Connectome analysis of a cerebellum-like circuit for sensory prediction

Authors

1. Krista E. Perks
2. Mariela D. Petkova
3. Salomon Z. Muller
4. Michael Genecin
5. Adishree Ghatare
6. Richard Schalek
7. Yuelong Wu
8. Michal Januszewski
9. Viren Jain
10. Jeff W. Lichtman
11. L. F. Abbott
12. Nathaniel B. Sawtell

Extended Data Fig. 1 ELL sample preparation and EM data examples.

All EM connectomic data are from a single biological specimen. A, uCT volume of a transverse section through the hindbrain containing the region of interest within the ventrolateral zone of the ELL (dashed rectangle). B, EM volume of the ELL obtained at 4 nm × 4 nm × 30 nm resolution. C, Examples of asymmetric, putative excitatory, synapses formed by granule cell axons (purple) onto apical dendritic spines in the molecular layer (blue: postsynaptic dendrite) (https://neuroglancer-demo.appspot.com/#!gs://efish-public/ng_states/excitatory_example.json). D, Example symmetric, putative inhibitory synapse formed by an MG cell axon (purple) onto a dendrite (blue) (https://neuroglancer-demo.appspot.com/#!gs://efish-public/ng_states/inhibitory_example.json). E, Additional example of asymmetric synapses formed onto somas (blue stars) by different ELL cell types (purple star: presynaptic axon of each cell type as labeled above). Related to Fig. 1.

Extended Data Fig. 2 Morphological characterization of ELL cell types.

See https://neuroglancer-demo.appspot.com/#!gs://efish-public/ng_states/Proofread_Classified_Cells.json for browsable renderings of all proofread and classified bodies of the main cell types analyzed in this study. A, Dendrogram visualization of the hierarchical clustering used to classify MG (n = 104) and SG (n = 545) cells. An additional 32 MG and 47 SG cells were manually categorized because they lacked sufficient soma in the volume and/or the majority of either their axon or dendrite exited the volume. B, Example MG and SG cells showing the variation in morphology across the population. Axons visualized in black. C, Reconstructed axons and basal dendrites of all MG cells showing the inverse relationship between axon and basal dendrite position (perpendicular to the layer plane in ELL) for MG− versus MG+ subtypes. D, Scatterplot of the median basal dendrite versus axon position for the 110/136 MG cells with sufficient axon and basal dendrite within the volume to classify. Dotted line shows the classifier obtained by linear regression on the dataset. Any cells shown in C with either axon or basal dendrite mostly out of the volume were excluded from this analysis and classified manually based on the location of the remaining in-volume axon/dendrite (n = 26 MG cells). E, Reconstructed axons and basal dendrites of all SG cells showing the inverse relationship between axon and dendrite position (perpendicular to the layer plane in ELL) for SG− versus SG+ subtypes. F, Scatterplot of the median basal dendrite versus axon position for the 559/592 SG cells with sufficient axon and basal dendrite within the volume to classify. Dotted line shows the classifier obtained by linear regression on the dataset. Based on these classification metrics, the SG− and SG+ subtypes are less distinct than MG+/MG− because of how much further ventrally SG− basal dendrites extend compared to MG−. Any cells shown in E with either axon or basal dendrite mostly out of the volume were excluded from this analysis and classified manually based on the location of the remaining in-volume axon/dendrite (n = 33 SG cells). G, Scatterplots (data points are randomly jittered along the x-axis within each category) overlaid on boxplots illustrating the soma diameter distribution of each cell type. Boxplots show the median (centre line), 25th and 75th percentiles (box bounds), and minima and maxima of the non-outlier data (1.5× IQR from the box edges); outliers are shown as individual data points in the scatterplot. Only cells with their full soma in the volume were included in this analysis. H, Same display as G for the distribution for soma location of each cell type. I, Same display as G for the distribution of the locations of synapses formed by each cell type. Related to Fig. 1.

Extended Data Fig. 3 The granule cell input pathway.

