Rod-cone signal interference impairs mesopic motion discriminability in a model circuit

26 Under mesopic conditions, such as dawn or dusk, signals from both rod and cone 27 photoreceptors contribute to percepti on. These parall el input s are combine d within the retina 28 before b eing se nt to subs equ ent visua l areas. The integration of thes e kineti cally-distinct parall el 29 signals pos es uniqu e challeng es for human vision. Though previous behavi oral studie s have 30 found that dim lighting conditions specifically impair motion perception in human subjects, the 31 origin of this dependence is uncl ear. I n the present study, w e create a mode l circuit that predicts 32 ganglion cel l respon se s to moving stimuli by incorporating electrophysiolo gically-deriv ed circuit 33 components into a Hassenst ein-Reichardt correlator, a classical motion-det ection model. The 34 model circuit demonstrate s that interactions betwe en rod- and cone-derive d signals ne gative ly 35 impact the encoding of a moving object’s direction under mesopic conditions. Furthermore , we 36 found that the model circuit could enhance its motion discriminability if it was only sen sitiv e to 37 the cone-activating components of th e stimuli. We conclud e that rod-cone signal interf erenc e 38 occurring at the lowest le ve l of vision has an impa ct on motion direction discrimination, a 39 higher-leve l task with rele vance for behavior. 40 41

42 In the transition betwe en dim and bright lighting, known as mesopic lighting conditions, 43 rods and cones are simultan eous ly activated. Me sopic conditions pose a unique challeng e to 44 human vision. Because me sopic vision relie s on paralle l str eams of visual i nput derive d from 45 both rod and cone photoreceptors, the retinal circuitry is in a state of transition, balancing the 46 gain of rod-mediated signals that are approaching saturation wit h the gain of cone-mediated 47 signals that are barely emerg ent (re viewed by Buck, 20 04 , Buck, 201 4, Gri mes et al ., 20 18 , 48 Stockman and Sharpe 2006). Post-photoreceptor circuitry processe s the rod and cone inputs in 49 paralle l before they are combine d to shape retinal outputs (Gouras and Lin k, 1966, Enroth-50 Cuge ll et al ., 1977). Cons eque ntly , it is not surprising that deficits in mesopic vision are a first 51 symptom in many visual diseas es (Pet zold and Plant 20 06, Arden and Hogg 1985). 52 Mesopic vision provides a rare opportunity to examine how parallel proces s ing (Kolb 53 and Nelson , 2003; Grim es et al , 201 8 ), a common computational schema i n neural circuits, 54 influence s human perception. In prev ious work, el ectrophysiological record ings have be en use d 55 to deve lop a model that predicts retin al ganglion cel l respon se s to rod and c one inputs (Songco-56 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 3 Aguas et al. , 2023, G rimes e t al. , 201 8 , Grimes e t al. , 2015) . The rod-cone re tinal model captures 57 two features of how parallel rod and cone pathways respond to contrived ti me-varying stimuli 58 such as Gaussian noise and sinusoida l waves: 1) ther e are distinct kinetics betwe en rod- and 59 cone-deriv ed signal s and 2) these paralle l signal s are combined prior to a shared rectifying or 60 thresholding nonlin earity. Thes e feat ures are consis tent with el ectrophysiological and human 61 psychophysical data showing an unexpected de structiv e interfer ence of rod - and cone-mediated 62 signals (Grime s et al ., 2 015 , MacLeod, 1972, review ed by Stockman and Sharpe, 20 06). 63 Models bas ed on physiology are a use ful tool for uncovering new mechanist ic 64 connections betw een paral le l signal p rocessing and visua l perception . Here we focus specifically 65 on motion. Though several human behavioral studies have found that lighting conditions affect 66 the accurate perception of motion (Bil ino et al., 2 00 8, Yoshimoto et al., 2 01 3, Yoshimoto et al., 67 2016, Mayeur et al ., 2 00 8, Gros sman and Blake, 1999, Sepulv eda et al ., 20 2 1), the origin of this 68 depend ence is uncl ear. In all of thes e cases, the impairments primarily affe ct the perception of 69 translational motion in the periphery. Compared to photopic or scotopic co nditions, human 70 observ ers had particular difficulty in accurately perceiving motion directio n for stimuli that 71 mimicked a walking human figure un der mesopic conditions (Bilino et al. 2 008) . Mesopic 72 conditions also unique ly e liminated motion priming in w hich the perceive d direction of motion 73 of an ambiguous stimulus dep ends on a prior pri ming motion stimulus (Yoshimoto et al., 2013, 74 Yoshimoto et al., 2016). Her e we te st the hypothesis that differences in the kinetics of rod- and 75 cone-deriv ed respon se s in mesopic conditions produce ambiguities in motion discrimination 76 that may contribute to these and re lat ed perceptua l phenomena (Songco-Aguas et al. , 2023). 77 There are se veral cla ssical mode l circuits for motion detection (Barlow and Levick, 1965, 78 Adel son and Movshon, 1982, re view e d in Borst and Egelhaaf, 1989). Funda mentally , thes e 79 models compute motion direction bas ed on the correlated spatial and t emporal changes in 80 brightness that occur when a moving object passe s in front of an array of p hotoreceptors. The 81 Hassens tein-Reichardt correlator consists of two neighboring simulat ed photoreceptors , or 82 more genera lly , two neighboring inpu t channels. The circuit correlates the i ntensity of 83 brightness b etw een the two input channel s after the output of one is passed through a temporal 84 low-pass filter that de lays its signa ls r elativ e to the other (Hassenst ein and Reichardt, 1956, 85 Reichardt, 1961, Borst and Euler, 2 01 1, Borst and Euler , 20 11). It is a relati vely simple mode l 86 that captures the spatio-temporal correlations und erlyin g motion; although it was originally 87 derive d from insect optomotor behavior, it is genera l and abstract enough to model visual 88 motion detection in other species (Bo rst and Helmstae dter , 201 5, Borst an d Egelhaaf, 1989, 89 Frechette et al , 200 5). 90 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 4 In the context of our study, the Hasse nstein-Reichardt model is simp ly a readout of 91 perceiv ed motion direction of a movi ng stimulus giv en rod-cone interactio ns. We examine how 92 paralle l rod and cone signal processin g influence s the retinal circuit’s abilit y to accurately 93 ascertain the direction of a moving stimulus. Sp ecifically, by incorporating t he rod-cone model 94 circuit into a Hassenst ein-Reichardt correlator, w e examine the impact of k inetic difference s 95 betwe en rod- and cone-deriv ed signal s on motion processing in mesopic conditions. Rod-96 derive d signals ar e slow er and have a stronger att enuation of low temporal frequenci es than 97 cone-deriv ed signal s (se e Fig. 1). We first demonstrat e the model’s abi lity to discriminate motion 98 directions across a range of motion speed s for different circuit architectures : circuits that are 99 exclusiv ely r espon sive to eithe r rod or cone input, and circuits that pro cess both rod and cone 100 mediated motion simultaneou sly . We hypothesize that the perceptual defici ts obser ved in 101 mesopic conditions are a consequenc e of the signal interfer ence b etwe en r od and cone mediated 102 signals at specific time scale s. In comparing the performance differences across thes e circuits, 103 we gain insight into how the p hysiologically-deriv ed mode l retinal circuit degrade s motion 104 discrimination ability and potentially contributes to the decr eased p erceptu al acuity observ ed in 105 human subjects. By te sting motion stimuli across a range of motion speeds (i.e. pu ls e delay s), we 106 examine how the known differences in signaling kinetics bet we en rods and cones influenc e 107 motion processing. Finally , we exami ne how the salience of motion stimuli impacts motion 108 direction discriminability. Altog ether, these findings expand our und erstan ding of how p arallel 109 rod-cone processe s unde rlie the p erce ption of stimuli that can affect hum an behavior. 