediate responses of single animals to dynamic stimuli with the statistical power necessary to analyze subtle changes in behavior. Briefly, populations of flies walking in glass test tubes were presented with a visual stimulus from below, and filmed from above (Figure 1A). Individual fly trajectories were tracked in real time (Figure 1B); all datasets represent 3,000–30,000 responses of single flies at each experimental condition. In this system, animals react to stimuli independently: isolated flies produce similar responses to motion as flies tracked within a population, and cross-correlation of responses in the population revealed no stimulus-specific interactions between animals (Suppl. Figure 2). While flies were free to move within the test tubes, we examined only those flies that were upside-down, observing the stimulus from a nearly constant distance, using the same dorsal portion of the eye. For these flies, optical distortion along the axis of stimulus motion caused by the curvature of the test tube is negligible.

Figure 1. Flies exhibit two distinct responses to motion.

(A) Schematic illustration of the experimental apparatus. Flies in test tubes view motion underneath, and are filmed from above. (B). A single fly trajectory. Fly heading is coded as either with (blue) or against (red) the axis of visual motion. (C) Trial Structure. Motion cues in alternating directions are presented in short epochs, interspersed with random dot movement (noise). Trajectories are then aligned precisely in time to stimulus transitions. (D) Schematic representations of a sparse stimulus and a dense stimulus, moving from left to right (arrows). (E) Changes in heading bias elicited by long presentations of sparse stimulus (black line). Noise (white area) was followed by two periods of coherent motion in alternating directions (arrows in shaded areas). (F) Changes in heading bias elicited by long presentations of a dense stimulus (gray line). Trial structure as in (E).

As a visual stimulus, we employed variants of the dynamic random dot stimulus (Newsome and Pare, 1988), adapted to the relatively poor spatial resolution of the fly (Figure 1D; Suppl. Movies 1, 2). In these stimuli, each dot moves only a short distance before disappearing (re-appearing at random on the screen), thus minimizing opportunities for flies to track single visual features without computing overall motion. Such stimuli allowed presentations of spatiotemporal frequency patterns different from those possible using periodic stimuli, permitting a wide range of conditions to be systematically explored. To verify that all stimuli we used are sampled appropriately by the fly eye, we modeled the first stage of the fly visual system. As our interest lay in direction-selective behaviors, we confirmed that the particulars of how the fly eye registers these stimuli cannot reverse the perceived direction of visual motion relative to its true direction (Suppl. Figure 3; Suppl. Discussion). Finally, we designed our experiments such that the behavioral responses to a test stimulus could be directly compared to a behavioral baseline identical in all ways except for the presence of coherent motion. Each experiment alternated epochs in which all dots were displaced in the same direction (designated ”motion”), with epochs in which dots were displaced in random directions, (designated “noise”; Figure 1C). Thus, the only change in the stimulus between baseline and experiment was the coherence of dot movement, and hence the motion signal; contrast and average luminance were held constant. In all subsequent analyses, we focus on changes in behavior caused by motion, relative to a baseline during the noise period.

Flies exhibit two distinct responses to visual motion

We reasoned that if multiple motion processing channels contribute to the fly's optomotor response, we should be able to excite them differentially under some stimulus conditions. We began by systematically varying the velocity, contrast, luminance, spatial density, and coherence of the visual stimulus. The sparsest or slowest stimuli caused flies to move against (opposite to) the direction of stimulus motion (Figure 1E; Figure 2A). This response direction was previously observed by moving stripe patterns underneath flies or surrounding them on all sides (Hecht and Wald, 1934; Kalmus, 1949; Gotz, 1970). Moreover, the response we observed was also consistent with previous studies in its tuning to critical motion stimulus parameters, and peaked at stimulus velocities the fly would experience itself while turning through a stationary world (Suppl. Figure 4, Figure 2C), suggesting that our stimuli activated neurons within a physiologically relevant range. For brevity, we refer to this behavior as the "against" response, and denote stimuli that cause it "sparse." Surprisingly, we also found stimuli that evoked movement in the same direction as the stimulus (a "with" response), marked by distinct kinetics and tuning properties (Figure 1F; Figure 2B). This response was strongest on dense, relatively fast stimuli ("dense" for brevity), and has not been previously reported in freely walking flies. These observations raised the possibility that a “with” response reflected the activity of a motion processing stream tuned differently than the stream underlying an “against” response.

Figure 2. Different changes in translation and rotation underlie each behavioral response.

(A, B) Heading bias ±2*s.e. elicited by a short (800ms) presentation of either a sparse stimulus (A, gray bar) or a dense stimulus (B, gray bar), moving from left to right (arrows). (C) Histogram of the full distributions of translational and angular velocities observed in the population. From this distribution of wild-type flies, two thresholds in translational velocity were set. The lower threshold defined stopped flies as those that move slower than 0.26 cm/sec. Another threshold is set at the mode of walking speeds (1.9cm/sec), defining the boundary between fast and slow walking speeds. Turning was defined as angular velocities above the 100°/sec. (D). A translation index (TI) reports change in the fraction of flies moving at fast walking speeds. A rotation index (RI) reports change in the fraction of turning flies. A turn direction index (TD) was computed from the fraction of flies turning with, relative to those turning against the direction of motion (arrow). (D, E) TI ±2*s.e. in time for the sparse (D) and dense (E) stimulus. Motion was presented moving from left to right (arrow), flies were categorized as facing either with (blue lines), or against (red lines), and the stimulus was either sparse (D, F, H) or dense (E, G, I). (F, G) RI ±2*s.e. in time. (H, I) TD ±2*s.e. in time.

