Overview of the cortical network model and calcium-based plasticity rule.

A: Rendering of the seven column subvolume of rat nbS1. (10% of the cells are shown). B: Representative morphologies for the 18 excitatory morphological types (m-types) and their typical firing pattern (top left). Neurite diameters are scaled (x2) for better resolution. C: Exemplary connections to layer 5 (thicktufted pyramidal cells, L5_TPC:A and L5_TPC:B) L5 TTPCs (top) and their short-term dynamics (bottom). Neurite diameters are uniform for better visibility. On the bottom: thin gray lines represent 20 individual trials, while the thicker black ones their means. D: Bouton density profiles of thalamocortical fibers, and locations of VPM (black) and POm (purple) synapses on neurons (in a 5 µm radius subvolume). The spatial scale bar on B applies to the whole figure. Panels adapted from Isbister et al. (2023); Chindemi et al. (2022); Reimann et al. (2022). E: Main variables of the plasticity model during coincident activation of the pre- and postsynaptic neurons for an exemplary synapse. Left: under in vitro-like conditions (adapted from Chindemi et al. (2022)). Right: same pair of neurons under in vivo-like conditions. Schematics on their lefts illustrate the difference between typical in vitro and in vivo protocols.

Changes in synaptic efficacy stabilize and they promote macroscale stable dynamics.

A: Top-down view of the spatial arrangement of the VPM fiber centers associated with the 10 input patterns. Bottom row 3rd: pyramid-like overlap setup of VPM patterns, 4th (bottom right) POm fibers (high-order thalamic input) associated with all stimuli. B: Raster plot of the microcircuit’s activity and the population firing rates below. The y-axis shows cortical depth. As cortical layers do not have the same cell density, the visually densest layer is not necessarily the most active. Panel adapted from Ecker et al. (2024). C: Firing rate of excitatory cells during the 10-minutes-long simulation. D: Plastic evolution of synaptic efficacy (ρ): L2 norm of changes in ρ across time. Inset shows the distribution of ρ values in the beginning (green; strickly 0s and 1s, see Methods) and end (purple) of the simulation. E: Evolution of mean ρ (aggregated over multiple synapses within a single connections). E1: L2 norm of changes in mean ρ (black) across time against STDP control in pink (inset, see Methods). E2: L2 norm of changes in ρ (similar to E1, but) compared to ρt=5minutes in black, STDP in pink, and their random walk controls (see Methods) with the same colors but dashed lines. F: Firing rates of excitatory cells from a simulation with higher [Ca2+]o and thus population bursts (see Supplementary Figure S6, Markram et al., 2015, and Methods) in gray and the baseline one in red as on panel C. G: Grand average mean ρ values across time for the synchronous simulation in gray and for the corresponding STDP control in pink. H: Same as E2 but for the synchronous simulation (calcium-based plasticity rule in gray instead of black).

Changes in synaptic efficacy are sparse.

A1: Propensity of changes in connections (mean ρ values) across layers. Connections are grouped into layers based on the soma location of the postsynaptic cell. Percentages are calculated within layers, i.e., 7.4% depression (in blue) for layer 2 (L2) means that among all connections that have the postsynaptic cell in L2, 7.4% depressed. A2: Plastic changes in mean ρ vs. mean pairwise firing rates of the pre- and postsynaptic cell of a given connection. B1: Same as A1 but for ρ (changes in synapses rather than connections). B2: Layer- and neurite type-wise (apical and basal dendrites) distribution of Δρ. Only synapses that underwent any plastic change (~ 15M) are shown. C1: Distribution of ĝAMP A in the beginning (in green) and end (in purple) of the simulation. C2: Plastic changes in ĝAMP A leading to the slight shift in the distribution on C1. D1: Same as B1, i.e., propensity of changes at the synapse level, but for synapses that cross the ρ∗ = 0.5 unstable fixed-point. D2: Cumulative histogram of ρ crossing the unstable fixed-point against time for depressing synapses in blue and potentiating ones in red.

Changes in synaptic efficacy are non-random and shaped by the networks topology and the input stimuli.

