Taking an electrical fingerprint of nerve cells
Figure legend see below. "Click" to view full-size image.
Electrophysiological recording of neurons with electrodes is a key technique in understanding neuronal circuits in the brain. Recent high-density multi-electrode arrays can provide very detailed information about extracellular action potentials, and about the neurons that generated them. However, there is a lack of methods for the automatized and efficient interpretation of these signals, for example, regarding the location and classification of neurons. In a collaborative effort, scientists from the Bernstein Center Freiburg (BCF) and from the Center for Integrative Neuroplasticity (CINPLA) in Oslo, Norway, have devised a new approach to achieve this. They created and exploited a kind of electrical fingerprint that is characteristic of different types of nerve cells, and that helps to map neural circuits more accurately and more reliably. Details about the new framework have now been published.
Neural circuits typically consist of many different types of neurons. This fact presents a challenge for the attempt to disentangle individual contributions of neurons to brain processes – a fundamental prerequisite for a better understanding of brain function. “Basic classification into excitatory and inhibitory neurons, and localization of cells based on extracellular recordings are already done now”, says Prof. Dr. Stefan Rotter, Director of the Bernstein Center Freiburg. “Current approaches, however, need a lot of human intervention, which makes them slow, biased, and potentially unreliable.”
The use of cutting-edge multi-electrode arrays, like the one developed within the NeuroSeeker consortium, opens up new possibilities that have not yet been fully exploited. “These electrodes give access to very detailed information, so we can ask more sophisticated questions about the function of brain circuits,” says Stefan Rotter. “For example, many researchers want to find out what is the exact contribution of certain subtypes of inhibitory interneurons to sensory information processing,” he explains. This requires filtering out the relevant information from the large amount of raw data generated by multi-electrode recordings, as effectively as possible. The research team has now proven a new concept, which learns to execute this task from realistic neuron simulations.
Combining biophysically detailed neuron models and machine learning techniques
“Using simulated data, we can demonstrate that it is possible to automatically localize and classify single neurons recorded with high-density extracellular electrodes,” explains PhD student Michael Kordovan from the Bernstein Center Freiburg. “In this setting, our new combination of biophysical modeling and machine learning techniques yields a much higher performance compared to state-of-the-art methods,” adds his colleague Alessio Buccino from Oslo, Norway, and San Diego, California, USA.
Furthermore, the new method also has the potential to differentiate subtypes of excitatory and inhibitory neurons. PhD student Benjamin Merkt from the BCF details: “Nerve cell types have very specific shapes, or morphologies. One could say: We have developed a framework with which we can take the electrical fingerprint of cells and use it to classify them.” The researchers are convinced that their research approach can contribute to mapping neural circuits more accurately and reliably in the future. The next step is to establish the performance of the procedure on small neural circuits instead of just single neurons, and to validate it with experimentally gathered ground-truth data.
Figure Legend
Graphical representation of the new method. The red arrows represents our approach for training (dashed lines indicate error backpropagation) and validating the Convolutional Neural Network (CNN) on simulated data. The green arrow indicates how the data analysis pipeline is used in an experiment. Starting with the red path, biophysically realistic simulations (A) are used to generate templates for the extracellular action potential (EAP; B), from which suitable features (e.g. amplitude and width) are extracted and fed to a Convolutional Neural Network (CNN; C) to localize and classify excitatory (blue) and inhibitory (red) neurons (D). When applied to experimental data (green arrow), recordings are first preprocessed (spike-sorting; E), then features are extracted from averaged waveforms (B), and finally the CNN trained on simulated data is used to localize and classify neurons (D). Figure: Michael Kordovan
Original publication
Buccino AP, Kordovan M, Ness TV, Merkt B, Häfliger PD, Fyhn M, Cauwenberghs G, Rotter S, Einevoll GT
Combining biophysical modeling and deep learning for multi-electrode array neuron localization and classification
Journal of Neurophysiology 120: 212-1232, 2018
Contact
Michael Kordovan
Bernstein Center Freiburg
Hansastraße 9a
79104 Freiburg
Tel.: +49 (0)761 203 9551
E-mail: michael.kordovan@bcf.uni-freiburg.de
Prof. Dr. Stefan Rotter
Professor of Computational Neuroscience
Managing Director, Bernstein Center Freiburg
Faculty of Biology
University of Freiburg
Hansastraße 9a
79104 Freiburg, Germany
Tel.: +49 (0)761 203 9316
E-mail: stefan.rotter@bio.uni-freiburg.de