Functional Clusters for Shape, Texture, and Motion Encoding in Macaque V2
- Taekjun Kim1,
- Rohit Kamath1,
- Gaku Hatanaka1,
- Tomoyuki Namima2,
- Celeste Dylla1,
- Wyeth Bair1, and
- Anitha Pasupathy1
- 1Department of Neurobiology & Biophysics and Washington National Biomedical Research Center, University of Washington, Seattle, Washington 98195
- 2Graduate School of Frontier Biosciences, The University of Osaka and Center for Information and Neural Networks, National Institute of Information and Communications Technology, Suita 565-0871, Japan
-
Author contributions: T.K., T.N., and A.P. designed research; T.K., R.K., G.H., T.N., C.D., and W.B. performed research; T.K., R.K., and G.H. analyzed data; T.K., R.K., G.H., W.B., and A.P. wrote the paper.
Abstract
Macaque primary visual cortex (V1) exhibits exquisite columnar organization, while midlevel area V4 does not. Here we investigated the functional organization and representational bases of intervening area V2 in three macaques (one male, two females) with high-density Neuropixels recordings and a variety of visual stimuli—shape, texture, drifting grating, and translational motion patches. We observed dense clusters of similarly tuned neurons often spanning ∼500 µm, consistent with a columnar structure. In terms of representational bases, V2 responses were largely explained by stimulus features based on local image statistics: shape tuning is well modeled by a linear combination of orientation filters, and direction selectivity is stronger with surface compared with object motion, in striking contrast to V4. Overall, our results support the progression from columns to sparse clusters as neuronal representations transform from encoding local features and feature conjunctions in V1/V2 to a high-dimensional object-based code in V4.
Significance Statement
By recording hundreds of neurons simultaneously across layers of macaque visual area V2, we show the first evidence of exquisite fine-scale functional clusters that encode higher-order shape, texture, and motion features, extending well beyond the classic orientation-selective columns seen in area V1. Comparative analyses with V4 further reveal distinct representational bases and organization patterns between adjacent cortical areas, demonstrating that columnar organization is preserved in early visual areas (V1 and V2) but is markedly attenuated in higher-order cortex such as V4.
Introduction
The second stage of cortical visual processing in the primate, area V2, occupies roughly 10% of macaque neocortex (Vanni et al., 2020) and is critical for processing both visual form and motion information. Studies have demonstrated selectivity for local orientation and spatial frequency (Foster et al., 1985; Gegenfurtner et al., 1996; Felleman et al., 2015), as in primary visual cortex (V1), but also for higher-order texture statistics (Freeman et al., 2013; Ziemba et al., 2016; Okazawa et al., 2017), stimulus form (Hegdé and Essen, 2000; Ito and Komatsu, 2004; Anzai et al., 2007), border ownership (Zhou et al., 2000), and stimulus motion contrast (Hu et al., 2018; Ma et al., 2021). Past studies using optical imaging methods have demonstrated remarkable organization across the cortical surface for features such as orientation, color, and motion (Ts’o et al., 1990; Lu and Roe, 2008; Felleman et al., 2015). Here we used high-density probes to study subpopulations of V2 neurons across laminae with a broad range of stimulus features of texture, form, and motion. Our studies are motivated by two goals. First, we wish to gain deeper insights into the representational bases and functional organization of V2. Prior work has demonstrated a dramatic shift in terms of functional organization and representational bases from macaque V1 to V4: columnar organization in V1 (Hubel and Wiesel, 1974; Obermayer and Blasdel, 1993) to sparse clusters in V4 (Namima et al., 2025) and basic feature detection in V1 (De Valois et al., 1982) to object-based encoding in V4 (Pasupathy et al., 2020). Revealing the representational bases and organization in the intervening area V2 for shape, texture, and motion stimuli will advance our understanding and constrain models of the ventral stream. Second, we aim to contribute to the debate on the origin and significance of columns (Horton and Adams, 2005). Some studies have posited that columns represent a critical functional unit (Mountcastle, 1957; Hawkins et al., 2017). On the other hand, demonstrations of their absence in some species (Girman et al., 1999; Adams and Horton, 2006) and in some brain regions (Johnson et al., 2000; Sato et al., 2009) raise the possibility that columns may instead arise as a by-product of representational expansion (Ringach, 2007; Jang et al., 2020), i.e., a small number of inputs project onto a large number of neurons. Moreover, columnar structure may also be less likely to develop when converging afferents exhibit high dimensionality (Purves et al., 1992). Our recent observation in area V4 (Namima et al., 2025), where shape and texture selectivity appeared less spatially clustered, supports this latter hypothesis and prompts further investigation into whether similar patterns exist in V2.
We used high-density, multicontact linear probes (Neuropixels, IMEC) to simultaneously record responses from dozens of neurons in area V2 of anesthetized macaque monkeys. We used a broad range of stimuli to study responses—2D shape silhouettes, naturalistic textures and their spectrally matched noise counterparts, sinusoidal drifting gratings, and second-order translational motion stimuli over a range of spatial displacements. To identify functional clusters of similarly tuned neurons, we first measured the similarity in tuning for each stimulus class by calculating the response correlations between all pairs of simultaneously recorded neurons and then performed cluster analyses. Comparison with previously collected data from area V4 using identical stimuli (Bigelow et al., 2023; Namima et al., 2025) reveals a striking distinction between the two areas in terms of functional organization and representational basis. V2, but not V4, exhibits extensive clustering for neurons encoding similar local features of an image. We also observed differences between V2 and V4 in terms of the representational basis for shape and motion stimuli. Additionally, our large-scale dataset offers new insights into the temporal and laminar dynamics of texture selectivity, highlighting distinct patterns in the processing of perceptual texture dimension of coarseness, directionality, and regularity.
Materials and Methods
Animal preparation
Electrophysiological recordings were conducted in visual area V2 of three anesthetized, paralyzed rhesus macaques (Macaca nemestrina: a 3-year-old 6.7 kg male, a 9-year-old 6.6 kg female, and a 5-year-old 5.5 kg female). The positioning of the craniotomy was guided by stereotaxic coordinates based on standard macaque brain atlases and confirmed by identifying cortical vasculature and sulcal landmarks after durotomy. This placement allowed access to both, the primary visual cortex (V1) and area V2. To accurately identify the border between areas V1 and V2, we characterized the ocular dominance structure using the widefield optical imaging technique described below. During widefield imaging, the exposed cortex was covered with a glass coverslip to stabilize the surface and minimize optical distortion. For subsequent electrophysiological recordings, the cortex was instead covered with an artificial dura to protect the tissue.
Experiments typically lasted 4–5 d. Anesthesia was maintained throughout the experiment using a continuous infusion of sufentanil citrate (4–6 μg/kg/h), and neuromuscular blockade was induced with vecuronium bromide (0.1 mg/kg/h). Both medications were administered intravenously in lactated Ringer's solution (8 ml/kg/h) supplemented with 2.5% dextrose. Throughout the experiments, artificial respiration was maintained, with the respiratory rate adjusted to keep expired carbon dioxide levels within a physiological range (3.8–4.0%). Body temperature was maintained at ∼37°C using a heating pad.
Intrinsic optical signal imaging
In two monkeys, we conducted intrinsic optical signal imaging to identify the V1/V2 border based on the ocular dominance pattern in V1. A square-wave drifting grating (4 s, 1.0 cycles/°, 4 Hz) was presented on the display (Display++, Cambridge Research Systems) at a distance of 127 cm. Eight directions were randomly presented in each session for 20 repeats for each condition. One of the eyes was occluded with a shield plate. The image was acquired with sCMOS camera (Zyla-4.2 USB3, Andor) equipped with a tandem lens configuration (two Nikkor 50 mm, Nikon) to achieve 13.3 × 13.3 mm field of view. A glass coverslip was placed on the cortex to prevent motion artifacts. The cortex was illuminated with an LED at 625 nm wavelength (M625L4, Thorlabs). Image acquisition was controlled by Micro-manager software (μManager; https://micro-manager.org; Edelstein et al., 2014). The response amplitude was calculated as a fraction of the image intensity from the baseline. First, baseline intensity R0 was calculated as an average over one second before the stimulus onset. Then the driven response (R1) was calculated as an average between 5 and 6 s after the stimulus onset. Finally, the response amplitude was calculated as difference over baseline (R1 − R0) / R0. The ocular dominance pattern was visualized as a t statistic between the left eye condition and the right eye condition. The vasculature pattern was also acquired with the same imaging setup under 530 nm illumination (M530L4, Thorlabs) to guide electrode penetration location.
