Highly Nonrandom Features of Synaptic Connectivity in Local Cortical Circuits

Sen Song, P. Jesper Sjöström, Markus Reigl, Sacha B. Nelson, Dmitri B. Chklovskii•View original

OverviewBalancedalloy voice

If you've ever stared at a tangle of cortical wiring and wondered if this is just chaos that works, or if there is a hidden blueprint, you're in good company. Song, Sjöström, Reigl, Nelson, and Chklovskii took that question head-on. They couldn't map every synapse in the cortex—no one can at that scale—so they did the next best thing: they measured enough of the local circuit to determine whether its statistics resemble a bag of dice or something more deliberate. They achieved this through hard work, with hundreds of simultaneous quadruple whole-cell recordings from thick-tufted layer five pyramidal neurons in rat visual cortex. This approach provided insights into who connects to whom, and how strongly, all in one snapshot. Here's the setup. They recorded from rats between postnatal day 12 and 20, ensuring the slices remained healthy and warm. They pushed four pipettes into four nearby pyramidal neurons at once. From each presynaptic cell, they evoked spikes and averaged the postsynaptic responses—excitatory postsynaptic potentials, or EPSPs—over about one to two minutes. Most pairs had between 45 to 60 trials; a few had as few as 10 but still met their quality standard. The signal-to-noise ratio was good enough to detect connections down to approximately one hundredth of a millivolt. Overall, they recorded 817 cells arranged into 83 triplets and 142 quadruplets, and across this dataset, the network was sparse: 931 connections out of 8,050 possible, which is about 11.6 percent. The first big result hits like a drumbeat: reciprocity is high. If this were a random network with the same overall connection probability, bidirectional links would be rare. Instead, Song and colleagues found about four times more reciprocal pairs than this simple model predicts, and Monte Carlo tests indicate that the probability of observing this excess by chance is below one in ten thousand. Why does that matter? Because it suggests the cortex isn't creating connections independently. Some pairs really prefer to wire up in both directions, creating feedback loops at the scale of two neurons. You can explore a result like that in various ways to ensure its validity. Are closely located cells more likely to be connected, and did they just happen to record more of them? They measured distances between labeled neurons and saw no systematic dependence of connection probability on distance; a chi-square test produced a p-value of 0.21. Maybe shallow recordings behave differently from deep ones? Across a depth of 10 to 100 micrometers below the slice surface, connectivity remained unchanged. What about age within that postnatal day 12 to 20 window? Neither the mean EPSP amplitude nor overall connectivity showed strong age effects, and repeating the analysis in a tighter window of postnatal days 14 to 16 yielded the same reciprocity increase. Even axonal cuts—inevitable in slices—did not correlate with stronger or more frequent connections. It's significant work, but it eliminates possible confounding factors one by one. Then they enlarged the question. If pairs aren't random, what about triads—three-neuron patterns? In this case, they exercised caution. They sorted all triplets into 16 classes, then constructed a null model that preserves the observed likelihood of unidirectional or bidirectional links between any two cells, but randomizes how those links combine into a triplet. It's akin to dealing cards from the same deck, but shuffling which cards end up together. Some motifs—the densely connected ones in particular—showed clear overrepresentation against that benchmark. When they increased the threshold, focusing only on stronger synapses, the clustering became even more pronounced. Strong triads tend to cluster together. For four-neuron patterns, the bar is set higher, as there are many more ways to connect four nodes. Thus, they constructed a larger random matrix, extracted countless four-by-four submatrices, and employed simulated annealing to match the triplet statistics of the real data. Even against that rigorous backdrop, one class of "chain-like" quadruplets appeared significantly more common—about 70 percent more frequent than expected, with a corrected p-value of 0.004. This is important as it indicates that the nonrandomness isn't merely a quirk of pairs or triads; some larger motifs are also making an appearance, even when you insist that the building blocks have the same pairwise statistics. Topology is only half the story. The other half is weight—the size of those EPSPs. The mean strength across all detected connections was 0.77 millivolts. However, the distribution's shape is particularly striking. Instead of a tidy bell curve or a quick exponential fall-off, the data display a heavy tail: strong synapses are far more common than those simpler models would predict. When modeled as lognormal, it fits; attempts with exponential or Gaussian models do not. This tail is significant because it carries the weight of the network. Connections above 1.2 millivolts—just 17 percent of the synapses—contributed about half of the total synaptic weight. A small minority of very strong connections dominate the overall weight. You can sense the logic starting to form. If some pairs are reciprocally connected more often than chance, and if a minority of edges are very strong, those two facts can overlap. They indeed do. Bidirectionally connected pairs had higher mean strengths than unidirectional pairs—0.95 