Intelligent in-cell electrophysiology: Reconstructing intracellular action potentials using a physics-informed deep learning model trained on nanoelectrode array recordings

Keivan Rahmani, Yang Yang, Ethan Foster, Ching‐Ting Tsai, Dhivya Pushpa Meganathan, Diego D. Alvarez, Aayush Gupta, Bianxiao Cui, Francesca Santoro, Brenda L. Bloodgood, Rose Yu, Csaba Forró, Zeinab JahedNature Communications 2025 · DOI

TL;DR

Researchers have developed a physics-informed deep learning model, PIA-UNET, that reconstructs high-fidelity intracellular action potentials (iAPs) using only non-invasive extracellular recordings. By training on a massive dataset of synchronized "nanocrown" electrode recordings, this tool enables high-throughput, long-term monitoring of heart cell health and drug toxicity without the need for invasive cell-puncturing techniques.

What Problem Does This Paper Address?

In cardiology and neuroscience, understanding the electrical "pulse" of a cell—the action potential—is the gold standard for diagnosing disease and testing new drugs. However, we face a "Goldilocks" problem with current tools:

  1. Patch-clamping is the gold standard for precision, but it is manual, invasive, and kills the cell quickly.
  2. Microelectrode Arrays (MEAs) are non-invasive and high-throughput, but they only record extracellular signals (eAPs). These are essentially "shadows" of the internal activity, lacking the detail needed to see how specific ion channels are performing.

The challenge is to get the rich, detailed data of an intracellular recording (iAP) using the non-invasive, high-throughput methods of extracellular sensors.

Methods

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

  • The Hardware: They used "nanocrown" electrodes—tiny, crown-shaped pillars that can record extracellular signals and, via a brief electric pulse (electroporation), gain temporary access to the inside of the cell to record ground-truth iAPs.
  • The Model: They adapted a UNET (commonly used in medical imaging) for 1D signal processing, adding "Attention" blocks to help the model focus on critical parts of the waveform, like the initial spike.
  • The "Physics" Ingredient: Purely data-driven models often hallucinate biologically impossible waveforms. To prevent this, the team integrated the Aliev-Panfilov model—a mathematical description of cardiac electrophysiology—directly into the AI's loss function. This acts as a "biological guardrail," forcing the AI to produce outputs that obey the laws of cardiac physics.

Datasets

The study utilized a unique, high-quality dataset of 2,364 synchronized eAP/iAP pairs from human induced pluripotent stem cell-derived cardiomyocytes (hiPSC-CMs).

  • Training Data: Included cells treated with various ion-channel blockers (like dofetilide and nifedipine) to create a wide variety of "abnormal" action potential shapes.
  • Validation: The model was tested against:
  • New devices not seen during training.
  • Cells treated with a completely different drug (Propranolol).
  • Data from commercial MEAs paired with traditional patch-clamp recordings to prove the model's generalizability beyond their custom lab hardware.

Key Insights and Ideas

  • The Extracellular "Fingerprint": The most significant "aha" moment was discovering that eAPs contain enough information to reconstruct iAPs with near-perfect correlation ($r = 0.99$). Features like eAP spiking velocity and decay rate are highly predictive of the internal state.
  • Physics as a Regularizer: The physics-informed loss function wasn't just a gimmick; it specifically corrected "super-repolarization" errors (where the model predicted the voltage dropping too low) and allowed the model to work on data from different types of sensors (MEAs vs. NEAs).
  • SHAP Analysis: By using SHAP (SHapley Additive exPlanations), the researchers could "peek inside the black box" to see which parts of the extracellular signal the AI was using to make its predictions, confirming that the model was focusing on biologically relevant features.

Key Results

  • High Accuracy: The model achieved a Mean Absolute Error (MAE) of only 0.041 on unseen test data.
  • Drug Sensitivity: The system successfully reconstructed the "prolonged" action potentials caused by cardiotoxic drugs, matching ground-truth measurements with high precision.
  • Scalability: The authors demonstrated the ability to reconstruct iAPs across 49 channels simultaneously, effectively turning a standard non-invasive array into a high-throughput intracellular screening tool.
  • Cross-Platform Success: The model successfully reconstructed iAPs from commercial MEA hardware, suggesting it could be deployed in existing labs without requiring custom nano-fabricated electrodes.

Limitations and Future Work

While a major step forward, the authors note a few hurdles:

  • Signal-to-Noise Ratio (SNR): The model’s performance degrades if the extracellular signal is too noisy. High-quality initial recordings are still a prerequisite.
  • Cell Type Diversity: This study focused primarily on cardiomyocytes. Future work is needed to see if the same physics-informed approach can be applied to neurons or other electrogenic cells.
  • Dataset Expansion: While 2,300+ pairs is large for this field, more diverse data (different cell lines, more drug types) will be required to build a truly "universal" reconstruction model.

Reference: Rahmani, K., et al. "Intelligent in-cell electrophysiology: Reconstructing intracellular action potentials using a physics-informed deep learning model trained on nanoelectrode array recordings." Nature Communications (2025). https://doi.org/10.1038/s41467-024-55571-6

Summary updated

Abstract

Intracellular electrophysiology is essential in neuroscience, cardiology, and pharmacology for studying cells' electrical properties. Traditional methods like patch-clamp are precise but low-throughput and invasive. Nanoelectrode Arrays (NEAs) offer a promising alternative by enabling simultaneous intracellular and extracellular action potential (iAP and eAP) recordings with high throughput. However, accessing intracellular potentials with NEAs remains challenging. This study presents an AI-supported technique that leverages thousands of synchronous eAP and iAP pairs from stem-cell-derived cardiomyocytes on NEAs. Our analysis revealed strong correlations between specific eAP and iAP features, such as amplitude and spiking velocity, indicating that extracellular signals could be reliable indicators of intracellular activity. We developed a physics-informed deep learning model to reconstruct iAP waveforms from extracellular recordings recorded from NEAs and Microelectrode arrays (MEAs), demonstrating its potential for non-invasive, long-term, high-throughput drug cardiotoxicity assessments. This AI-based model paves the way for future electrophysiology research across various cell types and drug interactions.
BibTeX
@inproceedings{rahmani2025intelligent,
  author = {Keivan Rahmani and Yang Yang and Ethan Foster and Ching‐Ting Tsai and Dhivya Pushpa Meganathan and Diego D. Alvarez and Aayush Gupta and Bianxiao Cui and Francesca Santoro and Brenda L. Bloodgood and Rose Yu and Csaba Forró and Zeinab Jahed},
  title = {{Intelligent in-cell electrophysiology: Reconstructing intracellular action potentials using a physics-informed deep learning model trained on nanoelectrode array recordings}},
  booktitle = {Nature Communications},
  year = {2025},
  doi = {10.1038/s41467-024-55571-6},
}
Figure