Electrophysiology
Extracellular recordings were carried out using Neuropixels 1.0 and Neuropixels 1.0 NHP (IMEC). Neural signals from the probe and non-neural event signals from other devices (sync pulse generator, photodiode) were acquired using a PXIe acquisition module (PXIe_1000, IMEC) and multifunction I/O module (PXI-6224, National Instruments-NI). These were transmitted to a data acquisition Windows machine via the PXI Chassis (PXIe-1071, NI). Action potential (AP) and local field (LF) potential signals from the 384 probe contacts (AP: 30 kHz sampling rate; LF: 2.5 kHz sampling rate) were amplified and bandpass filtered (AP: 0.3–10 kHz; LF: 0.5–500 Hz) and stored for offline analysis (spikeGLX; https://billkarsh.github.io/SpikeGLX/). Square-wave sync pulses generated with a microcontroller board (Arduino R3, Arduino) were stored in both the neural and non-neural event data recording streams so that spike timing and photodiode detection could be synchronized in offline analysis.
For probe insertion, we made a small slit (2–3 mm wide) in the artificial dura and passed the probe through it. The exposed cortical surface was maintained by the application of artificial cerebrospinal fluid to prevent drying. Probes were inserted using an oil hydraulic micromanipulator (Narishige, MO-97A) until cortical penetration was achieved and neural activity was detected. Additionally, a custom-built presser foot was used to provide tension to the cortical surface, providing additional stability during probe insertion.
For each individual recording site, binary data files collected across experiments were combined into a single binary file. Combined binary files were then processed with open-source automated spike sorting software (Kilosort 2.5/3; Pachitariu et al., 2024) with default parameters. The output was subsequently manually curated using the Phy2 interface (https://github.com/cortex-lab/phy) to ensure the quality of single-unit isolation. The detailed procedures for this curation are described elsewhere (Bigelow et al., 2023; Namima et al., 2025). Only well-isolated single units were included in the data analysis.
Following automated and manual spike sorting, a final preprocessing step was performed to align the different data recording streams. Because neural and non-neural event signals were recorded on separate streams with independent clocks, sync pulse edge times were linearly regressed to bring spike times, event timestamps, and photodiode signals into a common temporal reference frame. Subsequent analyses were then performed using custom Python scripts.
Receptive field characterization
Each penetration began with a hand mapping procedure to roughly identify the spatial receptive fields (RFs) of neurons recorded along the length of the probe. This preliminary RF mapping was performed for 2–4 channels that were well spaced along the probe's length.
To precisely characterize the RF locations, we conducted an automated mapping procedure on a 7 × 7 (or 9 × 9) grid (1° step) centered on the RF estimated by hand mapping. For this experiment, we used a random dot motion patch (diameter, 1.4°) and presented the stimulus for 300 ms, 10 times at each of the 49 (or 81) grid locations in random order. The RF center was then estimated by fitting a 2D Gaussian to the measured responses of each neuron (for more details, see Kim and Pasupathy, 2024). Based on prior data (Gattass et al., 1981), we estimated the V2 RF size as (RF diameter = 0.5 + 0.3125 × RF eccentricity). During each recording session, visual stimuli were centered on the aggregate of the fitted RFs across all recording channels.
Visual stimuli
During each session, we presented shape, texture, drifting grating, and translational motion patch stimuli in separate blocks. Each stimulus was presented 10 times in random order. All stimuli were presented against an achromatic background (17 cd/m2; x = 0.33 and y = 0.33 in CIE-xy chromaticity coordinates).
Texture stimulus set
The texture stimulus set was composed of 40 naturalistic grayscale textures, drawn from eight groups (five exemplars per group) defined by three perceptual dimensions (i.e., coarseness, directionality, regularity) known to affect human texture perception (Fig. 1A). Stimulus parameterization in terms of these perceptual dimensions has been described in detail previously (Kim et al., 2019, 2022). Briefly, coarseness and directionality indices, computed from the Fourier power spectrum of the images, capture spatial frequency distribution and orientation bias of each texture, respectively. Regularity was computed from the 2D auto-correlation map and reflects the repetitiveness of the texture pattern. Textures were grouped based on high and low values along each of these attributes (see diagrams in Figs. 1, 7), and neuronal responses were compared between the groups (e.g., coarse vs fine, directional vs nondirectional, regular vs irregular).
Each of the 40 texture images were presented in one of three versions, (1) the original image, (2) a contrast-reversed version in which pixel luminance was inverted, and (3) a spectrally matched noise version, yielding a total of 120 stimulus conditions. Spectrally matched noise images were created by removing the phase information of the original textures, by first computing the Fourier transform, randomizing the phase values, and then inverting the Fourier transform. This produces texture stimuli that retain the orientation and spatial frequency content of the original textures while removing the higher-order statistical dependencies (Freeman et al., 2013). To avoid any unintended associations between a texture category and specific local orientation energy, each texture was presented at a randomly chosen rotation on each repetition. To control overall luminance, texture images were adjusted to have the same mean RGB pixel values (100, 100, 100). All texture images were centered on the aggregate RF, presented through a circular aperture with a blurred boundary, and their size (8° diameter) was chosen to extend well beyond the estimated RF. Each stimulus was presented for a 300 ms duration, separated by 300 ms interstimulus interval.
Shape stimulus set
The shape stimulus set consisted of 50 parametric closed 2D shapes randomly selected for each recording site from the original 366 shape stimulus set used to characterize V4 neurons (Pasupathy and Connor, 2001; Fig. 1B); each shape was scaled to fit entirely within the RF boundary. Since simultaneously recorded neurons could have RFs that are spatially offset relative to each other, each shape in the stimulus set was presented at five spatial positions: the aggregate RF center and four additional locations shifted left, right, up, and down by half of the estimated RF diameter. For each neuron, the optimal stimulus position was identified post hoc as the position that elicited the strongest mean response averaged across all shapes, and responses at this position were used for all subsequent population analyses. This position optimization was not a concern for texture and drifting grating stimuli, because they are homogeneous and each stimulus covered a spatial extent substantially larger than the RF of individual neurons. Within a given recording session, the same set of 50 shapes was presented at all five positions. In most sessions, sets of 50 unique shapes were used, but in a few sessions (3 out of 20), larger sets of 120 shapes (multiple rotations of the same shape) were used. Stimuli were presented for a 300 ms duration separated by 300 ms interstimulus interval.
Drifting grating stimulus set
Drifting sinusoidal gratings (4 Hz temporal frequency, 1 s duration, 500 ms interstimulus interval) were presented at 12 directions of motion (in 30° steps) and five values of spatial frequency (0.25, 0.5, 1, 2, 4 cycles/deg; Fig. 1C). As with the texture stimuli, grating stimuli were presented through a circular aperture with a blurred boundary and were sized (8° diameter) to extend beyond the aggregate RF.
Translational motion stimulus set
Translational motion stimuli, like those used in Bigelow et al. (2023), comprised elongated noise patches presented at several sequential spatial locations centered on the estimated RF (Fig. 1D). At each spatial location, the stimulus was presented for a specific duration (dT) before instantaneously moving to the next location. These stimuli were designed to induce apparent motion in one of eight directions separated by 45° increments. We tested five spatial displacements (dX = 1/3, 1/6, 1/9, 1/18, or 1/36 × RF diameter), each paired with a corresponding stimulus duration (dT = 100, 50, 33.33, 16.66, or 8.33 ms) and number of spatial locations (7, 14, 21, 42, or 84), so that the motion trajectory and speed were held constant across all tested dX–dT combinations. The long axis of the elongated patches (aspect ratio, 3:1) was scaled to be 60% of the RF diameter, and the total stimulus duration was 700 ms. The mean luminance of the patches was matched to the background, and a unique noise pattern was randomly generated for each of the spatial locations on every presentation; thus, stimulus motion was defined by non-Fourier cues.
Data analysis
Normalized response matrix: quantification of neural response magnitude
To create a normalized response matrix for each visual stimulus set (such as the texture stimulus set), we generated response vectors for each unit. These vectors, with a length corresponding to M stimulus conditions (e.g., 120 texture images), were created by counting spikes within a window from stimulus onset to 100 ms after stimulus offset. We then normalized each vector by its maximum response. Finally, the normalized response vectors from simultaneously recorded neurons (ordered by channel positions along the recording probe) were aggregated to form the response matrix (M stimulus conditions × N units; Fig. S1).
Quantification of similarity in feature selectivity
To quantify the similarity in feature selectivity among simultaneously recorded neurons, we computed the Pearson's correlation coefficient between the normalized response vectors (described above) of all neuron pairs. These similarity values populated an N × N matrix, where N denotes the number of simultaneously recorded neurons. Values along the diagonal of this matrix, which are always 1 by definition, were replaced with NaNs; the off-diagonal values, ranging from −1.0 to 1.0, provided a measure of tuning similarity between neuron pairs.