millivolts compared to 0.61—and that gap was not a coincidence; the p-value was 3.1 times ten to the minus seven. Even more telling, the two directions within a reciprocal pair rose and fell together. The correlation of their strengths, measured on a logarithmic scale, was around 0.36 with a p-value below 0.0001. When examining unrelated pairs that share no neurons, the correlation essentially disappears. The strong connections cluster in the same areas rather than spreading out evenly. If you're curious why the weights are lognormal in the first place, Song and colleagues suggest a clear mechanistic explanation. Imagine that every time two neurons fire together, the synapse between them gets adjusted up or down by some factor—not a fixed amount, but a multiplying factor. If the strength after episode i is represented as w sub i, and each episode multiplies that by some factor F sub i, then after n episodes, you have w sub n equal to the initial strength times the product of all those F's. Taking the logarithm changes the product into a sum. If these log factors have some finite mean and variance, the central limit theorem comes into play, and the log of the weight becomes approximately Gaussian. Exponentiate it back, and you arrive at a lognormal distribution. It’s multiplicative noise through and through, and this is exactly the signature seen in the data. Now, pause for a moment. What emerges from all of this is a picture of the cortex that isn't simply random wiring with random weights. It's a sparse network where a handful of strong synapses account for a disproportionate share of the total. Reciprocal pairs are both common and powerful, and clustered motifs appear more frequently than by chance. In simple terms, it looks like a framework of strong connections embedded in a vast sea of weaker ones. The sea is extensive; it may play a significant role when neurons fire together. But the framework is what carries much of the load. The team even offers some intuition regarding how rare the very strongest links are. If you look for suprathreshold EPSPs—let's say, above ten millivolts—the lognormal fit indicates that they are extremely infrequent, on the order of one in a thousand connections. In a network of the size they considered, that translates to only a few dozen such exceptional connections. This underscores just how uneven the weight landscape is: many small ripples, with a few towering waves. All of this confidence relies on ensuring that the statistics aren't skewed by how the tissue is sampled. That's why those control analyses are significant. Distance between recorded neurons didn't clarify the patterns, nor did the depth of the cells in the slice. Proximity to an axonal cut end—a potential confounding factor in slices—did not predict connection probability or strength. Age, within that postnatal day 12 to 20 window, supported the key findings; even narrowing down to postnatal days 14 to 16 preserved the same overrepresentation of reciprocity and the same heavy tail. They also employed rigorous null models—preserving observed pairwise probabilities when testing triplets, and matching triplet counts when testing quadruplets—and corrected for multiple hypothesis tests to ensure that a few positive motifs did not emerge by chance alone. If you enjoy a bit of mathematics with your neuroscience, there's one more conceptual twist worth appreciating. In the simplest random graph, each possible directed edge is present with some fixed probability p, independent of every other edge. Under that model, the expected number of bidirectional pairs is the number of pairs times p squared, where N is the number of possible pairs. That's a straightforward baseline. The fact that the cortex has four times that expectation suggests that independence does not hold at the level of pairs. Something—developmental rules, activity-dependent plasticity, shared inputs—has driven certain pairs to connect in both directions, and often to be strong in both directions. This indicates structure, not noise. There are caveats. These are developing rats. The focus is a single cell type—thick-tufted layer five pyramids—in one cortical area. The recordings are taken from slices, with all the limitations this implies. And sampling, no matter how meticulous, is never all-seeing. Song and colleagues acknowledge these points. However, the core signals—excess reciprocity, overrepresented motifs, heavy-tailed strengths, and correlated weights within reciprocal pairs—persist through a long list of stress tests. They are not fragile. What does this wiring strategy offer the brain? The skeleton-backbone analogy serves as a useful guide. Strong, reciprocally connected clusters can ignite and sustain activity, producing stable patterns and rapid reverberations. Meanwhile, the vast network of weak connections can modulate, gate, and integrate, especially when activity is correlated across populations. Think of the backbone as the steel frame of a building, and the weaker links as the walls, doors, and plumbing. You need both, but they serve different roles in how forces flow. Where does this lead next? Two clear frontiers lie at opposite ends of the scale. On the micro side, denser reconstructions—either by using more simultaneous recordings or modern connectomics—can determine whether the same motif framework appears across layers, cell types, and species. On the macro side, in vivo recordings can test whether those strong, reciprocal clusters truly dominate dynamics during natural activity, as the weight statistics indicate. The map is beginning to sharpen. The next step is to observe this framework in action.