110 111 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 5 Figure 1 Rod-cone LN circuit model A: Rod-con e linea r-nonlinea r cascade mo del. A two- pron ged model of a n ON paras ol retinal ga nglion cell with rod and cone input. The linear fi lters capture the unique kinetics of rod- a nd L -cone-deri ved responses in meso pic conditions (blue : ro d; red: cone). After linear filtering, rod an d cone si gnals are summed, t hen passed to a sha red nonline arity and th en a rectifyin g nonlinearit y. T he final output of t he rod-cone retinal mod el repres ents t he s piking rate of a retinal gan glion cell. B: Rod and cone filter kinetics. T he rod fil ter (blue) is slow er a nd has an overal l mo re bip hasic sha pe than t he cone filter (re d). T he inset s hows that the dela y in time-to-pea k betwee n r od and cone filters is about 35 ms. 112 5 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 6

113 Retinal model 114 The rod-cone retinal model was de riv ed from electrophysiological r ecordings conducted 115 on whole mount preparations of isolated non-human primate retina, as previously d escribe d 116 (Dunn et al., 2 007; Trong and Rieke, 200 8). Thes e recordings we re taken f rom ON parasol 117 ganglion cel ls. The circuit components of the rod-cone linear-nonlinear mo del (Figure 1) wer e 118 derive d from voltage-clamp recording s of full-field white nois e (0-4 0 Hz bandwidth) using blue 119 LEDs (peak power at 460 nm) and re d LEDs (peak power at 640 nm) to iso late rod and L-cone 120 respons es , resp ectiv ely , as previous ly describe d (Songco-Aguas et al. , 2023 , Grimes et al ., 20 15, 121 Chichilnisky 2001). Circuit componen ts of the model were verified by takin g the model’s 122 explained variance from data with the same noise stimuli pr es ented . 123 The linear-nonlinear mode l re spons e s to motion stimuli, s(t) , w ere simu lat ed as follows . 124 First, rod and cone stimuli, s rod (t) and s cone (t), are convolv ed with their resp ective lin ear filter s, 125 h rod and h cone . 126 h i /g1499 s i (t) 127 Filtered sig nals are then s ummed. The summed signa l is passe d into a shar ed nonlin earity that 128 was fit from the ganglion cell respon s es with a third degree polynomial fun ction (see Fig. 1A, 129 middle panel); the output of this nonl inearity is an estimate of the cell’ s exc itatory synaptic 130 input. A second nonlin earity, a rectified linear function, r , convert s synapti c currents into retinal 131 ganglion cel l spike rates (s ee Fig. 1A , bottom panel). 132 r(x) = {0, if x<0 133 x, if x≥0/i8 134 Instantaneous curr ents b elow 0 amps do not result in spiking, but abov e that threshold, the 135 spike rate linear ly increas es with increasing current . 136 To define the shared nonlinearity b et ween rod and cone components of the linear-137 nonlinear mode l for a single cel l (Schwartz et al, 20 12; Dunn et al, 2 006), t he horizontal axis of 138 the measured cone non linearity is sca led re lativ e to the measured rod nonli nearity until they 139 overlap, and then the cone filt er is sca led accordingly, as pre vious ly de scribed (Songco-Aguas et 140 al., 2 023). 141 The rod-cone retinal circuit incorpora ted into the Hassen stein-Reichardt model (s ee 142 following s ection) was fitted with data from five ganglion cells . The linear a nd nonlinear 143 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 7 el ement s were fit to a respecti ve poly nomial function for each recorded ret inal ganglion cel l, 144 then the parameter fits for these poly nomial functions were averag ed acros s cell s. 145 Hassenstein-Reichardt correlator 146 The architecture of our Hassenstein- Reichardt model is illus trated in Figure 2 (se e 147 Supplem entary Fig. S1 for illu stration of classical Hassenst ein-Reichardt model). The input 148 channels interact with each other wit h a correlator time delay (Δt) . For example, the r eal-149 time respons e of the right input chan nel ( r right ) is multiplied with the time-delay ed respon se of 150 the left input channel ( ). By mirroring this process (multiplying the rea l-time respon se of the 151 left input channel, r left , with the time-delay ed respon se of the right input ch annel, ), then 152 subtracting the products from one another, we remov e the parts of the correlator’s re s p ons e tha t153 resu lt from direction-irrel evant correl ations in the stimulus (Borst and Ege l haaf 1989). The 154 final output of the Hassenstein-R eichardt model is calculated as: 155 () ( r right ) - ( )(r left ) 156 157 Figure 2 Motion correlator wit h rod- cone retinal mo del inpu t ch annels A: Hassenstein-Reich ardt cor relator with the r od-cone retinal circuit mod el as th e i nput cha nnels computes motion fr om r od-cone activating stimuli. Stimuli pass t hrou gh one input channel and then the ot her after some time dela y (Δs) corre sponding to motion vel ocity. Real -time a nd time-dela yed (Δt) 7 t .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 8 signals f rom o pposite input chan nels are multiplied. F or example, P 1 is t he product of a time-dela yed signal f rom t he left input chan nel a nd t he real-time signal from th e ri g ht input cha nnel. T he differ ence between th e products (P 1 and P 2 ) is com p uted. T he output is an expression of the motion of direction. B: Motion stimulus trace for a sin gle trial with a 10 ms pulse width and a 2 5 ms pul se dela y, in w hich the left input channe l is presente d with a 10 ms wide stimulus pulse, and t hen 2 5 ms later, the ri ght input chann el is stimulated with th e 10 m s pulse. Noise is adde d to these puls es bef ore t hey are processed b y t he rod-con e circuit. C: T he avera ge res ponse traces for leftwar d (top) and rig htward (bottom) motion ac ross 100 t rials. The preferred direction of motion elicits a pos itive peaked output. G ray cloud shows sta ndard de viation of mean at a given time p oint. 158 The pulse delay , Δs , is a free parameter that controls the time de lay betw een the arrival 159 of the stimulus at one input channel and its arrival at the other input chann el. In other words , 160 the pulse d elay corre sponds to the motion speed of the stimulu s. The pul se width , w , is 161 another free parameter and it control s the duration of the stimulus (i.e . the width of t he 162 stimulus pu ls e, or the overa ll angular size of the stimulu s). Theoretical ly, a Hassens tein-163 Reichardt circuit would be most sensi tive to the puls e de lays that match t he correlator time 164 delay. A s the puls e de lay ( Δs ) diverge s from the correlator time delay ( Δt ), wider and wider 165 pulse s widths ( w ) are ne eded for the circuit to register the stimul us as a moving stimulu s. This 166 takes the form of a piecewise linear re lationship. 167 Δt ≤ Δs : 168 w min > -Δt + Δs 169 Δt > Δs : 170 w min > Δt - Δs 171 We illustrat e this function in Supple mentary Figure S1. The piecewi se lin e ar relationship 172 describe s the Hasse nst ein-Reichardt correlator without the filter kinetics o r nonlinearitie s, but 173 once those are include d, this relation s hip is no longer guaranteed to be line ar due to the 174 nonlinearitie s inherent in the rod-con e retinal mode l circuit. 