Because the sampling distribution of Pearson's correlation coefficient is inherently skewed, particularly for strong correlations, we applied Fisher's Z-transformation to the correlation coefficients before performing any further analyses. All subsequent statistical operations, including computation of population means and baseline correlation subtraction, were conducted using the Z-transformed values.
Neuronal clustering analysis
We used an ad hoc clustering algorithm to identify neuronal clusters with similar feature selectivity. To ensure that our clusters represent spatially and functionally cohesive groups, we first excluded neurons that did not exhibit statistically significant correlation with at least five of their ten nearest neighbors. This step helped to filter out neurons with weak functional similarity with nearby neighbors. Next, we applied K-means clustering to the normalized response matrix of the remaining neurons. The K-means algorithm partitioned these neurons into distinct clusters based on their similarity in feature selectivity. To determine the optimal number of clusters, we used Silhouette analysis, which effectively balances the need to minimize within-cluster variance while maximizing between-cluster separation (Rousseeuw, 1987). K-means clustering often yields clusters that are very small in size. Therefore, as a final step, we implemented two postprocessing procedures to refine the clustering results. First, we calculated the mean response vector for each cluster and merged clusters when the correlation between mean response vectors exceeded a threshold (r > 0.6). Second, we removed clusters that included fewer than five neurons.
Orientation energy model
To determine whether a linear combination of V1-like oriented filter outputs can explain V2 shape responses, we computed the orientation energy of visual shapes using a steerable pyramid algorithm (Simoncelli and Freeman, 1995). Each shape image was decomposed using a steerable pyramid transform with four orientations and three scales, and the power of each filtered subband was calculated by squaring the filter responses. As orientation energy remained consistent across different scales, we focused on the power from the intermediate scale for subsequent analyses. We modeled shape responses as a linear weighted sum of four orientation filter energies. The overall significance of this model was assessed using the F statistic. To quantify the similarity of orientation model fits between neurons, we computed the Pearson's correlation coefficient of weight vectors for all neuronal pairs.
Direction selectivity index
For responses to drifting sinusoidal gratings and translational motion patch stimuli, we first identified orientation-selective neurons by performing a one-way ANOVA with stimulus orientation as the factor. Subsequently, for neurons exhibiting significant orientation selectivity, we quantified their direction selectivity using the direction selectivity index (DSI) given by DSI = 1 − (np / p), where p represents the stronger response and np the weaker response along the two opposite directions orthogonal to the preferred orientation. The DSI values reported in this study are based on baseline (calculated from no stimulus trials) subtracted responses. In our analysis, direction selective neurons are a subset of orientation-selective neurons. Neurons with DSI > 0.5 were classified as direction selective.
Throughout the manuscript, the term “direction” is used exclusively to refer to motion-related tuning properties (e.g., the preferred direction of translational motion stimuli), whereas “directionality” (and “directional”) is reserved specifically to describe a perceptual dimension of texture.
Orientation tuning curve fitting
Sinusoidal drifting gratings were tested with 12 motion directions (6 orientations), while translational motion patch stimuli
included 8 motion directions (4 orientations). To compare orientation tuning between drifting gratings and elongated translational
motion patch stimuli, we fitted the tuning curves using the von Mises function, a periodic generalization of the Gaussian
distribution:<mml:math display="block"><mml:mi>f</mml:mi><mml:mo stretchy="false">(</mml:mo><mml:mi>x</mml:mi><mml:mo stretchy="false">)</mml:mo><mml:mo>=</mml:mo><mml:mi>A</mml:mi><mml:mo>⋅</mml:mo><mml:mstyle
displaystyle="true" scriptlevel="0"><mml:mrow><mml:mfrac><mml:mrow><mml:mrow><mml:mi mathvariant="normal">exp</mml:mi></mml:mrow><mml:mrow><mml:mo>[</mml:mo><mml:mrow><mml:mi>κ</mml:mi><mml:mrow><mml:mi
mathvariant="normal">cos</mml:mi></mml:mrow><mml:mrow><mml:mo>(</mml:mo><mml:mrow><mml:mn>2</mml:mn><mml:mi>π</mml:mi><mml:mstyle
displaystyle="true" scriptlevel="0"><mml:mrow><mml:mfrac><mml:mrow><mml:mo stretchy="false">(</mml:mo><mml:mrow><mml:mi>x</mml:mi><mml:mo>−</mml:mo><mml:mi>μ</mml:mi></mml:mrow><mml:mo
stretchy="false">)</mml:mo></mml:mrow><mml:mn>180</mml:mn></mml:mfrac></mml:mrow></mml:mstyle></mml:mrow><mml:mo>)</mml:mo></mml:mrow></mml:mrow><mml:mo>]</mml:mo></mml:mrow></mml:mrow><mml:mrow><mml:mn>2</mml:mn><mml:mi>π</mml:mi><mml:msub><mml:mi>I</mml:mi><mml:mn>0</mml:mn></mml:msub><mml:mo
stretchy="false">(</mml:mo><mml:mi>κ</mml:mi><mml:mo stretchy="false">)</mml:mo></mml:mrow></mml:mfrac></mml:mrow><mml:mo>+</mml:mo><mml:mrow><mml:mi
mathvariant="normal">baseline</mml:mi><mml:mo>,</mml:mo></mml:mrow></mml:mstyle></mml:math>f(x)=A⋅exp[κcos(2π(x−μ)180)]2πI0(κ)+baseline,
where <mml:math display="inline"><mml:mi>x</mml:mi><mml:mo>,</mml:mo><mml:mspace width="0.25em"></mml:mspace><mml:mi>A</mml:mi><mml:mo>,</mml:mo><mml:mspace
width="0.25em"></mml:mspace><mml:mi>μ</mml:mi><mml:mo>,</mml:mo><mml:mspace width="0.25em"></mml:mspace><mml:mi>κ</mml:mi><mml:mo>,</mml:mo><mml:mspace
width="0.25em"></mml:mspace><mml:msub><mml:mi>I</mml:mi><mml:mn>0</mml:mn></mml:msub></mml:math>x,A,μ,κ,I0
are stimulus orientation (in degree), amplitude, preferred orientation, tuning width parameter, and modified Bessel function
of order 0, respectively.
Time course of texture-dependent modulation
Textures were grouped based on high and low values along each of perceptual dimension attributes (see diagrams in Figs. 1, 7), and neuronal responses to naturalistic textures (i.e., original and contrast-reversed) were compared between the groups (coarse vs fine, directional vs nondirectional, regular vs irregular). Additionally, we assessed the selectivity for naturalness by comparing each neuron’s mean responses to natural textures (i.e., original and contrast-reversed) versus their spectrally matched noise counterparts. For each single unit and stimulus condition, we extracted spike trains aligned to stimulus onset using 1 ms time bins and constructed peristimulus time histograms (PSTHs) by averaging responses across repeated trials and smoothing with a Gaussian kernel (σ = 5 ms). These PSTHs were then used to examine response dynamics over time. At each time point, we applied the Mann–Whitney U test to assess when responses to high- and low-value stimuli along each texture dimension significantly diverged. Alternatively, some analyses relied on mean responses within defined time windows (see legends of Figs. 7 and 8), rather than on responses at each time point.
Time course of selectivity index
To assess the temporal dynamics of neuronal selectivity for texture and shape, we performed a sliding-window receiver operating characteristic (ROC) curve analysis. This approach compared responses to textures grouped by high and low values along specific feature dimensions (e.g., coarseness, directionality, regularity, and naturalness), as well as to preferred (top 50%) versus nonpreferred (bottom 50%) shapes. For each neuron, and at each time point (in 1 ms steps), we quantified selectivity by calculating the area under the ROC curve (AUC) based on the distributions of PSTH values for the two comparison groups—either high versus low texture values or preferred versus nonpreferred shapes. AUC values range from 0.5 to 1, where 0.5 indicates no discriminability between conditions, values above 0.5 reflect greater responses to preferred stimuli.
LFP power analysis for cortical depth estimation
LFP power analysis was conducted using a 500 ms time window from −100 to +400 ms relative to visual stimulus onset. Signals
were bandpass filtered using a fourth-order Butterworth filter (0.3–250 Hz) to remove slow drifts and high-frequency noise.
