Ancient origin of the rod bipolar cell pathway in the vertebrate retina

Vertebrates rely on rod photoreceptors for vision in low-light conditions1. Mammals have a specialized downstream circuit for rod signaling called the primary rod pathway, which comprises specific cell types and wiring patterns that are thought to be unique to this lineage2–6. Thus, it has been long assumed that the primary rod pathway evolved in mammals3,5–7. Here, we challenge this view by demonstrating that the mammalian primary rod pathway is conserved in zebrafish, which diverged from extant mammals ~400 million years ago. Using single-cell RNA-sequencing, we identified two bipolar cell (BC) types in zebrafish that are related to mammalian rod BCs (RBCs) of the primary rod pathway. By combining electrophysiology, histology, and ultrastructural reconstruction of the zebrafish RBCs, we found that, like mammalian RBCs8, both zebrafish RBC types connect with all rods and red-cones in their dendritic territory, and provide output largely onto amacrine cells. The wiring pattern of the amacrine cells post-synaptic to one RBC type is strikingly similar to that of mammalian RBCs. This suggests that the cell types and circuit design of the primary rod pathway may have emerged before the divergence of teleost fish and amniotes (mammals, bird, reptiles). The second RBC type in zebrafish, which forms separate pathways from the first RBC type, is either lost in mammals or emerged in fish to serve yet unknown roles.

of the primary rod pathway may have emerged before the divergence of teleost fish and amniotes 48 (mammals, bird, reptiles). The second RBC type in zebrafish, which forms separate pathways 49 from the first RBC type, is either lost in mammals or emerged in fish to serve yet unknown roles.

RBC1 and RBC2 morphologies resemble mammalian RBCs 151
We next determined the morphological similarities of RBC1 and RBC2 with mammalian RBCs. In 152 mammals, RBC axons arborize in the innermost layer of the IPL 40 . By screening our zebrafish 153 transgenic lines, we identified two lines, Tg(vsx1:memCerulean) q19 (vsx1:memCer) and 154 Tg(vsx2:memCerulean) wst01 (vsx2:memCer), that each label BCs with axon terminals in the 155 innermost layer of the IPL (Fig. 2). Fluorescent in situ hybridization for the identified gene markers, 156 s100a10b and uts1, which are selectively expressed by RBC1 and RBC2 (Fig. 1c), revealed that 157 vsx1:memCer and vsx2:memCer label RBC1 and RBC2, respectively (Fig. 2a,b). We also 158 observed that dendritic arbors of both RBC1 and RBC2 cover the retina in a non-overlapping 159 manner, an arrangement called 'tiling' that is considered a hallmark of a BC type (Fig. 2e,f) 41 . 160 Therefore, both RBC1 and RBC2 represent single bipolar types that transcriptionally and 161 morphologically resemble mammalian RBCs. 162 We observed slight variations in morphology and molecular expression between RBC1 and RBC2. 163 The axon terminal of RBC1 is relatively spherical, similar to mammalian RBCs, in contrast to the 164 'flat-footed' axonal ending of RBC2 (Fig. 2c,d). RBC1 were immunoreactive for PKCα (Fig. 2c), 165 whereas RBC2 were not (Fig. 2d), consistent with the difference in their prkca expression (Fig.  166 1d). In addition, their abundance differed: RBC1s were more densely packed than RBC2s 167 (p=0.0052, Mann-Whitney two-tailed U test) (Fig. 2g,i). This difference in the densities is unlikely 168 due to regional variations as both RBCs are present in the dorsal and ventral-temporal retina at 169 similar densities (Fig. 2g,i and Fig. S2). The dendritic field sizes of the two RBC types were 170 inversely related to their cell density, consistent with their tiling arrangement ( Fig. 2g-j).

