Amacrine cells differentially balance zebrafish colour circuits in the central and peripheral retina

In vertebrate vision, the feature-extracting circuits of the inner retina are driven by photoreceptors whose outputs are already pre-processed. In zebrafish, for example, outer retinal circuits split “colour” from “greyscale” information across all four cone-photoreceptor types. How does the inner retina process this incoming spectral information while also combining cone-signals to shape new greyscale functions? We address this question by imaging the light driven responses of amacrine cells (ACs) and bipolar cells (BCs) in larval zebrafish, in the presence and pharmacological absence of inner retinal inhibition. We find that amacrine cells exert distinct effects on greyscale processing depending on retinal region, as well as contributing to the generation of colour opponency in the central retina. However, in the peripheral retina amacrine cells enhanced opponency in some bipolar cells while at the same time suppressing pre-existing opponency in others, such that the net change in the number of colour-opponent units was essentially zero. To achieve this ‘dynamic balance’ ACs counteracted intrinsic colour opponency of BCs via the On-channel. Consistent with these observations, Off-stratifying ACs were exclusively achromatic, while all colour opponent ACs stratified in the On-sublamina. This study reveals that the central and peripheral retina of larval zebrafish employ fundamentally distinct inhibitory circuits to control the interaction between greyscale- and colour-processing. Differential actions on the On- and Off-channels control the transmission of colour-opponent signals in the periphery.


circuits in the central and peripheral retina
Xinwei Wang 1$ , Paul A Roberts 1 , Takeshi Yoshimatsu 1 , Leon Lagnado 1*$ 4 and Tom Baden 1,2*$ 5 6 SUMMARY. In vertebrate vision, the feature-extracting circuits of the inner 7 retina are driven by photoreceptors whose outputs are already pre-8 processed. In zebrafish, for example, outer retinal circuits split "colour" 9 from "greyscale" information across all four cone-photoreceptor types. How 10 does the inner retina process this incoming spectral information while also 11 combining cone-signals to shape new greyscale functions? 12 We address this question by imaging the light driven responses of 13 amacrine cells (ACs) and bipolar cells (BCs) in larval zebrafish, in the 14 presence and pharmacological absence of inner retinal inhibition. We find 15 that amacrine cells exert distinct effects on greyscale processing 16 depending on retinal region, as well as contributing to the generation of 17 colour opponency in the central retina. However, in the peripheral retina 18 amacrine cells enhanced opponency in some bipolar cells while at the 19 same time suppressing pre-existing opponency in others, such that the net 20 change in the number of colour-opponent units was essentially zero. To 21 achieve this 'dynamic balance' ACs counteracted intrinsic colour 22 opponency of BCs via the On-channel. Consistent with these observations, 23 Off-stratifying ACs were exclusively achromatic, while all colour opponent 24 ACs stratified in the On-sublamina. 25 This study reveals that the central and peripheral retina of larval zebrafish 26 employ fundamentally distinct inhibitory circuits to control the interaction 27 between greyscale-and colour-processing. Differential actions on the On-28 and Off-channels control the transmission of colour-opponent signals in the 29 periphery.

INTRODUCTION
species remains unknown, we might expect ACs to contribute to both 102 chromatic and achromatic signalling [27][28][29] . To test if this is the case, we 103 surveyed light-driven signals of ACs in vivo across different eye regions of 104 the zebrafish retina. Surprisingly, this revealed that despite being highly 105 diversefor example in terms of kinetics and polarity -ACs were mostly 106 non-opponent and spectrally resembled linear combinations of UV-and 107 red-cone signals, which in zebrafish are associated with greyscale 108 processing 7 . Next, we imaged BC light responses in the presence and 109 absence of AC-mediated inhibition. This demonstrated that while BC 110 greyscale processing was profoundly altered across the entire retina, 111 colour processing was affected differently in different regions of the retina. 112 While in the central retina ACs served to set up a new "UV:yellow" axis of 113 colour opponency, in the peripheral retina the population representation of 114 spectral contrast 5,30 was essentially invariant to the removal of inhibitory 115 signals. However, this was not because opponency in individual BCs was 116 invariant to AC-block. On the contrary: ACs both routinely abolished and 117 generated spectral opponency at the level of individual BCs, but they did 118 so in approximately equal measure, such that the net change across the 119 population of BCs was essentially zero. To preserve the balance between 120 different chromatic and achromatic channels, ACs act near-exclusively 121 through On-circuits. 122 123 We conclude that distinct circuits serve to conserve the parsing of colour 124 information performed in the outer retina as the visual signal is transmitted 125 to central and peripheral BCs. The inhibitory interactions within the inner 126 retina that underly other visual processing tasks do not notably alter the 127 population representation of colour information. 128 129 RESULTS