Changes in heading bias emerge through distinct changes in fly translation and rotation

To distinguish contributions of potentially different motion processing streams, we asked what immediate changes in flies' movements lead to either response over time. Fly movement can be completely described by the translational and rotational velocities of each individual (Figure 2C). One might imagine that opposite responses to different stimuli simply result from reciprocal changes in movement. For example, flies could turn to follow one stimulus or turn to move against the other. Alternatively, opposite responses could emerge due to different effects of the two stimuli on translation, rotation, or both. To capture these effects, we defined three metrics that measured changes in these velocities across the population. As flies turned and walked forward at stereotyped angular and linear velocities, respectively, we defined a range in each velocity representing "turning" versus "not turning" and "fast" versus "slow" forward movement. Motionless flies were excluded from analysis, as they contributed no directional signal. To calculate indices from these distributions, we measured the stimulus-induced change in the frequencies of flies above and below each threshold, in each distribution, independently (see Materials and Methods). The translation index (TI) captured the relative increase or decrease in the speed of forward movement, with a TI of −0.2 indicating, for example, that the number of flies walking fast (as defined in Figure 2C) decreased by 20% after stimulus onset. Similarly, the rotation index (RI) measured whether turning increased or decreased relative to the pre-stimulus baseline. By analyzing our data using different thresholds to define high versus low velocities, we found that our description of behavior using these metrics was not sensitive to precisely where the thresholds were defined. Moreover, analysis of the stimulus-induced changes in the full translational and angular velocity distributions (see materials and methods) produced identical conclusions. Lastly, since turning could be in one of two directions, relative to the direction of stimulus motion, we also computed a turn direction index (TD) that measured the prevalence of turns into alignment with the stimulus direction, versus out of it, also as a fractional change from baseline (Figure 2C).

To test whether "with" and "against" responses were generated by changes in different aspects of movement, we asked how "sparse" and "dense" stimuli affected translation and rotation over time. In order to produce a direction-selective response, the behavior of flies oriented in the same and opposite directions as stimulus motion must be affected in different ways. Hence, we analyzed the translation and rotation indices at "with" and "against" orientations separately. This classification of flies into just two orientations relative to stimulus motion captured as much directional information as the flies themselves exhibited in their response (Suppl. Fig. 5). That is, flies did not align themselves precisely with or against either stimulus in the course of response, but only roughly biased their heading direction.

Sparse and dense stimuli, which biased headings in opposite directions, decreased forward movement at both orientations, although to different extents (Figure 2D,E). Translation of flies oriented in the "with" direction decreased three times as much on a sparse stimulus, as it did on a dense one, while translation at opposite orientations was affected almost identically on both stimuli. As a result, “with” flies slowed more than “against” flies on the stimulus that caused an “against” response, while “with” flies slowed less than “against” flies on a stimulus that caused a “with” response. Turning was also suppressed by both stimuli causing flies to move in straighter paths (Figure 2F, G). On both stimuli, turn suppression was greatest at “with” orientations. Compared to a threefold difference in TI, however, turns at "with" orientations were suppressed only about two times as strongly on the sparse stimulus as on the dense one. RI also followed a somewhat different time-course: on the sparse stimulus, following an initial peak of turn suppression, turning rebounded slightly and remained at the new level for the duration of the stimulus. At the “against” orientation, only relatively weak stimulus onset and offset responses were observed in RI (Figure 2F, G). Finally, when flies did turn, they tended to turn into alignment with the stimulus direction on both stimuli (TD>0, Figure 2H, I).

Hence, between two stimuli that cause opposite responses over time, only flies at the “with” orientation differed in their behavior. Notably, contrary to the models of response to motion in flies that assume optomotor response is aimed at matching stimulus velocity (Gotz 1975), the strongest direction-selective changes on either stimulus were inhibition of movement, that is, inhibition of forward movement, or of turning. At the same, “with” orientation, these two changes in behavior must have opposite consequences for the longer-term direction of migration. That is, preferential slowing in one direction will lead to the diffusion of flies in the other direction over time; while preferential straightening of trajectories at one orientation will cause flies to move longer in the same direction, by preventing them from turning into other directions. Other aspects of behavior being equal, the balance of TI and RI inhibition at the "with" orientation would determine in which direction flies displace over time. Indeed, the effects on TI and RI at the "against" orientation were nearly constant between the two stimuli, even though these two stimuli cause very different heading biases. In addition, while turning (TD) on the dense stimulus favored movement with the stimulus direction regardless of orientation, turn bias did not explain why movement reverses direction on the sparse stimulus, as TD was either positive or only weakly negative, depending on orientation, on the sparse stimulus (Figure 2H, I). Hence, the overall response direction must either be a consequence of how much flies slowed down at the “with” orientation relative to the stimulus-invariant slowing at the “against” orientation, or a consequence of how much flies slowed relative to how much they suppressed turning at the same, "with" orientation, or both. That is, response direction is either determined by relative changes in translation of “with” versus “against” facing flies, or by relative changes in translation versus rotation amongst “with” flies, or a combination of the two effects. In either case, the two stimuli had different effects on short-term changes in fly movement.

Rotational and translational responses are differentially sensitive to stimulus parameters

To determine which aspects of behavior are modulated systematically by stimulus parameters, leading to one or another response direction, we examined the transition from an "against" to a "with" behavior. As described above, changes in stimulus speed and density alone can determine the direction of response (Figure 3A,B). Broadly speaking, heading was biased most toward the "against" response at the slowest or sparsest stimuli, while denser or faster stimuli were biased more toward a "with" response (Suppl. Figure 6). Consistent with the fact that these stimuli affect different aspects of fly movement, "against" and "with" responses occur with distinct kinetics and are not mutually exclusive, as intermediate speed and density stimuli elicit both an early “against” phase as well as a later “with” phase (Figure 3B). Quantitatively, though, modulation of heading bias by stimulus parameters was not systematic (Suppl. Figure 6), reflecting the fact that an aggregate of changes in different aspects of behavior contribute to overall heading bias (Figure 2). Rather, it was modulation of fly rotation and translation that did vary systematically with stimulus parameters, although, these two aspects of movement were not sensitive to the same parameters, and their modulation was orientation-dependent. In particular, motion speed, but not stimulus density, strongly modulated translational movement (Figure 3C), while both stimulus parameters, in combination, modulated rotation in flies oriented in the same direction as stimulus motion (Figure 3D). By contrast, flies that experienced motion in the opposite direction did not modulate relative amounts of translation or rotation consistently with changing stimulus parameters, although some speed tuning is apparent in RI (Figure 3F, G). The monotonic, orientation-specific changes in two aspects of movement represent the effects of a wide range of stimulus conditions, demonstrating that neither change is specific to stimuli that evoke the strongest "with" or "against" responses. Unlike measures of the degree of translation and rotation, turn direction varied with stimulus density but not stimulus speed at "against" orientations, and showed a similar pattern at "with" orientations (Fig. 3E, H). Interestingly, turn direction at either orientation did not predict heading bias consistently, arguing that turn direction bias represents an additional aspect of behavior that contributes to, but does not determine, the response direction. On sparser stimuli, flies may orient to individual dots or dot clusters, including those that require turning against the stimulus direction. Such a response does not necessarily require the computation of stimulus motion. We therefore used stimuli of moderate density as the "sparse" condition in subsequent experiments to minimize this effect.