175 176 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 9 Supplem ental Figure S1 (S upplem ent to Fig. 2) Stimulus correlations in H assen stein-Reichard t circuit witho ut r od-cone kinetics We examined whet her th ere was o verlap (ie., correlation) for noiseless motion sti muli deliver ed between th e two input c hannels of a simp lified implementation of th e corr elator cir cuit, based on the time delay of the circuit and t he pulse wid ths and puls e de lays (re presentin g m otion vel ocity) of th e stimuli. Th e da rk squa res o n t he grid sign ify width and dela y combinations in w hich t here was no overlap, w hile t he w hite squa res indicate where ove rlap between the two in put c ha nnels was possible, and hence co rrelation could be computed by the circuit. F or example, with pulse wi dths of 0 ms, overlap between the input chan nels was i mpossible. 177 In our simulated experim ents , we chose a fixed Hassen stein-R eichardt corr elator time 178 delay of 50 ms (s ee Supp leme ntary Fig. S1), with the exception of the simul ated experimen ts 179 shown in Supplementary Figure S2 . For that experiment, we te ste d circuits with time delays of 180 25 ms and 75 ms to examine how t he time delay affected motion discriminability using our ful l 181 model circuits. We chose a correlator time delay of 50 ms b ecause this is comparable to the 182 temporal separation betw e en rod and cone filters (s ee Fig. 1) and hence corresponds to time 183 offsets in which rod and cone signals may interact. 184 Stimulus generation and simulated experiments 185 For each individual trial, a pulse of light was present ed to one of the linear- nonlinear 186 model input channels . Then, to simul ate spatial distance be twe en two cel ls, the other input 187 channel receiv ed the same pul se of lig ht after a pre-specified pulse d elay b e tween 0 ms and 100 188 9 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 10 ms, corresponding with stimulu s ve lo city (larger puls e de lays are slow er v el ocities while smal ler 189 pulse d elay s are faster v elocitie s). To mimic noise in the real respons es , Gaussian pink noise 190 with a mean of 0 and a standard deviation of 10% was scaled by the mean intensity of the 191 stimulus and added to the pul se s timulus . Noise was uncorre lated b etwe en rod-activating and 192 cone-activating components of the m otion stimulus and also uncorre lated betwe en stimu li 193 prese nted to the l eft and right input c hannels. The tru e direction of motion was based on which 194 input channel receiv ed the puls e first. 195 Classification and perfor mance 196 The model output was classified as a l eftward or rightward trial with a discr iminant 197 vector (Zhao et al, 2024) . The discriminant vector for a set of trials with a s pecific pulse de lay 198 was calculated by taking the mean res ponse traces of the model for 10 0 l eft ward trials and 100 199 rightward trials, then subtracting the mean leftward re spons e trace (non-preferr ed direction) 200 from the mean rightward response trace (preferred dir ection). This linear discriminant analysis 201 is effective un der the conditions of our simulations in which t he noise is independ ent of the 202 stimulus and uncorre lated across r es ponses . This process mirrors the Hassens tein-Reichardt 203 correlator calculation resu lting in the preferre d direction of motion having a positive polarity. 204 /g3404/g1731 /g1732 /g3398 /g1731 /g1732 To automatically label motion direction for a single trial, the re sponse trac e was 205 projected onto the discriminant vector by taking the dot product. A projecti on value of zero can 206 be interpr eted as a re sponse trace tha t was perfectly dis similar from both t he mean leftward 207 trace and the mean rightward trace. Projections with a value greater than ze ro were 208 automatically label ed rightward, whil e projections with a value le ss than ze ro were automatically 209 labe led l eftward. We ran 10 00 n ew tri als, half with leftward motion and hal f rightward motion. 210 The model’s ability for motion direction discrimination was measured as the percentag e of trials 211 where the automatic labeling capture d the true direction of the motion stimulus. S eparate 212 probability density function s were fitt ed to these the projection valu es of tr ue leftward and tru e 213 rightward trials to visualize the ov erla p in trials label ed l eftward or rightward. Confidence 214 interval s for the full mode l output res ponses w ere computed usin g the Clop per-Pearson int erval . 215 Implementatio n 216 All simulation s wer e run in MATLAB version 2 024b u sing a MacBook Pro with an Apple 217 M2 chip and 24 GB of memory. Code is publicly availab le on the Guti errez l ab’s motion-218 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 11 detection repo sitory, which can be fo und on https://github.com/Gutierrez -lab/motion-219 detection . 220 221

222 A motion discrimination circuit with rod- cone kinetics 223 To simulate retinal gang lion cel l spike rate traces, we fit a linear-non linear cascade 224 model using pre vious ly de scribed met hods (Songco-Aguas et al., 2 023, Gri mes et al ., 20 15, 225 Chichilnisky 2001). Voltag e-clamp recordings of ON parasol ganglion cell s from non-human 226 primate retina in mesopic conditions were us ed to fit the linear and nonlin e ar components of 227 our cascade model. This linear-nonlin ear model circuit captures the unique kinetics of rod-228 derive d and cone-deriv ed signal s in mesopic lighting conditions. Specifically , the rod filter has a 229 longer time-to-peak than the cone filter–with rod-deriv ed signal s delay ed b y approximately 33 230 ms relativ e to cone-deriv ed signal s (Fig. 1B). The rod filter is also more bip hasic than the cone 231 filter, meaning that the rod respon se has more of a negative respons e component after the initial 232 positive re spons e component (Songco-Aguas et al. , 2023, Grim es et al ., 2 01 8). As a 233 consequ ence, w e would exp ect that a rod-only linear-non linear mode l compared to a cone-only 234 linear-nonlin ear model wou ld be s low er in responding to the same stimu lus . 235 In addition to these kinetic filters, the linear-nonlin ear model circuit also contains a 236 nonlinearity that shapes ganglion cel l inputs. Becaus e the shape of this nonlinearity was found 237 to be consisten t betw een the rod-targ eted and cone-targe ted e lectrophysio l ogical recordings, we 238 interpret ed this phenomenon to mean that rod and cone signals are integr ated prior to this 239 nonlinear stag e, such as in the cone bipolar output synapse (Grime s et al ., 2 015 , Songco-Aguas 240 et al. , 201 8, Fain and Sampath 2018, Gouras and Link 196 6). Indeed, this model architecture 241 captures temporal interactions b etw e en flashed stimu li that sel ective ly activate rods or cones 242 (Grimes et al ., 2 015) . Models lacking a shared nonlinear component could not capture these 243 interactions. The output of the nonlin earity was converte d into retinal gang lion cell spike rates 244 through a rectified linear function. Altogether, w e refer to the se component s of our linear-245 nonlinear mode l coll ective ly as the rod-cone retinal mode l circuit (Fig. 1A; see Me thods). 