RBC1 receives rod inputs via mGluR6 receptors 240
We next asked whether zebrafish RBCs receive functional rod input via the metabotropic 241 glutamate receptor mGluR6 as seen in mammalian RBCs. We first investigated the expression of 242 mGluR6 in RBC1 and RBC2 dendritic tips at rod spherules. Super-resolution imaging of mGluR6 243 immunolabeling in vsx1:memCerulean and vsx2:memCerulean retinas showed that the dendritic 244 tips of RBC1, but not RBC2, robustly overlapped with mGluR6 immunoreactivity at contacts with 245 rod spherules (Fig. 4a,b). These findings are consistent with the transcriptional profiles, which 246 showed that RBC1 expresses higher mRNA levels of grm6a and grm6b, which encodes mGluR6, 247 than RBC2 (Fig. S1). 248 We then used electrophysiological recordings to ask whether mGluR6 mediates rod input to the 249 zebrafish RBCs. We prepared retinal wholemounts that preserve synaptic connections in the 250 outer retina, and performed whole-cell patch-clamp recordings on the axon terminals of RBC1 251 and RBC2 (Fig. 4c). Both RBC1 (n = 10) and RBC2 cells (n = 3) exhibited ON responses to a 252 cone-activating flash (red LED), confirming the successful patch-clamp recordings of light 253 responses in these BCs and demonstrating that both cell types are ON cells (Fig. 4d), consistent 254 with the position of their axonal arbors (Fig. 2c,d). 255 Although measuring rod-mediated responses from RBC2 was infeasible for technical reasons 256 (see Methods), we were successful in recording rod responses from RBC1. We were therefore 257 able to ask whether these responses mediated by mGluR6. We presented rod-isolating dim blue 258 flashes (10 ms) before and after introducing the mGluR6 receptor agonist 6-(2-259 aminopropyl)benzofuran (APB) to the perfusion solution. To isolate excitatory inputs to the cell, 260 all recordings were performed near the reversal potential for chloride-mediated conductances (~-261 60 mV) and in the presence of inhibitory receptor blockers, gabazine, strychnine, and TPMPA 262 ((1,2,5,6-Tetrahydropyridin-4-yl)methylphosphinic acid). Our results showed that nearly all rod 263 inputs were blocked in the presence of APB, indicating that, like mammalian RBCs, mGluR6 264 mediates rod input to RBC1 (Fig. 4e). 265 266

