Synaptic Pruning: The Adolescent Brain Deletes Half

Introduction

Somewhere around your second birthday, your brain held more connections than it would ever hold again. Not more neurons. The neuron count was already close to final. What you had in surplus were synapses, the tiny junctions where one cell passes a signal to the next, and you had built them at a rate that no engineer would call efficient. Then your brain started taking them apart.

That process is called synaptic pruning, and it is one of the strangest facts about how human beings become themselves. We do not build up to adulthood. We build past it, then carve back. A four-year-old's frontal cortex is denser with connections than a thirty-year-old's. The teenager who cannot plan a weekend is not working with an unfinished brain so much as an overstocked one, still deciding what to throw out [35].

Here is the part that made this article difficult to write. The headline above says the adolescent brain deletes half of what it built. You will find that claim, or something close to it, on nearly every page that currently ranks for this topic. Some say half the synapses vanish between ages two and ten. Some say half of everything is gone by adulthood. Some say half, without saying half of what.

Almost none of them cite the measurement.

So this article does something the popular explanations skip. Before anything else, it goes back to the papers where these numbers were first produced, checks what was counted, in which region, in how many brains, using which method, and then reports honestly how much of the famous figure survives contact with the original data. The short version: the number is real, but it does not mean what most people repeat, and the difference matters more than it sounds.

What follows is the whole story. The 1979 autopsy study that first counted synapses in human frontal cortex. The 1997 follow-up that showed different brain regions run on completely different schedules. The 2011 paper that pushed the endpoint of prefrontal pruning into the third decade of life. The immune molecules that mark a connection for removal, the cells that carry it away, and the molecular brake that stops them from taking too much. A January 2026 result that suggests adolescence is not only about deletion. And finally the question that matters most to anyone who studies for a living: does any of this have anything to do with why you forget what you learned last week?

Let us deal with this immediately, because everything downstream depends on it.

The claim that the brain loses about fifty percent of its synapses traces back most often to Peter Huttenlocher's 1979 study of human frontal cortex [1]. Open that paper and you find a specific sentence. The childhood peak in synaptic density sits about fifty percent above the adult mean.

Read that carefully. Above the adult mean.

If the peak is fifty percent higher than the adult value, then the peak is roughly one and a half times the adult level. Falling from one and a half back down to one is a decline of about a third, not a half. The famous fifty percent is a description of how far the peak rises, and somewhere in the retelling it flipped into a description of how much gets deleted. Those are different numbers. One is about one third larger than the other.

So is the fifty percent figure simply wrong? No. It exists, but it lives somewhere else.

Urte Neniskyte and Cornelius Gross, writing a review in Nature Reviews Neuroscience in 2017, state that activity-dependent synapse elimination reduces synaptic density by about fifty percent, and add that in primates roughly seventy percent of the axons crossing the corpus callosum are eliminated after birth [4]. That is a real, citable, peer-reviewed statement of the number. But notice what kind of statement it is. It is a review-level summary averaged across regions and studies. It is not the output of one measurement in one place.

And regional averaging hides an enormous amount. Reviews of adolescent brain development describe losses approaching half of the synaptic connections in some regions with very little decline in others. The brain does not prune uniformly. Talking about a single percentage for the whole organ is a bit like giving one average rainfall figure for an entire continent.

There is a third claim that gets folded into the same sentence, and it is about a completely different thing. Roughly half of the neurons a mammalian embryo generates do not survive to birth. That is programmed cell death, apoptosis, and it happens largely before you were born. It concerns whole cells, not connections. When a popular article writes that the brain loses about half of what it makes, it is often quietly borrowing the emotional weight of the neuron statistic and attaching it to the synapse story. They are separate processes with separate timelines.

So the honest version of the headline reads like this. The developing cortex builds far more synapses than it keeps. In some regions, close to half are eliminated between childhood and adulthood. In the specific frontal region where the counting was first done, the drop from peak to adult plateau looks closer to a third. And "adolescence" is the wrong window for a lot of it, because in some areas the heaviest elimination happens well before puberty and in others it continues for a decade after.

The title of this article is a hook. This section is the correction. Both can be true, and the second one is the reason to keep reading.

1979: the year someone actually counted

Before Huttenlocher, nobody had numbers.

The idea that the developing brain overproduces and then trims had been floating around since the 1960s, supported mostly by animal work and inference. What was missing was a human count. How many synapses per unit volume, at what age, in a real human cortex.

Huttenlocher went and counted them. The method was unglamorous and slow. Post-mortem tissue from twenty-one brains spanning newborn to ninety years, stained with phosphotungstic acid, which binds preferentially to the dense material at synaptic junctions, then examined under an electron microscope. He sampled layer three of the middle frontal gyrus. Tissue had to be collected within about thirty-six hours of death for the counts to hold up [1].