246 We incorporated our rod-cone retinal models as the two input channels of a 247 Hassens tein-Reichardt correlator (Figure 2A). The Hassen st ein-Reichardt correlator det ects 248 motion using a “delay-and-compare” computation (Hassenstein and Reichardt, 1956, Reichardt, 249 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 12 1961, Borst and Euler, 2 011 , Borst and Euler, 2 011) . It reli es on two input channels with identical 250 respons e propertie s–in our case two s imulated retina l ganglion cel ls– with a fixed spatial 251 separation betw een them . This spatial separation is implement ed as a temporal delay, Δt . The 252 direction of motion is given by the order in which eac h input c hannel recei ves the stimu lus , and 253 the velocity of motion is based on the delay in time betw e en when the first channel and the 254 second channel recei ve the stimu lus , which we refer to as the pulse de lay, Δs (Fig. 2B). The 255 output from the left channel is time-d elaye d then multipli ed with the real-time output of the 256 right channel, and vice versa . The res ulting polarity in the respon ses provi des the circuit’s 257 directional tuning. In this case, our ci rcuit’s preferred dir ection of motion i s rightward because a 258 rightward stimulus produces a re spon se trace with a positive peak whereas a leftward stimulu s 259 produces a negative p eak (Fig. 2C). A leftward-pref erring circuit would be i dentical, exc ept at 260 the subtraction stage , which would ha ve P 2 -P 1 instead (s ee Figure 2A) . 261 Our motion input consists of a noisy “moving” pulse —that is, a puls e of a gi ven 262 magnitude that activates one of the input channels, then after a period of time, Δs , a second 263 pulse that activates the other input channel. For thes e initial trials, w e chose our puls e 264 magnitude to have a 25% contrast over a fixed mean luminance. These pul s es contained both 265 rod-activating and cone-activating sti mulus components , and the stimulu s was scaled such that 266 both rod and cone wings of the rod-c o ne retinal mode l wer e equa l contribut ors to the spike rate 267 output. Indep endent Gau ssian pink noise was also added to each puls e, uni que to each trial. 268 Noise was uncorrelat ed bet we en stim uli pres ented to the l eft and right input channels, as we ll as 269 uncorrelate d betw een rod-activating and cone-activating components of th e motion stimulus . 270 Figure 2C shows that the mean output traces of the Hassenst ein-Reichardt correlator are 271 opposite in polarity betwe en stimu li moving in its preferred dir ection (rightward) and its non-272 preferre d direction (leftward) when o ther stimulus paramet ers (pul se widt h and pulse delay) are 273 kept fixed. 274 We created a systemiz ed method such that the model circuit responses coul d be decoded 275 and automatically label ed by the “per ceived” dir ection of motion for eac h tr ial. To do this, we 276 calculated mean re sponse trace s across l eftward and rightward trials of a gi ven pul se d elay (Fig. 277 3A; see Methods). We then calculat ed a discriminant vector by subtracting the mean leftward 278 respons e trace from the mean rightwa rd respons e trace. This process mirro rs the subtraction 279 stage in the symmetrical Hass enst ein -Reichardt model (Fig. 2A) that yields a preferred direction 280 of motion. We then took the output tr aces across 10 00 new trial s–half l eftw ard and half 281 rightward–and projected them along the discriminant (Fig. 3A). Automatic direction 282 discrimination was based on whether the projection value of a trial was greater than zero, in 283 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 13 which case it was label ed as rightward motion, or whether it was less than zero, in which case it 284 was label ed as l eftward motion (Fig. 3 B). We computed probability densiti e s of the projection 285 value s for the leftward and rightward trials to visualiz e the over lap betw een these two 286 distributions (Fig. 3C). This over lap determin es what fraction of trials would be erroneo usl y 287 identified (i. e. rightward trials with projection value s le ss than 0 and leftwa rd trials with values 288 greater than 0). 289 290 Figure 3 Automa tic l abeling of motion direction A: Discriminant vector for a rod -cone sti mulus with 25 ms pulse dela y a nd 10ms p ulse width ( Δs = 2 5 ms, w = 10 ms). The mean output trace of the Hassenstein-Reicha rdt circuit is com puted for both directions of motion and t he discriminant vector for the rig htward -prefe rring circuit is the mean rightwar d motion output minus th e mea n leftward motion output (see M ethods). B: Motion discrimination perfo rmance across a ran ge of puls e de lays. Accuracy fo r discriminating motion direction can var y betwe en 0 and 1, wh ere 0 is no accurac y, 1 is comp lete ac curacy, and 0. 5 is chance p erformance (indicated b y so lid li ne). Erro r bars re present the confidence interval with a minimum covera ge of 95 % of t he data, t h e Clo pper- Pearson interval. C: Probability distribution of pr ojections for trials wit h a 25 ms pulse dela y stimulus. Two distinguishable probability distributions arise from fitting t he projection v alues f ro m leftward (pur ple) and ri ghtward (teal ) trials. T he ove rlap be tween th e two distributions rep resents t rials in w hich leftward and ri ghtward motion are potentiall y mis classified. 291 With our autom ated labeling m ethod, we calculated the circuit’s motion discrimination 292 accuracy fro m the set of 1000 trial s: t he fraction of tri als in which the mod el’s automatic labe l 293 13 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 14 corresponded to the tru e direction of the presen ted stimu li (Fig. 3B). As to be expect ed, the 294 model performs at chance in the absence of a pulse de lay ( Δs = 0 ms) because the stimu lus 295 inputs are pres ent ed simultan eous ly t o the left and right input ch annels; he nce, for these trial s, 296 there was no actual motion in the stimulus. For puls e de lays of 20 ms, the model performed with 297 ~85% accuracy (i.e., it incorrectly lab el ed the direction of motion in ~15% of the trials). 298 Performance plateau ed for delays of 60-100ms , while smal ler d elay s resu lt ed in the circuit 299 mislabe ling a larger fraction of trials. The model’s inaccuracy with labeling stimuli with a 300 small er puls e de lay (i.e. , faster movin g stimuli) se emed to correspond to th e difference in the 301 time-to-peak of the rod and c one line ar filters. Our previou s work has demonstrated how 302 respons es to time-varying s timuli at s pecific temporal frequencie s can be attenuated du e to 303 interfer ence b etwe en rod and cone sig nals (Songco-Aguas et al ., 2 023). In the next section, w e 304 further investigat e the effect of rod-cone interfer ence on motion direction d iscrimination. 305 Disenta ngling the perceptual contributions of rod s and cones 306 Does interfer ence b etw een parall el ro d- and cone-derived signa ls hinder m otion 307 discrimination? We answered this qu estion by comparing the performance of the rod-cone 308 circuit with t he performance of a circ uit in whic h the input c hannels wer e e xclusive ly s ensitiv e to 309 the rod-activating component of the p res ented s timuli (rod-only) and a circ uit in whic h the 310 input channels were exclusi ve ly se nsit ive to the cone-activating component (cone-only). In 311 effect, thes e trials w ere simu lations of the peripheral retina with either a rod or cone knockout. 312 We repeated the sam e range of puls e delays (ie ., 0-10 0ms) as in our previo us test . The resu lts 313 are plotted in Figure 4 . 