Figure 3. Different behavioral responses are sensitive to distinct stimulus parameters.

(A) Stimuli spanning a range of speeds and densities, at constant contrast and luminance, between those optimal for "against" and "with" responses. Speed ranges as seen throughout the behaviorally-relevant vision field of a fly: magenta (420–920deg/sec), yellow (700–1540deg/sec) and green (980–2150deg/sec). Density is expressed as a percentage of the screen covered by moving dots. Lighter – higher, darker, lower density. (B) Heading biases ±2*s.e. evoked by 800ms of 3 stimuli (grey bar), separated by zero coherence noise of equivalent contrast and density (white regions). (C, F) TI (D, G) RI (E, H) TD (C–E). Flies oriented with the stimulus direction (blue). (GF–H). Flies oriented against the stimulus direction (red). Error bars: 2X S.E.M.

In sum, different stimulus parameters systematically modulate forward movement and turning at "with" orientations, while at opposite orientations, they do so inconsistently, weakly, or both. Therefore, while changes in behavior at both orientations contribute to the heading bias, we exclude "against" oriented flies from subsequent analysis. The systematic modulation of movement at "with" orientations by motion stimulus parameters that modulate overall behavior suggests that response at these orientations is the most sensitive readout of direction-selective, motion processing circuits. The measure of heading bias, on the other hand, while qualitatively informative, was at best a secondary readout of systematic stimulus effects on different aspects of movement. Taken together, these results show that the different components of the fly’s direction-selective response to motion, namely the translational and rotational responses of “with” facing flies, are sensitive to distinct stimulus features, raising the possibility that they are differentially controlled by the underlying neural circuitry.

Neural contributions to specific behavioral responses diverge early in visual processing

To identify circuits that may contribute selectively to different aspects of direction-selective behavior, we first asked whether inactivation of neurons at the earliest anatomical stages of the visual system affects all aspects of the behavioral response equally. We took a candidate-neuron approach, using available Gal4 drivers specific to single cell types to express the temperature sensitive synaptic silencer shibire^ts (Kitamoto 2001). As a control, we used a well-characterized Rh1-Gal4 driver to inactivate R1–R6 photoreceptors (Mismer and Rubin, 1987). Consistent with previous studies (Heisenberg and Buchner, 1977), R1–R6 provide critical signals for responses to motion, as silencing these cells severely compromised behavior on both sparse and dense stimuli (Fig. 4). Indeed, all measurable aspects of DS response to sparse (Figure 4A, C, E) and dense stimuli (Figure 4B, D, F) in translation and rotation indices were eliminated by R1–R6 silencing.

Figure 4. Genetic dissection of visual motion processing.

(A, B). T.I. (C, D) R.I. (E, F) T.D. An 800ms presentation of either a sparse (A, C, E) or a dense (B, D, F) stimulus was used. Control genotypes (Gal4/+ and shi/+; gray bars) and experimental genotypes (Gal4/+; UAS shi(ts)/+; blue bars). *** denotes p<.001; ** denotes 0.01>p>0.001. All significance comparisons were against the least favorable control. All experiments were done at the restrictive temperature (34°C). Error bars: 2X S.E.M.

Next, we silenced one of the immediate postsynaptic targets of R1–R6 photoreceptors, the L2 lamina monopolar cell (LMC), using a driver expressed specifically in those cells (Mollereau et al., 2000; Gorska-Andrzejak et al., 2005; Figure 6A). To our surprise, L2-silenced flies moved robustly "with" the direction of all stimuli tested, including sparse stimuli that normally cause an "against" response (data not shown). L2 silencing compromised translational response consistently on all conditions, but preserved different aspects of rotational responses depending on the specific stimulus used (Figure 4, Figure 7, Suppl. Fig. 7). On the sparse and dense stimuli, L2 silencing eliminated suppression of both forward movement and turning (Figure 4A–D), but did not eliminate a directional bias in turns (Figure 4E,F). In particular, L2-silenced flies displayed a robust, aberrant turn bias into alignment with stimulus motion, irrespective of response direction in controls (Figure 4E, F). This remaining directional aspect of behavior explained the long-term heading bias of L2-silenced flies. However, this finding failed to elaborate the contribution of L2 to translation and turn suppression, as both were equally compromised. To explore this issue further, we reasoned that while overall responses to sparse or dense stimuli may be dominated by either translation or rotation responses, the contribution of neurons toward one or the other behavior response might be detected more sensitively when the magnitudes of rotation and translation responses induced by the stimulus are approximately balanced. Hence, instead of comparing two conditions in which both the stimuli and aspects of behavior differ, we tested L2-silenced flies on a stimulus that evoked roughly balanced "with" and "against" responses, separated in time (Figure 3A, B). On such an “intermediate” stimulus, L2 silencing eliminated all response in translation as on other conditions, but preserved almost half of the normal turn suppression and retained a robust turning bias, as before (Suppl. Figure 7). Hence, while L2 was strictly required for translation response under all conditions, it was not absolutely required to suppress rotation under some conditions, or to turn directionally under any conditions. These findings imply that there must be at least one motion processing pathway downstream of R1–R6 that can provide directional signals to turning behaviors in the absence of L2’s activity. Conversely, direction-selective translational responses to motion stimuli are completely dependent on L2 function.

Figure 6. Foma-1 labels a small, specific group of neurons in the optic lobes and mushroom bodies.