The adult figure he arrived at was about 11.05 times ten to the eighth synapses per cubic millimetre, in the age range sixteen to seventy-two. In the oldest brains, seventy-four to ninety years, density had dropped to around 9.56 times ten to the eighth, a statistically significant decline based on four specimens. The peak sat at roughly one to two years of age, about fifty percent above the adult mean.

Twenty-one brains. Read that again.

This is the single most cited quantitative claim in the entire field, and it rests on twenty-one autopsy specimens, one cortical region, one cortical layer, sampled cross-sectionally. Each brain contributes one data point at one age. Nobody was followed over time, because you cannot follow a person over time using autopsy tissue. The curve everyone draws from this data is a curve fitted through single observations from different individuals, each of whom had a different genome, a different childhood, and a different cause of death.

Huttenlocher himself was careful about this. He extended the work through the 1980s, and by 1990 he had published a morphometric study covering broader aspects of human cortical development, still working within the same constraints [7]. The limitations were never hidden. They were simply lost somewhere between the journal and the internet.

Does that make the finding unreliable? Not at all. The overproduction-then-elimination pattern has been confirmed repeatedly, in other species, with other methods, by other labs. Jean-Pierre Bourgeois and Pasko Rakic did the equivalent work in macaque primary visual cortex, tracking synaptic density from fetal stages to adulthood and establishing the overproduction and elimination sequence in a primate model where developmental timing could be controlled far more precisely than in human autopsy material [6]. The phenomenon is real.

What is fragile is the precision. A number produced from twenty-one brains, one region, one layer, using a stain-and-count method with its own known biases, should be quoted with a range and a caveat. Instead it gets quoted as though someone measured the whole brain.

Not one clock, but many

The second Huttenlocher paper is the one that should have changed how everyone talks about this, and mostly did not.

In 1997, working with Arun Dabholkar, he applied the same electron microscopy approach to two different cortical regions and compared their timing [2]. The result was that the brain does not have a pruning schedule. It has several, running at once, finishing at different times.

Auditory cortex, sampled in Heschl's gyrus, reached maximum synaptic density around three months after birth. Prefrontal cortex, sampled in the middle frontal gyrus, did not peak until after fifteen months. Net elimination in auditory cortex was essentially finished by around age twelve. In prefrontal cortex it ran into mid-adolescence.

Three months versus fifteen months for the peak. That is a fivefold difference in timing between two regions of the same brain.

This staggered pattern, sometimes called heterochronous development, follows a logic that becomes obvious once you see it. Sensory systems settle first. They handle inputs that are relatively stable across all human environments, and there is an evolutionary advantage to getting them tuned early. Association cortex, the territory of planning and inhibition and social reasoning, stays open far longer, because what it needs to encode depends heavily on the specific world you happen to land in.

There is a detail in the 1997 paper worth flagging, because it undercuts a lot of confident writing about the teenage brain. The adolescent age range in that dataset contained only four data points, and the values were variable enough that the authors could not confidently define an adolescent trajectory from them. Much of the popular story about a dramatic adolescent synaptic collapse is extrapolated across a gap in the human data, filled in with animal work and MRI.

That is not a scandal. It is normal science working with the tissue it can ethically obtain. But when a health site tells you flatly that half your synapses disappear during your teenage years, the underlying human synapse counts for that specific window are thinner than the confidence of the sentence suggests.

Reviews of normal circuit development have made the same point for years, describing pruning as a regionally staggered, protracted process rather than a single event with a start and an end date [8]. The single-clock version is a simplification that survived because it fits in a headline.

The spines that refuse to grow up

Then, in 2011, Zdravko Petanjek and colleagues published a result that pushed the endpoint later than almost anyone expected [3].

They used the rapid Golgi method, which impregnates a small random subset of neurons with silver chromate so that individual dendrites and their spines become visible in full. Dendritic spines are the small protrusions where most excitatory synapses sit, so spine density serves as a structural proxy for synapse number. The sample covered thirty-two subjects from one week old to ninety-one years, all in dorsolateral prefrontal cortex.

Their central finding, stated in their own terms: dendritic spine density in childhood exceeds adult values by two to threefold, and begins to decrease during puberty.

Note what that is not. It is not a fifty percent figure. Petanjek reports a ratio between childhood peak and adult plateau, and the honest way to carry that number forward is as a ratio. Some later modelling papers have converted that ratio into a percentage decline, and those secondary estimates land somewhere in the forty to fifty percent range, but that arithmetic belongs to the modellers, not to Petanjek. Attributing a clean fifty percent to this paper misrepresents it.

The timing is what made the paper matter. In layer IIIc, spine density peaked between roughly two and a half and seven years. Decline in basal and proximal dendrites started around seven to nine years. Distal apical dendrites did not begin declining until about seventeen. And adult levels were not reached until around age thirty, after which they stayed stable.

