The logic of recurrent circuits in the primary visual cortex

Visual perception emerges from coordinated activity across the visual hierarchy, where recurrent circuits at each stage transform sensory representations. In primary visual cortex (V1), prior work suggests recurrent excitation can amplify weak signals to aid detection, whereas recurrent inhibition suppresses strong signals to improve discrimination. Disentangling recurrent influences from feedforward and feedback inputs is difficult, as all operate simultaneously in vivo. Prior studies isolated components by silencing cortex to measure thalamic feedforward input, or cooling higher areas to remove feedback. Here, the authors used targeted, high-resolution two-photon holographic optogenetics and simultaneous calcium imaging to causally probe the functional logic of local recurrent dynamics in layer 2/3 (L2/3) of mouse V1 in the absence of visual input. They hypothesized that recurrent network effects depend jointly on two organizing axes: physical space (distance-dependent connectivity) and feature space (orientation-dependent like-to-like connectivity). By designing photostimulation protocols that vary both the spatial distribution and orientation tuning of stimulated ensembles, they aimed to reveal when recurrence produces amplification versus suppression and to derive simple rules explaining the net effect of recurrent activity on nearby and distant neurons.

- Spatial dependence: E and I connectivity in L2/3 falls with physical distance, with most connections within ~200 µm. - Feature dependence: E-to-E connectivity is biased toward neurons with similar stimulus preferences (like-to-like). Many models consider spatial or feature wiring but rarely their interaction. - Prior perturbation studies: Single-neuron optogenetic perturbations in V1 reported strong like-to-like suppression among cotuned neurons, suggesting competition. Studies stimulating larger cotuned ensembles found like-to-like activation, but did not systematically manipulate spatial arrangement together with tuning. - Modeling predictions: To obtain strong like-to-like suppression, models required strong and specific E→I connections. Other work predicted network state or stimulus contrast could shift the net effect from suppression to activation. - Ensemble vs single-cell: Multicell photostimulation reveals diverse functional interactions; the impact of ensembles can differ qualitatively from single-cell effects. A gap remains in understanding how spatial arrangement and feature tuning jointly determine recurrent network responses.

Experimental approach - Species and expression: Adult mice (2–12 months) expressing GCaMP6s in excitatory neurons (tetO-GCaMP6s × Camk2a-tTA). Opsins (ChroME or ChroME2s) were expressed in excitatory neurons via Cre- or tTA-dependent AAVs; mRuby3 (nuclear) aided targeting. Specificity was validated by RNAscope. - Surgery and preparation: Headplate implantation and cranial window over left V1. AAV injections (200–300 nL) targeted V1. - Imaging and stimulation: Combined setup for 3D-SHOT two-photon holographic optogenetics and two-photon calcium imaging. Imaging at 920 nm across three planes (800×800 µm, planes spaced 30 µm) at 5.2–6.2 Hz. Photostimulation with temporally focused holograms generated by an SLM; stimulation laser synchronized to imaging to minimize artifacts. - 3D calibration: Automated multiplexed 3D calibration mapped imaging and stimulation volumes, correcting field curvature and registering hologram positions and diffraction efficiencies. Achieved high-quality optical and physiological point-spread functions (PPSF), enabling power compensation across 3D space and off-target exclusion criteria. - Photostimulation protocol: - Power test per cell to determine minimal ‘stimmable’ power, then scaled by 1.1–1.2× to ensure reliable spiking (~1 spike per 5 ms pulse). - Multitarget holograms (≥3 targets) used to stimulate ensembles; typical ensembles included ten excitatory L2/3 neurons, with 10 pulses at 10 Hz (100 total pulses; also systematically varied pulses, frequency, and ensemble size in control experiments). - Off-target exclusion: Excluded neurons within 15 µm radially on the same plane or 30 µm one plane away from a targeted cell; PPSF HWHM measured radial ~7.73 ± 0.37 µm, axial ~18.51 ± 1.69 µm; additional validation in sparse opsin mice showed no off-target light effects beyond exclusion zone. - Visual tuning and ensemble design: - Mice viewed randomized drifting gratings (8 directions) to compute tuning curves online; preferred orientation (PO), orthogonal response (OO), and OSI computed. - A discrete optimizer selected ensembles (10 cells) to span conditions: spatially compact vs diffuse (mean pairwise distance threshold ~200 µm), cotuned vs untuned (ensemble OSI >0.7 and mean member OSI >0.5 for cotuned; ensemble OSI <0.3 and mean member OSI <0.5 for untuned), minimizing cell reuse and matching power constraints. - Data acquisition and processing: Online source extraction (CalmAn OnACID) for real-time tuning; offline motion correction and segmentation with suite2p. Neuropil subtraction with cell-specific coefficient; ΔF/F computed relative to a moving F0 baseline. Distances computed from targeted coordinates. - Trial and data exclusion: Excluded trials with running >6 cm/s, ≥50% targeted cells failing to respond, or motion >4.7 µm. Excluded cells in off-target zones, recently stimulated, occluded by artifact, or not identified as cells. Ensembles excluded if too many targets unmatched in suite2p, high trial failure, or insufficient repetitions. FOVs excluded if low visual responsiveness, high running fraction, or too few detected cells. - Experimental conditions varied: Number of pulses (1–50 per cell), stimulation rates (3–50 Hz), and ensemble sizes (3, 10, 33 cells) while holding total pulses constant to test how input structure affects net recurrent effects. - Statistics: Predominantly nonparametric two-sided tests (except noted). Wilcoxon tests, ANOVA for parameter variations, Bartlett test for variances. Hierarchical nature reported (ensembles within FOVs within mice). Computational modeling - Model class: Linearized two-dimensional neural field model (rate-based) with excitatory (E) and inhibitory (I) populations in a 1400×1400 µm periodic domain. Optogenetic input modeled as weak point perturbations at stimulated locations. Low-gain, no-visual-input state allowed linearization around a steady state and a pathway expansion up to disynaptic terms. - Pathways considered: Monosynaptic and disynaptic excitatory (E→E, E→E→E) and disynaptic inhibitory (E→I→E) contributions to Δr_e. - Connectivity structure: Separation of spatial and feature dependencies. Spatial kernels as sums of narrow and broad Gaussians (parameters informed by Rossi et al. 2020 for broad components; narrow components adjusted to match observed near-field activation and surround suppression). Feature kernels h(θ) Gaussian in orientation difference; E→E feature parameters from Rossi et al.; inhibitory feature specificity tuned to match like-to-like suppression. - Free parameters: Effective strengths of E→E (w_e) and inhibitory pathway (w_ei w_ie, w_ei w_ii), narrow spatial parameters, and inhibitory feature rules. - Fitting and validation: Spatial scales adjusted to capture experimentally observed zero-crossing distances and activation/suppression magnitudes; compared conditions with and without feature-tuned E→I connections; predicted effects of ensemble spatial spread and tuning on responses of iso-, ±45°, and orthogonally tuned nontargeted neurons.