131
Surveying amacrine cell functions 132 To investigate how inhibitory microcircuits in the inner retina contribute to 133 processing we began by recording how the population of amacrine cells 134 (ACs) in larval zebrafish encode greyscale and colour information across 135 the inner plexiform layer (IPL) and across different parts of the eye. For 136 this, we used in vivo 2-photon imaging of SyGCaMP3.5 expressed under 137 the ptf1a promoter which targets the vast majority of ACs in zebrafish 25,31 138 ( Figure 1A-F). We recorded dendritic calcium responses of ACs to a 139 battery of widefield light stimuli testing basic visual processing tasks 140 (Methods): (i) an achromatic ("white") step of light (3 s On, 3 s Off, 100% 141 contrast) testing response polarity and kinetics ( Figure 1D); ii) a frequency 142 modulated chirp centred at 50% contrast testing frequency response 143 ( Figure 1D), (iii) steps of light (2 s On, 2 s Off, 100% contrast) at four 144 different wavelengths ('red': 592 nm; 'green': 487 nm; 'blue': 420 nm; 'UV' 145 382 nm) testing spectral sensitivity ( Figure 1E) and (iv) 'tetrachromatic 146 binary noise' (5 mins, 6.4 Hz, 100% contrast) which allowed us to extract 147 four 'spectral sensitivity kernels' per terminal to probe for spectral 148 opponency ( Figure 1F, Methods). 149   Figure S1A). In total, we recorded from n=927 ROIs in the central retina 172 and n=816 ROIs from the peripheral retina.

182
AC processes exhibited diverse responses across the tested battery of 183 stimuli ( Figure 1D-F). ROI 1, for example, consistently exhibited transient 184 Off-responses to both the white chirp stimulus ( Figure 1D) and when 185 probed with colour flashes ( Figure 1E), while ROI 2 exhibited sustained 186 On-responses. These different behaviours were also captured by the 187 spectral kernels ( Figure 1F), which additionally highlighted an overall 188 preference for long-over short-wavelength stimulation in both cases.  Figure S1B, Methods). This procedure simplified 210 complex responses into four components and their corresponding weights.

211
For example, ROI 2 ( Figure 1D) exhibited a relatively slow On-response 212 that was readily captured by a positively weighted Light-sustained 213 component alone (Supplemental Figure S1B,   complete overview in Supplemental Figure S2 and Supplemental Figure  247 S2 Extended 1). Regional information was not used to drive the clustering 248 (see Supplemental Figure S2

267 268
Amacrine cells are kinetically diverse but spectrally simple 269 Using the output of the joint clustering procedure, we compared how ACs 270 encode greyscale and colour information. As we detail below, this revealed 271 that ACs could be divided into two main spectral groups: a majority of 272 kinetically diverse but spectrally simple achromatic ACs, and a minority of  light (Figure 2A,B), a behaviour that was also captured in the spectral 287 kernels ( Figure 2C,D). For example, cluster C5 exhibited transient Off-288 responses to steps of light at any wavelength, while cluster C7 consistently 289 displayed sustained Off-responses. Correspondingly, each of the four 290 spectral kernels also indicated Off-behaviour ( Figure 2C), rendering these 291 clusters non-opponent. Similarly, kinetically distinct On-clusters C15 and C27 292 were also non-opponent. The remaining six clusters were more complex.

293
For example, C22 displayed transient On-Off responses at all tested 294 wavelengths, however with a notable Off-dominance during long-295 wavelength stimulation and On-dominance during short-wavelength 296 stimulation, rendering this cluster colour-opponent overall. Such 297 wavelength-dependent rebalancing of On-versus Off-amplitudesrather 298 than a 'classical' full polarity reversal as observed in cones 7 and bipolar 299 cells 18,19also rendered the remaining five clusters weakly opponent 300 overall.

302
The dominance of non-opponent responses amongst AC-clusters was 303 further illustrated by comparison of kernel amplitudes across different 304 wavelengths ( Figure 2F-H). For example, pairwise comparison of red-305 versus green-kernel amplitudes highlighted that most clusters exhibited 306 same-sign behaviour at both wavelengths ( Figure 2F). Only a minority 307 were red-green opponent (arrowhead), and these clusters notably also 308 exhibited the lowest kernel amplitudes overall. Qualitatively similar 309 behaviour was observed when comparing red-blue and red-UV 310 wavelengths ( Figure 2G,H).   Off-clusters, as well as five of the ten On-clusters followed a common, 330 long-wavelength biased and non-opponent spectral tuning function ( Figure  331 3A,B). Off-long-biased ROIs were mostly, though not exclusively located in 332 the IPL's traditional Off-layer ( Figure 3E), while On-long-biased ROIs were 333 only found in the On-layer ( Figure 3F).