Age thirty.

If you have ever seen the claim that the brain finishes developing at twenty-five, this is one of the datasets that quietly contradicts it. In human prefrontal cortex, the structural remodelling of excitatory connections appears to run through the entire third decade. Not intensely, not dramatically, but measurably.

There is a further nuance in that paper that rarely gets repeated. Different cortical layers behave differently. Layer IIIc carries more overproduced spines than layer V, both during development and in adulthood, and the deeper layers show more modest pruning. Even within a single region of a single cortex, the process is not uniform across depth.

The limitations here mirror the 1979 study. Post-mortem tissue. Cross-sectional. Thirty-two subjects across ninety-one years of lifespan, which means very few individuals per age band. The Golgi method impregnates only a fraction of neurons, and whether that fraction is representative is an assumption rather than a demonstrated fact. Spine density is a proxy for synapse number, not a direct count of synapses.

None of that invalidates the result. It does mean that the age-thirty endpoint should be held as a well-supported estimate from a small, careful human sample, rather than as a precise biological constant.

What brain scans can and cannot tell you

Here is where a large amount of popular writing goes wrong, and it is worth being blunt about it.

In 2004, Nitin Gogtay and colleagues published a longitudinal MRI study that has become one of the most reproduced images in developmental neuroscience [5]. Thirteen children, scanned every two years for eight to ten years, ages four to twenty-one. The resulting animation shows grey matter density decreasing in a wave that starts at the back of the brain and sweeps forward, with higher-order association cortex maturing after lower-order sensory and visual areas.

It is beautiful, it is genuinely longitudinal, and it is very frequently described as showing synaptic pruning happening in real time.

It does not show that.

MRI grey matter signal is a composite. It reflects synapses, yes, but also dendritic arbors, glial cell populations, vascular changes, and above all myelination, which increases through the same period and changes the tissue contrast that defines the grey and white matter boundary. When a scan shows cortical thinning, several biological processes could produce that signal, alone or together. Synapse elimination is one candidate. It is not the only one, and no MRI sequence in routine use resolves individual synapses.

So the correct statement is that MRI shows a maturational trajectory consistent with, among other things, synaptic pruning. The incorrect statement, which appears constantly, is that MRI shows the brain losing synapses.

The distinction matters because the two literatures get merged. Someone reads Huttenlocher's autopsy counts, then reads Gogtay's scan data, and combines them into a single confident story about the teenage brain shedding connections on camera. In fact one is a small cross-sectional count of actual synapses, the other is a longitudinal measure of bulk tissue signal, and the bridge between them is inference.

There is now a way to look at synaptic density in living human brains, using PET imaging with a tracer that binds SV2A, a protein found in synaptic vesicles. Ellis Onwordi and colleagues used it to show reduced synaptic density markers in people with schizophrenia [41]. That technique is a real advance, and it is the most plausible route to eventually watching developmental trajectories in living people rather than reconstructing them from autopsy. But it is a marker of vesicle protein, still an indirect measure, and it has not yet produced the longitudinal developmental dataset the field wants.

Here is a compact view of what each of the foundational measurements actually delivered.

Read the last column across all five rows and the problem becomes visible. There is no single number, because there was never a single measurement.

How a connection gets marked for deletion

Knowing that synapses disappear is one thing. Knowing how the brain decides which ones, and what physically removes them, took another three decades.

The answer, when it arrived, came from an unexpected direction. The immune system.

In 2007, Beth Stevens, working in Ben Barres's lab at Stanford, published a paper in Cell that reframed the field [9]. The subject was complement, an ancient cascade of proteins whose day job is tagging bacteria and damaged cells for destruction by immune cells. Stevens found that C1q, the protein that initiates the classical complement cascade, is expressed by developing neurons in response to signals from immature astrocytes, and that it localises to synapses throughout the postnatal central nervous system and retina.

Then came the test. In mice lacking C1q or C3, the refinement of connections between the retina and the thalamus failed. The paper describes large, sustained defects in central nervous system synapse elimination. Take away the tag, and the pruning does not happen properly.

The model that emerged is elegant. Weak or less active synapses get coated with complement proteins, essentially a molecular label reading remove this. Then something has to read the label.

That something is microglia, the brain's resident immune cells. For a long time they were thought of as passive caretakers that woke up only during injury or infection. Work in the late 2000s showed otherwise. Hiroaki Wake and colleagues found that resting microglia are anything but resting, constantly extending processes and directly contacting synapses, monitoring their functional state [13].

