(F-236) Convolutional neural networks for identifying electrogenic cell types and states based on spontaneous extracellular action potential waveforms

Introduction:: Phenotypic screening models using microelectrode arrays (MEAs) to record spontaneous extracellular action potentials (EAPs) from in-vitro iPSC-derived neurons is now a viable research method for studying drugs that target the nervous system in a preclinical setting [1]. However, in heterogeneous culture settings, it is difficult to identify and discriminate neuronal subtypes, which is necessary for so-called ‘single-unit pharmacology’. Currently, researchers rely on empirical pharmacological descriptions to discern functional subgroups from a recording session [2]. Capsaicin, for instance, temporarily evokes a substantial increase in mean- firing- rates for TRPV1-positive sensory neurons (i.e., presumptive nociceptors) [2]. However, capsaicin causes long-term desensitization and can even reduce cell viability [3]. Here, we present a convolutional neural network (CNNs) designed to learn normalized EAP shapes from pharmacologically- labeled data that correspond to subgroups within spontaneously firing mouse dorsal root ganglia (mDRG) and iPSC-derived sensory neuron cultures. CNNs serve as an ideal model to fit categorical EAP data because of the large number of parameters that capture complex patterns from the waveform, while also allowing multiple categories of waveforms per category [4]. Moreover, this architecture has been successfully applied to classify 1D electrophysiological data such as electrocardiogram (ECG) in such a way that correlates signals to heart disease states [6]. There is a need for autonomous electrogenic cell identification because the current techniques are not suitable for large scale 2D/3D neuron experiments – an increasing need for pharmaceutical companies.

Materials and Methods:: HiPSCs SN (Axol Bioscience,Anatomic) and astrocytes (BX-0600, Brainxell) were cultured on Cytoview 48-well MEAs (Axion Biosystems, Atlanta, GA). Additionally, mouse dorsal root ganglia (mDRG) were dissected and also cultured on MEAs. For spontaneous recordings, continuous data were recorded from all 48 wells (768 electrodes) simultaneously at 12.5 kHz per electrode and filtered using a 1-pole Butterworth band pass filter from 200-3000 Hz. Capsaicin (1 μM) was added to the cells during the final recording session to evoke neural subtype-dependent activity and extracellular action potentials (EAPs) were extracted from the time-series data using a 5.5σ adaptive threshold. The capsaicin responsiveness of individual neurons was used to initially classify the neurons into two groups. For this study, a Tensorflow (Microsoft, Redmond, WA) two-layer convolutional neural network was used to extract patterns from the EAPs, and a dense neural network of 256 artificial neurons was used to learn the capsaicin-responsiveness-dependent patterns. Four SN datasets were trained separately for 30 iterations at 10 epochs, with each iteration using a different assortment of cells in the training and evaluation set to ensure that the model was not overfitting to a specific subset. The initial binary groupings were empirically optimized by referencing the high confidence accuracy for unseen cells. After optimization, the training accuracy, evaluation set accuracy, and high confident evaluation set accuracy were recorded for each of the thirty iterations. The subsequent analysis was carried out using a paired t-test to compare the accuracies within each SN dataset.


Results, Conclusions, and Discussions:: MDRG and human stem cell derived neurons expressed TRPV1, making them responsive to 1 μm doses of capsaicin. We manually labeled subtypes by grouping cells by capsaicin-responsiveness (CR) and we ran the EAPs through the neural network 30 times at 10 epochs, changing the assortment of the training and unseen evaluation set each iteration. The mDRG sensory neurons exhibited capsaicin-responsiveness of 85.76%, and a 60.68% unseen data model accuracy with a 80.16% high confidence unseen data model accuracy. The Anatomic monocultures had a capsaicin responsiveness of 68.31%, with a 56.12% unseen data model accuracy and a 79.39% high confidence unseen data model accuracy. The Anatomic Co-culture was 36.81% capsaicin-responsive with a 65.04% unseen data model accuracy and a 90.72% high confidence unseen data model accuracy. Lastly, the Axol Coculture had a 58.96% capsaicin responsiveness and had a 60.02% unseen data model accuracy with an 70.40% high confidence unseen data model accuracy.

To the best of our knowledge, this is the first presented application of CNN models to distinguish neuronal cell types from heterogeneous populations in vitro. The accuracies yielded by the CNN strongly suggest that there are correlations between a sensory neuron’s capsaicin responsiveness and the associated waveform shapes. Sensory neural tissue is inherently heterogeneous, meaning there are naturally multiple different subtypes – even within the nociceptor subgroup [7]. Different types of sensory neurons will have varying membrane-bound density of TRPV1 [7], and TRPV1 density influences capsaicin responsiveness [8]. Since prior patch clamp studies have reported waveform differences between excitatory and inhibitory neuronal cell types, for example [9], it is reasonable that this method may be applied to other heterogeneous in vitro systems.






Acknowledgements (Optional): :

References (Optional): : [1] Cui, H. F., Ye, J. S., Chen, Y., Chong, S. C., & Sheu, F. S. (2006). Microelectrode array biochip: tool for in vitro drug screening based on the detection of a drug effect on dopamine release from PC12 cells. Analytical chemistry, 78(18), 6347-6355.

[2] Ao, Z., Cai, H., Wu, Z., Krzesniak, J., Tian, C., Lai, Y. Y., ... & Guo, F. (2021). Human spinal organoid-on-a-chip to model nociceptive circuitry for pain therapeutics discovery. Analytical Chemistry, 94(2), 1365-1372.

[3] Wall, P., Fitzgerald, M., Nussbaumer, J. et al. Somatotopic maps are disorganized in adult rodents treated neonatally with capsaicin. Nature 295, 691–693 (1982). https://doi.org/10.1038/295691a0

[4] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017). Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6), 84-90.

[5] Abdel-Hamid, A. -r. Mohamed, H. Jiang, L. Deng, G. Penn and D. Yu, "Convolutional Neural Networks for Speech Recognition," in IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 22, no. 10, pp. 1533-1545, Oct. 2014, doi: 10.1109/TASLP.2014.2339736.

[6] D. Li, J. Zhang, Q. Zhang and X. Wei, "Classification of ECG signals based on 1D convolution neural network," 2017 IEEE 19th International Conference on e-Health Networking, Applications and Services (Healthcom), 2017, pp. 1-6, doi: 10.1109/HealthCom.2017.8210784.

[7] Tavares-Ferreira, D., Shiers, S., Ray, P. R., Wangzhou, A., Jeevakumar, V., Sankaranarayanan, I., ... & Price, T. J. (2022). Spatial transcriptomics of dorsal root ganglia identifies molecular signatures of human nociceptors. Science translational medicine, 14(632), eabj8186.

[8] Szolcsányi, J. (2014). Capsaicin and sensory neurones: a historical perspective. Capsaicin as a Therapeutic Molecule, 1-37.

[9] McCormick, D. A., Connors, B. W., Lighthall, J. W. & Prince, D. A. Comparative electrophysiology of pyramidal and sparsely spiny stellate neurons of the neocortex. J. Neurophysiol. 54, 782–806 (1985). Introduces ‘regular-spiking’, ‘bursting’ and ‘fastspiking’ classifications of firing patterns of neocortical neurons, correlates firing patterns with spike shape, and identifies fast-spiking neocortical neurons as GABA-mediated interneurons