356
All remaining non-opponent On-clusters fell into a third group that was 357 spectrally "V-shaped" ( Figure 3C), with ROIs exhibiting an incomplete bias 358 to the centre of the IPL. This last group was overwhelmingly comprised of 359 ROIs from the central retina (Supplemental Figure S3A) which is known to 360 be heavily UV-dominated 18,19,33,34 . The spectral tuning of all 21 non-361 opponent ACs clusters was readily explained by inputs from red-and UV-362 cones (Supplemental Figure S3B-E), which are associated with achromatic 363 processing 7 . Strikingly, essentially all ROIs in the Off-layer and 364 approximately half in the On-layer were of this non-opponent population. 365 The remaining half of ROIs in the On-layer consisted of six AC clusters that 366 were both kinetically and spectrally complex, and weakly but consistently 367 colour opponent ( Figure 3D,H). The spectral behaviour of these colour-368 opponent clusters could not generally be explained without additional 369 inputs from the opponent 7 green-cones which are associated with colour 370 processing (Supplemental Figure S3B-E).

372
We next tested how these different distributions of amacrine cell functions 373 across the eye and IPL might be linked to spectral and temporal 374 processing in bipolar cells.

376
Bipolar cell signalling in the presence and absence of inhibition from 377 amacrine cells. 378 To investigate the effects of AC-mediated inhibition on the visual signal 379 transmitted through the inner retina, we combined pharmacology with in 380 vivo 2P imaging of BC synaptic terminals expressing the calcium biosensor 381 SyjGCaMP8m 18,42-44 ( Figure 4A-F, Methods). In each experiment, we first 382 scanned ~10˚ eye regions comprising typically 100-120 individual BC 383 terminals ( Figure 4A,B) and presented the same battery of stimuli 384 previously used to characterise amacrine cells ( Figure 4C-E). Next, we 385 injected a cocktail of gabazine, TPMPA, and strychnine into the eye to 386 pharmacologically block GABAA, GABAC and glycine receptors, 387 respectively 26 (Methods), which represent the major known sources of AC-388 mediated inhibition in the inner retina 15 . We then imaged the inner retina a 389 second time (e.g. Figure 4F-H) to compare the functions of BC terminals in 390 the presence or pharmacological absence of AC-mediated inhibition.

392
The efficacy with which this manipulation blocked inhibition in the inner 393 retina was evidenced by the increase in the gain of responses in BC 394 synapses ( Figure 5, Supplemental Video 1) and the decorrelation 13 of 395 these responses (Supplemental Figure S4A,B). We also evaluated the 396 effect of blocking inhibitory receptors on outer retinal function, where 397 horizontal cells spectrally retune the cone output 7 . Using existing cone-type 398 specific SyGCaMP6f lines 7 , we confirmed that the cones' spectral tunings 399 were invariant to the application of the drug-cocktail (Supplemental Figure  400 S4C-N). 401 402

429
In line with previous work 18, 19,45 BCs displayed a broad range of response 430 properties under control conditions, which included both On-and Off cells 431 with diverse temporal and spectral tunings. For example, in a scan from the 432 peripheral retina (Figures 4A,B) ROIs 1 and 2 displayed largely achromatic 433 Off-and On-responses, respectively, while ROIs 3-5 exemplified different 434 forms of colour opponency ( Figure 4C-E, Supplemental Video 2). A 435 substantial degree of functional diversity in BC-responses was also 436 observed following pharmacological AC-block, including the continued 437 presence of numerous colour opponent responses ( Figure 4F-H). 438 However, the nature and distribution of these disinhibited responses were 439 profoundly altered compared to control conditions, and in a manner that 440 systematically differed between the central and peripheral retina, as we 441 describe below.

443
Changes in greyscale processing caused by blocking inhibition 444 The most general effects of blocking inhibition from ACs were to make 445 responses in BCs larger and more transient, and this occurred across 446 retinal regions and for terminals of all polarities ( Figure 5). By fitting step 447 responses to the same four kinetic components previously used to fit ACs 448 we could account for >94% of the variance across BCs. Using the kinetic 449 weights, we automatically classified each BC-response as either 450 unresponsive, On, Off, or On-Off, and computed the average chirp-451 response traces for the latter three categories per retinal region and 452 condition ( Figure 5A,B). In both the central and peripheral retina, blocking 453 ACs reduced the number of Off-and unresponsive terminals and 454 unmasked the presence of 'intrinsically On-Off' terminals. Following block 455 of AC inputs, On-Off terminals were also observed in response to coloured 456 stimuli, most notably to red-and UV (Supplemental Figure S5A), but they 457 were never observed under control conditions. 458 459 Other effects of blocking inhibition were dependent on retinal region. 460 Unmasking of On-Off responses, for example, was much more common in 461 the peripheral compared to central retina. Moreover, on average, 462 peripheral BCs of all polarities followed the frequency-accelerating part of 463 the chirp for longer compared to central BCs, suggesting regional 464 differences in the modulation of temporal processing. Overall, while 465 blocking ACs mainly accentuated the pre-existing On-bias of the central 466 retina, the same manipulation yielded a more complex re-distribution of 467 response properties in the periphery.  19,33