In 2011, Rosa Paolicelli and colleagues published direct evidence in Science that microglia engulf synaptic material during postnatal hippocampal development, and that mice with disrupted microglial signalling showed transient deficits in synaptic connectivity [10]. Microglia were not bystanders. They were participants. Work from the same line of research went on to show that when neuron-microglia signalling is deficient, the cost is not confined to the synapse. It shows up as impaired functional brain connectivity and altered social behaviour [20].

The following year, Dorothy Schafer and colleagues closed the loop in Neuron [11]. Working in the developing visual system, they showed that microglia engulf retinal inputs in a manner dependent on both neural activity and the complement receptor CR3, and, critically, that they preferentially engulf the less active inputs.

That result deserves a moment. The slogan "use it or lose it" had been repeated for decades as a metaphor. Schafer's paper turned it into a mechanism. Activity determines the tag. The tag determines the engulfment. Less activity, more complement, more likely to be eaten.

A 2012 review by Alexander Stephan, Ben Barres and Beth Stevens laid out the full argument for complement as a developmental pruning system, and it remains the clearest synthesis of how an immune pathway ended up sculpting circuits [12].

Here is the pathway in its simplest form.

That diagram is a simplification of a system with many more components, but the logic holds. Activity protects. Silence marks. Immune cells collect.

The brake nobody talks about

A tagging system that only knows how to say remove this would be dangerous. Something has to say leave this alone.

In 2018, Emily Lehrman and colleagues identified that counter-signal [14]. Neurons display a surface protein called CD47, which binds to a receptor called SIRPalpha on microglia. The message is inhibitory. Do not eat.

Both proteins turn out to be enriched in the visual thalamus precisely during the peak period of pruning, which is exactly where you would want a brake if you were designing the system. In mice lacking CD47, microglia engulfed more, pruning went too far, and synapse numbers stayed reduced.

The elegant part is where CD47 sits. It localises to active synapses. So the same variable, activity, drives both arms of the system at once. Active connections put up a protective signal and avoid complement tagging. Quiet connections do neither.

This brake is not just a developmental curiosity. Xin Ding and colleagues showed in 2021 that losing microglial SIRPalpha promotes synaptic pruning in preclinical models of neurodegeneration [15]. A system built to sculpt a young brain can, if its brake fails later, start dismantling a mature one.

And microglia are not the only cells doing the eating. In 2013, Won-Suk Chung and colleagues showed in Nature that astrocytes also engulf synapses, using two phagocytic pathways called MEGF10 and MERTK [16]. Knock out either pathway alone and engulfment dropped by roughly half. Knock out both and it dropped by about eighty-five percent. Astrocyte phagocytosis peaked in early postnatal days and, importantly, continued into adulthood for both excitatory and inhibitory synapses.

So the brain has at least two independent cellular systems for removing connections, running on overlapping but distinct molecular machinery, with a built-in protective signal for the connections worth keeping.

Three ways to take a connection apart

Removal is not one operation. Structurally, there are several distinct ways a connection can be undone, and confusing them is a common error.

The first is axon retraction. The branch simply pulls back, withdrawing from its target without dying. The neuron survives intact and keeps its other connections.

The second is axon degeneration, a local breakdown resembling the process seen in injured nerves, where the branch fragments in place and the debris is cleared away.

The third is axosome shedding, described by Derron Bishop and colleagues in 2004 [17]. The retracting branch sheds membrane-bound packets of itself, which are then taken up by surrounding glial cells. Retraction, in other words, is not a clean rewind. It leaves litter, and the litter is collected.

Liqun Luo and Dennis O'Leary's 2005 review remains the standard treatment of how these modes differ and when each one is used [18].

All three of these differ fundamentally from apoptosis. In apoptosis, the entire neuron dies and every one of its connections goes with it. In pruning, the neuron lives. It loses branches the way a tree loses branches, and keeps growing.

This is the distinction that collapses in casual writing about the brain deleting half of itself. Losing half your synapses in a region and losing half your neurons are not remotely the same event, and they happen in different decades of life.

The uncertainty almost nobody mentions

Everything above is standard textbook material now. So here is something that is not.

In 2024, Marta Pereira-Iglesias and colleagues posed an uncomfortable question in Nature Neuroscience about the whole microglial pruning story [19]. The framing was hunters versus gatherers. Do microglia actively sever functional synapses, or do they mostly collect the remains of synapses that were already being dismantled by other means?

The reason the question is open is a gap in the evidence. Most of the case for microglial pruning rests on finding synaptic material inside microglia, on genetic knockouts that impair refinement, and on fixed-tissue imaging. What has proven remarkably hard to capture is live imaging showing a microglial process gripping an intact, functional synapse and cutting it away.

That does not overturn the model. The genetic evidence is strong, and the correlation between complement signalling and synapse loss is reproducible across labs and systems. But it does mean the causal step at the heart of the popular explanation, the moment of severing, is inferred rather than directly observed in most systems.