(A) Schematic depiction of the fly brain. (B). Image of the neurons labelled by Foma1 Gal4 (red), and photoreceptor terminals (green), in the medulla. Two groups of neurons are labeled in all Foma-1 flies: a large cluster of neurons in the mushroom body (arrows), and two small clusters of 3–4 neurons in each optic lobe, with projections confined to the lobula plate (arrowheads). (C, D). High magnification views of the optic lobe neurons labelled by Foma1 (red), combined with a neuropil marker (green), from two orientations. (E–G). Optic lobe neurons labeled by Foma1 (red), expressing the dendrite marker DSCAM-GFP (green). (H–J) Optic lobe neurons labeled by Foma1 (red), expressing the axon marker SNB-GFP (green). (K) Image of a Foma1/+, shi(ts)/+ brain, treated with HU. Foma-1 cell bodies in optic lobe (arrowhead) can be seen in one lobe (and are out of focus in the other); some of the processes (hash marks) can be seen in both lobes. This brain retains a few labeled cells in the central brain (arrows); many others had none.

Figure 7. Visual pathways underlying translational versus rotational behavioral responses diverge early.

(A–D) T.I. in time (A, C), or integrated (B, D), following a very brief (150ms, gray bar) presentation of a sparse stimulus. (E–H) control R.I. in time (E, G) or integrated by genotype (F, H). (I–L) control T.D. in time (I, K) or integrated by genotype (J, L). Contrast increment (A, B, E, F, I, J) or a contrast decrement (C, D, G, H, K, L).Significance of effect was tested by bootstrap (see methods). Error bars: 2X S.E.M.

Forward genetic analysis identifies neurons with highly specific motion deficits

To identify neurons required for motion processing, we conducted a forward genetic screen of 400 novel enhancer trap lines, using the GAL4/UAS system to express shibire^TS in randomly selected sub-populations of neurons. Flies were raised at permissive temperatures, and shifted to restrictive temperatures as adults, immediately prior to behavioral testing. By presenting visual stimuli to populations of flies bearing different enhancer trap lines, we identified lines that displayed behavioral deficits only when specific motion stimuli were presented, but which otherwise moved and behaved normally when presented with stationary visual stimuli. This behavioral selection enriched significantly for enhancer trap lines with optic lobe expression (data not shown). However, many such lines were broadly expressed and were excluded from further analysis. One line, designated Failed optomotor assay-1 (Foma1), had both a specific behavioral deficit, and was expressed in a very small number of neurons in the optic lobe.

At the restrictive temperature, Foma1 flies expressing shi^TS moved normally in the absence of motion stimuli, responded robustly to a sparse motion stimulus, but were significantly impaired in their response to the dense stimulus. In particular, Foma1-silenced flies failed to move "with" the dense stimulus (data not shown). Inactivation of neurons in Foma1 preserved 80% of response in translation on the sparse stimulus, and did not affect translation response on the dense stimulus (Figure 5A, B). By contrast, turns were suppressed more strongly on the sparse stimulus, and only weakly, if at all, on the dense stimulus relative to either control (Figure 5C, D). Finally, Foma1/+;shi^TS/+ flies lost almost all normally elicited directionality in turning, reversing the remaining small turn bias relative to their controls on both sparse and dense stimuli (Figure 5E, F).

Figure 5. Disrupting a small group of interneurons causes a specific deficit in behavioral responses to visual motion.

(A, C, E). Sparse stimulus. (B, D, F). Dense stimulus. (A, B) T.I. (C, D) R.I. (E, F) T.D. Control genotypes (Foma1/+ and shi/+; gray bars) and experimental genotypes (Gal4/+; UAS shi(ts)/+; blue bars). An additional experimental genotype was UAS shi(ts)/+ treated with hydroxyurea (HU). *** denotes p<.001; ** denotes 0.01>p>0.001. * denotes 0.05>p>0.01. n.s. not significant. All comparisons were against the least favorable control. Red asterisks/lettering denote comparisons between Foma1/+; UAS shi(ts)/+ compared to shi(ts)/+ HU. All experiments were done at the restrictive temperature (34°C). Error bars: 2X S.E.M.

Thus, Foma1 had a stronger effect on rotation than translation on the dense stimulus, and modest effects on both movement aspects on the sparse stimulus. This stimulus-specific, differential effect argues that the Foma1 deficits do not reflect non-specific impairments in visual function, such as contrast sensitivity, as stimulus contrast modulates translation and rotation proportionally in control flies (data not shown). In addition, as the turning response was greater in Foma-1 than controls on one stimulus condition, and smaller on another, the behavioral deficit cannot reflect a stimulus-independent impairment in the ability of flies to turn. Thus, Foma1 causes specific deficits in a direction selective response.

Foma1 was expressed in two dominant groups of neurons (Figure 6). In the visual system, the driver labeled a small cluster of 3 neurons per optic lobe, each with a broad arbor confined to specific layers within the lobula plate (Figure 6A–D), with at least one neuron extending a process to the contralateral optic lobe. Intriguingly, morphologically similar neurons known in larger Diptera are thought to be critical to visual motion processing (Egelhaaf et al., 2002), and have been activity labeled in Drosophila using optomotor stimuli (Bausenwein et al., 1990). In addition to this specific expression in the visual system, Foma1 also labeled 2 clusters of mushroom body (MB) neurons comprising mostly the alpha'/beta' cell types (Figure 6B). However, inactivation of these neurons is unlikely to contribute to the stimulus-specific Foma1 deficits, as chemical ablation of the entire MB had only a weak effect on translational responses (Figure 5B), and no effect on turning responses (Figure 5D, F), evoked by the dense stimulus. It also produced a distinct phenotype unlike that of Foma1 on the sparse stimulus (Figure 5A, C, E). By expressing lacZ from the Foma1-Gal4 driver in flies used for behavior, we confirmed that our HU treatment eliminated most of the driver's expression in the mushroom bodies, while preserving expression in the lobula plate (Figure 6K). Thus, HU treatment selectively eliminated most of the mushroom body neurons that would be silenced in the Foma1 line, without reproducing the Foma1 phenotype. Thus, inactivating the three lobula plate neurons in Foma1 likely causes its characteristic visual behavior deficits.