492
To analyse how these changes in BC function were distributed across the 493 IPL, we segregated terminals into ten strata and computed histograms 494 summarising the relative depth-distributions of On-, Off, and On-Off 495 terminals in each region and condition ( Figure 5C,D). Blocking inhibition 496 from ACs had distinct effects in the central and peripheral retina. In the 497 central retina, ectopic Off-responses in the On-layer were abolished 498 ( Figure 5C, arrowhead) but these were not affected in the peripheral retina 499 ( Figure 5D, arrowhead 1). In the central retina, blocking inhibition also 500 generated mixed On-Off responses in the Off-layer ( Figure 5C, arrowhead  501 2), while in the peripheral retina On-Off response appeared throughout the 502 IPL ( Figure 5D, arrowhead 2). These results demonstrate that ACs do not 503 simply regulate the gain and kinetics of the output from BCs, but also the 504 polarity. In the absence of inhibition, 19.3% of BC terminals in the 505 peripheral retina signal both On and Off transitions and these were 506 predominantly in the On-layer. This previously unrecognized function of 507 ACs -regulating the polarity of synaptic activity in BCs -was less 508 prominent in the central retina, where it was only evident in 7.6% of 509 terminals, and these were predominantly in the Off-layer. 510 511 Regional differences in the way that ACs interact with BCs were also 512 evident in the temporal domain ( Figure 5E,F). Responses tended to 513 become more transient after blocking inhibition but this effect was much 514 stronger in the periphery, where it involved an accentuation of both Light-515 and Dark-transient response components (Figure 5F, arrowheads). In 516 contrast, kinetic changes in the central retina were more moderate and 517 restricted to Light-transient and Light-sustained components ( Figure 5E, 518 arrowheads). 519 520 These results demonstrate that ACs interact with BCs in a highly regional 521 manner. In the central retina, ACs regulate the gain and speed of 522 responses, suggesting that here, ACs primarily serve as a gain-control 523 system 46 . But in the peripheral retina, ACs also regulate the segregation of 524 On and Off signals in the On-layer. Below we ask how these functional 525 reorganizations of the inner retina impact spectral processing.  Figure 5A,B, but here shown 530 separately for the four colour steps (cf. Figure 4D)  'currency' of colour vision. In zebrafish, BCs represent three types of 541 spectral opponency ( Figure 6A): Long-("red-green"), mid-("orange-blue") 542 and short-("yellow-UV"), with spectral zero crossings at ~523, ~483 and 543 ~450 nm, respectively 18, 19 . Of these, long-("red:green") and mid-544 wavelength opponency ("orange:blue") is already encoded at the level of 545 green-and blue-cones, respectively 7 , implying that ACs are not necessary 546 to establish these channels in BCs. In contrast, short-wavelength 547 opponency ("UV:yellow") is only weakly represented in UV-cones 7 , but 548 dominant amongst BCs 18 . The expectation, therefore, is that short-549 wavelength opponency requires the activity of ACs. This expectation was 550 confirmed in the case of BCs in the central retina but not the periphery. 551 552 553 554

558
Percentages of ROIs per spectral category (cf. A) based on BC kernels (cf. Figure 4E)

571
Opponency in BCs' was assessed from spectral kernels computed from 572 colour-noise responses, as used for classifying ACs (Figure 3A-D). 573 However, because BCs 18,19 are more spectrally diverse than ACs ( Figures  574  1-3), we classified each into one of eight (rather than four) groups: five 575 non-opponent groups (broad, long-biased, mid-biased, short-biased, V-576 shaped) and three opponent groups (long-opponent: Opp1, mid-opponent: 577 Opp2, short-opponent: Opp3, Figure 6A,Bfor full classification see 578 Supplemental Figure S6). This group allocation confirmed previous 579 results 18,19 that under control conditions, the distribution of spectral 580 response-types strongly differed across the two regions: While in the 581 central retina, spectral groups that required a strong UV-input accounted 582 for the vast majority of responses, the diversity of spectral response types 583 was much more evenly distributed in the peripheral retina, including greater 584 numbers of colour-opponent neurons (peripheral: 43.3%; central: 27.9%). 585 Strikingly, in both retinal regions, these spectral distributions remained 586 largely unchanged following AC-block: No major spectral response group 587 disappeared altogether ( Figure 6B, and Supplemental Figure S6), and the 588 abundance of many spectral groups was unchanged. For example, the 589 numbers of V-shaped non-opponent responses appeared entirely 590 unaffected in both retinal regions, while most colour-opponent groups 591 exhibited only marginal changes.