Microglial biology has also turned out to be more varied than early work suggested. Reviews of microglial signatures show these cells adopting different states in different regions and conditions [22]. Some of their functions have nothing to do with eating anything. Ana Badimon and colleagues showed in 2020 that microglia exert negative feedback control over neuronal activity itself, a regulatory role rather than a demolition one [21].

You will almost never see this qualification in a popular explanation of pruning. The standard summary hands you microglia eating synapses as settled fact with a tidy mechanism attached, and moves on. It is a strong hypothesis with substantial support and one important observational gap. That is a different thing.

What "use it or lose it" actually means

The phrase is everywhere, and it is doing a lot of work while explaining very little. What does activity dependence actually specify?

Not simply whether a synapse fires. What matters is the pattern, the relative timing, and the strength of activity, especially compared to neighbouring inputs competing for the same target. Yang Dan and Mu-ming Poo's work on spike-timing-dependent plasticity showed how precisely timing matters, with millisecond-scale differences in the order of pre- and postsynaptic firing determining whether a connection strengthens or weakens [29].

The cleanest place to watch this play out is not in the brain at all. It is in muscle.

Dale Purves and Jeff Lichtman's 1980 review in Science pulled together the evidence that reducing the number of axons contacting each target cell is a general feature of development, and suggested it might underlie the progressively restricted flexibility of the maturing nervous system [23]. At the neuromuscular junction, the process is unusually visible. A newborn muscle fibre receives input from several motor neurons at once. Over development, all but one are eliminated. Single innervation is the adult state.

Why is this the ideal model system? Because you can see every player. The synapse is large, accessible, and there is exactly one winner per fibre. You do not have to infer the outcome. You can look at it.

Rita Balice-Gordon and Lichtman then did the decisive experiment in 1994 [24]. They blocked receptors at one specific patch of a junction while leaving the rest functioning. The blocked region was eliminated. The active region survived. Same cell, same muscle fibre, same moment in time, different fates determined by local activity.

Later work from Stephen Turney and Lichtman added a twist that complicates the simple story [25]. Under the right experimental manipulations, the outcome of synapse elimination could be reversed, meaning the process is a competition with a decidable winner rather than an irreversible verdict handed down early.

Then there is the visual system, where this principle was first established in the most famous form. David Hubel and Torsten Wiesel's work on the period of susceptibility in the developing visual cortex showed that closing one eye during a specific window produced lasting reorganisation of cortical territory in favour of the open eye [26]. Outside that window, the same deprivation did far less. That is where the concept of a critical period comes from, and it earned a Nobel Prize in 1981.

The cerebellum offers another clean model. Purkinje neurons start life receiving input from multiple climbing fibres and end up with one. Work published in Science Translational Medicine linked failures of that specific pruning process to abnormal cerebellar oscillations and tremor, which is a useful reminder that pruning is not a vague background process but a step whose failure produces specific, measurable dysfunction [30].

Does weakening mean deletion?

Here is a question that sounds academic and turns out to be central to the learning question later in this article.

Synapses can be weakened without being removed. Long-term depression, the counterpart to long-term potentiation, reduces the strength of a connection. Robert Malenka and Mark Bear's review remains the standard account of both processes and how they relate [27]. If you want the strengthening side of the story in more detail, the mechanics of long-term potentiation are worth understanding on their own terms.

So does weakening lead to structural elimination, or is it a separate state a synapse can sit in indefinitely?

Simon Wiegert and Thomas Oertner addressed this directly in 2013 [28]. Their finding was more specific than a simple yes or no. Long-term depression triggered the selective elimination of weakly integrated synapses. Not all depressed synapses were removed. The ones that went were the ones that were poorly connected into the surrounding network to begin with.

That is an important qualification. Weakening alone does not guarantee removal. Weakening plus isolation does. A connection that has been depressed but sits within a well-integrated cluster of other active connections appears to be substantially more protected than one sitting alone.

Which points toward something with real practical weight. Context protects. A fact you learn in isolation, with no connection to anything else you know, may be structurally more vulnerable than the same fact embedded in a web of related knowledge. That is not a new idea in learning science, but it is interesting that the synaptic data points the same direction.

January 2026: the teenage brain is also building

For roughly forty years the adolescent story has been told as subtraction. Then, in January 2026, a paper appeared that complicates it.

Ryo Egashira and colleagues in Takeshi Imai's group at Kyushu University published in Science Advances a study of dendritic spine distribution in layer 5 extratelencephalic-projecting neurons [31]. They used SeeDB2, a tissue-clearing method the same group developed, combined with super-resolution microscopy, which allowed them to map spines across entire dendritic trees rather than sampling fragments [32].

What they found was a hotspot.