Both RBCs primarily synapse onto amacrine cells 267
To determine the synaptic targets of RBC1 and RBC2, we reconstructed their connectomes using 268 serial block face scanning electron microscopy (SBFSEM). In the reconstructions, we observed 269 an array of large BC axon terminals in the innermost layer of the IPL, which are characteristics of 270 RBC1 and RBC2 axons (Fig. S3a,b). To confirm that these large axon terminals belong to RBC1 271 and RBC2, we reconstructed dendrites of some of these BCs (Fig. 5a). Consistent with our 272 observations in light microscopic experiments ( Fig. 3a- We then reconstructed all the large axons in the SBFSEM image volume. To distinguish RBC2 288 from RBC1, we used the ribbon containing axonal distal bouton in the OFF layer as a proxy for 289 RBC2 (Fig. 3i,j and Fig. 5b). These reconstructions revealed the regular mosaic arrangements of 290 both presumed RBC1 and RBC2 (Fig. 5c), indicating that we identified most, if not all, presumed 291 RBC1 and RBC2 in the EM volume. Using this criterion, we also verified that dendritic tips of 292 RBC1 are doughnut shaped whereas those of RBC2 ended in a simple tip within the rod spherule, 293 consistent with our light microscopy data (Fig. 3c,d, 5a). 294 structures. We found that both RBC1 and RBC2 predominantly synapse onto ACs (Fig. 6a,f). The 308 majority of the postsynaptic processes received 4 or fewer ribbon synapses from one RBC1 or 309 RBC2, with an exception of one process, which received 14 inputs from one RBC1 (Fig. 6b,g). Taken together, these results demonstrated that RBC1 allocated the majority (91%) of synaptic 335 outputs to 3 types of AC: 14% to non-RS ACs, 37% to RS ACs, and 40% to one bi-stratifying ACs. 336 In contrast, the RBC2 we reconstructed synapsed primarily onto a mono-stratifying AC type (68%). 337 Among 24 ACs that we traced throughout the volume, only 5 were shared between RBC1 and 338 RBC2 (Fig. S5-9). Therefore, the downstream circuits of these BC types are largely separate at 339 least at the AC level. We found that zebrafish RBC1s synapse onto both wide field ACs with reciprocal synapses (the 364 RS ACs), resembling mammalian A17s, and a narrow field bi-stratifying ACs with extensive 365 synaptic connections, resembling mammalian A2s (Fig. 6b-e, Fig. S5,7). By marking synaptic 366 sites of RS ACs throughout their dendrites, we found that RS ACs are dedicated to the RBC1 367 pathway; synapsing predominantly (both input and output) with RBC1 and to a lesser extent with 368 RBC2 (Fig. S5), with no synapses with other BC types (n=7 cells). This synaptic specificity and 369 the reciprocal synapse arrangement in RS ACs mirror those of mammalian A17 ACs 31 (Fig. 7a,b). Next, we examined the synaptic arrangements of RBC1 with the A2-like ACs. First, we confirmed 393 that this type of AC is common to other RBC1s. By tracing the postsynaptic processes of 394 neighboring RBC1, we found another A2-like AC, which received a high number of ribbon inputs 395 from the neighboring RBC1 (Fig. 7c,d). We marked the locations of synapses with all BCs for 396 those two ACs (Fig. 7e). Gap-junctions are too small to be resolved in our SBFSEM images, but 397 as a proxy, we marked non-synaptic contacts (Fig. 7e). These revealed a striking similarity in the 398 distribution patterns of synapses and (potential) gap-junction sites across the IPL layers between 399 this AC type in zebrafish and mammalian A2 ACs (Fig. 7e,f), including the bouton structures in 400 the OFF layer that contain large presynaptic sites and mitchondria [45][46][47] (Fig. S4). Taken together, 401 we conclude that the circuit diagram among mammalian RBC, A2, and A17 ACs are conserved 402 in the zebrafish RBC1 pathway (Fig. 7g,h). 403 In contrast to the targets of RBC1, RBC2 formed synapses exclusively with wide-field ACs (Fig.  404 S8,9) and lack a synaptic partner with extensive synapses. Thus, RBC2 participates in a circuit 405 that differs from that of mammalian RBCs. 406

DISCUSSION 408
By combining scRNA-seq, electrophysiology, and light and electron microscopy circuit 409 reconstructions, we demonstrated that RBC1 shares many features with the mammalian primary 410 rod pathway (Fig. 7g,h), implying that the conserved rod pathway is evolutionarily ancient. 411

The number of BC types 412
In this study, we found 23 molecular types in adult zebrafish BCs. However, a previous This number of molecular types of bipolar cells (BCs) in zebrafish, as identified in this study, is 419 higher than that found in mammals investigated to date (14-15 across mammals) 26,27,29,53 , but 420 similar to that found in chick retina (22 molecular and 15 morphological BC types) 54,55 . The higher 421 number of BC types in zebrafish and chicken is not surprising, given that these species have 422 higher numbers of photoreceptor types: 5 in fish and 7 in chicks, compared to >=3 in mammals 56,57 . 423 We demonstrate here one source of the increase: a single type of BC carries most of the input 424 from rods in mammals, whereas zebrafish has two RBC types. 425