314 315 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 15 Figure 4 Motion dis criminability between simula ted knockout mo dels We com pa re t he disc rimina bility between t h ree diff eren t re t inal mod el ci rcuits as in pu t channel s to t he Ha s sens t ein -Reichar dt m ode l, each wi t h a fixe d cor rela t o r t ime d elay of 5 0m s ( ri ght ) . The r od - only ci rcuit (bl ue) is exclusiv ely res po nsive t o t he r o d-ac t iva t in g co mp onen ts of mo t io n st i muli, t he c one -onl y circuit ( re d) is exclu sively re s p onsive t o t he cone -ac t ivat in g com ponen ts, and t he ro d -cone circui t ( pur ple) is res po nsive t o b ot h . Iden t ical mo t ion s t imu li a re pas sed in t o t he se uniq ue ci rcuits, and di scri minabili t y i s c om pu t ed f or each circuit an d con dit i on. In se t : fo r mo t ion s t imu li between 2 0 -3 0 ms , t he c one -onl y circuit ou t p erf or ms bo t h t he c o mbine d an d ro d -on ly ci rcuits in mo t ion disc riminabili t y . Fo r s t imuli < 1 0 ms and ≥ 80 m s, pe rf or mance s b et wee n t he circuits a re c om para ble . Er ro r b ars re p res ent t he co nfidence in terval wit h a minimum cove rag e of 9 5% o f t he dat a (Clo ppe r -Pear so n in t e rval). S olid b lack lin e indicat e s chance (a m ot i on disc r iminat ion accuracy of 0.5, ie., 50 %). Asterisks 15 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 16 denote where statistical significance was reached in the comparison between the cone-only circuit performance and the combined rod-cone circuit performance (p ≤ 0.05). 316 For pulse de lays of 20 and 30 ms, we found a separation in motion discrimi nation 317 performance across the three circuits in which the cone-only circuit outperformed both the rod-318 cone model circuit and the rod-only ci rcuit (Figure 4 inset). It was unsu rprising that the cone-319 only circuit outperformed the rod-onl y circuit, since the cone-deriv ed respo nses ar e both faster 320 and more monophasic in their kinetic s than the rod-derived respon se s. U lti mately, a faster 321 respons e time would b e bet ter at enco ding faster moving stimuli . Additionally , due to the more 322 monophasic shape of the cone filter relativ e to the rod filter, more of the te mporal structure of a 323 given stimu lus woul d be pre ser ved in the output of the circuit. 324 We had hypot hesized that the rod- an d cone-derive d signals wou ld interf er e with motion 325 discriminability not only becaus e of their relative d elay in kinetics but b ecause of how they are 326 integrated in the retina . The observati on that the cone-only circuit outperfo rmed both rod-cone 327 and rod-only model circuits is in line with this initial hy pothesis; that rod-c one signal 328 interactions would impact motion co mputation. However , the rod-cone circuit had better 329 performance than the rod-only circuit which makes it diffi cult to determine the extent to which 330 impaired motion discriminability ma y not only be due to the characteristics of rod response 331 kinetics, but to the offset in time-to-p eak respons e kinetics for rods and cones . We test ed a 332 circuit with two cone inputs per chan nel wher e one cone filt er is shifted re l ative to the other 333 cone filter by the same amount as the rod filter would have b een (~33 ms). This circuit 334 performed as poorly as the rod-only circuit despite not having any of the broad and biphasic 335 features of the rod respon se – on ly the offset in time-to-peak. Thus, the offset b etwe en the rod 336 and cone respons e kinetics, and not merely the broad and biphasic kinetics of the rod response , 337 interfer es with motion discrimination . 338 Our choices for the model parameter value s can tune the prefer ence s of the model. A 339 prefere nce for direction of motion is built into the Hassens tein-Reichardt correlator circuit 340 during the subtraction stage (Fig. 2A) . At the same time, the choice of correlator time delay can 341 tune the circuit to prefer a range of pulse d elay s, which correspond to motio n velocity, to a 342 certain extent (Supp leme ntary Fig. S2, additional detail in Methods section “Hassens tein-343 Reichardt correlator”). For pulse de la ys that exactly match t he correlator time delay , the circuit 344 can accurately discriminate motion di rection for the largest range of pul se widths. In other 345 words, the mode l has optimal motion direction discrimination for moving s timuli over a wider 346 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 17 range of stimulus siz es when the d ela y betwe en two puls es , corresponding to motion velocity, 347 matches the Hassenst ein-Reichardt correlator’s tuning . For stimuli within the ran g e of p referre d348 pulse d elay s, the cone-only circuit per formance benefits the most . This is again because of the 349 faster-to-peak, more monophasic sha pe of the cone filter, compared to the rod filter. The cone 350 filter more clos ely r es embl es an impulse r espons e than the rod filter, thus e ncodin g the stimulu s351 more faithfully. Conv ers ely , the rod-only circuit consistentl y performs with the least accuracy 352 because the biphasic shape of the rod respons es across the l eft and right ch annels r educe s the 353 signal, making it more ambiguous. Taken together, the se r esu lts d emonstra te that rod-cone 354 signal interf erenc e interacts with both the directionally-tuned and the ve lo city-tuned 355 components of the motion correlator circuit. 356 357 Supplem ental Figure S2 (Supp lem ent to Figure 4) The correlator time del ay an d ve locity t uning We compare ho w motion direction discri minability is affected b y t he Hass enstein- Reichardt corr elator time delay ( /i1t ) in rod, cone, a nd rod-co ne model circuits. We com pare two time delays, 25 ms and 75 ms, in contrast to t he default time dela y we used in all ot her experiments, 50 ms. ( left) With a 2 5 ms time delay, t here is a clear s eparation between t he performa nces of the cone, rod, a nd combined rod- cone model circuits for stimuli with pulse dela ys betwee n 10- 30 ms. (ri g ht) With a 75 ms time d elay, there was n o statistically si gnificant separ ation in pe rformance for pulse dela ys bet ween 10- 30 ms. Error bars re present the confidence inter val with a minimum cov erag e of 9 5 % ( Clo pper-Pea rson interval). Solid black line indicates c hanc e (a motion discrimination accuracy of 0.5, ie., 50 %). Asterisks denote statistical significance was reach e d in t he com parison betw een t he cone-on ly circuit 17 d .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 18 performance and the combined rod-cone circuit performance ( p ≤ 0.05). 358 Stimulus salience and mo tion discriminability 359 Whic h properties of the motion stimulus its elf had an impact on the separation in 360 performance betwe en the thre e differ ent circuits? The size and re lative con trast of a stimulus 361 define its sa lience . A wide stimul us p uls e width repres ents a larg er, and thus more salie nt, 362 stimulus than one with a narrow pulse width. Likewise , a high contrast stimulus is more sa lient 363 than a low contrast stimulus. As show n in Figure 5, we varied the siz e and contrast of the motion 364 pulse whil e keeping al l other stimulu s parameters fixed. We found that motion direction was 365 more accurately classified by al l three variations of the circuit (rod-cone, rod-only, and cone-366 only) for stimuli with longer pul se wi dths and greater contrast (Fig. 5, top right panel) . In other 367 words, the mode l’s discrimination performance general ly increas ed with stimulus sali ence . 368 Performance suffer ed for all circuits pres ented with short puls e widths and low contrast stimuli 369 (Fig. 5, bottom left pane l). The stimul i with 150% contrast had an increased range of pulse 370 delays in which all circuits were 10 0% accurate at labeling motion direction (Fig. 5, right 371 column), compared to the 50% or 100 % contrast stimuli. On the other hand, varying the puls e 372 widths had a less pronounced impact on the circuit’s overall performance w ithin a given 373 stimulus contrast . 374 375 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 19 Figure 5 Motion dis criminability a nd sti mulus salien ce In gene ral, discriminability performa nce i mproves with increased salience–specifically with increas ed stimulus pulse width (ve rtical) or increas ed contrast (ho rizontal). E rror bars repre sent the confidence interval wit h a minimum cov era ge of 95 % (Clo pper- Pearson interval ). Solid black li ne indicates cha nce (a motion discrimination accurac y of 0.5, ie., 50 %). Asterisks denote statistical sign ificance was reach ed in the comparison between the cone-o nly circuit performance a nd t he combined ro d-cone circuit performance ( p ≤ 0.05 ) . 