The results published in “Spike-Train Communities: Finding Groups of Similar Spike Trains” (February 9, 2011 in the Journal of Neuroscience) provide a clear case for data sharing. The author of the paper, Dr. Mark Humphries, was motivated to develop a new algorithm and methodology to help identify similar spike train patterns generated by individual neurons to repeated stimuli or by large sets of simultaneously recorded neurons. Similar patterns can be evidence of common coding or for cell assemblies.
The author devised a new grouping method based on network theory (part of graph theory) that self-determines the maximum number of groups in any set of spike trains and groups them to maximize similarity within each group. The only free parameter used in the method is the timescale for pairwise comparison.
Note: The Spike-train Communities Toolbox for MATLAB includes the paper’s algorithm and control analyses on spike train time series and is publicly available from the Adaptive Behaviour Research Group’s code repository hosted by the University of Sheffield.
The author’s first task was to use synthetic spike trains with known groupings to test his algorithm and methodology. He used synthetic spike train data freely available from an online website (see note below) with known spike train groupings and showed:
- Using spike train statistics to help estimate the timescale of the pairwise comparisons works well.
- The algorithm reliably self determines the number of groups in each set of spike train data.
- The algorithm’s assignment of spike trains to each group is robust to noise.
With these testing results in hand, the author was ready to use his method on actual recordings from neurons.
Note: The synthetic spike train data the author used to test his clustering algorithm was created by Fellous, Tiesinga, Thomas and Sejnowski and was reported in “Discovering Spike Patterns in Neuronal Responses” (published March 24, 2004 in the Journal of Neuroscience). The data are available from here (there’s a small typo in the web address published in the paper).
Dr. Humphries tried his methodology on spike trains recorded from a single neuron. These data, recorded from the substantia nigra pars compacta, were obtained from a colleague at the University of Sheffield (see note below for associated paper) but are not available online. Earlier work led to the discovery of a direct connection from neurons in the superior colliculus to neurons in the substantia nigra. The data set analyzed with the new algorithm was from an experiment to see the effects of noxious stimuli on the activity of dopamine neurons in the substantia nigra and the role that the superior colliculus played in the effect. The original research team concluded their paper stating that “superior colliculus activity had relatively minor effects on the responses of dopamine neurons to noxious [stimuli].” The reanalysis using Dr. Humphries’ algorithm provided new insight that suggested the input from the superior colliculus could reset the timing of substantia nigra pars compacta neuron periodic activity to noxious stimuli.
Note: The single unit data was collected by Coizet, Dommett, Redgrave, and Overton and was reported in “Nociceptive responses of midbrain dopaminergic neurones are modulated by the superior colliculus in the rat” (published March 3, 2006 in Neuroscience).
The author analyzed multi-neuron data from visual cortex. These data were composed of spike trains collected simultaneously from spontaneously active neurons in primary visual cortex (V1) and V2. In V1 the algorithm found significant group structure during a little more than half the time examined. This suggested that groups were transient entities. The algorithm found two groups in V2 with in-phase population activity at about 0.8 to 1 Hertz and anti-phase activity at frequencies less than about 0.4 Hertz. The important interpretation of these data is that “individual neurons change their contribution to the population activity over time, and that, as a consequence, both ensemble and individual firing patterns are unique.” Further analysis also led to the conclusion that the organization of spontaneously active V1 and V2 cortex combined fluctuates over timescales of tens of seconds and each cortical area alters its internal correlation structure independently of the other.
Note: The multi-neuron data from cortical areas 17 (V1) and 18 (V2) were recorded by Tim Blanche in the laboratory of Nicholas Swindale, University of British Columbia, and downloaded from the National Science Foundation funded Collaborative Research in Computational Neuroscience data sharing website.
Finally, the author used his new algorithm and methodology to discover grouping dynamics embedded in a large scale computational model of the rat striatum. The model generated thousands of simultaneous spike trains that the algorithm used to detect cell assemblies. The model generated spontaneous cell assemblies in response to unstructured, background input that were shown by the algorithm to have complex dynamics organized on different timescales for different neurons. On one timescale, the neurons in two of the groups showed a distinct shift in their irregular repetition of a network state. On a different timescale, the neurons in two of the groups showed a regular repetition of a network state.
Note: The large-scale model of the rat striatum is available from the SenseLab ModelDB repository in MATLAB code. Go to the “Striatal GABAergic microcircuit, dopamine-modulated cell assemblies (Humphries et al. 2009)” record.
In summary, the new algorithm and methodology provided significant new insight into existing data (some available online; please see notes above) and existing computational models (both available from online repositories; please see notes above). The new insights into brain systems accomplished through the reanalysis of hard won data provide a strong case for continuing to make more data widely available. You can never tell who will come up with the methodology that will open the door to a flood of new insights. Dr. Humphries’
new method has the potential to help us directly observe dynamic cell assemblies.
Note: Method analyses included spectra and coherency (and their confidence intervals) computations using functions from the Chronux Analysis Software for MATLAB.