- Net population suppression from small ensemble activation: Stimulating ten L2/3 excitatory neurons typically decreased mean population activity across nontargeted neurons (mean −0.011 ± 0.0014 ΔF/F, P = 1.0×10⁻¹⁰; 160 ensembles, 18 FOVs, 13 mice). In sparse-opsin controls and no-stimulation/no-opsin controls, no significant effects were observed. - Heterogeneity in single-cell responses: After excluding off-targets, 2.34 ± 0.09% of nontargeted cells were significantly activated and 5.58 ± 0.19% significantly suppressed (99% CI, FDR 1%). - Input magnitude, not rate or ensemble size, drives suppression: Increasing total spikes (pulses) increased suppression (ANOVA P = 1.7×10⁻⁷). Varying stimulation rate (3–50 Hz) at fixed total pulses did not change net suppression (P = 0.74). Distributing fixed total pulses across different ensemble sizes (3, 10, 33 cells) yielded similar suppression (P = 0.78); smaller ensembles produced more variable responses (Bartlett P = 0.006). - Spatial profile: Nearby activation with surround suppression. Cells <30 µm from a target were activated (0.044 ± 0.005 ΔF/F, P = 1.1×10⁻¹⁰), while cells 50–150 µm away were suppressed (−0.013 ± 0.001 ΔF/F, P = 4.0×10⁻¹⁷). Modulation remained negative beyond ~50 µm and decayed toward zero with distance. This pattern persisted with visual stimulation present. - Ensemble spatial spread shapes surround suppression: Spatially compact ensembles produced stronger surround suppression than diffuse ensembles (slope 1.3×10⁻³ ΔF/F per µm of spread, P = 1.4×10⁻⁵), while nearby activation was independent of spread (P = 0.57). Interpretation: compact ensembles likely converge onto the same inhibitory neurons, driving stronger feedback inhibition. - Feature dependence in nearby cells for cotuned ensembles: For cotuned ensembles (ensemble OSI > 0.7, member mean OSI > 0.5), nearby iso-oriented nontarget cells were activated more than orthogonally tuned cells (P = 0.0035, Wilcoxon one-sided; n = 17 ensembles). Nearby orthogonally oriented cells were strongly suppressed. - Joint effects of space and feature: - Untuned ensembles: Both compact and diffuse untuned ensembles produced canonical nearby activation and surround suppression across all nontargets. - Cotuned ensembles: Spatially diffuse cotuned ensembles produced center activation and surround suppression; spatially compact cotuned ensembles eliminated nearby activation and instead caused nearby suppression. Subdividing by orientation preference revealed compact cotuned ensembles activated nearby iso cells but suppressed other nearby orientations; diffuse cotuned ensembles showed weaker nearby suppression across orientations. All ensembles suppressed distant cells. - Modeling requirements and predictions: A linear neural field model reproduced key phenomena only when including: - Highly local like-to-like E→E connectivity (narrow spatial scale). - Feature-specific convergence of cotuned E onto local inhibitory neurons (tuned E→I), yielding strong disynaptic inhibitory recruitment. - Trade-off with spatial compression: compact cotuned ensembles enhanced both excitatory and inhibitory pathways, but suppression (E→I→E) grew more, replacing nearby activation with suppression; orthogonally tuned neighbors were preferentially suppressed via stronger inhibitory recruitment. The final model matched all core experimental findings and decomposed recurrent logic into a balance of E→E vs E→I→E pathways.

