Filtering of Visual Information in the Tectum by an Identified Neural Circuit

The optic tectum in the vertebrate midbrain, called the superior colliculus in mammals, receives visual inputs from the retina and converts this information into directed motor outputs (1). In larval zebrafish, the tectum is divided into two main areas: the stratum periventriculare (SPV), which contains the cell bodies of most tectal neurons, and the synaptic neuropil area, which contains their dendrites and axons as well as the axons of retinal afferents (2–5). Neurons in the SPV, called periventricular neurons (PVNs), extend a single neurite, which branches extensively and may span the entire depth of the neuropil. Retinal axons mainly target the superficial layers of the tectal neuropil [i.e., the stratum opticum (SO) and the stratum fibrosum et griseum superficiale (SFGS); fig. S1] (5–8), where they make contact with the dendrites of periventricular interneurons (PVINs) that convey the visual information to other PVINs or to periventricular projection neurons (PVPNs). The axons of PVPNs exit the tectum in the deepest neuropil layer and project to premotor regions in the midbrain and hindbrain (2, 5, 6).

The tectum is required for the localization, tracking, and capture of motile prey, such as paramecia (9). Other visual behaviors (e.g., optomotor and optokinetic responses) are mediated by a different pathway not involving the tectum (10, 11). Consistent with a function in the detection of small objects, electrophysiology and optical imaging showed that single tectal neurons, in all vertebrates examined, often respond to small stimuli such as spots or bars, which occupy only a fraction of the neurons’ receptive fields (12–19). To reveal the neural substrate of this size filtering, we used Gal4 enhancer-trap lines (2, 20, 21) to drive the expression of the genetically encoded Ca²⁺ indicators GCaMP1.6 (22) and GCaMP3 (23). This allowed us to record visually evoked activity in dendrites and axons of specific classes of neurons.

We used the Atoh7:Gal4 transgenic line to drive expression of GCaMP1.6 in retinal axons, demarcating the superficial input layers in the neuropil (Fig. 1A). The fish’s retina was exposed to three visual stimuli, displayed on a miniature LCD screen (fig. S1): (i) a brief (25 or 50 ms) flash that filled the entire screen (horizontal visual angle 50°), (ii) a thin black bar (2°) moving at a speed of 0.25°/ms across the screen from anterior to posterior (A→P), and (iii) a bar of the same size and speed, but moving from posterior to anterior (P→A). The responses of the retinal axons did not differ significantly in amplitude and in time to peak between the large and the small stimuli (Fig. 1, B and C; maximum ΔF/F = 2.11 ± 0.19% for flash versus 2.08 ± 0.11% for A→P and 2.16 ± 0.13% for P→A; time to peak = 0.69 ± 0.03 s for flash versus 0.72 ± 0.11 s for A→P and 0.73 ± 0.05 s for P→A; P_amplitude = 0.31; P_time-to-peak = 0.54; n = 5). Indeed, responses were similar in amplitude across a range of stimulus sizes (Fig. 1D).

In the Gal4s1038t line, a small subset of PVPNs in the posterior-ventral quadrant of the tectum are labeled (2) (Fig. 1E). This population is activated by retinal stimulation, as surgical removal of one eye eliminated GCaMP1.6 responses in the contralateral tectum (fig. S1; maximum ΔF/F = 2.17 ± 0.23% in ipsilateral tectum versus 0.09 ± 0.14 in contralateral tectum; P_I-C < 3.16 × 10⁻⁴; n = 5). However, the response to the full-screen flash was weaker in the deep output layer than in the superficial input layer (Fig. 1, F and J; 3.01 ± 0.36% for superficial versus 1.65 ± 0.28% for deep; P < 10⁻⁴). Although the absolute fluorescence signals varied between fish, the deep-to-superficial response ratios were consistent (Fig. 1J and fig. S2; deep-to-superficial ratio = 0.48 ± 0.15; P < 0.01, n = 7). In contrast, small moving bars activated Ca²⁺ rises equally in the deep and the superficial layers (Fig. 1, G, H, K, and L; A→P: 2.95 ± 0.17% for deep versus 3.10 ± 0.22% for superficial, deep-to-superficial ratio = 0.95 ± 0.07; P→A: 3.2 ± 0.18% for deep versus 2.89 ± 0.19% for superficial, deep-to-superficial ratio = 1.10 ± 0.31; P_A→P = 0.16 and P_P→A = 0.23; n = 7). The tuning curve showed a systematic size-dependent reduction of the response (Fig. 1I), which suggests that large stimuli did not efficiently excite the cellular elements in the deep neuropil layer.

In the Gal4s1013t line, almost all tectal cells are labeled (2), allowing us to record Ca²⁺ responses across the entire visual field. The deep-to-superficial response ratios in response to a full-screen flash were not significantly different between the anterior and posterior halves of the tectum (0.41 ± 0.19 versus 0.36 ± 0.27; n = 3). Thus, there does not seem to be a topographic bias in size tuning across the visuotopic map

We used a mosaic labeling strategy to image the dendritic activity of individual PVNs. Two DNA constructs encoding UAS:GCaMP3 (23) and Dlx5/6:Gal4 (24) were co-injected at the two-cell stage, and larvae with only one or very few labeled PVNs were used for imaging (Fig. 1M). Of 38 PVNs recorded, 7 (18%) responded to small moving bars; the remaining cells did not respond to any of the stimuli. None of the PVNs was activated by the full-screen stimulus (Fig. 1, N to P, n = 7). In the seven PVNs sensitive to small moving bars, we could not detect significant differences in the Ca²⁺ response between the distal (superficial) and the proximal (deep) segments of their dendritic trees (P = 0.49), indicating the existence of a circuit that filters out low-frequency spatial inputs before they reach the PVN dendrites.

