Color vision requires comparison between photoreceptors that are sensitive to different wavelengths of light. In Drosophila, this is achieved by the inner photoreceptors (R7 and R8) that contain different rhodopsins. Two types of comparisons can occur in fly color vision: between the R7 (UV sensitive) and R8 (blue- or green-sensitive) photoreceptor cells within one ommatidium (unit eye) or between different ommatidia that contain spectrally distinct inner photoreceptors. Ommatidia exhibit unique fluorescence of their inner photoreceptors that appear either yellow (y) (70% of ommatidia) or pale (p) (the remaining 30%). The p-ommatidia contain UV-Rh3 in R7 and blue-Rh5 in R8, whereas y-ommatidia contain UV-sensitive Rh4 in R7 and green-sensitive Rh6 in R8. Photoreceptors project to the optic lobes: R1-R6, which are involved in motion detection, project to the lamina, whereas R7 and R8 reach deeper in the medulla. This paper analyzes the neural network underlying color vision into the medulla. The neural network in the medulla was reconstructed, focusing on neurons likely to be involved in processing color vision. The full complement of neurons in the medulla was identified, including second-order neurons that contact both R7 and R8 from a single ommatidium, or contact R7 and/or R8 from different ommatidia. Third-order neurons and local neurons were identified that likely modulate information from second-order neurons. Finally, highly specific tools are presented that will allow functional manipulation the network and test both activity and behavior. This precise characterization of the medulla circuitry will facilitate understanding of how color vision is processed in the optic lobe of Drosophila, providing a paradigm for more complex systems in vertebrates (Morante, 2008).

The medulla represents the major neuropil in the optic lobe. Despite this complexity, by using a series of TFs-Gal4 lines, the medulla network has been dissected at the single-cell level. ap- and ey-Gal4 are expressed in nonoverlapping populations, both in projection and local neurons. dll-Gal4 reveals expression almost exclusively in local neurons, although not all local neurons are marked by dll-Gal4. Through this extensive analysis it has been possible to reconstruct morphologically 38 types of projection neurons, 22 types of local neurons and 3 connecting neurons. Among these cell types, six new projection and four local neuron types not described before have been identified. Most of these new cell types (e.g., Tm_LM7 and Pm_LM7) include cells with ramifications exclusively in the lower medulla domain (Morante, 2008).

It is thus possible to define the elements represented in a 'column,' the medulla functional units: (1) two inputs from R7 and R8, (2) five ramifications from L1-L5 lamina neurons, (3) 11 types of columnar and 20 of noncolumnar projection neurons, (4) four types of columnar and 11 of noncolumnar local neurons, and (5) two types of columnar and one of noncolumnar lamina connecting neurons. Additionally, seven other noncolumnar local and seven projections neurons do not contact PRs. Overall, each R7/R8 termination pair is therefore surrounded by at least 54 different cell types, with 14 other cell types that do not contact PRs. The enormous number of cell types forming part of a column contrasts with the relatively small number of elements that feed into the medulla (R7-R8 and R1-R6 through L1-L5 neurons). Thus, the divergence in the flow of information between PRs and medulla neurons argues that much local processing occurs in the medulla. In contrast, in the deeper optic lobe (i.e., lobula and lobula plate), the number of cells and their wide-field ramifications argue for the convergence of information from medulla neurons (Morante, 2008).

This analysis of the color-vision network in the Drosophila medulla reveals the presence of two parallel routes carrying and processing visual information that coexist in the medulla neuropil: (1) a point-to-point pathway formed by columnar neurons that receive information from only a single ommatidium ('vertical integration'), and (2) a pathway with a broader receptive field composed by noncolumnar neurons that receive information from photoreceptors from several ommatidia ('horizontal integration') (Morante, 2008).

The first pathway integrates outputs from R8 and R7, likely allowing broad wavelength discrimination between UV- and blue or green: cells comparing signals from p-ommatidia might mediate better discrimination among short wavelengths, whereas those comparing outputs of y-photoreceptors should better mediate discrimination among longer wavelengths. This suggests that there are two independent and nonoverlapping retinotopic maps that separately process color information: one from p- and one from y-photoreceptors. These maps might be physically separated in higher brain centers, but specific p- or y-contacting neurons are not observed, are distinct projections to different layers in the lobula complex seen (Morante, 2008).

