Parkinson’s disease is a severe neurodegenerative disorder that affects an increasing number of people in the world. Understanding the underlying mechanisms and the neuronal basis of this disease is a key step in the endeavor to develop new therapies. This is the goal of DeepEphys project, which was launched jointly by the Swiss Data Science Center and the Bio Engineering laboratory at the Department of Biosystems Science and Engineering. For this project, we record neuronal cultures derived from human induced pluripotent stem cells with a known Parkinson’s disease mutation and apply machine learning techniques to find out whether there are systematic differences in the neuronal activity of diseased cells compared to healthy controls. Deciphering  and understanding these differences may help to set future research directions of Parkinson’s disease treatment and to screen for potential drug candidates.

Every day we perceive our environment and act within it thanks to billions of neurons in our brain. These neurons form a complex, highly interconnected dynamical system that continues to puzzle the scientific community. As for any system, understanding its behavior is particularly crucial when it is malfunctioning. In particular, neurodegenerative diseases are one instance where the brain’s processing is affected by neuronal cell death. A popular case is Parkinson’s disease (PD), which affected 6.1 million people in 2016 worldwide (GBD 2016 Parkinson’s Disease Collaborators, 2018) and imposes an increasing economical burden upon society (Yang et al., 2020)

Individuals suffering from PD show symptoms such as rigid motion, tremor, depression, and other forms of cognitive impairment. The pathology of PD is associated with cell death in specific regions of the brain and accumulation of proteins into Lewy bodies (Jellinger, 1989). However, the exact mechanisms underlying PD are a matter of ongoing research. 

Studying the cellular mechanisms and pathological alterations underlying PD in the natural setting (in vivo) has been challenging. The involved neurons and neural circuitry are located deep inside the brain and are hard to access. To study the relevant cellular system in vivo complicated invasive interventions would be required that would mean a substantial risk for the patient.

Using induced pluripotent stem cells to study Parkinson’s disease

The discovery of induced pluripotent stem cells (iPSCs) has paved promising new avenues to study aspects of neurodegenerative disorders (Torrent et al., 2015). iPSCs are re-programmed cells that can be derived from simple skin or blood cells, following minimally invasive interventions (Figure 1). iPSCs, in turn, can be guided to develop into literally any cell type. For example, different types of neurons can be derived from these cells: after a couple of days cultured in a dish (in vitro), iPSC-derived neurons start to show spontaneous electrical activity and form neuronal networks.

Figure 1. Studying aspects of Parkinson’s disease pathology in vitro. Somatic cells, for example skin or blood cells, are extracted from a human donor and then reprogrammed into induced pluripotent stem cells (iPSCs). In this study we then go on to study two different cell types: iPSCs-derived neurons with the A53T gene mutation, which has been associated with the PD, and a iPSC from a healthy control.

The iPSC technique allows studying aspects of PD in disease-relevant cells in vitro. This can be achieved, for example, by introducing specific genetic mutations that have been associated  with PD in previous studies (Kotzbauer, 2004). Here, we use neuronal cultures with a such a point mutation (A53T) and compare it to healthy control cultures. More specifically, we aim at determining whether there are systematic differences in the electrical neuronal activity that can be used to differentiate healthy and PD neuronal cultures.

In order to record the electrical signals of healthy and PD neuronal cultures, researchers of the Bio Engineering laboratory at the Department of Biosystems Science and Engineering (ETHZ-BSSE) in Basel use high-density microelectrode arrays (HD-MEA, Figure 2, left). HD-MEAs are versatile recording devices that comprise more than 26.000 micro-electrodes on an area as small as a tiny fraction of a fingernail (8 mm²). By growing iPSC-derived neuronal cultures on these devices, they are able to record the electrical activity from several thousands of neurons at high spatial and temporal resolution.

Figure 2. Recording neuronal activity with micro-electrode arrays. The iPSC neuronal culture can be plated on a high-density microelectrode arrays (HD-MEAs, panel on the left, courtesy of MaxWell Biosystems AG). After a couple of days the cultures start to show electrical activity, which can be recorded by the HD-MEA device. The signals of neurons, so-called spikes, are reflected as characteristic voltage deflections in the recorded signal, also known as waveforms (middle panel). We extract these time points and assign the particular waveforms, a process which is called ‘spike sorting’. In this way we obtain ‘spike trains’, which is time and cell-resolved neuronal activity (right panel).

Interpreting neuronal data with machine learning

HD-MEA electrodes record voltage potentials of their immediate surroundings. But how can we extract neuronal activity from these signals? We know that neurons’ principal communication is via stereotypical voltage discharges, so-called spikes. Spikes are essentially electrical signals sent to connected neurons; they form a unique footprint in the recorded voltage signal (Figure 2, center). 

Advanced algorithms have been developed to determine, when exactly and where a putative neuron spikes (Yger et al., 2018). These ‘spike sorting’ algorithms allow transforming the electrical signal of neurons recorded by the HD-MEA into so-called spike trains (Figure 2, right). A spike train is a collection of time stamps corresponding to the time when different neurons in the network spiked. Spike trains also allow to extract the position of neurons, as well as the shape of the spike, also referred to as its waveform. Different parts of the waveform are the result of different processes within the neuron and provide additional information about the nature of the underlying neuron type.