In adult mice, spines were not evenly distributed along the apical dendrite. They clustered densely in a middle compartment, the same region where dendritic calcium spikes are generated. In two-week-old mice, that clustering did not exist. Spines were spread evenly. Between roughly three and eight weeks of age, spanning what corresponds to childhood into adolescence, spine density rose sharply in that one compartment while the rest of the dendrite followed a different course.

The teenage brain, in this system, was not only removing. It was constructing, in a specific place, on a specific schedule.

There was a clinical hook as well. In mouse models carrying mutations associated with schizophrenia risk, early development proceeded normally but the adolescent hotspot failed to form. Which raises the possibility that some cases involve a failure to build rather than an excess of removal.

Now the caveats, because they matter enormously and the press coverage has been loose with them.

This is a mouse study. It examined one cortical area, primary somatosensory cortex. It examined one cell type, layer 5 extratelencephalic-projecting neurons. Whether the same compartment-specific construction occurs in primates or humans is not established. The authors themselves note the uncertainty. Nobody has demonstrated a spine hotspot in human adolescent cortex, and the methods used here cannot currently be applied to living human brains.

What the finding does is qualify a narrative, not overturn a body of evidence. Adolescence involves both removal and construction, at least in this system. The forty years of pruning data remain intact. What changes is the framing that adolescence equals net subtraction and nothing else.

This is also not the first evidence pointing that way. Work on experience-dependent spine dynamics has shown for years that cortical circuits both gain and lose spines in response to experience, with new spines forming and a subset stabilising into lasting structural traces [33]. Sonja Hofer and colleagues showed that experience leaves a lasting structural trace in cortical circuits, with spines added during one episode persisting and being reused later [34]. Myelination also continues to increase through adolescence and into the twenties even as grey matter signal declines. The picture was never purely subtractive. It was just told that way.

Here is how the field's understanding developed.

Nearly fifty years from the first human count to the first serious challenge to the subtraction-only framing. Science moves at the speed of its methods.

1982: a hypothesis about adolescence

If pruning peaks in adolescence, and several serious psychiatric conditions first appear in adolescence, the temptation to connect them is obvious.

Irwin Feinberg made that connection formally in 1982, in a paper asking whether schizophrenia might be caused by a fault in programmed synaptic elimination during adolescence [36]. The reasoning was structural rather than experimental. A process that reshapes cortical connectivity on exactly the timeline when the illness emerges is at minimum an interesting coincidence.

Gal Chechik, Isaac Meilijson and Eytan Ruppin approached it from a different direction in 1998, building computational models showing that pruning under limited synaptic resources could actually improve network performance rather than merely degrade it [37]. Deletion, done correctly, is optimisation. Done incorrectly, it is damage.

For decades that was where things stood. A plausible hypothesis, a suggestive timeline, and no molecular link.

Then in 2016, Aswin Sekar and colleagues published in Nature what is now the most cited paper connecting the two [38]. The strongest genetic association signal in schizophrenia sits in the major histocompatibility complex region of the genome, and nobody had explained why. Sekar's team showed that a substantial part of that signal comes from structural variation in the complement component 4 gene, and that alleles associated with higher predicted C4A expression in the brain associate with greater schizophrenia risk. They also showed that C4 mediates synapse elimination during postnatal development in mice.

Complement. The same tagging system Stevens had identified nine years earlier.

It is a genuinely exciting result and it deserves the attention it gets. But look closely at what it establishes, because this is where nearly every popular summary overreaches.

The reported effect size for the higher-expression alleles is an odds ratio of roughly 1.3. That is a modest association. It is not a switch that determines who develops schizophrenia. Most people carrying higher-expression C4 alleles never develop the illness, and many people with schizophrenia do not carry them.

What the paper demonstrates is an association in human genetics plus a plausible mechanism demonstrated in mice. What it does not demonstrate is that excessive pruning causes schizophrenia in humans. Those are different claims, and the gap between them is where scientific caution lives.

Later work has strengthened the mechanistic side. Yilmaz and colleagues showed in 2021 that overexpressing human C4A in mice promotes excessive synaptic loss and behavioural changes [39]. A 2022 study found elevated C4A in cerebrospinal fluid in first-episode psychosis patients who went on to convert to schizophrenia [40]. The convergence is real and it is accumulating.

It is still an association plus a mechanism, not a demonstration of causation in humans, and it will remain so until somebody can watch synaptic density change over time in living people and correlate it with symptom onset.

A 2025 review in Frontiers in Synaptic Neuroscience surveyed what could theoretically be done about excessive pruning, discussing minocycline for its effects on complement-mediated microglial activity, the C4 gene itself, and the astrocyte MEGF10 and MERTK pathways as potential targets [45]. The review is honest that most of these options remain hypothetical.

The other direction: too few deletions

If over-pruning is one failure mode, under-pruning is the other.