A, A GC axon (magenta) synapsing onto the spines of an ON (red) and MG− (blue) cell and onto the swelling of a smooth apical dendrite (gray) (arrows show synapse locations). B, Stacked histogram plot shows the number of synapses formed by each proofread and synapse-annotated GC axon pre postsynaptic cell type (n = 674 synapses from 64 GC axons onto 310 Output, MG, SG, and molecular layer interneurons (MLIs)). An additional 323 GC axon synapses innervated other non-spiny and unclassified elements. C, A 40 × 40 × 40 µm region of the molecular layer (~100 µm from the surface of the ganglion layer) capturing apical dendrites from Output, MG, and SG cells with soma located in a densely-reconstructed 60 × 80 × 60 µm central region of the ganglion, plexiform, and granular layers (see Methods).

Extended Data Fig. 4 Effects of synaptic connectivity on sensory cancellation.

A, Absolute peak membrane potential (relative to baseline) of ON (red) and OFF (blue) model Output cells during the first 4 min after turning an EOD mimic on for different levels of ‘selectivity’ of the effects of sensory input on broad spikes. A value of 1 (‘selective’) indicates that sensory input solely effects broad spikes and a value of 0 (‘non-selective’) indicates equal effects on broad and narrow spikes. When selectivity is low, MG cells substantially impair cancellation compared to not having MG input at all (dashed line). This occurs because sensory drive to narrow spikes amplifies EOD-evoked responses in the Output cells. MG recurrence further amplifies this effect, producing large depolarization in MG cells (~15 mV peak response) compared with no recurrence (2.25 mV). B, Membrane potential modulation of ON (red) and OFF (blue) model Output cells after 4 min of learning (quantified as variance normalized by the variance at zero MG synapse weight). Modulation is plotted as a function of MG:Output synaptic strength for different MG:Output connectivity patterns (left two plots), and as a function of MG:MG synaptic strength for different MG recurrence patterns (right two plots). For MG:Output connectivity, “reverse” denotes MG+ :ON and MG− :OFF, whereas “random” denotes both MG types synapsing onto both Output types. For MG:MG connectivity, “reverse” denotes MG+ :MG+ and MG− :MG− , whereas “random” denotes all-to-all connectivity. Vertical gray lines indicate default synaptic strength values used in the model. C, Absolute peak membrane potential (relative to baseline) of ON (red) and OFF (blue) model Output cells during the first 4 min of learning to cancel the EOD-mimic, shown for different levels of recurrence selectivity for broad spikes. A value of 1 (‘selective’) indicates that recurrence solely effects broad spikes and a value of 0 (‘non-selective’) indicates equal effects on broad and narrow spikes. When selectivity is high, the contribution of MG cells to cancellation is comparable to the condition without MG recurrence (dashed line), because negative images transmitted exclusively to broad spikes are not transmitted to Output cells. D, The reduction in MG contribution caused by increased broad spike selectivity can be compensated by increasing MG to Output synaptic strength, suggesting an alternative biophysical mechanism for maintaining an effective MG contribution of ~0.5 to the negative image in Output cells. E, Membrane potential modulation of ON (red traces) and OFF (blue traces) model Output cells in the first 4 min of learning, shown for different plasticity rates at GC to SG synapses. A value of 1 (“SG learning”) indicates that the ratio of post-synaptic EOD response to background activity matches that of MG cells (~60), resulting in faster depression of GC synapses in SG+ cells. Introducing SG learning slightly slows cancellation in ON cells because, as SG− synapses gradually cancel their sensory inputs, ON cells must reverse previously learned cancellation. This continual rebalancing limits the net contribution of SG plasticity to overall cancellation. In these simulations, SG cells are assumed to generate two spike types but lack sufficient electrical separation to decouple learning from signaling. Thus, unlike MG cells, SGs both receive and transmit sensory input. Learning at GC-SG synapses therefore cancels predictable EOD-evoked input at the level of narrow spikes, yielding a “cleaned-up” sensory signal in which predictable components are attenuated while unpredictable components (e.g. prey signals) are preserved. In this framework, SG cells enhance Output responses to unpredictable sensory input, whereas MG cells suppress Output responses to predictable input. F, Spike rate of ON (red) and OFF (blue) model Output cells after 4 min of cancelling the EOD-mimic for the different conditions described in Fig. 5d: without any MG connectivity, which effectively restricts learning to the Output cells (dashed), with MG-Output connections to Output cells, and thus two-site learning, but no recurrence (dotted), and with both MG to Output and MG-MG recurrent connections (fully connected model, solid). Grey lines denote the initial response. G, Ratio of peak sensory-evoked firing (measured during the first second after turning the EOD mimic on) relative to their baseline (equilibrium) firing rate for broad spikes for individual MG+ cells (n = 43) and ON cells (n = 24). In Fig. 5, population-averaged responses yielded a ratio of ~60:8 (MG+ :ON), comparable to that obtained here (~100:12) from single cells. Boxplots show the median (centre line), 25th and 75th percentiles (box bounds), and minima and maxima of the non-outlier data (1.5× IQR from the box edges). H, Contribution of MG cells to the negative image in model Output cells after 4 min of learning (left two plots) and at equilibrium when the EOD is fully cancelled (right two plots). After 4 min, approximately 90% of the negative image in ON cells (red traces) arises from GC-MG plasticity, with an even larger contribution in OFF cells (blue traces). These results replicate prior experimental findings in which Output cell plasticity was blocked using intracellular current injections⁸. At equilibrium, half of the negative image is contributed by MG input and half is from the direct GC input. Related to Fig. 5.