593
The only significant effect of blocking inhibition on colour opponency that 594 we could detect was a loss of short-opponent responses in the central 595 retina in favour of a corresponding gain in short-biased non-opponent 596 responses ( Figure 6B, arrow). This change was not observed in the 597 peripheral retina, where all three opponent groups persisted throughout the 598 pharmacological manipulation. 599 600 Having established that the short-opponent interactions between ACs' and 601 BCs were specific for eye-region, we looked more closely at inhibitory 602 circuits at different locations in the IPL. For simplicity, the distribution of 603 colour opponency was assessed by summing the three colour-opponent 604 groups into a single distribution per experimental condition ( Figure 6C,D, 605 individual distributions shown in Supplemental Figure S6B,D). The colour-606 opponent responses appearing after block of inhibition were both short-607 and long-wavelength opponent (Supplemental Figure S6D, arrowheads).

609
This finer analysis revealed that, despite the overall numerical conservation 610 of all three opponencies in the peripheral retina following block of inhibition 611 ( Figure 6B), this was made up of a loss of opponency in the Off layer and a 612 gain in the On layer ( Figure 6D).

645 ACs both create and mask colour opponency in individual BCs 646
To investigate the how blocking inhibition caused a redistribution of colour-647 opponency in BCs we tracked the same terminals across recordings. This 648 approach was not possible when expression of SyGCaMP was driven in all 649 BCs because the high density of terminals in the IPL made it difficult to 650 reliably identify the same terminal before and after the pharmacological 651 manipulation. We therefore performed a new set of experiments using a 652 different transgenic line where BCs expressed SyGCaMP3.5 sparsely 653 (Supplemental Figure S7A). This strategy allowed us to record from 20-30 654 individual terminals at a time, out of which 40-60% could be reliably 655 matched across control conditions and after blocking inhibition 656 (Supplemental Figure S7A-D). We sampled 182 terminals from 14 fish 657 covering the entire depth of the IPL. Because labelling was sparse, we 658 combined all paired data into a single eye-wide dataset. We also recorded 659 an equivalent but independent sham control dataset (n = 6 scans, n = 144 660 paired terminals), where we replaced the drug cocktail used to block 661 inhibition with an equivalent volume of non-pharmacologically active 662 vehicle. Sham injections had no significant effect on the functions we 663 analysed (Supplemental Figure S7 Extended, Methods).

665
As expected, blocking inhibition generally disinhibited BCs, resulting in less 666 selective, more transient, and larger amplitude responses ( Figure 7A-E).

667
As in our population dataset, changes in spectral processing were diverse. 668 In ROI-pair 1, responses to colour steps ( Figure 7B) were red-biased 669 during control conditions but responded to all four wavelengths after 670 blocking inhibition, and this spectral broadening was also observed at the 671 level of the kernels ( Figure 7C). It appears that in this case ACs were 672 masking an intrinsic short-wavelength response to set-up a long-673 wavelength biased BC. However, the effects on ROI-pairs 2 and 3 were 674 functionally opposite: ROI-pair 2 exhibited a green-UV colour-opponent 675 response during control conditions, which was abolished following AC-676 block, while vice versa ROI-pair 3 exhibited weak non-opponent response 677 during control conditions but green-UV opponency upon AC-block. 678 Accordingly, in ROI-pair 2, ACs were responsible for setting up BC-679 opponency, while in ROI-pair 3 ACs masked an intrinsic form of BC-680 opponency.

700
To systematically assess how BC colour opponency is generated and/or 701 destroyed by AC-circuits, we again allocated BC terminals into spectral 702 groups ( Figure 7F,G, cf. Figure 6B). Again, there was very little overall 703 change among colour-opponent groups, yet more than half of individual 704 terminals that exhibited colour opponency under control conditions lost 705 their opponency following AC-block (n = 29 of 49, 59.2%). At the same 706 time, an almost equal number of previously non-opponent BCs replenished 707 the population of colour opponent BCs (n = 24). This switching of opponent 708 BCs between conditions affected all three opponent groups: Only 2/8 709 (25%), 2/7 (29%) and 15/33 (45%) long-, mid-and short-wavelength 710 opponent terminals, respectively, maintained their opponency throughout following AC-block. The overall picture, therefore, is that ACs exert 718 different actions on colour-opponency in different BC terminals: in some 719 terminals, ACs contribute to the generation of colour-opponency but in 720 others they masked pre-existing opponency. These opposing effects of 721 ACs were exerted on all three colour-opponent channels in approximately 722 equal measure.