The study demonstrates that the sign and magnitude of recurrent modulation in L2/3 V1 depend jointly on the spatial arrangement and orientation tuning of the activated ensemble and of the postsynaptic neurons. While most patterns of modest ensemble activation evoke net population suppression, nearby neurons can be amplified, and this amplification is strongest for spatially diffuse ensembles and for nearby iso-oriented neighbors of cotuned ensembles. In contrast, compact cotuned ensembles preferentially recruit local inhibitory circuits, abolishing nearby activation and producing robust suppression, particularly for non-iso orientations. These findings address the central question of when recurrence amplifies versus suppresses cortical activity and reveal simple rules: space (distance and ensemble compactness) and feature (orientation match) jointly determine recurrent outcomes. The modeling shows that a combination of highly local like-to-like E→E connectivity and feature-tuned E→I convergence is necessary to reproduce the data, clarifying the relative roles of E→E amplification and E→I→E suppression. The framework reconciles divergent prior observations by highlighting differences in perturbation scale (single cell vs ensemble), ensemble geometry, and network state. It suggests that under modest, local perturbations without visual drive, V1 is in a low-gain regime where disynaptic inhibition often dominates, confining excitation to very short spatial scales and producing feature-selective suppression in the surround. These principles help explain how recurrent circuits can amplify weak, aligned inputs while enforcing competition across space and feature dimensions for stronger or more compact inputs.

This work defines a compact set of organizing principles for how recurrent circuits in V1 shape activity: - Nearby amplification with surround suppression is a generic signature of local recurrent dynamics even under modest ensemble activation. - Ensemble geometry and tuning jointly determine the balance between E→E amplification and E→I→E suppression: diffuse ensembles favor local amplification; compact cotuned ensembles recruit strong local inhibition and suppress non-iso neighbors. - A wiring rule combining highly local like-to-like E→E connectivity and feature-specific E→I convergence accounts for the observed dynamics. These insights provide a mechanistic logic for when recurrent feedback amplifies versus suppresses input, relevant for processing natural images with extended contours versus locally mixed textures. Future directions include probing faster temporal dynamics with higher-speed indicators (e.g., GCaMP8f), incorporating temporal nonlinearities into models, standardizing optogenetic perturbation parameters across studies, and extending spatial–feature continuous models toward unified frameworks that capture diverse stimulus and brain-state conditions.

- Isolation from external inputs: Experiments were performed without ongoing visual stimuli (except specific controls) to isolate recurrence; thus, evoked dynamics may differ from naturalistic conditions or stimulus-driven states. - Modest perturbations and low temporal resolution: Perturbations were small (∼10 APs in ∼10 neurons), enabling linear modeling but potentially underestimating nonlinear dynamics. Calcium imaging at ~6 Hz limits detection of rapid transients (e.g., brief excitation followed by suppression). - Parameter space and methodological constraints: 2P optogenetic perturbations have a vast parameter space (neuron selection, spacing, pulses, timing). Although off-target activation was carefully controlled and neuropil effects analyzed, residual uncertainties in exact spike timing and counts remain. - Generalizability across species/areas and states: Results are from mouse V1 L2/3 and may not directly generalize to other layers, areas, species, or behavioral states. - Modeling simplifications: Linearized, steady-state rate model with separable spatial and feature kernels omits temporal dynamics and nonlinearities; inhibitory feature rules were tuned to match data due to limited direct measurements.