Extended Data Fig. 5 Sensory input pathways.

A, Summary diagram of the sensory input pathways from EAFs to Output and MG cells (only connections >10% of total synapses to each cell type are shown). B, Pairwise cosine similarity of the input pattern per cell type for all cells in the analyses in Figs. 2 and 3. Cosine similarity was calculated on the total number of input synapses to each postsynaptic cell from each presynaptic cell type (after within-cell normalization to the total number of synapses for each postsynaptic cell). Postsynaptic cells are ordered along the x and y axes according to their cell types, and cell types are ordered according to putative ON/OFF-type physiology (based on connectivity and/or in-vivo data from previous work). C, Data: The conditional input probability (without z-scoring applied) for all Output, MG, SG, SP−, and Gr+ cells represented in Figs. 2 and 3. The color of each square represents the mean fraction (probability) of synapses from each presynaptic cell type (input) conditional on getting a synapse from a given presynaptic cell type (condition). Null: An example null model result (one of 100 iterations) used to calculate the z-score normalization of the observed (Data) conditional input probability matrix (see Methods). Cell types are grouped along the x and y axes according to their putative ON/OFF pathway associations. D, Histograms of presynaptic input location (relative to the ganglion-molecular layer boundary) for all inputs from each cell type. All cells used for the analyses in Figs. 2 and 3 are included here.

Extended Data Fig. 6 Feedforward and recurrent connectivity of MG cells.

See https://neuroglancer-demo.appspot.com/#!gs://efish-public/ng_states/Proofread_MG_Output_MGsyn.json for browsable renderings of all proofread and classified MG and Output cells along with the annotated MG postsynaptic target locations. A, Left: Conditional output analysis (applied to the 103 MG cells represented in Fig. 4) shows the observed connectivity from MG cells to Output and other MG cells compared with that expected from random connectivity (see Methods). Each entry corresponds to the probability (z-score normalized) of synapsing on the postsynaptic cell type shown on the horizontal axis (output) given that the presynaptic MG cell makes at least one synapse to the postsynaptic cell type shown on the vertical axis (condition). Cell types are grouped along the x and y axes according to their known electrophysiological response types (ON/OFF). Two clusters (illustrated with dotted lines) are observed among the postsynaptic cell types, corresponding to the ON and OFF pathways. Middle (Data): The conditional output probability (without z-scoring applied) for all MG cells represented in Fig. 4. The color of each square represents the mean fraction (probability) of an MG cell synapsing onto each postsynaptic cell type (output) conditional on synapsing onto a given postsynaptic cell type (condition). Right (Null): An example null model result (one of 100 iterations) used to calculate the z-score normalization of the observed (Data) conditional output probability matrix (see Methods). Cell types are grouped along the x and y axes according to their known electrophysiological response types (ON/OFF). B, Stacked histograms show the number of synapses onto individual Output and MG cells from different presynaptic cell types (indicated by colors) (n = 3347 synapses from 1002 MG, SG, SP−, Gr+, and EAF cells to the same 20 Output and MG cells represented in Fig. 3e,f; 5003 total synapses were annotated, with 33% of those synapses from unclassified elements). Cells are ordered along the horizontal axis by the number of inputs from SG+ (decreasing) and then SG− (increasing). Pie charts show the fraction (and total number) of synapses from each presynaptic cell type onto all ON, OFF, MG+ and MG− cells represented in the bar graphs. C, Divergence and convergence of MG connectivity to all Output cells (n = 103 presynaptic MG cells and 87 Output cells) and to all MG cells (n = 93 presynaptic and 115 postsynaptic MG cells). Boxplots show the median (center line), 25th and 75th percentiles (box bounds), and minima and maxima of the non-outlier data (1.5× IQR from the box edges); outliers are shown as individual data points. Related to Fig. 4.