The second pathway reflects more complex visual processing, with noncolumnar Tm neurons cells integrating information from several R8 and R7 photoreceptors. Because the p- and y-ommatidia each compare output of photoreceptors with widely different absorption spectra (UV and green, or UV and blue), 'horizontal integration' between y- and p-ommatidia for both R7 and R8 might allow a much more precise evaluation of the colored world, although at a reduced spatial resolution. Additionally, cells were found that might directly compare p- and y-R8 cells (e.g., contacting blue and green R8, but not R7). Similarly, cells that compare p- and y-R7 photoreceptors might integrate information from the two different types of UV-photoreceptors. It should be noted that the difference in peak of absorption between Rh3 and Rh4 (~20 nm) in R7 is sufficient to allow precise discrimination of UV wavelengths and is similar to the difference between M- and L- opsins in humans (30 nm). Thus, this horizontal integration should allow the convergence of several inputs from multiple R7 and R8 photoreceptors to a single medulla cell (Morante, 2008).

It is not clear how comparison between photoreceptors is achieved and whether local neurons expressing different neurotransmitters are involved to generate opposite outputs between R7 and R8 (for columnar neurons) or between p- and y-R7 or R8 (for noncolumnar neurons). It is likely that R7 cells do sum up their output to support the strong UV attraction that characterizes flies. However, inner photoreceptors are not involved in scotopic vision (dim light), and an organization in which neurons simply add their outputs makes little sense for what is known of the function and specialization of R7 and R8 cells in color vision. Therefore, it is likely that an opponent system exists in the medulla and that it is mediated by local neurons. Alternatively, color vision might not need an opponent system but might result from nonlinear interactions between R7 and R8. Their interaction with the same postsynaptic cell could be complex, e.g., Tm neurons might have synergistic responses to inputs from R7 and R8 or from photoreceptors from different ommatidia (Morante, 2008).

In the mammalian retina, horizontal and amacrine cells are interneurons that have a critical role in modulating retinal output. Horizontal cells provide lateral interactions between photoreceptor terminals, creating a 'center-surround organization,' enhancing the response of ganglion cells lying directly under the light stimulus and inhibiting their neighbors. Meanwhile, amacrine cells not only make inhibitory synapses on bipolar cells, thus controlling their output to ganglion cells, but also synapse onto ganglion cells and coordinate their firing. As in the mammalian retina, a great variety of neurons with local ramifications within the medulla might modulate the visual outputs carried by projection neurons. This modulation is accomplished by two kinds of local neurons. (1) Columnar local neurons with one Mi cell per ommatidium. Interestingly, the arborizations of many of these columnar local neurons resemble those of corresponding Tm columnar projection neurons that contact either one photoreceptor, or both R7 and R8. They intermingle with them in photoreceptor layers as well as in lower medulla layers. In this microcircuit, Mi cells appear to interact both pre- and post-synaptically with Tm ramifications. (2) Dm cells are noncolumnar local neurons that do contact photoreceptor layers, whereas Pm cells only have projections in lower medulla layers. Both classes ramify extensively over several columns. Interestingly, noncolumnar local neurons appear to be pre- and post-synaptic both in photoreceptors layers and in lower medulla layers. This suggests that Dm local neurons perform a first level of integration with Tm cells at the level of photoreceptor layers, whereas this information is further processed by Pm cells that act at a second level in lower medulla layers (Morante, 2008).

Therefore, columnar and noncolumnar projection neuron outputs could be modulated by columnar and noncolumnar local neurons, respectively. In the Drosophila antennal lobe, two kinds of local neurons exist: inhibitory local interneurons and a local excitatory population involved in processing projection neurons signals to downstream targets. A similar system might also exist in the medulla (Morante, 2008).

Whether Tm projection neurons are columnar or noncolumnar, they all arborize in lower medulla layers (M7-M10). For most Tm cells analyzed, these lower medulla ramifications appear to contain both presynaptic and postsynaptic terminations, suggesting that this region represents a second layer of integration in the color-vision pathway after the direct comparison between R7 and R8 (or between y and p R7/R8) (Morante, 2008).

A class of projection neurons, Tm_LM7 and Tm_LM8 only arborize in lower medulla layers, where they mostly exhibit postsynaptic arborizations. This suggests that they play a role as third-order neurons that collect more elaborate visual information already integrated by other Tm cells with ramifications in photoreceptor layers and presynaptic endings in lower medulla layers and then carry this processed visual information to downstream targets (Morante, 2008).

Like the mammalian visual cortex, the fly visual system is organized into retinotopic columns. A widely accepted biophysical model for computing visual motion, the elementary motion detector proposed nearly 50 years ago posits a temporal correlation of spatially separated visual inputs implemented across neighboring retinotopic visual columns. Whereas the inputs are defined, the neural substrate for motion computation remains enigmatic. Indeed, it is not known where in the visual processing hierarchy the computation occurs. Tbis study combined genetic manipulations with a novel high-throughput dynamic behavioral analysis system to dissect visual circuits required for directional optomotor responses. An enhancer trap screen of synapse-inactivated neural circuits revealed one particularly striking phenotype, which is completely insensitive to motion yet displays fully intact fast phototaxis, indicating that these animals are generally capable of seeing and walking but are unable to respond to motion stimuli. The enhancer circuit is localized within the first optic relay and strongly labels the only columnar interneuron known to interact with neighboring columns both in the lamina and medulla, spatial synaptic interactions that correspond with the two dominant axes of elementary motion detectors on the retinal lattice (Zhu, 2009).