Baseline correction was performed by subtracting the mean voltage during the 100 ms prestimulus period. For each channel and
trial, power spectra were computed using the periodogram method from Python's SciPy library (signal.periodogram), and then
trial-averaged power was calculated. To obtain the relative power map for each probe, power values were normalized across
channels at each frequency as follows (Mendoza-Halliday et al., 2024):<mml:math display="block"><mml:mrow><mml:mi mathvariant="normal">Relative</mml:mi><mml:mspace width="0.25em"></mml:mspace><mml:mi
mathvariant="normal">power</mml:mi></mml:mrow><mml:mspace width="0.25em"></mml:mspace><mml:mo stretchy="false">(</mml:mo><mml:mrow><mml:mi>c</mml:mi><mml:mo>,</mml:mo><mml:mi>f</mml:mi></mml:mrow><mml:mo
stretchy="false">)</mml:mo><mml:mo>=</mml:mo><mml:mstyle displaystyle="true" scriptlevel="0"><mml:mrow><mml:mfrac><mml:mrow><mml:mrow><mml:mi
mathvariant="normal">Power</mml:mi></mml:mrow><mml:mo stretchy="false">(</mml:mo><mml:mrow><mml:mi>c</mml:mi><mml:mo>,</mml:mo><mml:mi>f</mml:mi></mml:mrow><mml:mo
stretchy="false">)</mml:mo></mml:mrow><mml:mrow><mml:mrow><mml:mi mathvariant="normal">max</mml:mi></mml:mrow><mml:mo stretchy="false">[</mml:mo><mml:mrow><mml:mrow><mml:mi
mathvariant="normal">Power</mml:mi></mml:mrow><mml:mo stretchy="false">(</mml:mo><mml:mrow><mml:mn>1</mml:mn><mml:mo>:</mml:mo><mml:mi>n</mml:mi><mml:mi>C</mml:mi><mml:mi>h</mml:mi><mml:mi>a</mml:mi><mml:mi>n</mml:mi><mml:mo>,</mml:mo><mml:mi>f</mml:mi></mml:mrow><mml:mo
stretchy="false">)</mml:mo></mml:mrow><mml:mo stretchy="false">]</mml:mo></mml:mrow></mml:mfrac></mml:mrow><mml:mo>,</mml:mo></mml:mstyle></mml:math>Relativepower(c,f)=Power(c,f)max[Power(1:nChan,f)],
where c and f are channel and frequency, respectively.
From this two-dimensional matrix, we computed the average relative power within the alpha–beta (10–30 Hz) and gamma (75–150 Hz) bands as a function of channel depth. The point at which alpha–beta power minus gamma power crossed zero was interpreted as corresponding to layer 4 (Mendoza-Halliday et al., 2024). Channels with positive values were considered likely to be in superficial layers, and those with negative values in deep layers.
Experimental design and statistical analysis
Details of experimental procedure and visual stimuli are described above. For all statistical tests presented here, independent group comparisons were performed using a nonparametric Mann–Whitney U test, while the multiple-group comparison was done with one-way ANOVA. The statistical significance of the orientation energy model's goodness of fit was assessed using the F statistic. The strength of the linear relationship between pairs of variables was assessed by Pearson's correlation coefficient. A p value <0.05 was considered significant.
Results
We measured the responses of subpopulations of V2 neurons to four stimulus types—closed 2D shapes, natural textures, drifting sinusoidal gratings, and elongated object patches exhibiting translational motion (Fig. 1)—using high-density Neuropixels probes. These stimuli have been shown to drive neurons across the V1→V2→V4 hierarchy. In particular, spatial frequency × orientation gratings have been used extensively in studies of visual neurons (Hubel and Wiesel, 1968; De Valois et al., 1982), textures tap into V2's emerging naturalistic pattern selectivity (Freeman et al., 2013), and both closed 2D shapes and translational motion patches have revealed unique encoding properties in V4 not apparent in other early and midlevel visual areas (Pasupathy and Connor, 2001; Bigelow et al., 2023).
Visual stimuli and identification of the V1/V2 border. A, Texture stimuli. We utilized 40 naturalistic texture images categorized into eight groups (rows) based on three dimensions (C, coarseness; D, directionality; R, regularity; see texture groups legend) that influence human texture perception. Each dimension takes on two levels: C, coarse (orange) versus fine (blue); D, directional (orange) versus nondirectional (blue); and R, regular (orange) versus irregular (blue). For example, textures in the first row are coarse, directional, and regular, whereas those in the last row are fine, nondirectional, and irregular. Each texture was shown in three variations: the original, contrast-reversed, and spectrally matched noise. The fourth column of the texture groups legend represents the naturalness (N) dimension: the original and contrast-reversed variations are natural (orange) and spectrally matched noise is not (blue, data not shown). B, Shape stimuli. One rotation of the 2D shape set from (Pasupathy and Connor, 2001) is shown. A randomly selected subset (n = 50) from rotated versions of these stimuli (n = 366, see Materials and Methods) were used to study shape responses in 17/20 sessions; in the remaining sessions (3 out of 20), a subset of 120 stimuli was used. C, Grating stimuli. Drifting sinusoidal gratings (4 Hz, 1 s) were presented at 12 directions of motion (30° steps) and five values of spatial frequency (0.25, 0.5, 1, 2, 4 cycles/°). D, Translational motion stimuli. Elliptical noise patches with mean luminance matched to the background were displayed in sequence at seven locations centered on the RF (green circle). To ensure that the motion signal relied on non-Fourier cues, a unique noise pattern was randomly generated for each location on every trial. The long axis of the elliptical patch was oriented orthogonal to the direction of motion, which varied in increments of 45°. Five spatial displacement (dX) levels were tested while keeping motion speed constant (see Materials and Methods). The figure shown here corresponds to dX = 1/3 RF diameter. E, The exposed cortical surface (left) and imaging of ocular dominance (right) to identify V1/V2 border from one experiment. The brighter and darker stripes are right and left eye-dominant regions in area V1, respectively. V2 recordings were made posterior to the lunate sulcus but anterior to the ocular dominance bands of V1. Note that V1-V2 segregation based on the ocular dominance pattern aligns well with a transition from denser to sparser vascular patterns from V1 to V2.
Example sites: functional clusters for shape, texture, and motion stimuli
To quantify the similarity in tuning of nearby neurons, for each stimulus type—texture, shape, drifting grating, and translational motion patches (Fig. 1)—we computed the Pearson's correlation coefficient (r) between the response vectors of spiking activity of all pairs of simultaneously recorded neurons (see Materials and Methods). Figure 2 illustrates results from two example sites that reveal similar preference in tuning for nearby neurons across the various stimulus classes tested. In the first example (Fig. 2A–D), similar responses of nearby neurons can be visualized by horizontal streaks in the response matrices (Fig. 2A; also Fig. S1). For the drifting grating stimuli (Fig. 2A, third panel), two prominent horizontal bands of activity are evident in most of the simultaneously recorded neurons. Units 0–8 respond strongly to stimulus conditions 20 and 50 (motion direction 120 and 300°, green arrows), while units 11–37 show strong responses to stimulus conditions 5 and 35 (motion direction 30 and 210°, orange arrows). In both cases, the banding corresponds to two motion directions separated by 180°, indicating two sets of neurons with shared orientation selectivity. In the corresponding correlation matrix (Fig. 2B), which measures tuning similarity of pairs of simultaneously recorded neurons, two clusters with high similarity in tuning (red) are evident, aligning with the neurons’ orientation tuning profiles (Fig. 2A) and their spatial distribution along the probe (Fig. 2C).
Example recording sessions. A–D, Data from site 22 for texture, shape, drifting grating, and translational motion stimuli (columns) are shown. A, Each panel shows mean normalized activity (in the 0–1 range, see scale bar) of simultaneously recorded neurons (x-axis) across stimulus conditions (c-axis; see Fig. S1 for details of the normalization procedure). Texture includes 120 conditions (0–39, original; 40–79, contrast-reversed; 80–119, spectrally matched noise). Shape has 50 conditions (0–49). Drifting grating comprises 60 conditions (5 spatial frequencies × 12 directions spaced 30° apart, with 0–4 being spatial frequencies at direction 0°). Translational motion includes 40 conditions (5 spatial displacements × 8 directions spaced 45° apart, with 0–4 being displacements at direction 0°). A horizontal stripe pattern (green and orange arrows in drifting panel) indicates that multiple nearby neurons exhibit similar visual feature tuning. B, Pairwise tuning similarity between neurons is shown, with unit IDs on both axes. Color scale indicates the strength and sign of similarity, where red and blue represent positive and negative values for the Pearson’s correlation coefficient (r). Diagonal entries were replaced with NaNs and displayed in gray. C, The x-axis represents unit IDs, and the y-axis values correspond to the distance from unit# 0, defined as the neuron recorded closest to the probe tip. Neurons with similar tuning were organized into groups using clustering analysis (see Materials and Methods). The four panels display the same neurons, but colors represent different functional clusters for each stimulus type: orange marks the largest cluster. Neurons that do not belong to any cluster are represented in gray. D, Scatterplots comparing responses of a pair of example neurons. Data are the same as depicted in the corresponding columns of the normalized response matrix in A. These two units exhibited a positive correlation for texture responses but not for other stimulus classes (see pairwise tuning similarity matrix in B). E–H, Results from site 09. The conventions are the same as those in A–D.