A: Layer-wise distribution of changing connections. A1: Left: schematic of the changing subgraph. In black, connections whose efficacies are changing and their pre- and postsynaptic populations. In grey, connections whose efficacies do not change and neurons not partaking in a changing connection. Right: random control of the changing subgraph, generated by randomly selecting the same number of connections as the changing connections (red edges) between the same pre- and postsynaptic populations (black nodes). A2: Sankey plot of layer-wise distribution of changing connections. Thickness of lines is proportional to the number of changing connections between pre- and post node populations. A3: As A2, but for the random control (A1 right). B: Directed simplex counts in the changing subgraph and its random control. B1: Schematic of directed k-simplices and their counts in the schematic graphs on A1. Note that by construction the control must have the same number of 0- and 1-simplices (see Methods) which correspond to the number of cells and connections in the subnetwork. B2: Directed simplex counts across dimensions in the changing subgraph (black) and its random control (red). C: Input - output distance correlations. C1: Input distances as the Earth mover’s distance of the VPM fiber locations (see Figure 2A). Inset shows the overlap (based on Hamming distance, see Methods) of pattern fibers. C2: Output distances as Euclidean distance of mean ρ matrices in the 2-minute-long single pattern simulations. To its right: correlation of input and output distances. C3: Same as C2 but with Hamming distance. C4: Same as C2 but with Earth mover’s distance (on the output side as well).

Changes are more frequent within and across cell assemblies.

A: Schematics of the assembly detection pipeline. Significant time bins are clustered by cosine similarity and these clusters define functional assemblies: green, blue, light blue (see Methods). Bottom right, visual summary of the organization of pattern-specific assembly sequences of the two repetitions of pattern A. B: Activation of functional assembly sequences for multiple repetitions of the patterns presented (as in A bottom right). Each row within the 10 matrices corresponds to a single repetition of a given pattern. White: non-significant time bins. C: Number and location of neurons in each cell assembly. Constituent neurons are those for which the correlation of their spiking activity within the corresponding time bins is significantly stronger than chance (see Methods). Top: top-down view, bottom: depth-profile. D: Jaccard similarity (intersection over union) between cell assemblies. E: Initial mean ρ of within- and cross-assembly synapses. F: Propensity of depression and potentiation of within- and cross-assembly synapses in blue on the left and in red on the right respectively. Since assemblies are overlapping (see D) some synapses are counted in multiple pre- and postsynaptic assembly pairings.

Changes are structured by synaptic clusters.

A: Changes in clustered assembly synapse on an exemplary neuron from assembly 11 (A11). Neurite diameters are scaled (x2) for better resolution. (Synapse diameters are arbitrary.) B: Temporal evolution of the synapses on basal dendrites shown on A. Synapses are grouped by their dendritic sections (single stretches of non-branching dendrites) and ordered by their mean distance to the soma (closest on the bottom and furthest on top). C: Michelson contrast (Equation (15) in Methods) of the likelihood of plastic changes of synapses grouped into four categories: (non)-assembly and (non-)clustered) for each assembly. Negative values indicate that synapses belonging to the given category are unlikely to change. Depression on top (in blue) and potentiation below (in red). Black arrows indicate clustered assembly synapses. D: Same as C but for inital ρ values (see Methods).

Changes are stronger in central connections in the network.

A1: Schematics of edge participation in high-dimensional simplices in an exemplary –full— network on the top in black and in (assembly) –sub—-graphs on the bottom. Table summarizes edge participation values for the connection from node u to v across dimensions for all three cases. A2: Probability of changes (i.e., either depression or potentiation) vs. edge participation in high-dimensional simplices (see A1 and Methods). A3: Probability of depression (left) and potentiation (right) conditioned on plastic change (in any direction) vs. edge participation. B: Comparison with electron microscopy data. Total AMPA conductance (ĝAMP A) against edge participation across dimensions in our nbS1 model on the left. Total synapse volume (see Methods) against edge participation across dimensions in the MICrONS (2021) dataset on the right. Shaded areas indicate SEM. C: Probability of changes against edge participation in assembly subgraphs (see A1 bottom). Left, probability of changes in (2-minutes-long) pattern A simulations against edge participation in assembly subgraphs. Right, summary of the maximum probability values across patterns on the right. Arrow indicates the row shown in detail on its left and white crosses indicate the pattern-specific early assemblies (see Figure 5B).

Changes promote stimulus specificity.