Extended Data Fig. 7 Sensory cancellation in the model matches electrophysiological recording from MG and Output cells.

A, Schematic of the connectivity between MG and Output cells (left) and experiment timeline (see Results). B,C, Cancellation of the effects of turning on an EOD mimic (simulating a natural increase in EOD amplitude) on recorded (left) and model (right) neurons of various types. In the plot, 0 denotes the time of the EOD command. Responses are shown at equally-spaced time points from the start (dark) to the end (light) of the mimic on period. OFF cell spike rate (n = 55); OFF cell membrane potential (n = 13); MG− broad spike rate (n = 37); MG+ narrow spike rate (n = 24). Unlike broad spikes, narrow spikes rate is not driven to be constant but rather shows a modulated response to the corollary discharge input (black trace showing mean ± SEM of pre EOD-mimic response). Thus, the negative image transmitted by the narrow spikes is measured as the change from the response prior to turning the EOD mimic on⁸. To match model to data, we added to the model the initial data response to the EOD mimic. D,E, Same display as above, but after turning off the EOD mimic (simulating a natural decrease in EOD amplitude) for spike rate in OFF (n = 32) and ON (n = 42) Output cells and broad spike rate in MG− (n = 20) and MG+ (n = 29) cells. For MG− cells, we show model results with and without MG recurrence to demonstrate that the strong response observed in the recorded MG− cells can be explained by strong recurrence between MG cells. Related to Fig. 5.

Extended Data Fig. 8 Multiple sites of plasticity solve a continual learning problem.

A, Experimental time-course. B, Comparison of Output cell responses under different model conditions (black and brown) immediately after 10 min of noise consisting of random amplitude and polarity pulses delivered uncorrelated with the fish’s own EOD command. The EOD command occurs at time zero. The temporal structure of the effects of noise on Output cell responses observed in the model with GC-Output plasticity turned off (brown line), reflects the non-uniform temporal structure of the GC corollary discharge input (gray dashed line)¹⁴. Figure 6b shows similar results for correlated noise. C, Evolution of model Output cell responses during 10 min of noise exposure under identical experimental conditions as in Fig. 6b,c. D, Dynamics of the membrane potential of a model Output cell to excitatory GC input (grey) and inhibitory MG cell input during and after 100 min of noise exposure. Note that the changes in GC and MG synaptic input during the noise period oppose one another. Compare to Fig. 6d. Because ON and OFF cells exhibit similar responses, results throughout this figure are averaged across ON and OFF cells. E, Dynamics of the firing rate of a model Output cell during and after 100 min of noise exposure with plasticity at both GC to Output and GC to MG cell synapses (black) or only at GC to MG cell synapses (brown). Compare to Fig. 6e. F, Membrane potential modulation of Output cells immediately after 10 min of noise exposure as a function of strength of MG recurrence. Stronger recurrence results in reduced Output cell modulation. G, Peak firing rate of the model Output cells during and after 100 min of noise exposure with MG:Output synaptic strength of 0.5 (black trace), the value fitted to data, or strength of 1 (grey trace). Dotted horizontal lines denote equilibrium spike rate. H, Similar to G, but comparing different effective learning rates of Output cells. The black trace corresponds to a modeled ratio of sensory response to equilibrium firing of 8 (the value fitted to data) and the gray trace to a ratio of 16. Related to Fig. 6.

Cite this article

Perks, K.E., Petkova, M.D., Muller, S.Z. et al. Connectome analysis of a cerebellum-like circuit for sensory prediction. Nature (2026). https://doi.org/10.1038/s41586-026-10690-6

Download citation

• Received: 03 July 2025

• Accepted: 21 May 2026

• Published: 02 September 2026

• Version of record: 02 September 2026

• DOI: https://doi.org/10.1038/s41586-026-10690-6