Molecular genetic techniques were used to manipulate neural circuits that mediate two well-known visual behaviors in freely behaving fruit flies: motion-dependent optomotor responses and stationary light-elicited phototaxis responses. To efficiently analyze large numbers of individuals and multiple fly lines, a simple yet robust high-throughput assay was devised that tracks the real-time spatial distribution of up to 100 walking flies responding with either optomotor reflexes to panoramic image movement or fast phototaxis toward a stationary narrow-band light source. To generate apparent motion, an array of LED panels fashioned into a three-sided visual 'hallway' displays a computer-controlled centrifugal-centripetal cycling motion stimulus. Dark stripes against a bright background continuously drift from the center of the hallway toward opposite ends and then switch direction to converge at the center. Flies are contained within a clear acrylic tube in the center of the hallway and tend to distribute themselves evenly along the length of the tube prior to visual stimulation. Wild-type flies move against the direction of image motion such that, in response to centrifugal motion, flies rapidly converge at the center; after a switch to centripetal motion, they segregate equally to the two ends of the hallway arena. Similarly, flies exhibit positive phototaxis and rapidly converge upon a high-intensity LED in the center of the hallway. The spatial distribution of the population over the length of the hallway is calculated online for each video frame in real time and thus provides rapid spatiotemporal measurements of the group walking behavior (Zhu, 2009).

The walking assay in a combined histological and behavioral screen was used to identify neuronal components that mediate either light-directed or motion-elicited orientation. The UAS/Gal4 system was used to perform a confocal microscopy-based histological screen for Gal4 enhancer trap lines that are expressed in specific groups of neurons of the visual system. These circuits were manipulated with tetanus neurotoxin light chain (TNT), which cleaves neuronal synaptobrevin to suppress synaptic transmission and thus was used to study the behavioral consequences of circuit inactivation (Zhu, 2009).

In the first visual relay, the lamina, each retinotopic column contains five large monopolar (L) cells: L1-L5. From a screen of more than 300 enhancer trap lines, one (termed Ln-Gal4) showed expression tightly restricted to the lamina. Morphological criteria, analyses of single-cell MARCM clones, and labeling of L4s and L5s with anti-BSH antibody together reveal that Ln strongly labels L3s and L4s and is expressed at lower level in L2s and L5s. Expression in L1 was undetectable with standard GFP staining but labeled weakly with multiple copies of UAS-mCD8-GFP and UAS-N-sybGFP. The noteworthy feature of this driver is its specificity for lamina cells (Zhu, 2009).

Inactivating this complement of lamina neurons produced flies that were completely insensitive to visual motion cues in any visual assay. In the walking arena, both control progeny of Ln-Gal4 flies mated with wild-type Canton-S (Ln-Gal4/+), and UAS-TNT/+ flies responded normally to motion signals by converging at the center of the hallway and then dispersing upon motion reversal. Normal optomotor behavior is apparent both in the full spatiotemporal distribution of flies along the hallway and in the temporal dynamics of those flies converging upon the origin of the drifting patterns at the center of the hallway. In contrast, crossing Ln-Gal4 with UAS-TNT eliminated all motion responses. The loss-of-motion responses in the Ln-Gal4/UAS-TNT flies persist under all combinations of spatial, temporal, and contrast conditions tested and also in a standard optomotor flight assay in which these flies failed to respond to either progressive or regressive motionInactivating this complement of lamina neurons produced flies that were completely insensitive to visual motion cues in any visual assay. In the walking arena, both control progeny of Ln-Gal4 flies mated with wild-type Canton-S (Ln-Gal4/+), and UAS-TNT/+ flies responded normally to motion signals by converging at the center of the hallway and then dispersing upon motion reversal. Normal optomotor behavior is apparent both in the full spatiotemporal distribution of flies along the hallway and in the temporal dynamics of those flies converging upon the origin of the drifting patterns at the center of the hallway. On the other hand, crossing Ln-Gal4 with UAS-TNT eliminated all motion responses. The loss-of-motion responses in the Ln-Gal4/UAS-TNT flies persist under all combinations of spatial, temporal, and contrast conditions tested and also in a standard optomotor flight assay in which these flies failed to respond to either progressive or regressive motion (Zhu, 2009).