The number of RBC types 426
Previous morphological characterization of zebrafish BCs found only one BC type that connects 427 rods and red cones. Axons of these BCs terminate in the innermost layer of the IPL 48 . We 428 speculate that this type actually includes both RBC1 and RBC2, which were combined owing to 429 their striking morphological similarity (Fig. 2,3). Consistent with this hypothesis, studies in goldfish 430 have reported two morphologically distinct "mixed" BC types that receive dominant inputs from 431 rods 22 . They have large axon terminals at the bottom of the IPL, but the axon of one mixed BC 432 type contains a smaller axonal distal bouton in the OFF layer, similar to RBC2 58 . Immunostaining 433 for PKC only labels mixed BCs without an axonal distal bouton, similar to RBC1 59 . The presence 434 of these features in goldfish suggests that RBC1 and RBC2 are conserved among teleost fish. 435 A2-and A17-like ACs may also be conserved in goldfish. Paired elecrophysiological recordings 436 between goldfish RBC1 and ACs revealed that RBC1 provides synaptic inputs to two 437 morphological types of ACs: wide-field mono-stratifying and narrow field bi-stratifying AC types 60 . 438 The dendrites of the bi-stratifying ACs wrap around the RBC1 axon terminals 60 , similar to A2 ACs 439 in zebrafish and mammals (Fig. 7c). Goldfish RBC1 receives GABAergic reciprocal feedback at 440 the axon terminals, like mammalian RBC 61,62 . Taken together, although it remains unknown 441 whether these two AC types in goldfish exhibit similar synaptic connectivity patterns to those of 442 mammalian A2 and A17 ACs, the findings in goldfish are consistent with the idea that the primary 443 rod pathway, including A2 and A17 ACs, is conserved in goldfish. Some differences in 444 physiological properties between mammalian RBC and goldfish RBC1 were also found. First, 445 goldfish RBC1 receives GABAergic lateral inhibition 61 . The exact cell types that provide this 446 inhibitions are unclear, but it is likely coming from the wide-field mono-stratifying ACs, as their 447 dendrites extend laterally. In contrast, mammalian A17 ACs do not provide lateral inhibition onto 448 RBCs, as each varicosity of A17 ACs at the RBC axon terminals operates independently of each 449 other 63 . Second, goldfish RBC1 exhibits spikes 64,65 , whereas the spikes are only found in cone 450 BCs in mammals. Nonetheless, the absolute visual sensitivity of goldfish is comparable to that of 451 mammals 66 , suggesting that the primary rod pathway we discovered in teleosts is capable of to identify BC cells that connect with all rods in their dendritic field 54 . Unlike the species mentioned 462 above, the presumed rod bipolar cells, which have light sensitivity close to that of rods, in sea 463 lamprey are OFF type 70 . However, their connectivity with rods is unknown. Thus, it remains 464 unclear whether RBC2 orthologs are present in species other than zebrafish. 465 In mammals, morphological, molecular, and functional studies have identified only a single RBC 466 type 23,26,27,71 . Therefore, we speculate that either RBC2 evolved after the divergence between 467 teleost fish and mammals, or mammals lost this pathway. 468

Roles of cone inputs in RBCs 469
Cone inputs onto rod-dominant mixed BCs have been proposed to broaden the dynamic range of 470 light intensities to which they can respond 72 . Consistent with this idea, we found that both RBC1 471 and RBC2 are selective for red cones, which, with their broad spectral sensitivity, are suited for 472 encoding achromatic luminance information 73 . Because rods evolved from cones 1 , we speculate 473 that RBCs may have emerged from red-cone specific CBCs. The red cone selectivity is also 474 conserved in at least one of the mixed rod dominant BC types in goldfish 74 . Although cone 475 selectivity is unknown in Salamander rod-driven mixed BCs, their spectral sensitivity curve is 476 broader at longer wavelengths than that of rods, indicating that they may connect to red cones 14 . 477 Electrophysiological recording from rod-driven BCs in Giant Danio, a teleost fish species, showed 478 that rod and cone inputs onto rod-dominant BCs are mediated by different mechanisms: rod inputs 479 through mGluR6, whereas cone inputs through both mGluR6 and EAAT (excitatory amino acid 480 transporter) 72 . In this BC type, mGluR6 and EAAT-mediated inputs suppress each other, likely to 481 allow this cell to respond to both rod and cone dynamic ranges 72 . Electrophysiological recordings 482 in zebrafish found that some of ON BCs responded to glutamate via both mGluR6 and a 483 glutamate-gated chloride conductance increase mechanism, which is likely through EAATs 75 . 484 However, the nature of EAAT contributions for cone responses in RBC1 and RBC2 is unknown. 485 While the study in Giant Danio suggest that mixed inputs expand the dynamic range of rod-486 dominant BCs, electrophysiological recordings in goldfish and salamander have found that the 487 dynamic range of rod dominant BCs is similar to that of rods and that cone contributions to the 488 light response are small 14,76 . Therefore, the roles of red-cone inputs to RBCs remain to be 489 determined. 490

