Intelligent In-Cell Electrophysiology: Reconstructing Action Potentials from Nanoelectrode Data Using Physics-Informed Deep Learning.

Zeinab Jahed, Keivan Rahmani, Ethan Foster, Ching-Ting Cai, Yang Yang, Aayush Gupta, Rose Yu, Bianxiao Cui, Francesca Santoro, Csaba ForróResearch Square 2024 · DOI

TL;DR

Researchers have developed a physics-informed deep learning model, PIA-UNET, that reconstructs high-fidelity intracellular action potentials (iAPs) from non-invasive extracellular recordings (eAPs). By combining nanoelectrode array technology with cardiac excitation physics, this approach enables high-throughput, long-term, and precise monitoring of cell health and drug toxicity without the need for invasive physical penetration of the cell membrane.

What Problem Does This Paper Address?

In electrophysiology, researchers face a classic trade-off: precision versus throughput.

The "gold standard" for measuring the electrical activity inside a cell is the patch-clamp technique. It provides incredibly precise intracellular action potential (iAP) data but is manual, invasive, and can only handle a few cells at a time. On the other end of the spectrum, Microelectrode Arrays (MEAs) are non-invasive and high-throughput, but they typically only record extracellular action potentials (eAPs). These eAPs are essentially "blurred" versions of the internal signal, lacking the fine detail needed to detect subtle changes in ion channel activity—changes that are critical for assessing drug safety (cardiotoxicity).

While Nanoelectrode Arrays (NEAs) can record both, getting a stable intracellular signal usually requires electroporation (using electric pulses to poke tiny holes in the cell). This is disruptive to the cell's physiology and prevents long-term, continuous recording. The authors sought to eliminate this disruption by using AI to "translate" the non-invasive extracellular signal into its intracellular counterpart.

Methods

The authors introduced a deep learning architecture called Physics-Informed Attention-UNET (PIA-UNET).

  1. Hardware: They used specialized "nanocrown" electrodes that can switch between extracellular and intracellular recording modes.
  2. The Model: The core is an Attention-Residual-Block UNET. It uses an encoder-decoder structure to compress eAP features and reconstruct the iAP waveform.
  3. Physics Integration: Unlike a standard "black box" AI, PIA-UNET is "physics-informed." The researchers integrated the Aliev-Panfilov model—a set of differential equations describing cardiac excitation—directly into the model's loss function. This ensures the reconstructed waveforms aren't just statistically likely, but biologically plausible.
  4. Quantile Regression: They also developed a version (Q-PIA-UNET) that outputs confidence intervals, allowing researchers to see the uncertainty in the AI’s predictions.

Datasets

The study utilized a unique, high-quality dataset of 1,910 synchronized eAP/iAP pairs collected from human stem-cell-derived cardiomyocytes.

  • To ensure a diverse range of waveforms, the researchers used Dofetilide, an ion-channel blocker that alters action potential duration (APD) in a dose-dependent manner.
  • Data was segmented into 1.6-second windows (8,000 data points) for training and evaluation.
  • The dataset is notable for its high signal-to-noise ratio (S/N > 90 dB), providing a clean baseline for the deep learning model.

Key Insights and Ideas

  • Hidden Correlations: The analysis revealed strong correlations between eAP amplitude and iAP spiking velocity ($r=0.903$) that were previously unquantified. This suggests that the extracellular signal contains far more "hidden" information about the cell's internal state than previously thought.
  • Physics Matters: Adding the Aliev-Panfilov physics constraints improved the model's performance by 11.7% compared to a standard Attention UNET. The physics term acts as a regularizer, preventing the model from generating "impossible" biological signals.
  • Fidelity Requirements: The model's accuracy is sensitive to the quality of the input. Low eAP spike amplitudes or slow decay rates (often caused by poor cell-electrode coupling) lead to higher prediction errors, highlighting the importance of the "nanocrown" hardware in capturing high-fidelity raw data.

Key Results

  • High Accuracy: The PIA-UNET achieved a Mean Absolute Error (MAE) of 0.028 ± 0.006 on the test set and a coefficient of determination ($r^2$) of 0.99 for membrane potential reconstruction.
  • Precise Timing: The model accurately predicted Action Potential Duration (APD) metrics, with errors for APD70 and APD90 (key markers for heart health) staying below 2.6%.
  • Drug Screening Potential: The system successfully reconstructed the "prolongation" effect of Dofetilide. This demonstrates that the AI can detect drug-induced changes in cell behavior entirely from non-invasive recordings.
  • Scalability: The authors demonstrated the model's ability to reconstruct iAPs across a 49-channel array simultaneously, proving its potential for high-throughput screening.

Limitations and Future Work

  • Cell Type Specificity: Currently, the model is trained specifically on cardiomyocytes. Expanding this to neurons or other electrogenic cells would require new, synchronized training datasets.
  • Dataset Size: While 1,910 pairs is a strong start, deep learning models generally benefit from even larger, more diverse datasets encompassing different disease states and a wider variety of drug interactions.
  • Hardware Dependence: The success of the reconstruction relies on the high-quality signals provided by the nanocrown electrodes; it remains to be seen how well the model generalizes to data from standard, lower-resolution MEAs.

The authors suggest that future work will focus on expanding the dataset to include more cell types and refining the physics-informed loss to account for even more complex biological variables.

Summary updated

Abstract

<title>Abstract</title> Intracellular electrophysiology is utilized across different scientific and medical disciplines, including neuroscience, cardiology, and pharmacology, due to its pivotal role in exploring and comprehending the electrical properties of cells within diverse biological systems. Traditional methods for intracellular electrophysiology, such as patch-clamp, are highly precise but suffer from being low-throughput and invasive. In contrast, Nanoelectrode Arrays (NEAs) present a promising alternative, enabling simultaneous intracellular action potential (iAP) and extracellular action potential (eAP) recordings with high throughput. However, accessing intracellular potentials with NEAs remains a challenge. This study introduces a technique for intracellular electrophysiology supported by artificial intelligence (AI) that capitalizes on thousands of synchronous eAP and iAP pairs collected from stem-cell-derived cardiomyocytes on NEAs. Our analysis on this unique dataset uncovered strong correlations between specific features of eAP and iAP waveforms, such as amplitude and spiking velocity, that were not detected before, suggesting that extracellular signals could be reliable indicators of intracellular signals. By developing a physics-informed deep learning model trained on these datasets, we show successful reconstruction of iAP waveforms from extracellular recordings. We then demonstrate our model’s potential for non-invasive, long-term, and high-throughput assessments of drug cardiotoxicity. Although we show one specific application of the AI-based model, it opens the door for future research in electrophysiology, with the promise of expanding the dataset to encompass various electrogenic cell types and drug interactions.
BibTeX
@inproceedings{jahed2024intelligent,
  author = {Zeinab Jahed and Keivan Rahmani and Ethan Foster and Ching-Ting Cai and Yang Yang and Aayush Gupta and Rose Yu and Bianxiao Cui and Francesca Santoro and Csaba Forró},
  title = {{Intelligent In-Cell Electrophysiology: Reconstructing Action Potentials from Nanoelectrode Data Using Physics-Informed Deep Learning.}},
  booktitle = {Research Square},
  year = {2024},
  doi = {10.21203/rs.3.rs-4266976/v1},
}

Loading related content...

Figure