Guomei Tang and colleagues published a study in Neuron in 2014 examining post-mortem cortical tissue from people with autism spectrum disorder [42]. They found increased dendritic spine density in layer V pyramidal neurons, with evidence of reduced developmental spine pruning. The mechanism they identified involved hyperactivated mTOR signalling and impaired autophagy, the cell's system for degrading and recycling its own components.

The mouse experiments were where it got interesting. In a genetic model with the same pathway disrupted, rapamycin, which inhibits mTOR, corrected the spine pruning deficit and improved autism-like behaviours. The correction depended on autophagy being functional.

So the paired story is appealing in its symmetry. Too much pruning associates with schizophrenia. Too little associates with autism. One brain over-edits, the other under-edits.

Be careful with that symmetry. It is a useful organising idea, not an established fact. The human evidence in Tang's study is correlational, drawn from post-mortem tissue in a limited sample. Autism is not one condition with one biology, and spine density findings are not consistent across every study or every brain region examined. The mouse work is genuinely strong. The extrapolation to human autism generally is where confidence should drop.

Claims about ADHD and pruning are weaker still. Philip Shaw and colleagues showed in 2007 that ADHD is characterised by a delay in cortical maturation, with peak cortical thickness reached several years later than in controls [43]. That is a real and replicated MRI finding. But as established earlier, cortical thickness is not a synapse count. Interpreting delayed thinning as delayed pruning is inference layered on inference.

For dementia, the picture is different again. Soyon Hong and colleagues showed in 2016 that complement and microglia mediate early synapse loss in Alzheimer's disease models, with C1q and complement activation preceding plaque deposition [44]. The developmental machinery, reactivated inappropriately in an ageing brain. That mouse evidence is strong. Whether the same pathway drives human Alzheimer's synapse loss to the same degree remains an active research question, and broader reviews of microglia in neurodegeneration describe a field still assembling the picture [52].

Since the popular coverage rarely distinguishes between these, here is a plain grading.

Not one of these rows says "established." That is the accurate state of the field, and it is worth saying plainly because almost nobody does.

Pruning is not forgetting

Now to the question that probably brought a fair number of readers here. If the brain deletes unused connections, is that why you forget things?

Mostly, no. And the distinction is worth getting right.

Developmental synaptic pruning is structural, meaning connections are physically removed. It runs largely on a maturational schedule rather than a moment-to-moment one. It is regionally specific, with different areas following different clocks. And its heavy phase is concentrated from early childhood through, in prefrontal cortex, roughly the third decade of life.

Everyday forgetting in an adult is a different animal. When you cannot recall a name you knew last week, the dominant explanations are retrieval failure, interference from competing memories, and synaptic weakening, rather than the literal developmental deletion of a connection. The trace is often still there. You cannot reach it. This is why active recall works at all, and why a cue that failed a moment ago can suddenly succeed with a different prompt.

That said, honesty requires flagging where this line blurs. Adult brains do show ongoing structural spine turnover. Connections are made and lost throughout life, not only during development. There is also an active, enzyme-driven side to forgetting that goes beyond passive decay, which is a genuinely separate mechanism from developmental pruning but shares the property of being deliberate rather than accidental. Readers interested in that thread will find the story of active forgetting and the enzymes that erase memories useful.

So the boundary is real but not absolute. Developmental pruning and adult forgetting are different processes on different timescales driven by different machinery. They are not entirely unrelated, because both ultimately involve synaptic strength and stability. They are simply not the same thing, and using one to explain the other produces confident nonsense.

There is a related point about ageing worth separating out too. The declines in memory that come with age involve their own set of mechanisms, and reading them as a continuation of adolescent pruning is another common conflation. The story of how ageing changes memory runs on different biology.

What your brain does every night

There is one process that genuinely does resemble pruning in miniature, happens to you constantly, and is directly relevant to learning. It happens while you sleep.

Giulio Tononi and Chiara Cirelli proposed the synaptic homeostasis hypothesis, laid out most fully in a 2014 Neuron review [46]. The argument runs like this. Being awake means learning, and learning means net synaptic strengthening. Strengthening costs energy and space, and it degrades signal-to-noise if it continues without limit. Sleep, on this account, renormalises. It scales synapses back down so the system can start the next day with room to work.

For years this was a hypothesis with indirect support. Then in 2017 two papers landed in the same issue of Science.

Luisa de Vivo and colleagues did the anatomy the hard way [47]. They measured 6,920 synapses in mouse motor and sensory cortex using three-dimensional electron microscopy, comparing sleep and wake. The axon-spine interface, a structural measure of synaptic size, was about eighteen percent smaller after sleep than after waking.

But the scaling was not uniform. Roughly twenty percent of synapses, the largest and strongest ones, were largely spared. The remaining eighty percent, the weaker and more plastic population, were the ones that shrank.

