News & Events

Novel method for inferring neuronal assemblies


Both spontaneous and stimulus-evoked neural activity can lead to synchronous activation of neuronal assemblies in many areas of the brain.

The idea was first proposed by Donald Hebb who theorised that neural circuits are designed to enable such synchronous and sequential activity through groups of co-active neurons which fire together - Hebbian assemblies. His predictions have been shown to be correct through recent advances in neuroscience such as calcium imaging and neuropixels probes – a hallmark of population activity is indeed the organisation into assemblies which correlate with a diverse range of brain functions.

These observations have generated a new paradigm in neuroscience where neuronal assemblies are viewed as ‘units of brain computation’. In particular, through the characterisation of the structure and dynamics of these assemblies, crucial insights could be yielded into how brain computations are distributed across neural networks.

Characterising the properties of these assemblies is not straightforward however - neurons can fire independently of their assemblies and not all neurons within an assembly are necessarily recruited each time the assembly fires, for example, adding noise. Previous methods used to characterise neuronal assemblies did not allow for a rigorous statistical inference of their features, in particular how many assemblies can be identified from the data, which potentially can lead to erroneous conclusions. .

In their paper however, Giovanni Diana, Thomas Sainsbury and Martin Meyer, from the Meyer lab, describe their method which is specifically tailored to neuronal data. Noise, within-assembly synchrony and assembly activity are all directly estimated from the data and are used to cluster neurons into assemblies in their model, offering a biologically faithful representation of the data. Their method is based on a hierarchical model of neuronal activity and grounded in Bayesian inference, which allows neurons to be grouped into assemblies based on the most likely scenario and it provides statistical confidence for the conclusions drawn from the model outperforming a range of existing techniques..

The authors of the new paper have used their model to characterise neuronal assemblies from large-scale functional imaging data from the zebrafish tectum and mouse visual cortex, and from neuropixels probes in the mouse visual cortex, hippocampus and thalamus. Their new technique can dissect complex neuronal population data, revealing features such as firing pattern and the degree of synchronous and asynchronous firing of constituent neurons in an assembly and offer insights into its behavioural and physiological relevance.