We observed similarly clustered tuning preferences for shape and translational motion patch stimuli (Fig. 2B, second and fourth panels). With texture stimuli, we observed a notably different profile: tuning similarity was consistently high across the entire length of the probe even for neurons #0 and #43 separated by 1.6 mm (Fig. 2B, first panel). For each stimulus type, we used an ad hoc clustering algorithm to group neurons into clusters based on their tuning similarity (Fig. 2C; also see Materials and Methods). In the case of texture stimuli, nearly all cells exhibited similar tuning and were classified as a single group. In contrast, for shape, drifting grating, and translational motion stimuli, at least two distinct groups emerged with different tuning characteristics. Figure 2D illustrates the responses of two example units (#0 vs #35) that belong to the same group for texture selectivity but assigned to different groups for other stimulus types; their responses are positively correlated for texture stimuli but negatively correlated for all other stimulus types. Figure 2, E–H, shows similar overall trends from another recording site. Together, these findings suggest that in V2, neurons exhibit clustered selectivity for a variety of stimulus classes, with texture stimuli leading to a more widespread, coordinated response across neurons.
Population results
We quantified the prevalence of functional clusters for different visual stimulus types using data from three macaque monkeys: 24 sites (1,489 neurons) for texture, 20 sites (1,138 neurons) for shape, 7 sites (600 neurons) for drifting grating, and 12 sites (802 neurons) for translational motion patch stimuli. Studied neurons had RF eccentricities largely within the central 10° of the visual field (mean ± SD: 6.87 ± 2.07°; range: 2–12°; see also Fig. S2A,B). At the population level, we observed functional clusters in response to all stimulus classes; however, as with the examples in Figure 2, texture stimuli consistently evoked more extensive clustering than other stimulus types (Fig. 3A). To explore this further, we examined how response similarity varies as a function of interneuron distance for each stimulus type (Fig. 3B). For texture responses, mean similarity between neurons decreased gradually with increasing distance, remaining substantially high (mean Z-transformed r = 0.22) even at 2 mm. In contrast, for shape, drifting grating, and translational motion patch stimuli, similarity was markedly lower over the same distance. Consequently, at neuronal distances in the 1–2 mm range, mean similarity values for these stimuli were 0.04, 0.09, and 0.06, respectively. Notably, for the drifting and translational motion stimuli, tuning similarity did not decay linearly; instead, it exhibited a sharp initial decay followed by a plateau at larger distances (typically beyond 500 μm). This plateau pattern was not apparent for texture or shape stimuli, which showed a more continuous decline.
Population data: Neuronal clusters in area V2. A, Results from clustering analysis (see Materials and Methods) across 24 recording sites (x-axis) for the four stimulus classes. In each recording site, individual neurons are represented by dots positioned at the depths of the corresponding contact along the probe. A depth of 0 corresponds to the location of the deepest neuron recorded in each session. Distinct clusters of neurons with similar tuning are color-coded, while neurons not assigned to any cluster appear in gray. Orange indicates the largest cluster within each site, followed by green for the second largest, and red for the smallest. Note that colors are meaningful only within a recording site; the same colors across different sites do not indicate shared feature selectivity. Sites 22 and 09 correspond to the example data shown in Figure 2C,G. B, Dependence of tuning similarity on interneuronal distance. In each panel, thin gray lines correspond to individual recording sites, with each data point representing the mean value of tuning similarity between pairs of neurons from 200 μm bins positioned every 100 μm. The black line indicates the average value across recording sites. Blue lines indicate the baseline correlation derived from shuffled neuronal pairs across different penetrations using identical stimuli. Such a baseline correction is not applicable to the shape condition, because stimulus sets were randomly selected across sessions. C, Baseline-subtracted version of the data shown in B. Red lines highlight two specific penetrations (sites 09 and 10) where RF shifts along the probe were minimal (see top panels in D). D, Shift in RF position along the length of the probe. For four representative penetrations, RF eccentricity (x-axis) is plotted against the distance along the probe (y-axis). E, We quantified the relationship between functional similarity and spatial displacements of RF by measuring the mean Z-transformed correlation across all neuronal pairs recorded within a penetration against the change in RF eccentricity (from 0.1 to 0.9 percentile) across neurons recorded along the length of the probe. We found that penetrations with larger RF shifts tended to exhibit lower functional similarity. Colors correspond to penetrations identified in D.
One possible explanation for the larger clusters observed in response to textures is that some of the texture stimuli may not have driven the neural activity strongly; this could induce uniformly low responses for a subset of stimuli across the population, thereby inflating tuning similarity. In our recordings, most neurons were indeed more strongly modulated by coarse textures (associated with lower spatial frequencies) than by fine textures and by natural textures rather than noise textures (see below, Neuronal clusters encoding different perceptual texture dimensions). To address this potential explanation, we performed three analyses. First, we asked whether larger neuronal clusters for texture were observable even when responses to fine or noise textures were excluded from the response similarity calculations. Figure S3, A–C, pairs of neurons across different recording sites, reasoning that if response similarity is inherent to our stimulus design, the strong similarity will remain between responses of randomly paired neurons. Across 10,000 random pairs, the mean similarity based on texture stimuli was 0.18 ± 0.002 (mean ± SEM; Fisher Z-transformed baseline correlation; see Fig. 3B, top panel, blue line)—nearly reaching the values observed between neurons separated by 2 mm within the same recording site. Thirdly, in addition to analyses focused on texture stimuli, comparative analyses involving shape stimuli confirmed that driven firing rates and response variability were similar between texture and shape stimuli (Fig. S3D–E). These findings indicate that the choice of texture stimuli contributes to a higher baseline of response similarity. Subtracting this baseline reduces the initial difference between texture and other stimulus categories (Fig. 3C). However, the residual clustering for texture remains robust in magnitude, and our results reflect a genuine functional organization that cannot be explained solely by stimulus-intrinsic factors or limited drive by a subset of stimuli.
A limitation of the distance-based analysis described above is that lateral and interlaminar separations between neurons were treated as equivalent. This arises because interneuron distance was defined as the distance between recording channels along a linear probe, which was inserted at varying angles relative to the cortical surface across recording sites. Although we could not determine the exact numerical angle for each penetration (i.e., recording site), we used RF alignment across layers as a proxy for the probe trajectory and then analyzed the relationship between the magnitude of RF shift and the functional similarity of neurons recorded simultaneously from the same penetration (Fig. 3D,E). We found a clear negative relationship: penetrations with larger RF shifts (indicating a trajectory more parallel to the layers) showed significantly lower functional similarity among neighboring neurons. Conversely, when the RF shift was minimal (indicating a trajectory more orthogonal to the layers), functional similarity remained high. However, even in these orthogonal cases, we still found that similarity declined as the distance between neurons increased (Fig. 3C, red lines). This confirms that distance-dependent changes occur interlaminarly, though they are less pronounced than the changes seen across the lateral extent of the cortex.
Our results in V2 provide a striking contrast to the sparse functional clustering we recently observed in V4 with the same shape and texture stimuli used here (Namima et al., 2025). To directly compare the results from V4 and V2, we calculated, for each neuron, the mean functional correlation (Z-transformed, baseline corrected) among neuron pairs separated by <500 μm. We found that ∼66% of V2 neurons exhibited statistically significant functional correlations within this distance (p < 0.05, based on bootstrapping tests; Fig. 4). This proportion was similar for texture and shape stimuli. In contrast, our parallel analyses of area V4 revealed significant correlations in only 35% of neurons. Furthermore, the magnitude of these correlations was substantially larger in V2 than in V4. For textures versus shapes, the median Fisher's Z-transformed correlation values in V2 were 0.34 versus 0.10 prior to correction (0.17 vs 0.10 following correction), whereas in V4, those values were only 0.08 versus 0.03 (0.04 vs 0.03 following correction). Taken together, these reductions in both the proportion of correlated neurons and correlation magnitude suggest a substantial change in functional organization between V2 and V4.