Several lines of evidence suggest that, for the Ln-Gal4/UAS-TNT phenotype, signaling is preserved through the L1 pathway and quite possibly through both L1 and L2. First, using TNT to inactivate these cells has little influence on motion responses. Second, phototaxis responses to wavelengths that stimulate photoreceptors R1-R6 and, hence, both L1 and L2 postsynaptically are fully intact in the Ln-Gal4/UAS-TNT flies. Indeed, the phototaxis responses in these animals are significantly stronger than for the intact controls, particularly for green light. Signaling through the lamina would account for the robust phototaxis behavior because without R1-R6 signaling in ninaE flies, phototaxis responses are compromised (Zhu, 2009).

The opposite influence upon optomotor and phototaxis behavior of inactivating the Ln circuit raises the question of whether this is a property unique to the Ln circuit. Motion and phototaxis responses were compared in the walking assay and it was found that any manipulation to L1 or L2 resulted in reduced phototaxis responses to UV. However, phototaxis responses to green light were not significantly affected. This is in direct contrast to the Ln circuit, which enhances phototaxis responses, particularly for green light (Zhu, 2009).

The remarkable specificity of the Ln driver for lamina circuits coupled with the stringent behavioral phenotype -- motion blindness and enhanced phototaxis -- suggests that functional segregation of the two behaviors occurs early in the visual pathway. In classical physiological experiments, lamina neurons have been shown to function in tandem with photoreceptors as light-level adaptive high-pass filters. Recent electrophysiological analyses have further suggested that the ON-OFF transient response properties of lamina projection neurons may underlie the remarkably high-performance discrimination of small objects embedded within a background of visual clutter by downstream target-detecting interneurons. The current results provide additional genetic and functional evidence for complex processing within peripheral visual circuits (Zhu, 2009).

Which neurons implement these computations? TNT does not influence motion sensitivity in L1+L2, suggesting that this circuit does not mediate the motion-blind Ln phenotype. By using a combination of selective genetic inactivation and selective rescue experiments with L1 and L2, Rister (2006) concluded that other lamina neurons (L3 and L5) are neither necessary nor sufficient for motion detection, though both receive either direct or indirect input from the photoreceptors. Due to the conflicting results of L1+L2 manipulation, the current experiments neither confirm nor refute any potential role of L3 and L5, which are labeled by the Ln driver. However, L4 is strongly labeled by Ln and is the only lamina cell that provides regular synaptic connections between neighboring optic columns. Spatiotemporal correlation of light signals from two neighboring visual columns is a hallmark of elementary motion detection. Thus, the hypothesis is compelling that inactivation of L4, in part, underlies the elimination of motion responses in our assays (Zhu, 2009).

The topology and ultrastructural organization of L4 within the lamina implicates this neuron for elementary motion computations. In Drosophila, physiological and behavioral studies have disclosed that the spatial separation of elementary motion detector (EMD) inputs corresponds to that of the ommatidia lattice, indicating that adjacent visual columns function as paired input arms of the EMD. In addition, a classic study revealed two sets of primary EMDs with approximately equal strength and oriented across the hexagonal array of the compound eye at -30° (-X direction) and +30° (+Y direction) with respect to the equator]. According to SEM reconstructions, there are only two cellular connections between visual columns in the lamina: an irregular amacrine network not involved in motion processing and an ordered L4 network that projects between posteroventral and posterodorsal columns. By aligning the coordinate systems of the functional and anatomical studies, it eas found that the topology of L4 connections precisely matches that of the required interconnection of EMD arrays; through two sister collaterals, each L4 projects to two L2s in neighboring posteroventral (-X) and posterodorsal (+Y) columns (Jay, 2009).

If L4 receives direct input from L2 in the lamina and if L4 is critical for motion coding, why then does inactivating L2 not produce motion blindness? One possibility is that input to L4 from an amacrine cell [6] could be amplified under conditions in which L2 input is removed. Feedback-dependent mechanisms have been shown to amplify photoreceptor output upon inactivation of postsynaptic histamine channels. The anatomical organization of L4 columnar collaterals is observed within both the lamina and the medulla and is conserved across species. Therefore, another possibility is that synaptic connections to L4 within the medulla, which are presently enigmatic, may carry the requisite inputs (Jay, 2009).

Definitive characterization of the specific cell circuit that is responsible for the remarkable behavioral phenotype of Ln-Gal4/UAS-TNT flies will require advanced histological reagents and electrophysiological analyses. The results presented in this study lay the groundwork by highlighting lamina and medulla circuits that are vital for conditioning early motion signals. These data emphasize the important role that lamina circuits play in ultimately orchestrating complex visual behaviors such as color vision, phototaxis, and motion-dependent optomotor behaviors (Jay, 2009).