Synaptic pruning · Lobeworks/17

Synaptic pruning is the elimination of synapses during development, after an early overproduction, which removes the connections that are little used and keeps those that activity confirms, so that a brain is shaped as much by what it removes as by what it grows.

Synaptic pruning. Synaptic pruning is the elimination of synapses during development, after an early overproduction, which removes the connections that are little used and keeps those that activity confirms, so that a brain is shaped as much by what it removes as by what it grows.

Peter Huttenlocher counted synapses in post-mortem human frontal cortex in 1979 and found a rise through early childhood to well above adult density, then a long decline. Counts of dendritic spines in the prefrontal cortex later showed childhood density two to three times the adult value, falling from puberty and still being trimmed into the third decade of life (Petanjek et al., 2011). The selection is use-dependent, in the spirit of Hebbian learning: weak, rarely co-active inputs lose, strong ones stay. Part of the removal is done by microglia, which engulf synapses tagged with complement proteins (C1q and C3), a pathway shown in the developing mouse visual system.

More synapses is not better. A circuit with every connection kept is noisy and slow to read; removing the weak contacts sharpens maps and lowers the cost of running them, a gain in signal against noise and in energy.

Errors in either direction are research lines for disorders. Variants that raise the complement protein C4 increase the risk of schizophrenia and drive more pruning in mice, which fits the illness's onset in late adolescence (Sekar et al., 2016); post-mortem temporal cortex in autism has shown higher spine density with less developmental pruning (Tang et al., 2014). Both are lines of evidence, with mouse models and correlations behind them, not explanations of either condition.

Machine learning borrows the idea. Model pruning deletes the weights of a trained network that contribute least, and the network often keeps its accuracy at a fraction of the size; the brain prunes while it is still learning, and by activity instead of by a score computed after training.

Development builds too much and then edits.

Experience decides which contacts survive the edit, so the adult circuit is the residue of what was used.

Questions: What does synaptic pruning have to do with autism research? One research line proposes that too little pruning leaves some autistic brains with too many synapses. Post-mortem temporal cortex from autistic people showed more dendritic spines on layer V pyramidal cells than controls, with less of the decline seen across development, and the excess went with an overactive mTOR pathway and impaired autophagy, the cell's recycling of its own parts (Tang et al., 2014). In mice with the same overactive mTOR, the pruning defect and the social differences were reversed by the drug rapamycin. It is one line among several, built on small post-mortem samples and mouse models, and autism is too varied for one mechanism to explain it. Who removes the synapses during pruning, and how are they chosen? Microglia, the brain's resident immune cells, do much of the removing: they engulf presynaptic terminals and digest them. The synapses to go are marked with complement proteins, C1q and then C3, the same tags the immune system puts on bacteria, and microglia recognise the C3 tag through their receptor CR3. In the developing mouse visual system the tagging follows activity, so the weaker, less active inputs are the ones marked and eaten, and blocking the pathway leaves extra synapses in place.