Comparison between V2 and V4. A, B, We assessed the local functional organization for texture stimuli by calculating, for each neuron, the average baseline-corrected Z-transformed correlation with all neuronal pairs within a 500 µm distance. The same color scale is used for both top (V2) and bottom (V4). B, Histogram of local texture-tuning similarity. Histograms show the baseline-corrected Z-transformed correlations within 500 µm for areas V2 and V4. Black bars denote neurons with statistically significant positive correlations. The median value for V2 (0.17, red line) is significantly higher than that for V4 (0.04) (Mann–Whitney U test, p < 0.001). C, D, Same comparison analyses applied to shape tuning similarity in area V2 (top) and area V4 (bottom). Baseline correction was not applicable to the shape condition, because stimulus sets were randomly selected across sessions. The median value for V2 (0.10) is significantly higher than that for V4 (0.03) (Mann–Whitney U test, p < 0.001). Note that for both shape and texture stimulus types, tuning similarity of neurons in V2 is more spatially extensive than in V4, as reflected by higher correlation values within 500 µm. In A–D, panels are based on reanalysis of previously published V4 data (Namima et al., 2025).
Dissecting the representational bases of clustered encoding in V2
Prior studies in V2 have demonstrated selectivity for basic visual features such as local orientation, spatial frequency, and motion direction (Foster et al., 1985; Levitt et al., 1994; Gegenfurtner et al., 1996; Felleman et al., 2015), as well as a preference for naturalistic textures over spectrally matched noise (Freeman et al., 2013; Ziemba et al., 2016; Okazawa et al., 2017). Optical imaging has further revealed columnar organization for orientation and motion direction (Lu et al., 2010; Hu et al., 2018). Here, we examine whether the selectivity and spatial clustering observed in response to our more complex stimuli can be accounted for by tuning and clustering based on lower-level visual features. Specifically (1) we asked how orientation energy contributes to the clustering of shape responses, (2) we compared direction and orientation selectivity for drifting versus translational motion, and (3) we characterized texture selectivity of neuronal clusters in terms of human perceptual dimensions.
Clustering of shape responses in V2 can be explained by a weighted sum of orientation energy
Area V2 is known to encode combinations of orientations (Ito and Komatsu, 2004; Anzai et al., 2007), potentially supporting the representation of more complex visual features such as angles, contours, or textures. We sought to investigate whether the shape tuning of V2 neurons, and their clustered shape preferences, could be described in terms of weighted combinations of V1-like oriented filter outputs.
For each neuron, we determined the weights for four orientation-tuned filters to best fit the shape responses using linear regression (see Materials and Methods). Our analysis showed that the orientation energy model provided a statistically significant fit to the shape responses in approximately half of the V2 neurons tested with shape stimuli (517 out of 1,138 neurons, 45.4%; Fig. 5A,B). Furthermore, we observed a strong correspondence between similarity matrices derived from orientation weight vectors (four filters) and those based on the full set of shape responses (50 shapes; Fig. 5C). This finding suggests that for a substantial portion of V2 neurons, shape selectivity and clustering based on shape responses can be largely explained by a linear combination of V1-like oriented filter outputs. When the same analysis was applied to V4, ∼35% of neurons (804 out of 2,295) exhibited a statistically significant fit, with the overall explained variance being significantly lower than in V2 (Mann–Whitney U test, p < 0.001). However, since the number of shapes (n = 50) used was limited, a more precise comparison between the two areas would require a larger and more diverse set of stimuli to fully capture the differences in shape selectivity and underlying neural coding mechanisms (Oleskiw et al., 2014).
Orientation energy model fit to shape responses. A, B, For each neuron, shape responses were predicted using an orientation energy model based on linear regression. Black dots in A and black bars in B represent neurons with statistically significant goodness-of-fit (r) as determined by an F test (p < 0.05), while gray dots and bars represent neurons that did not reach statistical significance. In most sessions, we used a shape set of 50 stimuli, but in a few sessions (3 out of 20), a larger set (n = 120) was used, which led to statistical significance even with smaller correlation values (r < 0.4). The vertical line represents the median (0.40). C, A comparison of the tuning similarity matrices derived from measured (top) versus predicted (bottom) responses from the orientation energy model. Matrices exhibit qualitatively similar structure.
Clustering of responses to motion stimuli: contributions of surface and object-based motion computations
Previous studies have predominantly used drifting gratings and fields of dynamic dots to probe motion processing in visual cortex (Britten et al., 1992; Snowden et al., 1992; Livingstone, 1998; Gur et al., 2005). These surface motion stimuli, where a visual pattern drifts within a static aperture, have been valuable for measuring the selectivity and clustering of direction selective neurons in V1, V2, and MT (Albright et al., 1984; Hawken et al., 1988; Lu et al., 2010). Recently, using translational motion stimuli, we identified a subgroup of neurons in V4 that encode motion direction only when defined by the displacement of a stimulus patch across the RF. These “object motion” sensitive neurons encode motion over a large range of spatial displacements even when defined by non-Fourier cues (Bigelow et al., 2023) unlike neurons in the quintessential motion area MT (Churchland et al., 2005; Hedges et al., 2011). Motivated by these findings, we used the same stimuli (Fig. 1C,D) to examine the clustering of motion responses in V2 and identify whether area V2 shows distinctive sensitivities for drifting and translational motion.
Across our dataset, ∼44% (179 out of 408) of orientation-tuned V2 neurons were direction selective in response to drifting motion stimuli, and a smaller proportion (∼29%; 141 out of 486) were direction selective for translational motion stimuli for the largest displacement level (dX = RF/3) we tested (Fig. 6A–D). For both stimulus classes, direction selective neurons were not randomly distributed but exhibited a clustered organization within specific recording sites (e.g., sites 7, 9, 20; Fig. 6A,C), suggesting localized cortical domains of direction selectivity. This clustering occurred even though drifting and translating stimuli were defined by Fourier and non-Fourier cues, respectively. Overall, clustering was more prominent with drifting than with translational motion, but in both cases, neurons tuned to two opposite directions (i.e., 180° apart) were intermixed (Fig. 6I).
Comparing direction selectivity for drifting and translational motion. A, Individual neurons from seven recording sites tested with drifting motion stimuli are shown as dots and colors denote orientation-selective neurons (orange), direction selective neurons (green), and neither (gray). B, Histogram of the direction selective index (DSI) for orientation and direction selective neurons (using the same color scheme as A) computed at the optimal spatial frequency determined for each neuron. C, D, Results from 12 recording sites (the seven sites shown in A plus five others) tested with translational motion stimuli are shown using the same convention as A, B. DSI values were computed at the largest spatial displacement (dX = RF/3). E, Normalized response matrix for drifting grating stimuli (60 conditions = 12 directions × 5 spatial frequencies) from one session (site 20), following the same convention as in Figure 2A. A horizontal stripe pattern indicates that multiple nearby neurons exhibit similar visual feature tuning. F, Direction tuning curves from two example neurons from the session in E. The corresponding neurons are marked with the same colors in panel E. Error bars represent the standard error of the mean. G, Normalized response matrix for translational motion stimuli for the same session in panel E. H, Tuning curves for translational motion stimuli for the same two neurons in F. Red/blue tuning curve shapes are similar (compare panels F and H), but direction selectivity for drifting motion differs from that for translational motion. I, The optimal direction (color) and DSI value (arrow length) for each stimulus type and each neuron from three recording sites (sites 07, 09, 20) are shown. Clustered preferences for direction and orientation can be observed. J, DSI for drifting grating stimuli (x-axis) and translational motion stimuli (y-axis). The data revealed no correlation between the two measures (r = −0.02, p = 0.96). Red dotted lines indicate median DSI values for translational (0.304) and drifting (0.423) motion, respectively. K, Histogram of peak difference between drifting and translational motion tuning curves across 308 neurons that exhibited orientation selectivity for both stimulus types. The histogram reveals a bimodal distribution, indicating that the difference in preferred direction is typically either near 0 or 180°.
For individual neurons, direction tuning curves with drifting and translational motion stimuli were similar in overall shape (Fig. S4A, population data), although the peak and strength often differed. For example, in Figure 6E–H, the neuron shown in blue exhibited stronger direction selectivity with drifting motion stimuli; the neuron in red preferred opposite directions (90° and 270°) when tested with drifting versus translational motion stimuli. Across the population, motion direction selectivity was stronger for drifting than translational motion (Fig. 6I,J) and the difference in peak showed a bimodal distribution, i.e., tuning for drifting and translational motion were in the same or opposite directions (Fig. 6K). These results suggest that direction selectivity for drifting and translational motion may arise from differential modulation of an underlying orientation tuning curve. Unlike V4, we found no evidence in V2 for a distinct object-based motion processing system that is sensitive to translational but not drifting motion (Fig. S4A,B; Bigelow et al., 2023). V2 is also unlike MT, which is insensitive to non-Fourier motion cues (O’Keefe and Movshon, 1998) and exhibits a loss of direction selectivity for translational motion when spatial displacement is large (Fig. S4C,D; Mikami et al., 1986; Churchland et al., 2005). With respect to the translation motion stimuli used here, it is possible that even in the absence of traditional “Fourier cues,” high-contrast elongated noise stimulus patch may activate orientation-selective filters. Nevertheless, the distinct direction tuning properties we observed in V2 suggest that the underlying neural mechanisms for processing motion direction from these two types of stimuli are not identical. Overall, our results support the hypothesis that V2 exhibits clustered selectivity for surface motion defined by Fourier and non-Fourier cues over a broad range of spatial displacements.