A: Firing rates and pairwise spike correlations extracted from non-plastic simulations before plasticity, i.e., in the naive circuit vs. after the 10-minute-long plastic simulation. B1: Spike correlation of connected cells vs. edge participation in high dimensional simplices (see Methods) before and after plasticity on the left and right respectively. B2: Spike correlation of connected cells vs. edge participation in the MICrONS (2021) dataset (see Methods). Shaded areas indicate SEM. C: Jaccard similarity of assemblies detected before vs. after plasticity on the left. Shared early assembly of patterns B and E splitting into two stimulus-specific ones after plasticity on the right. (Indicated by red arrow and rectangle on its left). See more detailed plots in Supplementary Figure S10B. D: Reliability of assembly sequences (see Methods and Supplementary Figure S10B). Significance of increases was determined with Kruskal-Wallis test: *: p 0.05, **: p 0.01, ***: p 0.001, ****: p 0.0001. E: Spike time reliability (see Methods) of single cells to the different patterns before and after plasticity. (The lack of stars means no significant increases after plasticity.) F: Short-term dynamics of an exemplary potentiated L5 TTPC connection before (in green) and after plasticity (in purple). At in vitro [Ca2+]o on the left, and in vivo on the right. Thin lines represent 20 individual trials, while the thicker ones their means. G: Interspike interval (ISI) distribution of all excitatory neurons before and after plasticity on the left. Zoom in on low ISIs (≤ 50 ms) on the right.

Physiology of excitatory cells and E to E connections.

A: Distribution of ion-channel densities in the excitatory (cADpyr) electrical type (etype). B: Validation of dendritic physiology of the cADpyr etype on L5 TTPC mtypes. B1: Validation of back-propagating action potential (bAP) amplitude for basal (teal) and apical (blue) dendrites. Reference data (in orange) comes from Stuart and Sakmann (1994); Larkum et al. (2001) (apical) and Nevian et al. (2007) (basal). Lines show exponential fits for the in silico (teal and blue) and in vitro (orange) data. Color bar indicates dendritic diameter. B2: Validation of EPSP attenuation. Reference data comes from Berger et al. (2001) (apical) and Nevian et al. (2007) (basal). Lines and color bar same as in B2. Data taken from (and partially shown in) Reva et al. (2022). C: Anatomy and physiology of E to E connections. C2: Connection probability and number of synapses per connections for all E to E connections. White boxes indicate non-feasible connections, or on the left panel: no pairs found within the 200 µm intersomatic distance used. C2: Mean (over 100 pairs) PSP amplitude (left) and CV (std/mean on the right) of all E to E connections. Data taken from (and shown in) Isbister et al. (2023). C3: Initial synaptic physiology parameters. From left to right: ρ, ĝAMP A, and USE.

STDP and synapse-specific parameters of the plasticity model.

A: Spike timing, pairing frequency, and pathway dependence of plasticity. B: Layer- and neurite type-wise distribution of measured [Ca2+]i peaks (used to derive parameters of the plasticity model shown in C). Synapses are grouped based on the soma location of the postsynaptic cell. 10% of all synapses are shown. Schematics on their lefts illustrate the measurement protocols. C: Layer- and neurite type-wise distribution of depression and potentiation thresholds (θp and θp) of the plasticity model. Synapses grouped and shown as in B. D: Correlations of the parameters shown in B and C.

Calibration of the in vivo-like network state.

A: Same as Figure 2B (i.e., raster plots of the microcircuit’s activity) under different synapse setups. The microcircuit equipped with the plasticity model of Chindemi et al. (2022) only resembles that of the non-plastic network’s of Isbister et al. (2023) when VDCCs (voltage-dependent calcium channels) are blocked (last row). B: Re-calibration of the in vivo-like state using the plasticity model. B1: Left: Euclidean distance of the measured percentages of firing rates (PF Rs) from the target ones in different iterations of the calibration process. Right: Validation of network states after the final (4th) iteration. Dashed gray line along the diagonal indicated perfect match. B2 Left: Injected Ornstein-Uhlenbeck (OU) conductances in the non-plastic model of Isbister et al. (2023) vs. the plastic one for PF R = 40% (the state used in the current article). Dashed gray line along the diagonal indicated perfect match. Right: Layer-wise (absolute) firing rates of excitatory (E) and inhibitory (I) subpopulations at PF R = 40%. Legend on the bottom applies to the last three panels in B.

Activity of the thalamic fibers.

Raster plots of VPM fibers forming each of the 10 input patterns (Figure 2A) for the stimulus stream (i.e., from pattern A at 2000 ms to pattern J at 6500 ms). Bottom row shows the same for non-specific POm fibers. (Adapted from Ecker et al. (2024).)

Changes in synaptic efficacy during plasticity.

