Artificial intelligence-based method and system for rapid drug screening using intracellular electrical recordings

Abstract

Disclosed are methods and systems for characterizing intracellular electrophysiology supported by artificial intelligence (AI). The disclosed techniques and systems capitalize on large amounts of intracellular action potential (iAP) and extracellular action potential (eAP) recordings, with high throughput collected from cells, using microelectrode arrays (MEAs) and / or nanoelectrode arrays (NEAs). In some aspects, an AI-supported method for characterizing an intracellular electrophysiological response of cells includes receiving, at a data processing module operating on a computing device, a data set comprising electrophysiological signals obtained from one or more cardiomyocyte cells on a sensor device; providing, to an artificial intelligence (AI) system comprising a trained machine learning (ML) model operating on a computer system, the data set; and reconstructing, by the trained ML model, an intracellular action potential (iAP) data set from the data set that predictively determines iAP waveforms or iAP characteristics of the one or more cardiomyocytes.

Description

ARTIFICIAL INTELLIGENCE-BASED METHOD AND SYSTEM FOR RAPID DRUG SCREENING USING INTRACELLULAR ELECTRICAL RECORDINGSCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This patent document claims priority to and benefits of U. S. Provisional Patent Application No. 63 / 717,739, titled “ARTIFICIAL INTELLIGENCE-BASED METHOD AND SYSTEM FOR RAPID DRUG SCREENING USING INTRACELLULAR ELECTRICAL RECORDINGS” and filed on November 7, 2025. The entire contents of the aforementioned patent application are incorporated by reference as part of the disclosure of this patent document.STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

[0002] This invention was made with government support under grant FA9550-23- 1-0090 awarded by the Air Force Office of Scientific Research (AFOSR). The government has certain rights in the invention.TECHNICAL FIELD

[0003] This patent document relates to techniques for intracellular signal sensing.BACKGROUND

[0004] Intracellular signaling is an important mechanism by which cells respond to their internal and external (extracellular) environment. Intracellular signals are used to sense the cellular environment, including gene expression and protein expression and modifications.

[0005] One method for measuring intracellular signaling is the patch clamp technique. The patch clamp technique is a highly sensitive electrophysiological method for measuring electrical currents across the membranes of cells, particularly electrically excitable cells like neurons. Patch clamp typically involves a technician to place a micropipette at the cell membrane of a cell and apply pressure (e.g., suction) to form a high-resistance seal with a cell membrane (e.g., GQ seal), allowing for the recording of currents through ion channels. Depending on the goal of the measurement, the patch clamp technique can be performed in various configurations where the cell membrane is left intact or is ruptured. For example, whole-cell patch clamp allows for the recording of both membrane potential and ionic currents across the entire cell but ruptures the cell membrane in the process. Cell-attached patch clamp, on the other hand, records ion channelactivity from a patch of the cell membrane without rupturing it, which is useful for studying single channels.

[0006] While the patch clamp technique is considered as an accurate and reliable method for measuring intracellular signals, it suffers low-throughput and is non-scalable to meet the growing demands, particularly toward the transition into personalized medicine.SUMMARY

[0007] Disclosed are methods, systems, and devices for characterizing intracellular electrophysiology supported by artificial intelligence (Al). The disclosed techniques and systems capitalize on large amounts of intracellular action potential (iAP) data and extracellular action potential (eAP) data, including iAP and eAP recordings collected from cells with high throughput using microelectrode arrays (MEAs) and / or nanoelectrode arrays (NEAs), which are used to train a machine learning model to create an exemplary Al system in accordance with the disclosed technology. As demonstrated by example implementations described herein, the Al system is able to predictively determine intracellular electrophysiological signals (e.g., iAPs) of cell types that the model has been trained on in response to varieties of stimuli and conditions, such as drug response.