In an effort to characterize the lobula plate neurons labeled by Foma1 further, we costained the Foma1 driver with dendritic and pre-synaptic markers (Figure 6E–J). Both major and minor branches of the lobula plate arbors are labeled in a punctate pattern with the dendritic marker Dscam::GFP (Figure 6E–G; Wang et al., 2004), while the presynaptic marker Snb::GFP could only be resolved in the major branches of these arbors (Figure 6H–J; Estes et al., 2000). Although the morphology of lobula plate interneurons in Drosophila is not completely characterized, these findings are consistent with lobula plate interneuron properties in other Diptera, in which wide-field arborizations in the lobula plate can represent both dendritic and axonal processes, either of which can make electrical contacts with other neurons (Farrow et al. 2005).

On the whole, Foma1 deficits complement those of L2. Whereas L2 was strongly required for stimulus-specific slowing under all circumstances, Foma1 was weakly required for slowing on some conditions, but not necessary for slowing at all on others. Across a wide range of stimuli, L2 silencing preserved directional response only in aspects of rotation: either turn bias, turn inhibition, or both. L2 is thus required less strictly for rotation than for translation. Foma1, by contrast, is required more strictly for rotational aspects of response in a stimulus-specific manner. As photoreceptors R1–R6 are required for all aspects of behavior under all stimulus conditions, these phenotypes argue that rotational and translational responses are at least partially separable by genetic criteria, diverging downstream of R1–R6. Most surprisingly, L2, a neuron type immediately post-synaptic to photoreceptors, already carries information more essential for some aspects of response to motion than for others.

Different motion processing streams respond to different visual information

R1–R6 photoreceptors make synaptic connections with three lamina monopolar cells, designated L1–L3 (Meinertzhagen and Hanson, 1993). Although L1 and L2 receive anatomically identical photoreceptor inputs, all three cell types differ in their post-synaptic partners. The stimulus-dependence of the rotational response in L2-silenced flies could therefore reflect the differential processing of each stimulus toward different aspects of response either at the level of L2, or by downstream neurons along pathways that split in the lamina. As lamina monopolar cells are thought to extract information about stimulus contrast (Srinivasan et al., 1982), we tested the contributions of R1–R6 and L2 to the sparse stimulus response under different contrast conditions. We reasoned that if motion pathways diverged in the lamina and had different contrast sensitivities, they would contribute differentially toward translation and rotation under some contrast conditions. In this case, inactivating L2 should have different effects on behavior under specific contrast conditions, using otherwise identical stimuli. If, on the other hand, contrast processing in the lamina does not contribute toward differentiating stimuli for one aspect of response versus another, inactivation of L2 under two different contrast conditions where the responses of control flies are comparable should affect both translational and turning behaviors, under both conditions, equally.

To do this experiment, we systematically varied the contrast of the visual stimulus to find two contrasts that produced responses of comparable magnitude in both translational and rotational aspects of response. We chose two identically "sparse" stimuli, one of moving dots that were darker than background (contrast decrement), and the other, of dots that were lighter than the background (contrast increment). The magnitudes of both translational and rotational responses were roughly matched between the two stimuli (Figure 7). To ensure that potentially separable aspects of response were not confounded during the course of a longer stimulus presentation, we presented flies with brief motion pulses which revealed distinct phases in all response metrics (Figure 7A, C, E, G, I, K). We found that L2's contribution was not specific to any one phase of response (data not shown), and therefore present summary data from just the first, representative response phase.

As a control, we confirmed that inactivation of R1–6 photoreceptors abolished nearly all response components under both stimulus conditions, leaving only a small, aberrant turning response on the decrement stimulus that may reflect incomplete inactivation (Figure 7B, D, F, H, J, L). L2, however, was required differentially between the two contrast conditions, for different aspects of response. The contrast decrement stimulus produced the stronger phenotype. Here, silencing L2 abolished almost all responses, producing a phenotype similar to that of R1–R6 silenced flies (Figure 7D, H, L). However, on the contrast increment stimulus, silencing L2 abolished all translation response, but left almost half of the normal turn suppression intact (Figure 7B, F). Moreover, L2 silencing did not diminish directionality in the remaining turns, enhancing it somewhat instead (Figure 7J). Although the magnitudes of control responses were not perfectly matched between increment and decrement stimuli, it was the stimulus that elicited a stronger response in controls that was most broadly compromised by L2 silencing. This demonstrates that the differential effects of L2 on translation versus rotation do not reflect quantitative differences in the strength of response itself, but rather are a function of the specific stimulus conditions used. Thus, L2 participates in the motion processing stream that guides rotational responses to different extents depending on contrast polarity, but is required for translational responses under all circumstances. This finding suggests that visual contrast information already differentiates the contributions of LMCs to particular behavioral responses.

Discussion

We have developed a high-throughput system that is amenable to forward genetics, provides a sensitive, quantitative analysis of behavior, and easily permits detection of very subtle behavioral changes. Using this system, we can detect small deficits in behavior due to neuronal disruption in the course of a short experiment. Moreover, we can parameterize the stimulus-behavior relationship efficiently, allowing neural function to be described precisely in the context of a freely behaving animal. Combining a forward genetic approach, targeted neuronal disruption, and extensive stimulus parameterization, we have demonstrated by double-dissociation that at least partially non-overlapping circuits diverge in the early stages of the visual system to affect translational or rotational responses to motion. This finding is incompatible with the notion that a common motion processing step guides all DS behaviors, and with the more general view that sensorimotor decisions are typically informed by sensory cues that are first generically processed without regard for their consequences to behavior. The fly visual system is likely to be one of the first contexts in which the organization of complete pathways, from sensory input to behavior, is understood at the level of identified cells and circuits. Our results provide a behavioral framework for constraining the neural computations that underlie motion vision, and raise the possibility that many, if not all, visual stimuli undergo behaviorally specialized processing early in the visual stream. It will be interesting to examine whether this general principle also holds true in more complex nervous systems.