Neuronal clusters encoding different perceptual texture dimensions
Previous studies have demonstrated that responses of V2 neurons can be modeled in terms of higher-order texture statistics (Okazawa et al., 2017), but it remains unclear whether this selectivity can be explained in terms of tuning to specific perceptually defined dimensions of texture as observed in V4 (Kim et al., 2022) and whether spatial clustering exists for these dimensions. To address this, here we examined the selectivity of individual neurons along four dimensions—the perceptual dimensions of coarseness, directionality, and regularity (Kim et al., 2022), and a naturalness metric (see Materials and Methods)—and asked how this preference compared across neurons within a recording site. Results from one recording site are illustrated in Figure 7A (the same recording site as Fig. 2A). Across recorded neurons, strong responses represented by dark horizontal strips in the response matrix (Fig. 7A, left) mirror the stimulus grouping along the coarseness dimension (Fig. 7A, right). Responses were also stronger to both the original (top) and contrast-reversed (middle) naturalistic textures than to spectrally matched noise counterparts (bottom). Temporal dynamics (Fig. 7B) also reveal a preference for coarse, irregular, and natural textures across most neurons in this recording site.
Neuronal clusters with different texture selectivity. A, Normalized response matrix (x-axis, unit ID; y-axis, texture ID) for an example session and the temporal response profile along the four texture dimensions (right). Responses to three texture variations are shown: original (top), contrast-reversed (middle), and spectral noise (bottom). The texture dimensions, C, D, and R follow definitions in Figure 1. The naturalness dimension (N) takes on two values: original and contrast-reversed are natural (orange) while spectral noise is not (blue). In the spectral noise condition, C, D, and R dimensions are marked in green and black rather than orange and blue, since they were not included in the computations shown in B. B, For each neuron, time points with statistically significant differences in mean responses between the two levels along each texture dimension (e.g., coarse vs fine) were identified and marked as orange or blue corresponding to the greater value (see Materials and Methods). For C, D, and R dimensions, only responses to naturalistic textures were used in comparisons. Note that nearly all neurons exhibit similar texture selectivity associated with a preference for coarse, irregular, and naturalistic features. C, D, The same analyses as in A, B were performed for site 10. Nearly all neurons demonstrate similar texture selectivity, showing a preference for coarse, nondirectional, and naturalistic features. E, F, Results from site 00. Nearly all neurons demonstrate similar texture selectivity, showing a preference for coarse, directional, and naturalistic features. G, H, Results from site 20. Clusters of neurons with slightly different preferences for directionality and regularity were identified across different depths, but overall texture selectivity remained similar. I, Across all recording sites, the proportions of neurons significantly modulated by coarseness (left) were calculated from mean responses within the 0–400 ms poststimulus window, and their positions were mapped along the probe (right). Gray dots indicate neurons with no preference. J, L, The same analyses as in I for directionality (J), regularity (K), and naturalness (L) features.
A preference for coarse and naturalistic texture features was commonly observed across recording sites (Fig. 7I,L). However, selectivity for other texture attributes varied substantially across individual neurons and sites. For example, neurons recorded at site 10 showed a preference for coarse, nondirectional, and naturalistic textures with minimal influence from regularity (Fig. 7C,D). Data from other example sites are shown in Figure 7E,F,G,H, further demonstrating that V2 contains specialized subpopulations tuned to different texture attributes, while maintaining a general bias toward naturalistic features (for more detailed examples of single unit activity, see Fig. S5).
Distinct temporal dynamic patterns are differently distributed across cortical layers
In previous sections, our analyses of functional clustering in V2 neurons were based on time-averaged responses without considering their temporal dynamics. To investigate whether incorporating temporal dynamics reveals additional structure, we analyzed the evolution of feature selectivity over time and assessed whether distinct temporal dynamics are systematically organized across layers.
As a first step, we clustered neurons based on the similarity of their temporal response profiles (PSTHs). Our analysis used a K-means clustering approach and was constrained to a 300 ms window (−100 to 200 ms) around stimulus onset. This identified three clusters with distinct response dynamics. In Figure 8A, neurons in the black cluster exhibited a short-latency (∼65 ms) transient peak that rapidly decayed. In contrast, neurons in the blue cluster showed a slightly delayed peak at ∼90 ms, and rather than decaying quickly, responses were maintained as sustained activity persisting well beyond 100 ms after stimulus onset. Lastly, neurons in the green cluster showed a slower rise in activity without a distinct peak, but responses converged with those of the blue cluster after 130 ms. Importantly, the temporal dynamics of individual neurons within each cluster were highly consistent across texture and shape PSTHs (Fig. 8A, left vs right), although peak responses were stronger for shape. Neurons exhibiting sustained activity patterns (blue, green) demonstrated significantly stronger selectivity for both shape and all texture features compared with those with transient responses (Fig. 8B, black), suggesting that these neurons may play an important role in supporting both texture and shape processing in V2.
Temporal dynamics of texture selectivity and its relationship with cortical depths. A, All recorded V2 neurons with clear visual responses were grouped into three clusters using K-means algorithm, based on the similarity of their temporal dynamics from −100 to 200 ms relative to stimulus onset. The PSTHs in the black cluster are characterized by a fast transient peak, whereas those in the blue and green clusters exhibit stronger sustained activity, persisting 100 ms after stimulus onset. B, For each PSTH cluster, selectivity indices (i.e., area under the ROC curve; see Materials and Methods) for four texture dimensions and shape are plotted as a function of time. Selectivity for coarse, naturalistic texture features and shape emerged earlier than that for directional and regular texture features. Notably, neurons with sustained responses (blue and green clusters) exhibited stronger overall selectivity. C, For each PSTH cluster, cortical layers of individual neurons were estimated by the difference of relative LFP power between gamma (50–150 Hz) and alpha–beta (10–30 Hz) frequency ranges. Positive and negative values indicate neurons located in the superficial and deep layers, respectively, with values near 0 corresponding to layer 4. D, E, The relative proportions of neurons with statistically significant modulation for each texture attribute. Two distinct time windows were analyzed to contrast early (D, 50–100 ms poststimulus onset) and later (E, 100–150 ms poststimulus onset) neuronal activity. Neurons with significant modulation are represented by orange and blue colors, with color schemes consistent with those in Figure 7.
To assess whether these neuronal groups also differed in their laminar distribution, we estimated the laminar positions of individual neurons based on the relative differences in local field potential (LFP) power spectra between the alpha/beta and gamma frequency ranges (Mendoza-Halliday et al., 2024; see Materials and Methods and Fig. S6). While both sustained and transient neurons were distributed across cortical layers, sustained neurons were more frequently found in superficial layers (Fig. 8C). In contrast, transient neurons were slightly more common around layer 4, suggesting a possible association with initial feedforward processing. These findings suggest that texture and shape processing in V2 is organized not only in time but also across cortical depth, with distinct layers contributing preferentially to early versus sustained encoding of visual features.
Our recent V4 study found that response modulation related to directionality and regularity is significantly delayed than those for coarseness and contrast (Kim et al., 2022). To determine whether similar results are observed in V2, we quantified the proportion of neurons exhibiting statistically significant modulation for each texture attribute within two poststimulus time windows: early (50–100 ms) and late (100–150 ms) phases (Fig. 8D,E). In the early phase, a large proportion of neurons were modulated by coarseness and naturalistic features (large orange slices), while fewer neurons responded to directionality and regularity. In the later phase, the proportion of neurons responding to coarseness and naturalistic features remained largely unchanged, but there was a marked increase in neurons sensitive to directionality and regularity (indicated by the larger blue slices). This analysis showed that selectivity for coarse and naturalistic features emerges rapidly, whereas sensitivity to directionality and regularity develops more gradually during the response. We did not observe a clear relationship between these response dynamics and cortical layers.
In summary, V2 neurons form large clusters that jointly encode multiple perceptual texture dimensions. Both texture and shape selectivity emerge gradually across time and cortical space, with sustained dynamics in superficial layers further refining the representation.