Unifying mixed BCs and RBCs 491
In mammals, it was initially thought that RBCs exclusively synapse with rods 77 . However, several 492 recent studies have demonstrated convincingly that RBCs also receive synapses from cones, at 493 least in mice and rabbits 8,77,78 . Indeed, mouse RBCs contact the majority of M-cones (~80%), 494 which are analog of zebrafish red-cone, in their dendritic territories 8 . RBCs were likely thought to 495 be exclusive to rods because of the high ratio of rods to cones in the outer nuclear layer in mice 496 and rabbits 79,80 . As a consequence, only a few cones, generally three or fewer, synapse on a 497 mouse RBC, compared to inputs from ~35 rods 8,78 . 498 The rod-driven BCs in non-mammals are classically called "mixed" BCs because they connect 499 with both rods and cones. However, as argued above, this mixed connectivity is conserved in the 500 mammalian RBCs. Moreover, the dendritic specificity and connectivity of RBCs are conserved in 501 mice and zebrafish. In both species, RBCs connect with all rods and the majority of red-cones (or 502 M-cones) in their dendritic fields (Fig. 3f,h). Therefore, although the coverage of cones in mice is 503 still lower than that in zebrafish, converging rod and red-cone inputs is likely a conserved feature 504 of RBCs in all vertebrates even if the proportion and number of cone inputs may vary across 505 species. This leads us to propose that the non-mammalian mixed BCs and the mammalian RBCs 506 represent a single class of neurons, RBCs. Finally, taken together with the striking similarity in 507 the downstream circuitry of RBCs between zebrafish and mammals, we conclude that zebrafish 508 RBC1 is transcriptomically, anatomically, and functionally equivalent of mammalian RBC and that 509 they share the same evolutionary origin. 510

ACKNOWLEDGEMENTS 512
We thank the Vision Core at the University of Washington for processing zebrafish retina samples 513 and acquiring serial images for SBF-SEM. We thank Rachael N. Swanstrom for helping with cell 514 tracing of the EM volume. Funding was provided by the MCSA fellowship ("ColourFish" 748716) 515 from the EU Horizon 2020 to TY, the NIH EY14358 to ROW, U01MH105960 and R01 EY022073 516 University of Sussex. For all experiments, we used adult zebrafish (age 6-18 months) of either 542 sex that were kept at 28°C in a room with a normal 14/10 light cycles. 543 The following previously published transgenic lines were used: Tg(vsx1:GFP) nns5 83 , 544 Tg(vsx1:memCerulean) q19 84 , Tg(trb2:tdtomato) q22 85 . In the larva Tg(vsx1:memCerulean) q19 labels 545 a subpopulation of OFF layer stratifying BCs 84 . In adults, while OFF stratifying BCs are still weakly 546 labeled, Cerulean is now strongly expressed in RBC1 (Fig. 2a,c). In addition, 547 Tg(vsx2:memCerulean) wst01 line was generated by injecting pBH-vsx2-memCerulean-pA plasmid 548 into single-cell stage eggs. Plasmid was diluted in 1x Danieau's solution to a concentration of 50 549 ng/ml. Plasmid solution was loaded into a pulled-glass micropipette, mounted to a 550 micromanipulator (Narishige), and pressure-injected via attachment to a Picospritzer II (Parker). 551 Injections were made at 10 psi for durations from 100 to 200 ms. Injected fish were raised and 552 out-crossed with wild-type fish to screen for founders. Positive progenies were raised to establish 553 transgenic lines.