Understanding the behavioral mechanisms underlying the optomotor response

The optomotor response has long been a paradigm for examining the neural mechanisms underlying motion detection. Both behavioral and electrophysiological studies have used this behavior, and the stimuli that promote it, to define one of the central models of motion processing, the Hassenstein-Reichardt correlator model (Reichardt and Poggio, 1976; Buchner, 1984). However, this response varies substantially in different contexts, having opposite sign, depending on precisely how the stimulus is presented, and on how the animal is allowed to move (Hecht and Wald, 1934; Kalmus, 1949; Gotz, 1970; David, 1979). Here we uncover novel aspects of this classical behavior, synthesizing previously discordant observations in the context of a single behavioral paradigm. We find that in freely walking flies, the long-term direction of migration is a secondary consequence of rapid, direction-selective changes in specific aspects of individual animal movements (Figure 8A). We find that motion seen in the same direction as stimulus motion, back-to-front across the retina, effects two types of immediate, sustained changes in movement, depending on stimulus characteristics. In particular, sparse, relatively slow stimuli strongly suppress both forward movement, and turning. The net effect of these changes is dominated by suppression of forward movement in one direction, as flies move in the opposite direction over time. Conversely, straightening of trajectories by turn suppression dominates on different stimuli, stabilizing the headings of flies facing in the same direction as the stimulus. In addition, all but the sparsest stimuli bias remaining turns in the same direction as stimulus motion. All of these effects are elicited in different proportions by different stimuli, spanning a wide range of stimulus conditions, suggesting that most motion cues normally encountered by the fly will modulate multiple behavioral outputs simultaneously (Figure 8B). These results suggest an explanation for the previously observed differences in the optomotor responses in other experimental systems, in that the precise balance between the animal’s translational and rotational responses, and how they were modulated by the stimulus, likely varied with experimental context.

Figure 8. Summary and Model.

(A) A directional response over time results from selective slowing of flies at one orientation (top), or selectively stabilized headings at a different orientation (bottom) with respect to the long-term response direction. Relative to the direction of stimulus motion, both changes take place at the same orientation, where back-to-front motion is seen across the retina. B. Translation and rotation changes in behavior are modulated differently by two stimulus parameters. C. Preferential activation of an L2-dependent pathway promotes translational changes in response to visual motion; activation of a Foma1-dependent pathway promotes rotational changes preferentially. Both pathways depend on photoreceptor type R1–6, but diverge immediately postsynaptic to these cells.

The mechanisms of response we have described are different from mechanisms thought to guide the optomotor behavior of immobilized animals, either walking or flying, even though these behaviors follow the same relationships with stimulus parameters as in our paradigm (Buchner, 1984; Gotz and Wenking, 1973; Gotz, 1975). This discrepancy is significant, as the dominant model of optomotor response in freely moving animals was extrapolated from responses to rotating visual patterns by immobilized animals. In particular, this model proposed that a goal of the fly motion system is to achieve "equilibrium" of rotational cues between the two eyes (Gotz, 1975). This model predicted that the only direction selective responses in freely walking flies would be limited to rotations aimed at balancing bilateral optic flow. Our findings do not bear out this prediction. We find that both rotational and translational aspects of movement are modulated in a direction-selective way, and that this modulation is due to distinct neural processes, as inactivation of specific neurons leads to deficits that affect either translational or rotational responses to motion in different ways. We also find that while most stimuli bias the direction of turns into alignment with the stimulus, a change in behavior consistent with the equilibrium model, a prominent, independently tuned effect of all stimuli is to suppress turning in all directions. This response is not consistent with a control mechanism to achieve a bilateral equilibrium of velocities, as such a model would predict that turns that increase optic flow are suppressed, while those that decrease it are promoted. In sum, our ability to observe direction-selective changes in all aspects of behavior of freely moving animals has revealed prominent features of response to motion that were not detected in earlier studies using immobilized animals. Moreover, while a different approach examined optomotor response in free-walking flies, DS changes in all degrees of freedom of movement were not examined, focusing only on the aspect of response we measure as turn direction bias (Strauss et al., 1997). Indeed, using this metric we find the same tendency of flies to turn with most stimuli as previously described. Our results are also consistent with earlier studies that found strong behavioral responses to back-to-front motion in freely walking, and freely flying Drosophila (Hecht and Wald, 1934; Kalmus, 1949; David, 1979). Thus, our findings are not at odds with previous results, although they do not support the model derived from them.

Defining the structure of the motion detection system

The precise nature of the circuit underlying motion detection is unknown. Electrophysiological and behavioral studies in immobilized animals have proposed a circuit in which a single, common elementary motion detection step early in the visual system extracts local motion cues, which are then assembled at later neural stages into complex, ethologically relevant patterns that control specific behavior outputs in different contexts (Hildreth and Koch, 1987; Reichardt and Poggio, 1976; Buchner, 1984). In their simplest forms, models that posit a single, common motion detector predict that distinct, direction selective motor outputs should be similarly sensitive to each parameter of motion stimuli, and that disrupting the function of a neuron involved in elementary motion processing should affect all motor outputs similarly, for the same stimulus. Our observations suggest a more complex view. In particular, our results argue that multiple motion-sensitive pathways contribute to behavioral responses to motion in different contexts, informed by signals that segregate early in the visual system (Figure 8C). Intriguingly, anatomical evidence suggests that motion sensitive behaviors may be executed in some animals independently of lobula-plate neurons, activity labeling studies detect up to three retinotopically organized pathways in the medulla, and electrophysiological evidence has suggested that luminance ON and OFF signals may contribute to motion pathways through separate DS channels (Franceschini et al., 1989; Sinakevitch, 2003; Bausenwein et al., 1992; Bausenwein and Fischbach, 1992). Our work provides the first evidence that this underlying complexity of the visual system subserves specific behavioral roles.