Individual ρ traces (10 per panel) during plasticity. Sampled traces have |Δρ| ≥ 0.1 and mean over the 10 minutes within [0.2, 0.8].

Synchronous (unstable) network activity.

A: Same as Figure 2B (i.e., raster plots of the microcircuit’s activity) under a different network state, characterized by higher [Ca2+]o and thus population bursts (see Markram et al. (2015)). A1 in the beginning of the 10-minute-long simulation, while A2 at the end. B: Firing rate of excitatory cells (zoom-ins on the beginning and end (corresponding to the raster plots above) are shown in Figure 2F). C and D: Same as Figure 2D and E1 (i.e., L2 norm of changes in ρ (synapses) and mean ρ (connections) across time).

Changing connections in plastic control simulations.

A: Same as Figure 2B and 3A2 (i.e., raster plots of the microcircuit’s activity and plastic changes in mean ρ vs. firing rates under different conditions). The last row of A2 is not an exact replica of Figure 3A2 as these simulations were run for 2 minutes. B: Similar to Figure 3A1 (i.e., layer-wise propensity of changes in mean ρ) but split across conditions. C: Similar to Figure 2E1 (i.e., L2 norm of changes in mean ρ values for all conditions.)

Changes in cross-assembly synapse clusters.

A: Distribution of within-assembly Δρs across the four categories introduced in Figure 6 (non)-assembly and (non-)clustered). Black arrow indicates clustered assembly synapses. Boxes show all values, while black dots are 1000 samples from each. Significance test was run on the balanced samples (1000 each): 2-way ANOVA and post-hoc Tukey’s test: *: p 0.05, **: p 0.01, ***: p 0.001, ****: p 0.0001. B: Schematics of cross-assembly synapse clusters. Postsynaptic neuron from the green assembly, presynaptic ones from the orange assembly. C: As Figure 6 C and D, i.e., Michelson contrast (Equation (15) in Methods) of the likelihood of plastic changes of synapses grouped into categories (not only four as there are several presynaptic assemblies) on the left, and initialization of ρ values on the right. Depression in blue and potentiation in red. Black arrow indicates clustered assembly synapses. Assembly 12 (A12) was chosen as the postsynaptic assembly as it appeared the latest in the sequence (Figure 5B) and thus had the most presynaptic assemblies which were active before it in time. D: Same as A, but for cross-assembly synapses (data from several postsynaptic assemblies, not only A12 shown on C).

Structural comparison with MICrONS and probability of change vs. pattern-indegree.

A: Image taken from MICrONS (2021) (under CC-BY 4.0 license). Distributions of edge participation in high dimensional simplices (see Methods) across dimensions in the MICrONS (2021) dataset and in our nbS1 model (Reimann et al., 2022). B: Probability of change vs. pattern-indegree (number of connections from VPM patterns). B1: Probability of changes (in the 2-minute-long single pattern simulations) vs. pattern-indegree of the presynaptic (left) and postsynaptic (right) neurons. B2: Probability of changes (i.e., either depression or potentiation) against the pattern-indegree of both pre- and postsynaptic neurons on the left (for pattern A only). Probability of potentiation conditioned on plastic change (in any direction) against pattern-indegrees on the right. C: Pattern-indegree of pre- and postsynaptic cells (for the four base patterns only as the remaining ones are based on these four).

Layer-wise changes and assemblies detected before and after plasticity.

A: Layer-wise distribution of changes (after - before) from left to right in: firing rate, pairwise spike correlation, pairwise spike correlation of connected cells, and ρ values. B: Same as Figure 5B (i.e., sequential activation of assemblies over repetitions of patterns) but for assemblies detected from 2-minute-long simulations (leading to fewer presentations of patterns) before and after plasticity on top and bottom respectively. C: Davis-Bouldin index (Davies and Bouldin, 1979) used to determine optimal number of assemblies before (in green) and after (in purple) plasticity. D: Reliability of assembly sequences (see Methods) for selected pattern E (as that is the one with a stimulus specific early assembly after plasticity with increased reliability) for different number of assemblies detected before and after plasticity. Red rectangle indicates data shown on Figure 8D. As on the main Figure p-values from Kruskal-Wallis test. Missing entiries indicate either a decrease or a non-significant (p-value > 0.05) increase in assembly sequence reliability. E and F: Different number of assemblies before plasticity. E1: As B, but for 14 assemblies. E2: Ward’s linkage tree and the threshold that results in 14 assemblies. F: As E, but for 15 assemblies.