724
ACs modulate BC-spectral processing via the On-channel 725 The dominance of colour-opponent AC-circuits in the On-layer ( Figure 3H) 726 suggests that these ACs have a role in determining how the output from 727 BCs is spectrally tuned. To test this idea, we analysed the colour-step inhibition. This apparent On-dominance in AC-dependent spectral re-tuning 736 was particularly striking in the example shown in Figure 8D: this BC-737 terminal exhibited spectrally non-selective Off-responses under control 738 conditions, but block of inhibition unmasked spectrally selective short-739 wavelength On-responses (arrowheads), making this terminal colour-740 opponent overall. 741 742 To systematically assess the role of On-and Off circuits for shaping 743 chromatic and achromatic circuit functions, responses to each colour-step 744 were again fitted with a weighted sum of four kinetic building-blocks. We 745 reasoned that any achromatic effects of ACs on BCs should lead to a high 746 degree of covariation across wavelengths, while any spectral 'retuning' of 747 BCs should manifest in some wavelength-responses being affected more 748 than others. To test this, the degree of response covariation across 749 wavelengths was assessed for each terminal. The plot in Figure 8E shows 750 how blocking inhibition from ACs changed responses in On-and Off-751 terminals during red stimulation (dRed) plotted against the corresponding 752 amplitude changes in response to UV stimulation (dUV). The equivalence 753 line at 45 o is the case where responses to red-and UV covary perfectly. 754 Most Off-responses (black dots) fell near the equivalence line but On-755 responses showed a mixture of behaviours with a second population falling 756 on or near the 0˚ line (arrowhead) representing response amplitudes 757 changing to UV steps but not red. Only few points fell on the 90˚ line, 758 indicating a notable absence of On-responses that were modulated in red 759 without also being modulated in UV.

782
To summarise this behaviour, we computed the corresponding angular 783 histogram ( Figure 8F), which showed a single peak around 45˚ for Off-784 responses indicating mostly co-variation, but two main peaks for On-785 responses: one at 45˚, and another at 0˚. This general pattern was stable 786 for all possible colour combinations ( Figure 8G, individual colour pairs 787 shown in Supplemental Figure S8A-F). In the On-channel, but not in the 788 Off, shorter wavelength responses were modulated more strongly than 789 long wavelength responses.

797
This analysis provides further evidence that most spectral modulation of 798 BCs by ACs occurs via the On-channel, with short-wavelength circuits 799 being key targets of this modulation. 800 801 DISCUSSION 802 Investigations of visual processing have usually dealt with stimulus 803 dimensions of space and time, grey-scale processing, separately to colour. 804 Here we have shown that grey-scale and colour processing interact 805 through inhibitory circuits in the inner retina that vary between different 806 zones (Figures 1-3). Blocking inhibition from ACs increased gain in BCs 807 and made responses more transient, as well as unmasking mixed On-Off 808 that were much more common in the peripheral retina compared to the 809 center (Figures 4,5). Simultaneously, ACs contributed to the generation of 810 colour-opponency in the central retina but in the periphery, there was a 811 mixture of effects, enhancing colour-opponency in some BCs while 812 suppressing pre-existing opponency in others (Figures 3,6,7). ACs 813 counteracting intrinsic colour-opponency of BCs acted with a high degree 814 of specificity through just one of the two fundamental channels for grey-815 scale processing -the On-pathway (Figure 8). We conclude that the 816 central and peripheral retina of larval zebrafish employ fundamentally 817 distinct inhibitory circuits to control the interaction between greyscale-and 818 colour-processing. 819 820 The role of ACs in colour vision 821 In larval zebrafish, two forms of colour-opponency are established at the 822 level of cone outputs 7 , but three are observed at the level of the 823 downstream BCs 18 . Accordingly, the expectation is that the 'third' form of 824 opponency ('UV:yellow') is set up by ACs. This was found to be the case in 825 the central retina ( Figure 6B,C) but in the periphery the population 826 representation of colour-opponency was remarkably invariant to 827 pharmacological removal of inner retinal inhibition ( Figure 6B). At the level 828 of individual BCs, however, ACs could either enhance or suppress colour 829 opponency. These opposing effects were balanced across BCs by a 830 combination of factors. First, a majority of ACs was essentially achromatic 831 ( Figure 3A-C), indicating that they do not alter spectral processing in a 832 direct manner. Second, a minority of chromatic ACs appear to implement a 833 switch, by which they mask pre-existing colour opponency in some BCs, 834 while at the same time generating qualitatively equivalent information 835 elsewhere (Figure 6D, 7F,G). This switch was implemented mostly by On-836 circuits (Figure 8), and correspondingly the dendrites of ACs that exhibit 837 spectral opponency were located in the traditional On-layer ( Figure 3H). 838 839 While it seems intuitive to suggest that colour opponent ACs underpin 840 colour processing in BCs, it may also be that non-opponent ACs are 841 involved in the same task. In principle, combining a spectrally broad AC 842 with a spectrally narrow BC would lead to opponency, while combining an 843 intrinsically opponent BC with a spectrally narrow or V-shaped AC might 844 abolish the opponency. Indeed, blocking inhibition from ACs affected BCs' 845 colour opponent signals across the entire IPL, including in the Off-layer 846 where colour opponent ACs are absent.

