henhaobio
Sign in

NatureNeurosciencePeer-reviewed journal article

Functional connectomics reveals general wiring rule in mouse visual cortex

Zhuokun DingiD, Paul G. FaheyiD, Stelios PapadopoulosiD, Eric Y. WangiD, Brendan CeliiiD, Christos PapadopoulosiD, Andersen ChangiD, Alexander B. KuniniD, Dat Tran, Jiakun FuiD, Zhiwei DingiD, Saumil Surendra PateliD, Lydia NtanavaraiD, Rachel FroebeiD, Kayla PonderiD, Taliah MuhammadiD, J. Alexander BaeiD, Ágnes L. BodoriD, Derrick BrittainiD, JoAnn BuchananiD, Daniel J. BumbargeriD, Manuel CastroiD, Erick Cobos, Sven DorkenwaldiD, Leila ElabbadyiD, , Andreas Savas ToliasiD

PDFDOIPreprintData
In-depth full-text digest available · Jump to digest ↓

Abstract

Understanding the relationship between circuit connectivity and function is crucial for uncovering how the brain implements computation. In the mouse primary visual cortex (V1), excitatory neurons with similar response properties are more likely to be synaptically connected, but previous studies have been limited to within V1, leaving much unknown about broader connectivity rules. In this study, we leverage the millimeter-scale MICrONS dataset to analyze synaptic connectivity and functional properties of individual neurons across cortical layers and areas. Our results reveal that neurons with similar responses are preferentially connected both within and across layers and areas -- including feedback connections -- suggesting the universality of the like-to-like connectivity across the visual hierarchy. Using a validated digital twin model, we separated neuronal tuning into feature (what neurons respond to) and spatial (receptive field location) components. We found that only the feature component predicts fine-scale synaptic connections, beyond what could be explained by the physical proximity of axons and dendrites. We also found a higher-order rule where postsynaptic neuron cohorts downstream of individual presynaptic cells show greater functional similarity than predicted by a pairwise like-to-like rule. Notably, recurrent neural networks (RNNs) trained on a simple classification task develop connectivity patterns mirroring both pairwise and higher-order rules, with magnitude similar to those in the MICrONS data. Lesion studies in these RNNs reveal that disrupting like-to-like connections has a significantly greater impact on performance compared to lesions of random connections. These findings suggest that these connectivity principles may play a functional role in sensory processing and learning, highlighting shared principles between biological and artificial systems.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 3 (15 Oct 2024) · CC BY 4.0

Functionally similar neurons in mouse visual cortex connect preferentially across layers and areas, including feedback, and feature tuning, not receptive-field position, predicts synapses beyond axon-dendrite proximity.

Earlier studies of like-to-like wiring were mostly confined to small volumes, usually a single layer of V1, so it was unclear whether the rule extends across layers, areas and feedback pathways. Using the densely reconstructed MICrONS volume, unconnected controls with nearby axons and dendrites, and a factorized digital twin, this work separates axonal-scale from synaptic-scale selectivity and tuning content from receptive-field location.

Key results

  1. ShownLike-to-like wiring by signal correlation holds within V1, within higher visual areas, and in feedforward and feedback projections. It appears at both the axonal and synaptic scales and rises in a graded way with co-travel length, synapse density, synapse size and synapse count.

    Graph of 148 proofread presynaptic and 4811 postsynaptic neurons (144 and 3920 passing inclusion). Connected pairs exceed ADP and same-region controls, and ADP exceeds same-region (paired t-tests, BH-corrected) in V1→V1, HVA→HVA, V1→HVA and HVA→V1; cleft volume r = 0.032, p < 0.001; replicated with in vivo signal correlations. · MICrONS functional connectomic dataset; Functional similarity is enhanced at both the…; Model prediction performance threshold; Statistical analysis of mean signal correlations

  2. ShownSplitting digital-twin tuning into 'what' and 'where' components shows that feature preference predicts synapse density, multiplicity and cleft volume. Receptive-field center proximity tracks axon-dendrite co-travel but not synapse density, apart from an HVA→V1 exception noted in a figure legend.

    Feature similarity is the cosine of 512-length readout weight vectors; RF distance is the visual angle between readout centers. Both scale with co-travel distance (Ld), but RF distance is uncorrelated or, in V1, anti-correlated with Nsyn/mmLd (legend notes an HVA→V1 exception); in area-by-layer GLMMs RF effects vanish at synaptic scale. · Factorized in silico functional representation; A like-to-like rule for feature similarity, but…; Like-to-like rule generalizes across joint layer…; In silico feature weight similarity and…

  3. SuggestedPostsynaptic neurons that share one presynaptic input are more functionally similar to one another than a pairwise like-to-like model predicts, which points to higher-order organization.

    Expected post-post signal correlation came from a Poisson GLMM of synapse counts on signal correlation, feature similarity and RF distance with projection-type interactions; observed similarity exceeded it in three of four projection types, at axonal and synaptic scales. · Neurons with common input are functionally…; Functional similarity among all postsynaptic…; Estimation of ρ expected by pairwise…

  4. SuggestedA vanilla recurrent network trained on image classification develops stronger like-to-like connectivity and post-post similarity than before training. Removing its like-to-like weights hurts accuracy more than removing random weights.