We next showed that spatial filtering is achieved by feedforward inhibition. Local application of the GABA_A receptor antagonist bicuculline (Bicu) to the tectum increased responses in the entire neuropil, but not uniformly. In the deep layers of Gal4s1038t, UAS:GCaMP1.6, Ca²⁺ signals rose by more than a factor of 15, whereas in the superficial layers the increase was by only a factor of 5 (Fig. 2, E to G; superficial, 8.19 ± 0.36% for Bicu versus 1.74 ± 0.19% for control; deep, 10.69 ± 0.41% for Bicu versus 0.69 ± 0.09% for control; P_SUP < 1.4 × 10⁻⁷ and P_DEEP < 9.9 × 10⁻⁸, n = 4), inverting the normal ratio (Fig. 2H and fig. S3; deep-to-superficial ratio for Bicu = 1.38 ± 0.10 versus 0.43 ± 0.07 for control; P < 3.1 × 10³; n = 4). Bicu administration to the tectum had no detectable effect on the strength of retinal inputs (Fig. 2, A and B; 1.99 ± 0.18% for Bicu versus 2.52 ± 0.19% for control; n = 3). In contrast, intraocular Bicu injection produced a robust increase in the Ca²⁺ response (Fig. 2, C and D; 4.96 ± 0.70% for Bicu versus 2.16 ± 0.24% for control; n = 4).

Gal4s1156t drives expression in a specific population of neurons whose cell bodies are located in the SO (Fig. 3, A to C) (2). Antibody staining showed that most, or all, neurons in this layer expressed the GABA markers Gad67 and Reelin (Fig. 3, A and B, and figs. S4 and S5) (94.71 ± 0.6%; 229 cells counted in n = 3 larvae). Furthermore, almost all Gal4s1156t-expressing cells were GABA-positive (54 of 56 cells in n = 4 larvae). Labeling of single cells by mosaic expression of cytoplasmic DsRed and synaptophysin fused to green fluorescent protein (Syp-GFP) (25) revealed that these cells extend a broad, regularly branched axonal arbor, containing many pre-synaptic specializations (Fig. 3, D and E). Cells with similar morphologies have been described in other vertebrate species (26). Strikingly, these superficial interneurons (SINs) showed a robust response only to the full screen, not to small moving bars (Fig. 3, F and G; 2.27 ± 0.32%; P < 1.34 × 10⁻⁴; A→P: 0.21 ± 0.14%; P→A: 0.09 ± 0.16%; P = 0.42; n = 6). We conclude that SINs are tuned to large stimuli.

The SINs may provide feedforward inhibition of PVNs. If so, their removal from the circuit should alter the tuning of PVNs and should impair a behavior that relies on this circuit property. We blocked synaptic transmission in the SINs by driving tetanus toxin light chain fused to cyan fluorescent protein (TeTxLC-CFP) (27) in Gal4s1156t. Double-transgenic larvae captured far fewer paramecia than controls (Fig. 4A), whereas their optomotor behavior was unaffected (fig. S6). Blocking transmission from the small number of PVPNs in Gal4s1038t did not reduce prey capture rates. Using the pan-tectum Gal4s1013t line (2), we generated a fish expressing both the genetically encoded photosensitizer KillerRed (28) and GCaMP1.6 in both PVNs and SINs. To selectively kill the SINs, we illuminated the SO with an intense green laser (563 nm). Terminal deoxynucleotidyl transferase– mediated deoxyuridine triphosphate nick end labeling (TUNEL) staining and in vivo annexin V labeling showed apoptotic cells only in the targeted region (9.5 ± 1.8 TUNEL⁺ cells per section on the ablated SO versus 1.0 ± 0.6 TUNEL⁺ cells per section on the control side, P < 0.05; 4.5 ± 1.0 TUNEL⁺ cells in the SPV of the ablated tectum versus 4.0 ± 0.4 in the control SPV, P = 0.5) (Fig. 4, B to D). After photo-ablation of SINs, Ca²⁺ responses in the PVNs to a full-screen flash were equalized across the neuropil layers (Fig. 4E; deep-to-superficial ratio = 0.47 ± 0.8 before illumination versus deep-to-superficial ratio = 0.98 ± 0.11 after; P < 10⁻³ after illumination, n = 4). No significant change in response ratios was observed in the tectum contralateral to the illumination (before: deep-to-superficial ratio = 0.55 ± 0.8; after: deep-to-superficial ratio = 0.61 ± 0.12; P = 0.38).

Together, our findings support a contribution of SINs to the neural mechanism that filters visual inputs in the tectum. In one possible scheme (fig. S7), which is supported by the morphologies of PVN cell types (2, 5), SINs make GABAergic contacts with some PVINs, which in turn convey this information to the dendritic arbors of PVPNs. Thus, the visual information flows from superficial to deep through a neural filter that subtracts low-frequency spatial information. This circuit may support prey capture by allowing the animal to track a moving object against a background that changes uniformly in brightness or is composed of low spatial frequencies. Given that the mammalian superior colliculus has similar layer-specific spatial filtering properties (1, 12), it seems likely that this circuitry is evolutionarily conserved among vertebrates.