They achieved this through hard work, with hundreds of simultaneous quadruple whole-cell recordings from thick-tufted layer five pyramidal neurons in rat visual cortex. This approach provided insights into who connects to whom, and how strongly, all in one snapshot.

Here's the setup. They recorded from rats between postnatal day 12 and 20, ensuring the slices remained healthy and warm. They pushed four pipettes into four nearby pyramidal neurons at once.

From each presynaptic cell, they evoked spikes and averaged the postsynaptic responses—excitatory postsynaptic potentials, or EPSPs—over about one to two minutes. Most pairs had between 45 to 60 trials; a few had as few as 10 but still met their quality standard. The signal-to-noise ratio was good enough to detect connections down to approximately one hundredth of a millivolt.

Overall, they recorded 817 cells arranged into 83 triplets and 142 quadruplets, and across this dataset, the network was sparse: 931 connections out of 8,050 possible, which is about 11.6 percent.

The first big result hits like a drumbeat: reciprocity is high. If this were a random network with the same overall connection probability, bidirectional links would be rare. Instead, Song and colleagues found about four times more reciprocal pairs than this simple model predicts, and Monte Carlo tests indicate that the probability of observing this excess by chance is below one in ten thousand.

Why does that matter? Because it suggests the cortex isn't creating connections independently. Some pairs really prefer to wire up in both directions, creating feedback loops at the scale of two neurons.

You can explore a result like that in various ways to ensure its validity. Are closely located cells more likely to be connected, and did they just happen to record more of them? They measured distances between labeled neurons and saw no systematic dependence of connection probability on distance; a chi-square test produced a p-value of 0.21.

Maybe shallow recordings behave differently from deep ones? Across a depth of 10 to 100 micrometers below the slice surface, connectivity remained unchanged. What about age within that postnatal day 12 to 20 window?

Neither the mean EPSP amplitude nor overall connectivity showed strong age effects, and repeating the analysis in a tighter window of postnatal days 14 to 16 yielded the same reciprocity increase. Even axonal cuts—inevitable in slices—did not correlate with stronger or more frequent connections. It's significant work, but it eliminates possible confounding factors one by one.

Then they enlarged the question. If pairs aren't random, what about triads—three-neuron patterns? In this case, they exercised caution.

They sorted all triplets into 16 classes, then constructed a null model that preserves the observed likelihood of unidirectional or bidirectional links between any two cells, but randomizes how those links combine into a triplet. It's akin to dealing cards from the same deck, but shuffling which cards end up together. Some motifs—the densely connected ones in particular—showed clear overrepresentation against that benchmark.

When they increased the threshold, focusing only on stronger synapses, the clustering became even more pronounced. Strong triads tend to cluster together.

For four-neuron patterns, the bar is set higher, as there are many more ways to connect four nodes. Thus, they constructed a larger random matrix, extracted countless four-by-four submatrices, and employed simulated annealing to match the triplet statistics of the real data. Even against that rigorous backdrop, one class of "chain-like" quadruplets appeared significantly more common—about 70 percent more frequent than expected, with a corrected p-value of 0.004.

This is important as it indicates that the nonrandomness isn't merely a quirk of pairs or triads; some larger motifs are also making an appearance, even when you insist that the building blocks have the same pairwise statistics.

Topology is only half the story. The other half is weight—the size of those EPSPs. The mean strength across all detected connections was 0.77 millivolts.

However, the distribution's shape is particularly striking. Instead of a tidy bell curve or a quick exponential fall-off, the data display a heavy tail: strong synapses are far more common than those simpler models would predict. When modeled as lognormal, it fits; attempts with exponential or Gaussian models do not.

This tail is significant because it carries the weight of the network. Connections above 1.2 millivolts—just 17 percent of the synapses—contributed about half of the total synaptic weight. A small minority of very strong connections dominate the overall weight.

You can sense the logic starting to form. If some pairs are reciprocally connected more often than chance, and if a minority of edges are very strong, those two facts can overlap. They indeed do.