848
In the future, the different anatomical distributions of colour coding BCs in 849 the presence and absence of AC-inputs ( Figure 6D) may provide an 850 important handle for studying the diverse AC-BC circuits that contribute to 851 this overall spectral balancing. Further, understanding if and how these 852 correlative observations are causally linked will likely require the use of 853 more specific transgenic lines that allow more selectively interfering with 854 specific types of BCs and ACs. The same strategy should also help to 855 decipher those BC circuits where ACs mask a pre-existing opponency.

857 A special role of On-circuits in zebrafish colour vision? 858
Most AC-mediated spectral tuning in BCswhether leading to changes in 859 opponency or simply a rebalancing of non-opponent spectral tunings -860 were predominately implemented via the On-rather than the Off-channel 861 ( Figures 3H, 6D, 8). This observation adds to a growing body of evidence 862 that zebrafish generally leverage On-rather than Off-circuits to compute 863 diverse aspects of colour-information. For example, both at the level of the 864 retinal output 33 , and within the brain 32,47,48 , most spectral diversity is 865 represented in the On-channel. In contrast, the spectral tuning function of 866 the brain's overall Off-response essentially resembles the spectral tuning 867 function of red-cones in isolation, which also corresponds with the mean-868 spectrum of natural light in the zebrafish natural habitat 7,32 . From here, it is 869 tempting to speculate that zebrafish generally use the Off-channel as an 870 'achromatic reference', while On-circuits can, where required, provide 871 spectrally biased points of comparison to serve spectral and colour vision.

872
A predominant use of one rather than both polarities for encoding spectral 873 information could also be advantageous in that it might permit largely 874 unaltered travel of the red-cones' achromatic signal to the brain: by 875 restricting the bulk of spectral computations to the On-strata of the IPL, 876 circuits within the Off-strata can operate in an essentially achromatic 877 manner. In agreement, the vast majority of ACs in the Off layer exhibited 878 such achromatic tunings ( Figure 3A,E). In the future it will be interesting to 879 test if such an On-dominance amongst spectral computations is also a 880 feature in other species.

882
The possible link between polarity and distinct spectral functions also 883 raises the question how On-Off responses are used for spectral coding. 884 Among ACs, we observed a strong link between colour opponency and 885 On-Off responses ( Figure 2B, Supplemental Figure S2). Correspondingly, 886 among BCs, intrinsic On-Off responses were routinely unmasked by 887 blocking ACs, particularly in the central retina where they were present 888 throughout the On-layer. (Figure 5A,B,E,F, Supplemental Figure S5A). In 889 the future it will therefore be important to probe more directly to what extent 890 On-Off processing in ACs and BCs can be causally linked to specific 891 aspects of colour processing. Another unanswered question is the source 892 of BCs' On-Off responses. Opponency in cones alone is unlikely to explain 893 this observation. This is because cone-inversions from their intrinsic Off-894 response to an HC-mediated overall On-response occurs exclusively at 895 long-wavelengths 7 ; however, many unmasked On-Off BCs were short-896 wavelength biased (Supplemental Figure S5A) implicated in motion processing [53][54][55] . Similarly, another key neuron 937 implicated in mammalian motion processing is the starburst amacrine cell 938 (SAC) which co-releases acetylcholine alongside GABA 56,57 . However, the 939 functional role of SACs outside mammals remains sparsely explored 58 .