376 For most of these conditions, there w as still a s eparation of accura cy betwe en the cone-377 only circuit relative to rod-only and r od-cone circuits across 20-30 ms puls e delay s, that was 378 p revious l y obs erv ed for 10 ms p uls e w idths at 100% contrast. For the 20 ms p ulse widths at 50 %379 contrast, however , there was no signif icant separation in the performances betwe en the mode ls 380 at pulse de lays b etwe en 20-30 ms . 381 19 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 20 Ultimately , our model d emonstrat es t hat interactions betwe en rod- and co ne-deriv ed 382 signals n egative ly impact the comput ation of motion direction. W hen we pull apart rod-cone 383 signal interactions by creating a circuit that is only responsiv e to cone inputs (ie., es sential ly a 384 rod knockout cir cuit), we find that the circuit is able to discriminate motion direction with 385 greater accuracy. For stimulus conditions which furt her decrease d the salie nce of the motion 386 stimulus , we obs erv e a greater ov eral l degradation in motion discriminabili ty relativ e to 387 discriminability for more salient stim uli. 388 389

and Conclusions 390 Unlike photopic lighting conditions w here cone photoreceptors dominate t he visual 391 respons e and scotopic lighting where rod photoreceptors dominate the visu al respons e, me sopic 392 lighting activates both rods and cones (review ed by (Buck, 200 4, Buck, 20 1 4, Grime s et al ., 393 2018)) . We hypothesized that destruc tive interf erenc e betw een rod- and cone-mediated sig nals 394 in the retina would hinder motion perception given that it hinders the ability to perceive 395 stationary flickering stimuli of certain frequencie s (Songco-Aguas et al , 202 3). Alternativ ely , it is 396 possibl e that rod and cone signals could constructive ly contribute to motion discrimination by 397 averaging out the indepen dent nois e i nherent to each rod and cone input p athway, thus 398 improving the signal to noise ratio. To examine this, we incorporated a rod-cone retinal circuit 399 model trained on e lectrophysiology d ata into a directionally-sensitiv e motion-detection circuit. 400 We found that motion discri minability was decrease d by the integration of rod- and cone-401 derive d respon se s for a specific range of pulse de lays– namely 2 0-30 ms–co rresponding with the 402 relativ e de lay in kinetics betwe en rod- and cone-derived re spons es . Les s sali ent stimuli l ed to 403 greater lo ss es in motion discriminability. 404 Simulating motion corre l ation as a post-retinal pro cessing stage 405 Our goal was to inves tigate how inter actions betwe en parall el rod- and cone-deriv ed 406 signals impacted the fidelity of retina l outputs, with motion discrimination as an important 407 example. We us ed a classical motion correlator, the Hass enst ein-Reichardt model as a readout 408 of motion. It was originally deve loped as a model of motion encoding in the insect visual s yst em, 409 and it relies on the corre lations betw e en two stimulu s inputs to compute the pres ence of motion 410 for a given stimulus (Reichardt 1961). The classical Hassenst ein-Reichardt model has been 411 shown to be a useful mode l for motion perception gen erally , including for applications in 412 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 21 computer vision (Qiu et al ., 20 25, Bir koben et al. , 2023 , Yan et al. 2022 , Basch et al. 201 0). For 413 our purposes , it serv es as a conveni en t readout to estimate the encod ed motion information 414 from two ch annels of the retinal mod el. To examine the impact of early reti nal processing on 415 motion encoding, we enhanced the H assen stein-R eichardt model by integr ating a simulated 416 rod-cone circuit into the c hannels. The rod and cone respons e kinetics and nonlinearitie s wer e 417 fit using el ectrophysiology data collec ted from non-human prim ate retina. By including rod and 418 cone linear kinetics and shared nonli near processing , we exp licitly study h ow signal interactions 419 among those components impact a ca nonical motion computation. 420 While the respons e filter s for the rods and cones in our model are physiolog ically 421 realistic, the corre lator time-de lay and multiplication components do not h ave a corresponding 422 physiological interpretation in the primate retina nor further downstream. However , the full 423 model is an abstract emulation of the directionally-tun ed and ve locity-tune d visual proces sing 424 that occurs in t he visual syst em, down stream of the retina. The directional t uning is based on 425 the subtraction stage prior to the final output of the Hassenstein-R eichardt model. The v elocity 426 tuning comes from the value of the sp ecific correlator time delay. This kind of directional and 427 velocity tuning wou ld be part of post-retinal computations, such as those that occur in t he 428 middle temporal (MT) visual area (Maunsel l and van Ess en, 1983). 429 Other studi es have obs erv ed perceptu al deficits in motion processing that a re attributed 430 to post-retinal proces sing. For example, subject s have a difficult time reporting the location of a 431 moving object at high speeds, perhaps becaus e of a deficit in the high-level binding of the 432 position of the object and the cue to r eport the position (Linares et al, 2 009 ). Signals trav eling 433 through the parvocellular pathway ar e slow er than those through the magn ocellu lar pathway 434 (Maunsel l et al, 1999), allowing for the possib le interpr etation of the correl ator time delay in the 435 Hassens tein-Reichardt model in term s of this post-retinal processing feat ur e; however , our 436 study exclu sive ly concerns motion dis crimination, whic h is generally handl ed by the 437 magnocellu lar pathway (Merigan et al., 1991). Our stu dy probed the rol e of specific retinal 438 interactions in perceptual d eficits rela ted to motion discrimination, but future studie s may 439 expand on our model to investigat e how post-retinal proces sing interacts w ith rod-cone 440 interactions. 441 Rod-cone signal interference degrades motion disc riminability 442 In our present stu dy, we found that motion processing is degrad ed by de str uctive signal 443 interfer ence, as evid enced by the wor se performance of the rod-cone mode l circuit compared to 444 its cone-only counterpart . Poor motion discriminability was to be expected from the rod-only 445 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 22 circuit because the slow kinetics of the rod input degrade the fine t emporal structure r equired 446 for the accurate encoding of the time- varying stimulus (Son gco-Aguas 202 3). Similar to the 447 classic visual s timulus of a flickering s pot (MacLeod 19 77, Stockman and Sh arpe 2006, Songco-448 Aguas et al. , 2023), mov ement or change of any kind requires the stimu lus to vary in time. The 449 lag betw een rod-mediat ed respon se s and cone-mediated respon se s and su bsequ ent d estructiv e 450 interfer ence ultimat ely b lurs the final temporal structur e of the stimulus , preve nting accurate 451 stimulus encoding within a ch annel. I n contrast to a flickering spot, our mo ving stimulu s also 452 varied in space. The los s of fidelity in encoding within a c hannel propagates through the cir cuit, 453 where a comparison across both c han nel s takes place with a built-in time delay. 