    Vanilla RNN, 1000 tanh hidden units, 20 time steps, MNIST inputs, 10 epochs, batch 512, Adam. Connection = weight > 0.01 (top 35th percentile). Like-to-like ablation candidates (weight > 0.01, signal correlation > 0.2) were about 10.5% of weights; their removal lowered test accuracy more than random ablation. Accuracy values not given in text. · Like-to-like connectivity and its functional…; RNN model; Figure legends

  5. ShownDigital-twin signal correlations computed over 250 new movie clips match a 30-clip in vivo benchmark better than in vivo estimates from the 6 clips available in MICrONS. Model estimates of orientation tuning and receptive-field position also agree with in vivo measurements.

    In the figure's example comparison, correlation with the 30-clip benchmark was 0.69 in silico vs 0.40 for 6-clip in vivo, and in silico was better in all three validation mice (p < 0.001); 95% of gOSI > 0.25 cells had in silico vs in vivo preferred-orientation difference below 9.77°; readout location tracked sparse-noise RF centers. · Factorized in silico functional representation; Validation of In silico signal correlations; Validation of orientation tuning; Figure legends

Methods, data and reproducibility

Two-photon calcium imaging and dense EM reconstruction from one mouse were coregistered; digital-twin (CvT-LSTM) signal correlation, feature-weight similarity and RF center distance were related to axon-dendrite co-travel and synapse counts via paired t-tests and generalized linear mixed models against proximity and same-region controls.

  • MICrONS coregistered two-photon calcium imaging and EM volume (V1, LM, AL, RL; layers 2 through 6); 115,372 functional units, estimated 75,909 excitatory neurons · mouse (single animal, GCaMP6s in excitatory neurons) · n = 13,952 manually matched excitatory neurons
  • Proofread connectivity graph: manually proofread presynaptic neurons (84 full, 64 partial axon proofreading) and NEURD-cleaned postsynaptic partners · mouse (MICrONS animal) · n = 148 presynaptic, 4811 postsynaptic (144 and 3920 pass inclusion)
  • Additional mice expressing GCaMP6s via Slc17a7-Cre and Ai162 used for validation scans; the orientation-tuning scans targeted the V1/HVA junction to mimic the MICrONS animal · mouse · n = Ten mice (3 female, 7 males) · JAX 023527; JAX 031562
  • Signal-correlation validation: 6 clips x 60 repeats, then 36 clips x 10 repeats split into a 30-clip benchmark and a 6-clip MICrONS-equivalent set · mouse · n = two mice (repeat calibration); three mice (benchmark)
  • Validation scans with 57.6 minutes of sparse noise (receptive fields) and 40 additional minutes of Monet stimuli (orientation tuning) · mouse · n = three scans each
  • Natural-movie scans used to train the shared digital-twin core network · mouse · n = 8 scans from 8 mice
  • MNIST images used as classification inputs for the vanilla RNN
Methods & tools
Digital twin, cvt-lstm (convolutional vision transformer + LSTM) model of Wang et al. (2024), conv-lstm digital twin variant for orientation tuning, Mesoscopic two-photon calcium imaging and electron microscopy, NEURD (automated proofreading of postsynaptic and control neurons), pcg_skel (presynaptic axon skeletonization), CAVE annotation tables (synapses_pni_2, nucleus_detection_v0), SciPy (KDTree query_ball_tree, paired t-tests), glmmTMB (Tweedie and Poisson GLMMs), emmeans (emtrends) and performance (Nakagawa's R squared) R packages, statsmodels (Benjamini-Hochberg correction), DeepLabCut (pupil tracking), PyTorch (RNN training)
Reagents & environment
Custom pipelines in Matlab, Python and R (glmmTMB, emmeans, performance, statsmodels, SciPy, PyTorch); Docker and Kubernetes used for development and deployment; no package versions stated.
Reproducibility gaps
Analysis code, digital-twin architecture, learned model weights and in silico similarity metrics are only promised for release by journal publication; until then, access is through the corresponding authors.
Limitations (6)
  • All MICrONS data come from a single mouse, and the analyses rest on a small number of manually proofread neurons. The authors note that more proofreading, matching and extension are needed to complete the graph and reduce possible connection biases. · stated by the authors
  • The digital twin is treated as a black box, and validation experiments are still needed for new stimulus domains. Behavioral state and task variables, which explain much of the neural response, are not yet included in the model. · stated by the authors
  • Unlike earlier studies, orientation-preference like-to-like within V1 L2/3 excitatory neurons was not significant against region-matched controls. The authors attribute this to a local orientation bias where the volume sits. · stated by the authors
  • Results are correlational, and pairwise like-to-like may only partly capture more complex structure-function rules. The authors call for theory-driven experiments to establish causality and for tests of higher-order motifs. · stated by the authors
  • The first 40 presynaptic neurons were chosen unblinded to function, and axon proofreading for 64 neurons targeted feedback branches only. The presynaptic sample may therefore not represent the population evenly across projection types. · noted in this summary
  • The only correlation strength printed in the text is small (cleft volume vs signal correlation r = 0.032), and most per-projection statistics sit in supplementary tables not included here, so effect magnitudes are hard to judge. · noted in this summary

Read this if you study how cortical wiring relates to neuronal tuning, analyze functional connectomics or EM datasets like MICrONS, or use digital-twin models to test structure-function rules across visual areas.

Citations
103
FWCI
42.46
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISMuir '17Papadopouli '24Klavinskis- '25Biswas '26Tang '24Buendía '24
76% semantic matchNeuroscience
Dylan R. Muir, Patricia Molina-Luna, Morgane Roth+2 more·PLoS Computational Biology·Dec 14, 2017· 9 citations