941
The role of green-and blue-cone circuits in supporting inner retinal 942 colour processing 943 Unlike red-and UV-cones, zebrafish green-and blue-cones provide 944 strongly colour opponent outputs due to feedforward signals from the 945 HCs 7 . Accordingly, these cones might directly support colour opponency in 946 BCs. In support of this hypothesis, the spectral zero-crossings marked by 947 these two cones remain represented within BCs, both in the presence and 948 in the absence of ACs. However, only a minority of green-and blue-cone-949 like BCs retained their specific opponency upon AC-block ( Figure 7A). This 950 suggests that while the signals from green-and blue-cones can be directly 951 used to support colour opponency in BCs, this motif is by no means 952 dominant when considering the complete circuit. Instead, most BC circuits 953 that represent these two spectral opponencies required inputs from ACs. In 954 zebrafish, green-but not blue-cones provide cone-type-exclusive drive for 955 at least two anatomically distinct types of BCs 12 , providing a possible 956 neural substrate for the minority of green-cone-like BCs that were 957 unaffected by AC-block. These might account for some of the unmasked 958 green-cone-like BCs when ACs were blocked. Possible green-cone-959 exclusive BCs might also link with the observation that green-light 960 stimulation could result in long-wavelength biased spectral effects on BCs 961 (Supplemental Figure S9D,E), and that most opponent ACs seemed to be 962 partially built from green-cone inputs (Supplemental Figure S3D,E).

964
In contrast, the possible roles of blue-cones in zebrafish colour vision 965 remain much more elusive. A blue-cone-exclusive BC is not known to 966 exist 12 , which leaves the origin of any intrinsic blue-cone-like BC-tunings 967 unclear. Further, we found no evidence of any major involvements of blue-968 cones in AC-processing (Supplemental Figure S3D,E). On the other hand, 969 the four LEDs used in the present study were not optimally placed to 970 disambiguate blue-from UV-cone contributions (Supplemental Figure  971 S3B,C Stimulator intensity was calibrated to be spectrally flat at 30 mW per LED, deviation projection across the tetrachromatic noise data. The ROIs from 1096 control condition and drug condition were drawn separately. Terminals 1097 were paired across the two conditions using the experimenter's best 1098 judgment, which we found to be more reliable than automated procedures. 1099 The matching of terminals across conditions was greatly facilitated by the 1100 sparse expression strategy, and throughout we tried to be as conservative 1101 as possible to only match terminals when we were certain that they are the 1102 same ones (i.e. minimising false positives, at the expense of false 1103 negatives). For cone recordings, the ROIs were drawn manually from the 1104 standard deviation projection across time.

1105
In all scans within the IPL layer, IPL boundaries were drawn by hand using 1106 the custom tracing tools provided in SARFIA 67 . The IPL positions were then 1107 determined based on the relative distance of a ROIs' centre of mass 1108 between the IPL boundaries and mapped to the range 0% to 100%. 1109 Fluorescence traces for each ROI were z-normalised, using the time 1110 interval 2-6 seconds at the beginning of recordings as baseline. A stimulus 1111 time marker embedded in the recording data served to align the Ca 2+ 1112 traces relative to the visual stimulus with a temporal precision of 1 ms. 1113 Responses to the chirp and step stimuli were up sampled to 1 kHz and 1114 averaged over 5 trials. For data from tetrachromatic noise stimulation we 1115 mapped linear receptive fields of each ROI by computing the Ca 2+ 1116 transient-triggered-average. To this end, we resampled the time-derivative 1117 of each trace to match the stimulus-alignment rate of 500 Hz and used 1118 thresholding above 0.7 standard deviations relative to the baseline noise to 1119 the times ti at which Calcium transients occurred. We then computed the 1120 Ca 2+ transient-triggered average stimulus, weighting each sample by the 1121 steepness of the transient: 1122

1123
Here, is the stimulus ("LED" and "time"), is the time lag (ranging 1124 from approx. -1,000 to 350 ms) and M is the number of Ca 2+ events. The 1125 resulting kernels are shown in z-scores for each LED, normalised to the 1126 first 50 ms of the time-lag. To select ROIs with a non-random temporal 1127 kernel, we used all ROIs that exceeded a standard deviation of ten in at 1128 least one of the four spectral kernels. underestimates their full diversity, which is likely part related to the 1238 necessarily incomplete sampling of the full stimulus space and/or possible 1239 incomplete labelling of the ptf1a driver. 1240 1241 Fitting of AC-cluster spectral tuning functions with cones. Spectral 1242 tuning functions of AC clusters means were matched to those of previously 1243 recorded cones (cf. Supplemental Figure 3B,C) based on the four relative 1244 kernel amplitudes (as shown in Figure 3A-D). Fitting was done as follows: 1245 For each tested cone-combination (e.g. all cones, or R+U only, etc.) we 1246 computed 10 6 possible combined tunings at random by summing the 1247 respective reduced cone tuning functions (Supplemental Figure 3C) with 1248 random scaling between -1 and 1 each. We then computed the linear 1249 correlation coefficient between each AC-cluster's tuning function, and each 1250 of the randomly generated combined cone-tunings, in each case choosing