Discussion
We used high-density Neuropixels probes to study populations of single neurons and determine the functional organization and the representational bases of area V2. Our results provide the first documentation of clustered selectivity across cortical layers in V2 for higher-order form, texture, and motion stimuli. The organization we observed rivals those previously demonstrated in V1 and is strikingly different from the sparse clusters observed in V4 with similar stimuli (Namima et al., 2025). V2 clusters based on responses to texture stimuli were spatially more extensive than those based on motion stimuli, suggesting that texture information may be integrated over broader cortical networks than those processing spatially localized stimulus orientation or moving direction. In terms of representational bases, V2 encodes visual stimuli largely in terms of surface characteristics, supporting the hypothesis that object-based coding emerges first in V4 (Pasupathy et al., 2020). Overall, these results provide new insights into the representational transformations across the early and midlevel stages of visual processing, offering constraints to refine future hierarchical models of the ventral stream.
Functional organization: relationship to prior studies
V2 exhibits a well-defined modular organization characterized by repeating patterns of thin, thick, and pale stripes, identifiable histologically through cytochrome oxidase (CO) staining (Horton, 1984). Functional maps from optical imaging have demonstrated the existence of distinct functional clusters for features (e.g., orientation, color, motion) corresponding to these stripes (Ts’o et al., 1990; Lu and Roe, 2008; Felleman et al., 2015). These approaches focused on properties across the cortical surface and neuronal subpopulations at a coarse level (although a few studies achieved finer resolution; see Hubel and Livingstone, 1985; Shipp and Zeki, 2002). Our study expands this understanding by demonstrating robust clustering at the single-neuron scale across cortical layers for higher-order stimulus dimensions.
With regard to the size of clusters, we found that similarly tuned neurons spanned ∼500 μm along the length of the probe for shape, drifting grating, and translational motion patch stimuli. We may not have seen larger clusters encompassing all cortical layers, because our probes were often inserted at an oblique angle traversing multiple orientation domains (Figs 2, 3, 6). The number of columns crossed was variable across penetrations and dependent on the penetration angle (ranging from perpendicular to parallel to the cortical layers). With texture stimuli, we observed a trend toward broader functional clusters, with response similarity decreasing more gradually and remaining relatively high even beyond 1 mm. This higher correlation was observed even when we subtracted the baseline correlation to address stimulus-driven effects—preference for naturalistic and coarse stimuli across the V2 population, and differences in local contrast across stimuli. Larger clusters based on texture responses may arise because texture selectivity is built by integrating higher-order correlations of stimulus features across space. Indeed, orthogonal coding of orientation and spatial frequency has been reported (Zhang et al., 2023). The texture clusters we identified, which prefer coarse features and extend across multiple orientation columns, may in part reflect the fact that neuronal groups tuned to low spatial frequencies necessarily encompass neurons with diverse orientation preferences. However, weaker responses to spectrally matched noise textures with identical spatial frequency content and varied preference for higher-order features (e.g., directionality, regularity) across recording sites indicate that spatial frequency alone is insufficient to explain these effects.
Our current results are strikingly different from our previous work in V4, where neurons tuned to specific shapes and textures formed very sparse clusters; clearer clustering was observable only at a coarser stimulus category level (e.g., horizontal versus vertical orientations) and at the population level, as revealed with optical imaging (Tanigawa et al., 2010; Li et al., 2013) or by pooling activity over spatial scales of hundreds of microns (Namima et al., 2025). Thus, our results highlight a dramatic shift in modular structure from V2 to V4.
Columnar organization in the context of area size and feature complexity
Our findings are pertinent to the longstanding debate on cortical columns as fundamental units of computation (Horton and Adams, 2005). The contrasting columnar structure we observed in V2 and V4, when viewed alongside the differences between these areas in terms of their size, inputs, and physiological properties, sheds light on when columns might arise. V2 is the second largest visual area in the macaque brain, ∼80% the size of V1 (Vanni et al., 2020). V4 is much more compressed, approximately half the size of V2, yet it is considered to be the most interconnected visual cortical area in the macaque brain (Felleman and Van Essen, 1991). Functionally, V2 exhibits tuning for spatially homogeneous texture patterns, which may be computed by integrating basic features extracted from V1 (Freeman et al., 2013) and pooling neurons with similar feature preferences over spatial locations. In contrast, V4 processes more complex combinations of visual features, such as shape, texture, and color (Bushnell and Pasupathy, 2012; Okazawa et al., 2017; Kim et al., 2019) to build object-based representations (Pasupathy et al., 2020). The higher dimensionality of these representations, resulting from a convergence of diverse inputs onto V4, may require a more distributed coding strategy, leading to weaker spatial clustering (Purves et al., 1992). Thus, orientation-based columns in V1 are largely maintained in similarly sized V2. But a massive representational compression from V2 to V4, due to V4’s smaller size, combined with high input convergence onto V4 may prove to be unconducive for column formation.
Properties of V2 encoding: implications for future studies
Prior studies have shown that V2 neurons encode combinations of orientations, such as angles and junctions (Ito and Komatsu, 2004; Anzai et al., 2007), possibly through the integration of inputs from V1 neurons with different orientation preferences. Consistent with this hypothesis, we found that V2 responses to shape stimuli could be largely modeled in terms of image characteristics as weighted combinations of orientation filter responses (Fig. 5). This is quite unlike V4, where responses to shape stimuli cannot be explained in terms of a spectral RF model (Oleskiw et al., 2014). Instead, V4 responses reflect the characteristics of a shape boundary in object-centered coordinates. While future studies should confirm these findings with an expanded set of stimuli, our results indicate that a critical transformation may occur in the ventral visual pathway, where representations based on locally defined orientation combinations in V2 are converted to globally defined, object-centered representations in V4 (Pasupathy and Connor, 2001).
With regard to motion direction selectivity, neurons in both V2 and V4 appear to encode information about drifting and translational motion. However, while V4 shows distinct tuning between these two motion types (Bigelow et al., 2023), V2 exhibits a moderate correlation between the tuning curves. This raises the possibility that tuning for both motion types may arise from surface-based encoding in V2 (Adelson, 2001), whereas in V4, translational motion tuning may relate to object-based encoding (Pasupathy et al., 2020). Future studies could explore these V2/V4 differences in motion direction encoding and the underlying circuitry further. We also found that motion direction selective neurons in V2 were not evenly distributed across recording sites but were instead concentrated at specific locations (Fig. 6). This may relate to prior work linking V2 thick stripes to motion processing (Sincich and Horton, 2005; Lu et al., 2010).
Our analysis of temporal dynamics revealed that V2 neurons with sustained responses display stronger selectivity for both shape and texture features and are mainly distributed in superficial cortical layers. This laminar-specific pattern resonates with prior work suggesting that preference for naturalistic texture arises de novo in V2 (Ziemba et al., 2019). Future studies combining high-resolution anatomical techniques with neurophysiological recordings will be needed to study this process more precisely.
One limitation in directly comparing our current V2 results with previous V4 findings is the difference in experimental conditions. Specifically, the V4 data were collected from awake animals with a median of five repetitions per condition, while the current V2 data were obtained under anesthesia with 10 repetitions per condition. We consider it unlikely that anesthesia fundamentally alters the spatial clustering of similarly tuned neurons, but it might influence the magnitude, synchronization, and temporal dynamics of neuronal responses (Lee et al., 2021). Future studies using matched stimuli, recording methods, and animal states are needed to enable more rigorous comparisons across visual areas.
Conclusion
Studies over decades have investigated the functional organization and the representational bases of various stages along the visual hierarchy, typically focusing on one area at a time with customized stimuli. Here, by comparing two successive visual areas with identical stimuli and recording methods, our results identify distinct properties of areas V2 and V4, reveal general insights into why columns might arise in the visual cortex, and offer constrains to improve future models of the ventral stream.
Footnotes
- Received October 23, 2025.
- Revision received February 24, 2026.
- Accepted February 25, 2026.
-
We thank the lab members for providing helpful discussions and comments on the manuscript, and the WaNBRC veterinary staff and Instrumentation services for their support. This work was supported by National Eye Institutes grant R01 EY018839, EY029601 to A.P.; National Institute of Neurological Disorders and Stroke Grant U01 NS131810 to A.P. and W.B.; NEI Center Core Grant for Vision Research P30 EY01730 to the U.W.; National Institutes of Health/Office of Research Infrastructure Programs Grant P51 OD010425 to the WaNBRC.
-
The authors declare no competing financial interests.
-
This paper contains supplemental material available at: https://doi.org/10.1523/JNEUROSCI.1994-25.2026
- Correspondence should be addressed to Taekjun Kim at taekjun{at}uw.edu or Anitha Pasupathy at pasupat{at}u.washington.edu.
- Copyright © 2026 the authors