Single cell transcriptomics data analysis 573
We performed the initial preprocessing using the cellranger software suite (version 2.1.0, 10X 574 Genomics), following steps described previously in our study of Zebrafish RGCs 86 . The 575 sequencing reads were demultiplexed using "cellranger mkfastq" to obtain a separate set of 576 fastq.gz files for each of 8 samples, which were distributed across Y biological replicates. Reads 577 for each sample were aligned to the zebrafish reference transcriptome (ENSEMBL zv10, release 578 82) using "cellranger count" with default parameters to obtain a binary alignment file and a filtered 579 gene expression matrix (GEM) for each sample. To account for intronic reads, the binary 580 alignment files were processed using velocyto with default parameters 87 , producing a loom file 581 containing a GEM for exonic reads and a separate matrix for intronic reads. The matrices were 582

Preprocessing and batch integration 590
The combined GEM was filtered to remove genes expressed in fewer than 25 cells, and cells 591 expressing fewer than 50 genes resulting in 25,233 genes and 19,492 cells. Briefly, each cell was 592 normalized to a total library size of 10,000 and the normalized counts were log-transformed using 593 the function Seurat::NormalizeData. We used Seurat::FindVariableFetures with option 594 selection.method = "vst" to identify the top 2000 highly variable genes (HVGs) 88 in each batch. 595 Next, we performed scRNA-seq integration. We used Seurat::FindIntegrationAnchors and 596 Seurat::IntegrateData, both with options "dims=1:40" to perform Canonical Correlation Analysis 597 (CCA)-based batch correction on the reduced expression matrix consisting of the HVGs. The 598 "integrated" expression values were combined across replicates, and used for dimensionality 599 reduction and clustering. 600

Dimensionality Reduction, Clustering and Visualization 601
To remove scale disparities between genes arising from differences in average expression levels, 602 the integrated expression values for each HVG were z-scored across the cells using 603 Seurat::ScaleData. Next, we performed Principal Component Analysis (PCA) on the scaled matrix, 604 and used Seurat::ElbowPlot to select principal components (PCs). Using the top 20 PCs, we built 605 a k-nearest neighbor graph using Seurat::FindNeighbors and identified transcriptionally distinct 606 clusters using Seurat::FindClusters, using a resolution parameter of 0.5. 607 Using the top 20 PCs, we also embedded the cells onto a 2D embedding using Uniform Manifold 608 performed using pvclust with parameters method.hclust = "complete" and method.dist = 620 "correlation". The resulting output was visualized as a dendrogram. 621

Immunostaining and light microscopy imaging 630
Adult zebrafish were humanely euthanized in ice-chilled fish water. After decapitation, retinal 631 tissues were dissected from the enucleated whole eyes by removing cornea, lens, and epithelial 632 layer in 1x in phosphate-buffered saline (PBS). The tissue were immediately fixed in 4% 633 paraformaldehyde (Agar Scientific, AGR1026) in PBS for 20 min at room temperature (RT) 634 We obtained the cell density of RBC1 and RBC2 by counting the axon terminals of these cells 659 within regions of interest from confocal image stacks of the dorsal and ventro-temporal retina. 660 RBC2 was counted from images of the Tg(vsx2:memCerulean) wst01 line. Because not all RBC1 661 labeled in the Tg(vsx1:memCerulean) q19 line express mCerulean, we quantified the density of 662 PKC labeled cells with axon terminals in the bottom layer of the IPL. Counts were obtained from 663 3-4 retinas from 3 animals of each line. For RBC1, axons were quantified within an area between 664 37,000 and 85,000 µm 2 , and for RBC2, the areas were between 11,000 -22,000 µm 2 . 665

Dendritic field 666
The dendritic field was defined by tracing the extent of a given cell's dendrites with the polygonal 667 select tool, and removing any concavity using FIJI (see Figure S2). The dendritic arbor area was 668 then obtained by calculating the area enclosed by the polygon. Because the dendritic tips of some 669 neighboring cells of the same type overlapped and could not be distinguished readily, one 670 investigator repeatedly traced (3 to 4 measurements for a single cell) the dendrite boundary, and 671 obtained the respective area for a given cell until at least three measurements were within ± 2.5% 672 of the average of all previous measurements for that cell. Confocal images from three fish were 673 used, with images of RBC1 and RBC2 cells acquired from the dorsal and ventral regions of the 674 retina: 10 to 17 cells per fish were measured for each location and cell type, resulting in a total of 675 33 to 41 cells measured for each location and cell type.