Machine learning can help to find patterns in complex datasets by applying classification methods. Having extracted a range of features from the recorded neuronal activity (such as intensity of neuronal firing, irregularity, and many more) we train various models to differentiate the PD cultures from healthy cultures  (see Figure 3, top). In a next step, we then ask whether we can predict whether the recorded activity is coming from a PD or healthy control culture. Finding systematic differences among neuronal cultures with specific PD-relevant mutations might be informative for the development of medical treatments for PD and future drug screening efforts. 

 

Predicting Parkinson’s mutation with neuronal activity

Preliminary results indicate that we can differentiate between healthy and PD cultures with high accuracy. The effect of the PD mutation seems to be strongly reflected in the electrical activity of the cultured neurons. However, which features do the classifiers rely on to distinguish between healthy and PD cultures? Our analysis indicates that the differences are mainly related to neuronal bursting, that is, activity patterns that during which most neurons of an ensemble spike with high frequency (see Figure 3, left). We find that PD cultures show bursting earlier in the development and that they burst more frequently compared to their healthy counterparts. However, the average duration of bursts is shorter in Parkinson’s cultures than in the healthy cultures. Questions, such as, whether the Parkinson’s mutation causes neurons to “run out of breath” earlier or whether the difference in activity do reflect differences in neuronal connectivity are currently investigated.

Figure 3. Analyzing spike train data with machine learning. Different features can be extracted from the spike trains of healthy and Parkinson’s cultures. Such features include, for example the frequency of bursts (increased network activity), the firing rate of neurons, and other metrics. Based on these features we train classifiers to differentiate between healthy and Parkinson’s cultures. This allows us to identify which features are most predictive to detect differences between the two conditions. In the future we will develop a similar pipeline to establish whether neuronal connectivity allows differentiating healthy from PD cultures.

In the near future we will expand on the current analyses and aim at investigating whether the studied PD mutation also impinges on the neuronal connectivity observed in vitro. However, since we do not know the anatomy of the underlying circuitry of our cultures, we will make again use of machine learning approaches to approximate the neuronal connectivity from the recorded electrical activity. We will therefore use methods that detect statistical dependencies among neuronal activity to provide estimates for the underlying network. Based on these connectivity estimates, we will train classifiers to see whether healthy and Parkinson’s cultures can be differentiated (see Figure 3, bottom). 

 

Concluding remarks

After introducing briefly the concepts of PD, human iPSC, and how to record neuronal activity, we showed that:

  • It is possible to study aspects of PD in disease-relevant cellular models in vitro using human iPSC-derived neurons and HD-MEA recordings.
  • The neural activity of iPSC cultures contains valuable information that allows to classify healthy vs. PD neurons. 
  • Important differences between PD and control neurons are the duration and frequency of network bursts.
  • Further investigation is needed to characterize the mechanisms underlying observed phenomena, for example, by analyzing the network structure of the neuronal cultures.

There is still a long way to fully understand what is causing Parkinson’s disease and thus development of effective treatments and drugs to mitigate its effects. With the combined effort of the SDSC and ETHZ-BSSE we will contribute to this important goal.

 

References

GBD 2016 Parkinson’s Disease Collaborators. (2018). Global, regional, and national burden of Parkinson’s disease, 1990-2016: a systematic analysis for the Global Burden of Disease Study 2016. Lancet Neurology, 17(11), 939–953.

Jäckel, D., Bakkum, D. J., Russell, T. L., Müller, J., Radivojevic, M., Frey, U., Franke, F., & Hierlemann, A. (2017). Combination of High-density Microelectrode Array and Patch Clamp Recordings to Enable Studies of Multisynaptic Integration. Scientific Reports, 7(1), 978.

Jellinger, K. (1989). Alzheimer pathology in Parkinson’s disease. In Neurology (Vol. 39, Issue 6, pp. 874–874).
https://doi.org/10.1212/wnl.39.6.874-b

Kotzbauer, P. (2004). Fibrillization of alpha-synuclein and tau in familial Parkinson’s disease caused by the A53T alpha-synuclein mutation. In Experimental Neurology (Vol. 187, Issue 2, pp. 279–288).
https://doi.org/10.1016/j.expneurol.2004.01.007

Torrent, R., De Angelis Rigotti, F., Dell’Era, P., Memo, M., Raya, A., & Consiglio, A. (2015). Using iPS Cells toward the Understanding of Parkinson’s Disease. In Journal of Clinical Medicine (Vol. 4, Issue 4, pp. 548–566).
https://doi.org/10.3390/jcm4040548

Yang, W., Hamilton, J. L., Kopil, C., Beck, J. C., Tanner, C. M., Albin, R. L., Ray Dorsey, E., Dahodwala, N., Cintina, I., Hogan, P., & Thompson, T. (2020). Current and projected future economic burden of Parkinson’s disease in the U.S. NPJ Parkinson’s Disease, 6, 15.

Yger, P., Spampinato, G. L., Esposito, E., Lefebvre, B., Deny, S., Gardella, C., Stimberg, M., Jetter, F., Zeck, G., Picaud, S., Duebel, J., & Marre, O. (2018). A spike sorting toolbox for up to thousands of electrodes validated with ground truth recordings in vitro and in vivo. eLife, 7.
https://doi.org/10.7554/eLife.34518

Authors

Christian Donner, SDSC
Philipp Hornauer, BSSE
Taehoon Kim, BSSE
Damian Roqueiro, BSSE
Manuel Schröter, BSSE