That selectivity is the whole point. Sleep does not turn the volume down on everything. It turns the volume down on most things, and leaves the strong ones alone. Which means the relative prominence of yesterday's important connections goes up, not down.

Graham Diering and colleagues, in the companion paper, worked out a molecular mechanism [48]. An immediate-early gene product called Homer1a drives homeostatic scaling-down at excitatory synapses during sleep, removing AMPA receptors, and the process is gated by falling noradrenaline levels. Wakefulness suppresses it. Sleep permits it.

Now the part that keeps this from being a simple story about sleep as a demolition crew.

Guang Yang and colleagues showed in 2014 that sleep promotes branch-specific formation of dendritic spines after learning [49]. Not just pruning. Building, in specific locations, tied to what was learned before sleep. And Wei Li and colleagues showed in 2017 that REM sleep selectively prunes and maintains new synapses in both development and learning [50]. Prunes and maintains. In the same sentence, in the same process.

The pattern is now familiar. Removal and construction are not opposing forces. They are two halves of the same editorial operation. If you want the applied version of this, the mechanics of how sleep consolidates spaced learning follow directly from these findings.

What this means if you are an adult who studies

Time to answer the practical question directly, and to answer it honestly rather than conveniently.

You will occasionally see the claim that spaced repetition prevents your synapses from being pruned. As a statement about developmental synaptic pruning, that is wrong. Developmental pruning runs on a maturational schedule shaped by genetics, region, and early experience. It is not waiting to see whether you reviewed your flashcards. If you are twenty-eight years old, the heavy phase of prefrontal spine reduction is happening or finishing regardless of your study habits, and no review schedule pauses it.

The framing is wrong. That does not mean the underlying intuition is worthless. It means the correct version is different, and more interesting.

Here is what the evidence actually supports. In an adult brain, repeated and appropriately spaced retrieval drives long-term potentiation and the consolidation processes that follow it. Those processes stabilise specific synapses. Min Fu and colleagues showed that repetitive motor learning induces the coordinated formation of clustered dendritic spines, and that clustered spines are more likely to persist than isolated ones [51]. Hofer's work showed that experience leaves structural traces that survive and get reused [34].

And recall Wiegert and Oertner's finding: it was the weakly integrated synapses that got eliminated after depression, not simply the depressed ones [28]. Integration protects.

Put those together and the defensible claim looks like this. Spaced retrieval does not prevent developmental pruning. What it does is drive potentiation and consolidation that strengthen and structurally stabilise the particular connections carrying a memory, and it embeds them into clusters of related activity, making them more resistant to ordinary turnover and interference.

That is a weaker claim than "studying saves your synapses from deletion." It is also true, which is the relevant difference. And it has a practical consequence the stronger claim does not: if integration is what protects, then how you connect new material to what you already know is not a study-technique nicety. It is closer to the mechanism itself.

The evidence on retrieval practice and the brain lines up with this, as does the broader story of neuroplasticity across the lifespan. Adult brains change. They change less dramatically and more slowly than developing ones, and they change according to what gets used, connected, and consolidated.

One more implication worth stating. If sleep is where selective scaling and selective stabilisation happen, then sleep is not the thing you sacrifice to make room for more study. It is part of the study. The 2017 findings make that fairly hard to argue with.

What would actually settle this

A short list, because the honest position on most of this topic is that key questions are open.

The field needs longitudinal measurement of synaptic density in living human brains, tracked across adolescence and into the twenties. Everything we currently know about the human developmental trajectory is reconstructed from autopsy tissue sampled cross-sectionally, mostly from small numbers of brains. SV2A PET imaging is the most promising route, and it is still an indirect marker rather than a synapse count.

It needs direct live observation of microglia removing intact functional synapses, in more systems, to close the hunter-versus-gatherer question [19].

It needs the 2026 construction finding replicated across regions, cell types, and ideally in primate tissue, before anyone rewrites the adolescent story around it [31].

And it needs a way to test causation in humans for the clinical links, which is the hardest problem of all, because the direct experiment is not one anyone can ethically run. What is realistic is convergence: genetics, imaging, stem-cell-derived neurons, and treatment trials all pointing the same direction.

Until then, the accurate summary is this. The developing human cortex builds far more connections than it retains and eliminates a large fraction of them on a schedule that varies by region and runs, in prefrontal cortex, into the third decade. That elimination is driven by activity, executed at least partly by immune signalling and glial cells, and balanced by a protective brake and, in at least one system, by simultaneous construction. Its failures are plausibly linked to several psychiatric conditions, with the strongest link amounting to an association plus a mechanism rather than proof.

And the number in the title of this article? It comes from a review-level regional average, not a whole-brain measurement, and the study most often blamed for it never said it.

That correction is not a footnote. It is the most useful thing on this page.