454 In our previous res earch, we paired e l ectrophysiology with human psychop hysics to 455 investigat e how the kinetic difference betwe en rod- and cone-mediated sign als influe nces vi sual 456 perception in certain conditions (Grimes et al ., 20 15, So ngco-Aguas et al. , 2023). In one of these 457 studie s, we examined how interferenc e betw een rod- and cone-mediated si gnaling led to the 458 inability to perceiv e flickering stimuli in the peripheral retina (Songco-Aguas et al. , 2023). We 459 found evidenc e for destructiv e interf e rence be twe en thes e two signal pathways and that we 460 could affect the interferenc e by alteri ng either the temporal frequ ency of the flickering stimulu s 461 or the relative phase d elay be twe en the rod-activating and cone-activating c omponents of the 462 stimulus (Songco-Aguas et al. , 2023). 463 That said, our motion-discriminati on circuit is not only dependen t on the respons es 464 from a single channel – it also ne eds t o compare the signals from two ch an nel s receiving 465 stimulus input with a delay that repre sent s motion across the visual field . O ur interpretation for 466 our model circuit’s performance drop in discriminability is that 1) rod-cone signal interf erenc e 467 happens to some extent within a chan nel , degrading the ov erall output of si ngle channel s, and 468 that 2) this signal degradation is exac erbated by the motion correlator circuit in its 469 multiplication stage. This is akin to h ow the linear stage s of the rod-cone retinal mode l initially 470 cause the interfer ence , while the cons eque nt nonlinear s tages heighte n the effect of the signal 471 interfer ence downstr eam. Sp ecifically , the multiplication stage of the correl ator is a nonlinear 472 process that emphasizes correlation s in a time-varying motion stimulus (Dror, O'Carrol l, and 473 Laughlin 2000 , Suarez and Koch 198 9). Since the rod-cone retinal model d egrade s the temporal 474 structure of the stimulu s, the Hass ens tein-Reichardt correlator “los es” the t emporal information 475 it needs to corre late the signa ls be twe en the two input channels and encode them accurately as 476 motion. 477 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 23 Visual pe rception in peripheral retina and natural behavior 478 We note that the rod and cone kinetic data was fit to d ata collected from peripheral 479 neurons in primate retina. In many of the psychophysics experiments of motion detection in 480 mesopic conditions, subj ects are instr ucted to discriminate motion that is p eripheral to a 481 fixation point in t heir field of view (Bilino et al. , 20 08 , Yoshimoto and T akeuchi, 2013, 482 Yoshimoto et al., 2016, S epul veda et a l., 2 021) . One can imagine that in a na tural settin g, an 483 animal without any behavioral constr aints would det ermine the direction o f a moving object by 484 turning to look at the object directly (ie., fov eating it). Becaus e our model es sential ly simulat es 485 two nearby ON parasol ganglion ce lls in the peripheral retina, it is not suite d to make 486 predictions about foveal vi sual proces sing. Despit e this, our model i s stil l wel l-suite d to examine 487 how parallel processin g at the lowest l eve l of human vision might imp act a more complex 488 behavior. Studi es show that subjects perceiv e the motion of a stimulus as rever sing und er low 489 light conditions when a blank frame i s intersp ers ed in the stimulu s, but onl y when the stimulus 490 is in their periphery (Takeuchi and D e Valois, 20 09). This is likely du e to a first-order biphasic 491 filtering of the stimulu s consist ent with rod-derived respon se s (Snowden et al, 1995). Our mode l 492 ultimately capture s a known visual phenomenon , namely me sopic perceptu al impairments in 493 peripheral vision, and it demon strate s how interaction between rod- and cone-mediated sig nal 494 kinetics directly impact motion discri mination. 495 Specific perceptual deficits in mesopi c motion processing have been r eported in previous 496 literatur e with implications for every day tasks such as driving (Wood, 2019) . Motion stimuli that 497 emulate a locomoting figure among a background of inco herent dots wer e u niquely difficult for 498 human observer s to perceiv e in meso pic conditions, compared to ph otopic or scotopic 499 conditions (Bilino et al., 2 00 8, S epul v eda et al. , 2021) . Although it has been posited that rod-500 cone interactions occurring within th e peripheral retina may under lie the p henomenon (Bilino 501 et al. , 200 8), ther e has not been a dir ect examination of responses at the l e vel of the neura l 502 circuit. Our model re sult s are in agree ment with previous studi es which hav e describ ed 503 perceptual d eficits for motion, specifically in dim lighting conditions (Bilino et al., 2 00 8, 504 Yoshimoto et al., 2013, Yoshimoto et al., 2 016, Mayeur et al ., 2 008 , Gros s man and Blake, 1999, 505 Sepul veda et al ., 2 021) . 506 507 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 24

508 Electrophysiological recording s use d to train the model were coll ected b y Fred Rieke and 509 William Grimes. Fe edback and comm ents on study from Rieke lab mem ber s, including t est 510 readers A lison Web er and Alex White. This work was supported by the Nati onal Institute s of 511 Health, Bethe sda MD, grants K22NS1 041 87, EY0281 11, and 5R90DA03346 1, as wel l as the UW 512 Institute of Neuro engine ering Washington Research Foundation Innovation Post-Baccalaureate 513 Fellow ship. 514 515 Contributions 516 FR, GJG, and ASA design ed the simul ations and the computational study. FR collected the data 517 used to fit the model circuits. ASA cre ated the model . ASA and GJG implemented the 518 simulations, analy zed the simu lated d ata, and drafted the manuscript. All authors participated 519 in editing the manuscript. 520 521 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 25

522 1. Adelson, E. H., & Mo vsh on, J. A. (1982). Phenome nal co here nce of movin g visual patterns. 523 Nature, 300( 58 92), 523 -52 5. 524 2. Arden, G. B., & Hog g, C. R. (198 5). Rod-c one interactions an d anal ysis of retinal di sease. Brit ish 525 Journal of Op hthal mology , 69 ( 6), 40 4–4 15. https ://doi.org/ 10.113 6/bjo.6 9.6. 404 526 3. Barlow, H. B., & Le vick, W. R. (1 9 65). Th e mecha nism of directionally selective uni ts in rabbit's 527 retina. T he Jour nal of p hysio logy , 178(3 ) , 477. 528 4. B a s c h , M . E . , C r i s t e a , D . G . , T i p o n u t , V . , & S l av i c i , T . ( 2 0 1 0 ) . E l a b or a t e d m o t i o n de t e c t o r b a s e d on 529 Hassenstein-Reic hardt cor relator model. Latest T rends o n Systems, 1, 1 92-1 95. 530 5. Billino, J. , Bremme r, F., & Geg enfurtner, K. R. (2008). Motion processing at low li ght lev els: 531 Differential effects o n t he perception of s pecific motion types. Jou rnal of Vis ion , 8 (3), 1 4. 532 https://doi.or g/10.11 67/8.3.1 4 533 6. Birkoben, T. , Hansen, M ., I gnato v, M., Zi egler, M ., & Koh lstedt, H. (2023). A Bio-inspired 534 Perceptual Decision-Makin g Circuit Bas e d on the Hassenstein-Reicha rdt Dir ection Detector. In 535 Bio -Insp ired Inform ation Pat hways: Fr o m Neur oscience to Neurotr onics (pp . 3 17 -333). C ham : 536 Springer Inte rnational Publishin g. 537 7. Borst, A., & Eg elhaaf, M. (1 989 ). P rinciples of visual motion detection. Trends i n N eurosciences, 538 12(8), 29 7–30 6. https ://doi.org/10. 1016/ 0166-22 36(8 9) 90010-6 539 8. Borst, A., & Euler, T . (2011 ). Seein g T hin g s in Motion: Mode ls, Circuits, and Mec ha nisms. 