An extremely rich repertoire of bursting patterns during the development of cortical cultures

A small dish of neurons sits in an incubator. They've just been plated, taken from the cortex of a rat embryo, separated from each other, and set down on a grid of electrodes. They don't know each other yet. They're not doing anything. Over the next five weeks, something will happen. And for a long time, researchers thought they knew what it was. They were wrong about almost all of it. Dissociated cortical cultures are one of neuroscience's most useful tools. You take cortical tissue from an embryonic rat, break it apart enzymatically, and plate the cells onto a multi-electrode array, or MEA, which is essentially a small chip embedded with fifty-nine extracellular electrodes, each thirty micrometres across, arranged in a square grid with two hundred micrometres between them. The electrodes record action potentials from nearby neurons and can also deliver electrical pulses to stimulate them. The result is a living neural network you can watch form in real time, manipulate pharmacologically, and record from at dozens of sites simultaneously. Scientists use these cultures to study network formation, spontaneous activity, plasticity, and disease. They're simpler than an intact brain— controllably so. The problem was that most prior work studied very few cultures, focused on narrow slices of activity, and arrived at a reasonably tidy picture. Population bursts— brief windows when many neurons fire together, sending the array-wide spike rate sharply above baseline— were known to be a dominant feature. Different groups described different signatures: avalanche-like distributions, calcium transients, increasing burst complexity after two weeks, a shift toward shorter sharper bursts after a month. But each study was working with limited samples and inconsistent terminology. What if the picture wasn't tidy at all? That's the question Daniel Wagenaar, Jerome Pine, and Steve Potter set out to answer with brute empirical force. They followed fifty-eight cultures grown from eight different dissections over nine months, at five different plating densities ranging from three thousand to fifty thousand neurons on areas of thirty to seventy-five square millimetres. They recorded for five weeks, collecting nine hundred sixty-three half-hour sessions and thirty-six overnight recordings. To detect bursts reliably, they used an algorithm called SIMMUX, which searched each electrode trace for sequences of at least four spikes with inter-spike intervals below a density-dependent threshold, then grouped any overlapping multi-electrode sequences into a single burst event. This was careful, systematic work at a scale the field hadn't attempted. Plating density turned out to matter enormously. The aggregate spike detection rate scaled linearly with density— straightforward because more cells near the electrodes means more recorded spikes. But the timing of everything else was density-dependent in ways that weren't obvious. In dense cultures, individual neurons started firing around four to five days in vitro, and array-wide population bursts appeared by five to seven days. Sparser cultures lagged behind on both counts. During those bursts, the array-wide spike detection rate could jump a hundredfold over baseline. These weren't subtle fluctuations— they were massive, synchronized events. Stimulation-response measurements showed why: in dense cultures, functional projections grew rapidly in the first week and spanned the entire one point seventy-two millimetre electrode array within fifteen days. Sparse cultures wired up more slowly. A denser network is a faster network— not just in its activity, but in its physical formation. Two other developmental signatures changed with age regardless of density. Average burst duration fell from about one second when bursts first appeared to under two hundred milliseconds after twenty days in vitro. And burst onset time— the window between when the first electrode fires and when the whole array is recruited— shrank from roughly three hundred milliseconds down to about thirty milliseconds. The network was getting better at coordinating itself. Now here's where the story departs from what anyone expected. After about two weeks, most cultures were dominated by population bursts. Previous reports suggested this was where development more or less settled— a mature, recurring rhythm. Wagenaar and colleagues found the opposite. Development kept going. Burst patterns continued to change throughout the five-week observation period, and they were extraordinarily varied. To get a handle on that variety, the team built a classification system. Any burst spanning fewer than five electrodes was called tiny and set aside. For larger bursts, they defined a reference count— the number of spikes in the third-largest burst in a recording, which they called N-star— and sorted bursts into large, medium, and small based on their spike counts relative to that value. Bursts were flagged as long-tailed if the array-wide firing rate stayed at least fifty percent above baseline for three or more seconds after the burst peak. A recording was labeled long-tailed when at least half its large and medium bursts had that property. Burst-rate variability was also quantified: if the highest burst rate in a recording differed from the lowest by a factor of ten or more, that recording was classified as highly variable. The most striking pattern they identified was the superburst. These are tight clusters of bursts— typically four to twelve in a row— separated from each other by several minutes of quiet tonic firing. The bursts inside a superburst come in rapid succession; the gaps between superbursts are at least ten times longer than the gaps within them. Superbursts appeared in roughly half the cultures. Within superbursts, the internal shape mattered too: some showed spike counts decaying across successive bursts, which the team called normal, while others showed spike counts growing, which they called inverted. To capture how bursty a recording was with a single number, they computed what they called a burstiness index. They counted spikes in non-overlapping one-second windows, found the fraction of total spikes contained in the most active fifteen percent of those windows— a value they called f-fifteen— then subtracted zero point fifteen and divided by zero point eighty-five. The result is zero for continuously firing cells and one when all activity is concentrated in bursts. It's an elegant compression of a complex signal. Inter-burst intervals across all recordings ranged from one second to three hundred seconds. Different cultures showed different amounts of temporal clustering. Some were highly regular; others were wildly variable. And these patterns weren't fixed— they shifted over the course of development, sometimes dramatically from week to week in the same culture. Which raises an obvious question: how much of that variability is real, and how much is just noise? Wagenaar and colleagues measured this directly. They computed a difference index, or DI, from pairwise comparisons of recordings, and built three comparison types: day-to-day DIs from the same culture on consecutive days, sister-culture DIs between cultures from the same plating batch recorded on the same day, and cross-plating DIs between cultures from different batches at the same developmental age. The finding was clear. Sister cultures were only marginally more different from each other than the same culture was from itself on consecutive days. But cultures from different platings were substantially more different— and that gap was statistically significant at almost every age examined. Batch identity is the dominant source of variability in these systems, and it grows with age. The methodological implication is stark. If you study a handful of cultures from a single preparation, you are not characterizing the phenomenon. You are characterizing one idiosyncratic sample of it. Wagenaar and colleagues put it plainly: report not just how many cultures you used, but how many platings they came from. This is a critique that applies to a substantial fraction of the existing literature. The authors close by pointing outward. They suggest that the richness of these dynamics, rather than being a complication, is precisely what makes cultures useful. By mapping the full range of activity patterns in vitro and comparing them with healthy and diseased brain development in vivo, researchers can build more relevant models. To make that possible, the team released their entire dataset publicly— forty-five gigabytes of spike waveforms, four gigabytes of reduced timestamp files, and example MATLAB code. This was an early and generous act of data sharing in systems neuroscience. Open questions remain. Why do older cultures produce fewer stimulus-evoked bursts than younger ones? Why does the number of spikes in a burst scale exponentially rather than linearly with the number of participating electrodes? Why do sister cultures follow more similar trajectories than non-sisters? The recordings exist. Others can look. The dish of neurons was not doing something simple. It was doing something the field didn't yet have the vocabulary to describe. Wagenaar, Pine, and Potter spent five weeks and nearly a thousand recordings building that vocabulary— and what they found was not a tidy story, but a genuinely strange and varied one. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.