Our model advances understanding of the specific behavioral functions of lamina monopolar cells, LMCs, and provides a context for considering their electrophysiological responses. By measuring all aspects of behavioral response in freely walking animals to a wide range of stimuli, we discovered that DS changes in translation depend more critically on contributions from one of these neurons, L2, than do DS changes in rotation. This differential effect is profound, making the difference between following a stimulus and moving against it. Moreover, our finding that behavioral responses elicited by presentations of contrast increments and decrements are different, and are affected differently by disruption of L2, provides a context for understanding the specialization of LMC responses to different pattern contrast regimes. In particular, Dipteran LMCs respond proportionately to stimulus contrast (Srinivasan et al., 1982), display asymmetric ON and OFF transients, and are sensitive to the spatial structure of the stimulus (van Hateren, 1992). As we find that translation and rotation are sensitive to different combinations of stimulus density and speed, the specialized behavioral contributions of L2 may reflect the differential sensitivities of individual LMC types to spatiotemporal structure. Recent work has examined the effect of ablating L2 and other lamina monopolar cells on optomotor behavior, analyzing only the turning responses to rotational motion in immobilized flies (Rister et al. 2007). This work found that ablation of L2 does not affect turning bias evoked by moving gratings, except at very low stimulus contrasts. This result is consistent with our data, as we see a strong “with” turning bias using all stimuli in L2 silenced flies, even when this response is aberrant. However, because these studies could not measure either DS changes in translation or turn suppression, the differential contributions of L2 to these processing streams was not detected.

Our behavioral data also complement electrophysiological studies of LPTCs. Detailed examination of connections amongst a subset of lobula plate neurons in large insects has defined microcircuitry capable of pooling direction-selective signals, given inputs from motion detectors for small regions of visual space (Haag and Borst, 2004). However, the limitations of electrophysiology in this system have hampered our understanding how direction selective signals arise in the lobula plate. Hence only indirect evidence suggested that the array of detectors estimating local motion does so uniformly for all motion-sensitive behaviors (Buchner, 1984). Moreover, the diverse response properties of LPTCs have led to a range of predictions about behavioral contexts where their signals may be required (van Hateren et al., 2005; de Ruyter van Steveninck, 2001). Using a forward genetic approach that began only with the requirement that motion-evoked behavioral responses be specifically altered, we identified a small group of cells in the lobula plate as playing critical roles. In our system, these cells adjust the extent of turning in response to different stimuli, with a weaker effect on translation on some stimuli. This observation suggests that different LPTCs participate in distinct microcircuits, each of which is tuned to particular stimulus parameters, and is coupled to particular behaviors. Thus a complete description of any neuron’s role in motion processing can only come through examining the effects of a diverse stimulus set on all of the direction selective behavioral responses of the animal.

Behavioral specialization in the visual system

Our studies suggest that specialized processing of visual information to guide specific behavioral outputs is a broader feature of visual system organization than has been previously appreciated. Specialization in the context of motion vision has been previously explored in an ethological context, by comparing the morphology or tuning properties of lobula plate tangential cells in different insect species that have distinct motor behaviors (Sinakevitch et al., 2003; O’Carroll et al., 1996, O’Carroll et al., 1997). Based on these approaches, specialization has been proposed to arise far downstream of the initial stages of visual processing. Our finding that behavioral specialization is an organizational principle for motion processing, beginning with the core computations, changes the framework in which the tuning properties of visual neurons ought to be considered. Since the behaviorally relevant divergence between motion processing pathways occurs “upstream” of the lobula plate, generic operations on these different inputs could be sufficient to inform distinct behaviors. Alternatively, the entire motion processing circuitry, along each of the streams, could be different. It is perhaps surprising for a compact nervous system to implement parallel motion processing streams, where a single circuit may have been more efficient from an engineering perspective. Such a solution could reflect either the way in which the fly visual system evolved from simpler forms, or may reflect functional limitations of the circuits themselves. That is, by having at least two circuits, it may be possible to optimize their response properties for particularly relevant behavioral conditions.

Forward Genetic Dissection of Visual Circuitry

The growing set of genetic tools for manipulating neuronal activity has raised the possibility of dissecting neural processes using increasingly precise techniques. In flies, these tools have previously been applied to the examination of the behavioral roles of “candidate” neurons, pre-selected for analysis based on anatomical or molecular considerations, combined with either high-resolution studies of single animals, or relatively low-resolution studies of populations (Suh et al., 2004; Demir and Dickson, 2005; Manoli et al., 2005; Rister et al., 2007). While powerful, such applications are limited in their ability to identify new circuit components and, in the case of richly structured sensory processes like vision, limited in their capacity to capture behavioral responses to diverse stimuli, one animal at a time. By developing a forward genetic approach, combined with a high-throughput, quantitative behavioral analysis, we have overcome both limitations. Indeed, our study has opened the possibility that an unbiased, “saturation” genetic screen for neurons can now be conducted, resulting in an extensive description of many circuit elements in a particular neural pathway. In the context of motion vision, the unbiased approach is predicated on applying behavioral criteria to select lines in which motion processing has been compromised, and then identifying the sets of neurons that, when disrupted, are responsible for the defect. The choice of motion stimuli to elicit behavior thus sets the range of processing for which the visual system is probed; the resolution of behavioral observation limits the sensitivity to changes in processing. Our approach has extended the notion of a forward screen to specific circuit elements, in the context of a comprehensive, real-time description of the behavior of freely moving animals reacting to complex stimuli. We anticipate that the combination of sensitivity, temporal resolution and parameterization we have introduced will be critical to identifying and characterizing the many different neuron types that make specific contributions to behavior, or alternatively, play complex roles in processing stimuli. These behavioral analytic methods can be generalized to other sensorimotor integration tasks in flies, as well as other animals amenable to genetic analysis, and thus are likely to be of broad application. Our work also demonstrates that the fruit fly visual system provides a model in which well-defined computations can be linked to behavior and analyzed using forward genetic techniques. As many basic neural processes are widespread in the animal kingdom, and potentially evolutionarily ancient, an increased understanding of their circuit implementation in the fly will broadly inform our understanding of their function in other organisms.

Experimental Procedures

Animal Husbandry

All strains and stocks were maintained under 12:12 light dark cycles, at 25°C, 45–60% humidity and transferred to fresh food on alternate days. Flies for behavior were collected on the day of eclosion and sorted into fresh molasses vials, with 33 flies/vial. 2–4 day old flies were used for most experiments. The following stocks were used: Oregon R, UAS shi^TS (on III), Foma1Gal4, D21Gal4, Rh1Gal4 (on X), UAS synaptobrevin GFP (on II), UAS DSCAM-GFP (on III or X) and UASmCD8GFP (on III). All wild type experiments were carried out at the restrictive temperature for shibire ts, 34°C.