540 Neuron, 71 ( 6), 97 4– 99 4. https ://doi.or g/ 10.1016/j.neuron.201 1.08.031 541 9. Borst, A., & Helmstaedter, M. (201 5). Co mmon circuit desig n in fly and mammali an motion 542 vision. Nature N euroscience, 18(8 ), 10 67 –1076. https ://doi.or g/10.1038/nn. 40 50 543 10. Buck, S. L. (200 4). Ro d-cone interactions. In L. M. C halupa & J. S. We rner ( Eds.), The Visual 544 Neurosciences ( pp. 863 -878). Boston : MI T P ress. 545 11. Buck, S. L. (201 4). T he interaction of rod and cone signals : pathwa ys and ps ychop h ysics. In J. S. 546 W. A. L. M. C halupa ( Ed.), The Ne w Visu al Neu rosciences (p p. 485 -4 97). Bosto n: MIT Pr ess. 547 12. Chichilnisk y, E. J. (2001). A simple w hite noise a nalysis of n euronal lig ht response s. Networ k: 548 Computation in Neu ral Systems , 12(2), 1 99–2 13. https ://doi.or g/10.1080/713 663 221 549 13. Dror, R. O., O' Carroll, D. C., & Laug hlin, S. B. (2001). Accurac y of ve locity estimati on by 550 Reichardt corr elators. Jour nal of the Opti cal Society of Ame rica. A , Optics, image science, and 551 visio n , 18(2) , 2 41–2 52. https ://doi.org/ 1 0.136 4/josaa.18.000241 552 14. Dunn F A, Doan T, Sampat h AP, Rieke F ( 2006) C ontrollin g t he Gain of Ro d-Media ted Signals in 553 the Mammalian Retina. J Ne urosci 26 :3 9 59– 39 70. 554 15. Dunn, F. A., Lank heet, M. J., & Rieke, F. ( 2007). Li ght adaptation in cone vision in volves 555 switching betwee n r eceptor and post-rece ptor sites. Nature, 4 49 (7162 ), 603- 606. 556 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 26 16. Enrot h-Cu gell, C., Hertz, B. G., & Le nnie, P. (1 977). Con ver gence of ro d and cone si gnals in t he 557 cat’s retina. T he J ourna l of P hysiol ogy , 26 9 (2), 2 97–318. 558 https://doi.or g/10.11 13/jp hysiol.1 977.sp0 11903 559 17. Fain, G., & Sampath , A. P. (2018). Rod an d cone inte ractions in t he retina. F1000R esearch , 7 , 560 657. https://doi.or g/10.12 688/f1000rese arch.14 412.1 561 18. Grossman, E. D., & Blak e, R. (1 99 9). Perc eption of co here nt motion, biological mot ion and fo rm-562 from-motion und er dim -lig ht conditions. Vision resea rch, 39 (22), 3721– 3727. 563 https://doi.or g/10.101 6/s0042- 6 989( 99) 00084-x 564 19. Hassenstein, B., & Reicha rdt, W. (1 95 6). Systemtheor etische a nalyse de r zeit-, rei h enfolgen -und 565 vorzeich enauswertun g bei der bewe gun gs perzeption des rüsselkäf ers c hloro phanus . Zeitsc hr ift 566 für Naturfo rschu ng B , 11( 9-10), 5 13- 524. 567 20. Gouras, P., & Link, K. (1 96 6). Rod and co ne interaction in dark -adapted monke y g anglion cells. 568 The Jou rnal of P hys iology , 18 4 (2), 4 99. 569 21. Grimes, W. N., Grav es, L. R., Summers, M. T., & Rieke, F. (201 5). A simple retinal mechanism 570 contributes to perce ptual interactions between ro d- a nd cone -mediated responses i n primates. 571 eLife , 4, e0803 3. https ://doi.org/ 10.75 54 /eLife.08033 572 22. Grimes, W. N., Son gco- Aguas, A., & Riek e, F . (2018). Paral lel Processing of Rod a n d Cone Signa ls: 573 Retinal Fu nction and Hum an P erception. A nnual Review of V isi on Sc ience, 4 (1), 1 23–141 . 574 https://doi.or g/10.11 4 6/annure v-vision- 091517 -03 405 5 575 23. Linares D, Holcombe AO, White A L (2009) W here is t he movin g object now? Jud g ments of 576 instantaneous position s how poor tempor al precision ( SD = 70 ms). Journal of Visi on 9 : 9. 577 24. MacLeod, D. I. A. (1 972). Rods Cancel Co nes in Flicker. Na t u r e , 23 5 , 173 –174 . 578 25. Maunsell, J. H., & Van Essen, D. C. (198 3 ). Functional p roperties of ne urons in mi ddle tempo ral 579 visual area of the macaque mo nke y. I. Sel ectivity for stimulus direction, s peed, and orientation. 580 Journal of neurop hysio logy , 49 (5 ), 1 127 –1147. https://doi.or g/10.11 52/jn.1 983.4 9.5.1 127 581 26. Maunsell J HR, G hose G M, Assad J A, Mc AD AMS CJ, Boudreau C E, Noera ge r BD ( 199 9) Visual 582 response latencies of ma gnocellular and parvocellula r LG N neurons in macaque monkeys. Visual 583 Neuroscience 1 6 :1–1 4. 584 27. Mayeu r, A., B rémond, R., & Bastien, J. M . C. (2008). Effect of Tas k an d Eccentricity of the Tar get 585 on Detection T hres holds in Mesopic Visio n: Implications for Road Li ghtin g. Hum a n Factors: T he 586 Journal of the Hum an Factors and Ergo nomics Soc iety, 50 (4), 712– 721. 587 https://doi.or g/10.1 518/001872008X312 260 588 28. Meriga n W H, B yrne C E, Maunsel l J H (1 9 91) Does primate motion pe rception d epe nd on the 589 magnocellula r pat hwa y? J Neurosci 11: 34 22–342 9. 590 29. Petzold, A., & Plant, G. T. (2006). Clinical disorders affecting mesopic vision. Op ht halm ic a nd 591 Physio logica l Optics , 26 (3), 32 6– 341. https://doi.or g/10.1111/j.1 47 5-131 3.2006.00 417.x 592 30. Qiu, Z., Yan, C., C hen, T., Hua, Y., Todo, Y., & Tan g, Z. (202 5). A motion dir ection detecting 593 model for colored ima ges based on th e H assenstein–Reichardt model. Mac hi ne V i sion and 594 Applicati ons , 36 (1), 2 9. 595 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint 27 31. Raphael, S., & MacLeo d, D. I. A. (2011). Mesopic luminance assessed with minim um motion 596 photometr y. Jour nal of Vis ion , 11( 9), 1 4– 14. https ://doi.org/ 10.1167/ 11.9.1 4 597 32. Reichardt, W. (19 61). Autocorrelation, a p rinciple for evaluation of se nsor y informa tion by the 598 central ner vous system. In Senso ry com municatio n ( pp. 303– 317). M IT Press. 599 33. Schwartz G W, Okawa H, Dunn FA, Mo rg an J L, Kersch ensteiner D, W ong R O, Rie ke F (20 12) T he 600 spatial structure of a nonlinear rece ptive f ield. Nature N euroscience 1 5 :15 72–1 580. 601 34. Sepulveda, J. A., Anderso n, A. J., Woo d, J. M., & McKend rick, A. M. (2021). Motio n perce ption at 602 mesopic lig ht l evels : Effects of p h ysiologi cal agein g and eccentricity. Op hth almic a nd 603 Physio logica l Optics , 41(2), 447 –4 56. https://doi.or g/10.1111/o po.12783 604 35. Suarez, H., & Koc h, C. ( 198 9). Linkin g lin ear th res hold u nits with quadratic model s of motion 605 perception. Neural Computat ion, 1(3), 31 8-320. 606 36. Snowden RJ, Hess RF, Waug h SJ (19 9 5) The p rocessing of tem poral modulation a t different 607 levels of r etinal illuminance. Vision Resea rch 35 :77 5–78 9. 608 37. Songco- Aguas, A., Grimes, W. N., & Riek e, F . (202 3). Rod -cone si gnal interfere nce in t he retina 609 shapes perce ption in primates. Fr ontiers in Opht halm ology , 3, 1230084. 610 https://doi.or g/10.3 389/fop ht.2023.12 3 0084 611 38. Stockman, A., & S harpe, L. T. (2006). Int o the twilig ht z one : Th e compl exities of m esopic vision 612 and luminous efficiency. Opht halm ic a nd Phys iolo gical Optics , 26 (3), 225 –239. 613 https://doi.or g/10.11 11/j.147 5-1 313.2006 .00325.x 614 39. Takeuc hi T, De Valois K K (200 9) Visual motion mechanisms under low retinal ill uminance 615 revealed by motion rev ersal. Vision Resea rch 49 :801–80 9. 616 40. Tron g, P. K., & Rie ke, F. (2008). Ori gin o f correlated activity betwe en parasol retin al ga nglion 617 cells. Nat Neurosci, 11(11) , 1 34 3-13 51. 618 41. Wood JM (2020) Ni ghttime d riving : visu al, lig hting and visibility c hallen ges. O pht halmic and 619 Physiolo gical Optics 40:187 –201. 620 42. Yan, C., Todo , Y., Koba yashi, Y., Tan g, Z., & Li, B. (2022). A n artificial visual syste m for motion 621 direction detection based on t he Hass enst ein–Reichardt co rrelator model. Electron ics , 11(9 ) , 622 1423. 623 43. Yoshimoto, S., Okajima, K., & Ta keuchi, T. (201 6). Motion perce ption unde r m eso pic vision. 624 Journal of Vis ion , 16( 1), 1 6. https ://doi.o rg/10.11 67/1 6.1.16 625 44. Yoshimoto, S., & Tak euchi, T. (2013). Vis ual motion priming re veals wh y motion p erception 626 deteriorates durin g mesopic vision. Jo ur nal of Vis ion, 13(8), 8–8. https ://doi.or g /10.1167/13.8.8 627 .CC-BY-NC-ND 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted December 27, 2025. ; https://doi.org/10.64898/2025.12.26.696616doi: bioRxiv preprint