Bidirectionally connected pairs had higher mean strengths than unidirectional pairs—0.95 millivolts compared to 0.61—and that gap was not a coincidence; the p-value was 3.1 times ten to the minus seven. Even more telling, the two directions within a reciprocal pair rose and fell together. The correlation of their strengths, measured on a logarithmic scale, was around 0.36 with a p-value below 0.0001.

When examining unrelated pairs that share no neurons, the correlation essentially disappears. The strong connections cluster in the same areas rather than spreading out evenly.

If you're curious why the weights are lognormal in the first place, Song and colleagues suggest a clear mechanistic explanation. Imagine that every time two neurons fire together, the synapse between them gets adjusted up or down by some factor—not a fixed amount, but a multiplying factor. If the strength after episode i is represented as w sub i, and each episode multiplies that by some factor F sub i, then after n episodes, you have w sub n equal to the initial strength times the product of all those F's.

Taking the logarithm changes the product into a sum. If these log factors have some finite mean and variance, the central limit theorem comes into play, and the log of the weight becomes approximately Gaussian. Exponentiate it back, and you arrive at a lognormal distribution.

It’s multiplicative noise through and through, and this is exactly the signature seen in the data.

Now, pause for a moment. What emerges from all of this is a picture of the cortex that isn't simply random wiring with random weights. It's a sparse network where a handful of strong synapses account for a disproportionate share of the total.

Reciprocal pairs are both common and powerful, and clustered motifs appear more frequently than by chance. In simple terms, it looks like a framework of strong connections embedded in a vast sea of weaker ones. The sea is extensive; it may play a significant role when neurons fire together. But the framework is what carries much of the load.

The team even offers some intuition regarding how rare the very strongest links are. If you look for suprathreshold EPSPs—let's say, above ten millivolts—the lognormal fit indicates that they are extremely infrequent, on the order of one in a thousand connections. In a network of the size they considered, that translates to only a few dozen such exceptional connections.

This underscores just how uneven the weight landscape is: many small ripples, with a few towering waves.

All of this confidence relies on ensuring that the statistics aren't skewed by how the tissue is sampled. That's why those control analyses are significant. Distance between recorded neurons didn't clarify the patterns, nor did the depth of the cells in the slice.

Proximity to an axonal cut end—a potential confounding factor in slices—did not predict connection probability or strength. Age, within that postnatal day 12 to 20 window, supported the key findings; even narrowing down to postnatal days 14 to 16 preserved the same overrepresentation of reciprocity and the same heavy tail. They also employed rigorous null models—preserving observed pairwise probabilities when testing triplets, and matching triplet counts when testing quadruplets—and corrected for multiple hypothesis tests to ensure that a few positive motifs did not emerge by chance alone.

If you enjoy a bit of mathematics with your neuroscience, there's one more conceptual twist worth appreciating. In the simplest random graph, each possible directed edge is present with some fixed probability p, independent of every other edge. Under that model, the expected number of bidirectional pairs is the number of pairs times p squared, where N is the number of possible pairs.

That's a straightforward baseline. The fact that the cortex has four times that expectation suggests that independence does not hold at the level of pairs. Something—developmental rules, activity-dependent plasticity, shared inputs—has driven certain pairs to connect in both directions, and often to be strong in both directions. This indicates structure, not noise.

There are caveats. These are developing rats. The focus is a single cell type—thick-tufted layer five pyramids—in one cortical area.

The recordings are taken from slices, with all the limitations this implies. And sampling, no matter how meticulous, is never all-seeing. Song and colleagues acknowledge these points.

However, the core signals—excess reciprocity, overrepresented motifs, heavy-tailed strengths, and correlated weights within reciprocal pairs—persist through a long list of stress tests. They are not fragile.

What does this wiring strategy offer the brain? The skeleton-backbone analogy serves as a useful guide. Strong, reciprocally connected clusters can ignite and sustain activity, producing stable patterns and rapid reverberations.

Meanwhile, the vast network of weak connections can modulate, gate, and integrate, especially when activity is correlated across populations. Think of the backbone as the steel frame of a building, and the weaker links as the walls, doors, and plumbing. You need both, but they serve different roles in how forces flow.

Where does this lead next? Two clear frontiers lie at opposite ends of the scale. On the micro side, denser reconstructions—either by using more simultaneous recordings or modern connectomics—can determine whether the same motif framework appears across layers, cell types, and species.

On the macro side, in vivo recordings can test whether those strong, reciprocal clusters truly dominate dynamics during natural activity, as the weight statistics indicate. The map is beginning to sharpen. The next step is to observe this framework in action.

Related lectures