Intelligent In-Cell Electrophysiology—Reconstructing Intracellular Action Potentials Using a Physics-Informed Deep Learning Model Trained on Nanoelectrode Array Recordings

Intracellular electrophysiology, crucial in neuroscience, cardiology, and pharmacology, explores and provides insights into cellular electrical properties. Traditional techniques like patch-clamp are precise but invasive and low-throughput, necessitating the development of alternative methods that offer similar precision with higher throughput and less invasiveness.<br/>Nanoelectrode Arrays (NEAs) comprise nano-scale electrodes capable of recording extracellular signals and, through methods such as electroporation, gain transient intracellular access. NEAs offer a high-throughput alternative for recording intracellular (iAP) and extracellular action potentials (eAP). However, accessing intracellular potentials with NEAs remains a challenge due to their limited control.<br/>In this study we introduce a technique for intracellular electrophysiology supported by artificial intelligence (AI) that leverages thousands of synchronous eAP and iAP pairs collected from stem-cell-derived cardiomyocytes on NEAs (under revision in Nature Communications). Our method of data collection utilized neighboring channel pairs on the NEA, capturing their corresponding eAP and iAP. We validated our method by first comparing iAP waveforms simultaneously collected from NEAs and patch-clamp from the same cells. Despite some discrepancies in amplitude, normalization confirmed the accuracy of NEA recordings. Furthermore, by analyzing action potentials from two closely positioned channels on an NEA, we confirmed that these recordings are almost identical, effectively representing each other. This finding underpinned our strategy of using eAP and iAP pairs from adjacent channels as representatives of each other. These eAP-iAP pairs were obtained by administering various ion-channel-affecting drugs such as propranolol, dofetilide, flecainide, and nifedipine. Analyzing the collected pairs revealed high correlations between eAP and iAP waveform features like amplitude and spiking velocity, indicating that eAPs could effectively predict intracellular shapes.<br/>Next, we used the collected data to train the Physics-Informed Attention-UNET (PIA-UNET). Our model integrates a modified Aliev-Panfilov model into its loss function to ensure the physiological plausibility of the reconstructed signals. This hybrid approach combines the strengths of deep learning with the physics of electrophysiological modeling, allowing for accurate and robust reconstruction of iAP waveforms from eAP recordings. The PIA-UNET model demonstrated high accuracy, with a coefficient of determination (r2 = 0.99) between predicted and actual iAP values, and a mean absolute error (MAE) of 0.028 ± 0.006 on the test set.<br/>We demonstrated our model's potential for non-invasive, long-term, and high-throughput assessments of drug cardiotoxicity. Our model accurately reconstructs iAPs from non-invasive eAP recordings, enabling detailed drug effect studies on cardiac cells—key for cardiotoxicity screening. In experiments, we showed its capability to assess changes in iAP waveforms over time when exposed to a shape-altering drug like propranolol, not included in the training data. Our model reliably detected changes in action potential durations (APDs) across multiple cells, providing population-level drug effect insights.<br/>Although this demonstrates one specific application, it opens the door for future research in electrophysiology. The approach can be extended to various electrogenic cell types and drug interactions. Expanding the dataset to encompass a broader range of cellular responses and pharmacological compounds will enhance the model’s robustness and applicability. The integration of AI with NEA technology marks a significant advancement, offering a non-invasive, high-throughput, and precise tool for intracellular electrophysiology studies. This holds promise for advancing our understanding of cellular electrical properties and improving drug safety assessments.