Behavioral Apparatus

For each experiment, 3–7 test tubes containing 100–233 flies total were placed inside a clear temperature-controlled chamber (Peltier device: Melcor; controller: Alpha-Omega Instruments). Stimuli were presented on a 21" Totoku ProCalix monitor (Irving, TX) directly underneath the TC chamber, facing upward, controlled by an ATI Radeon graphics card, at a refresh frequency of 200Hz. Flies were illuminated in the near-IR by 2 Advanced Illumination LED Line Light arrays (880nm) at opposite ends of the TC chamber, and filmed from above by an IR-sensitive camera (Sony XC-EI50) through an IR-pass filter (Lee Filters USA, filter #87C). Video signal was acquired through the Bitflow Raven frame grabber at 30Hz, and analyzed by a custom-written real-time tracker (Suppl. Materials and Methods).

Trial Structure

Each trial consisted of a motion epoch, in which most dots were displaced coherently in one of two opposite directions along the long axis of the test-tubes (98% coherence, τ_½ = 171 ms.), and a noise epoch, in which each dot was displaced independently of all others in a random direction (0% coherence). Motion and noise epochs alternated. In all figures except 1 and 7, the motion pulse duration was 800 ms, which was sufficient to reach 50–80% of the maximal response to any stimuli. The motion period in Figure 1 is 7 sec, and in Figure 7 a 150ms pulse of motion was presented. Within each trial, stimulus luminance was clamped at a constant value; to achieve desired contrast between dot and background at constant average luminance when the number of dots (dot density) changed, the intensity of both dots and background was adjusted appropriately. Stimulus contrast was calculated as the Michelson contrast of luminance values: (dot-background)/(dot+background). Stimuli in which dots were lighter than background were named 'contrast increment stimuli,' while stimuli in which dots were darker than background were named 'contrast decrement.' Gray-level luminance values were measured using the PR-650 Colorimeter (Photo Research Inc, Chatsworth, CA). The monitor was periodically calibrated with this same colorimeter.

Behavioral Analysis

Raw data from the real-time tracker was processed with scripts written in MATLAB (Mathworks, Natick, MA).

Heading bias: each stimulus epoch was classified by the direction of stimulus motion, and fly orientation was rotated accordingly into stimulus coordinates. The fraction of moving flies in each orientation, with and against stimulus motion, was calculated at each time point relative to stimulus onset. This metric was summarized by integrating HB over the entire period of response.

Translation and rotation indices: joint distributions of angular and translational velocities, VR and V_T, conditioned on a fly's orientation with respect to the stimulus, were constructed at each time point. The mean and variance of the marginals were then examined as a function of time, fly orientation, and direction of turn relative to stimulus (into or out of alignment with the motion direction). The variance of each distribution was examined to verify consistency with the relative changes in the mean. Following this procedure, translation and rotation indices were calculated as follows. Two thresholds were placed in V_T, one containing the mode of near-zero velocities defining stopped flies, and one just below the mode velocity of normal forward movement (1.9 cm/sec), thus separating arbitrarily defined "fast" versus "slow" forward movement. One threshold was placed between the maxima in V_R (at 100 deg/sec). Stationary flies (those below the first V_T threshold) were excluded from analysis, as they did not contribute to the directional response, having no rotation or forward movement. The fraction of the remaining flies in the "fast" or "turning" zones as denoted in Figure 2C were calculated at each time point. TI and RI were calculated relative to the pre-stimulus period as follows:

$$\text{TI}(\mathrm{t})=\{1/(\text{TI}({\mathrm{t}}_0))*\#\text{moving}/(\#\text{moving}+\#\text{slowing})\}-1$$

$$\text{RI}(\mathrm{t})=\{1/(\text{RI}({\mathrm{t}}_0))*\#\text{turning}/(\#\text{turning}+\#\text{not turning})\}-1$$

where TI(t₀) and RI(t₀) represent metric averages in the 1 sec before motion onset.

In addition, a turn direction index was calculated by separating turns into those into and out of alignment with the stimulus direction, as illustrated in Figure 2C, as follows:

$$\text{TD}(\mathrm{t})=\{1/(\text{TD}({\mathrm{t}}_0))*\#(\text{turning with})/(\#\text{turning})\}-1$$

These metrics were summarized by averaging each metric during the period of stimulus motion, after excluding the initial stimulus onset response. For the 800ms pulses of motion, this constituted the last 433 ms of the motion pulse. For the 150ms motion pulses, the data was summarized by integrating each of the different phases of the response separately. The location of each threshold was varied substantially, producing qualitatively similar conclusions. Thus, the metrics were not sensitive to the precise location of any of the thresholds, and reproduced conclusions reached by analyzing raw velocity data in every case.

Statistical tests

To test the significance of neuronal effects on behavior in Figures 4 and 5, we performed a two-tailed paired sample t-test on the time trace of each measure of behavior as indicated in main figure legends. Each time-evolving metric was paired between experimental and control genotypes by the metric's time index. To test the significance of each neuronal manipulation in Figure 7, we assumed the null hypothesis that the behavior of Gal4/+ heterozygotes for each driver is indistinguishable from the behavior of Gal4/+; UAS shi(ts)/+ flies, and tested this assumption by bootstrap.

Histology and Imaging

Brains were dissected from adult flies using standard methods (Lee et al., 2001), stained with the synaptic marker nc82 (1:30; Wagh et al., 2006) rat anti mouse CD8 (1:100; Lee and Luo, 1999), mouse anti-GFP (Promega) and rabbit anti lacZ (Ebens et al., 1993). Secondary antibodies anti rat Alexa 488 and anti mouse Alexa594 (Molecular Probes) were used at 1:200. Samples were imaged using either a Leica TCS SP2 or SP5 confocal using a 20X immersion or air or 40X immersion lens. Images were processed using Huygens (SVI), rendered using Imaris (Bitplane), and assembled using Photoshop (Adobe).