Beyond correlations in MEA recordings - how far can we go?

Understanding interactions among neurons and the behaviour of whole neuronal populations is a necessary step to understanding brain function. Recent advances in experimental techniques (multielectrode array recordings, MEA) provide a wealth of data that could yield insight into the collective activity of large groups of neurons. However, most conventional analysis techniques, such as firing rates or correlations, are insufficient to draw far-reaching conclusions. In search for measures able to capture more complex phenomena, over the last decade a new approach has been developed - pairwise maximum entropy modelling (MaxEnt). It is a statistical model that fits two sets of parameters to explain the probability of spiking patterns in the network: individual neuron parameters that could be interpreted as excitability; and pairwise interaction parameters that could be interpreted as connection strength between neurons. Successful application of this model to a variety of recordings has shown that, despite low correlation values, neuronal interactions play an important role in shaping network activity (Schneidman et al., 2006). Additionally, the failure of MaxEnt in certain cases could be an indicator of higher-order interactions between neurons (Ohiorhenuan et al., 2010).
In present work we examine the performance of the MaxEnt model in application to novel, high-density recordings. Neural activity from primary neuron cultures was recorded with the 4096 channel Active Pixel Sensor (APS) MEA, allowing for reliable isolation of single unit activity at near-cellular resolution (Berdondini et al., 2009). Such data allows us to explore the stability of the model, its performance on varying spatial scales and utility in providing information about the network. To this end, large numbers of four-unit MaxEnt models are constructed for randomly chosen neurons on two spatial scales. Results of the fits indicate that the model is able to detect a difference in interaction strengths between groups of nearby neurons and those further apart. Additionally, it appears that the advantage of MaxEnt over independent model is more apparent in close range. Those results could be interpreted as an indicator of local clustering.