[0008] In some aspects, a method for characterizing an intracellular electrophysiological response of cells supported by artificial intelligence includes receiving, at a data processing module operating on a computing device, a data set comprising electrophysiological signals obtained from one or more cardiomyocyte cells on a sensor device; providing, to an artificial intelligence (Al) system comprising a trained machine learning (ML) model operating on a computer system, the data set; and reconstructing, by the trained ML model, an intracellular action potential (iAP) data set from the data set that predictively determines iAP waveforms or iAP characteristics of the one or more cardiomyocytes.

[0009] In some aspects, a system for characterizing intracellular electrophysiology supported by artificial intelligence includes a sensor device to acquire extracellular action potential (eAP) signals from one or more cardiomyocyte cells; and a computing system in communication with the sensor device, the computing system comprising one or more computers each including a processor, and a memory coupled to the processor and storing instructions that, when executed by the processor, cause the computing system to perform operations comprising: receiving a dataset including the eAP signals, and reconstructing, by a machine learning (ML) model operating on the computing system, intracellular action potential (iAP) signals from the eAP signals, wherein the reconstructing the iAP signals predictively determines iAP waveforms or iAP characteristics of the one or more cardiomyocytes.

[0010] In some aspects, a method for characterizing an intracellular electrophysiological response of cells supported by artificial intelligence includes receiving, at a data processing module operating on a computing device, a data set comprising electrophysiological signals obtained from one or more cardiomyocyte cells on a sensor device; providing, to an artificial intelligence (Al) system comprising a trained machine learning (ML) model operating on a computer system, the data set and an instruction comprising a condition parameter; and reconstructing, by the trained ML model, an intracellular action potential (iAP) data set from the electrophysiological signals that predictively determines iAP waveforms of the one or more cardiomyocytes in response to the condition parameter.

[0011] In some aspects, a method for training a machine learning (ML) model for predicting intracellular response signals includes simultaneously acquiring, at a sensor device, pairs of intracellular action potential (iAP) signals and extracellular action potential (eAP) signals from cardiomyocyte cells on one or more electrodes of the sensor device; processing, at a data processing module operating on a computing device, the simultaneously acquired pairs of iAP and eAP signals based on a signal-to-noise (SNR) threshold indicative of a giga-ohm (GΩ) seal of the one or more cardiomyocyte cells to the electrode of the sensor device, the processing including selecting electrophysiological signal data that meets the SNR threshold; providing the selected electrophysiological signal data to a machine learning (ML) model operating on a computer system; and training the ML model to reconstruct an iAP data set from the selected electrophysiological signal data by recognizing patterns indicative of relationships between eAP waveforms and iAP waveforms of the selected electrophysiological signal data.

[0012] The subject matter described in this patent document can be implemented in specific ways that provide one or more of the following features.BRIEF DESCRIPTION OF THE DRAWINGS

[0013] FIG. 1A shows a diagram of an example embodiment of an Al-based system for determining intracellular signal responses, in accordance with the present technology.

[0014] FIG. IB shows a diagram illustrating an example embodiment of the data processing module that can be included or implemented with the example embodiments of the disclosed AI-based systems and methods, in accordance with the present technology.

[0015] FIG. 1C shows a diagram of an example embodiment of a method for characterizing an intracellular electrophysiological response of cells supported by artificial intelligence, in accordance with the present technology.

[0016] FIG. ID shows a diagram of an example embodiment of a method for training a machine learning (ML) model for predicting intracellular response signals, in accordance with the present technology.

[0017] FIGS. 2A-2H show images, diagrams, and data plots depicting example embodiments of nanoelectrode-based extracellular action potential (eAP) and intracellular action potential (iAP) data collection and pre-processing techniques, in accordance with the present technology.

[0018] FIG. 3A-3I shows data plots and data tables depicting example results showing quantitative relationships between eAPs and iAP waveform features based on example implementations of the disclosed technology.

[0019] FIG. 4A-4G shows a diagram and data plots associated with an example embodiment and exemplary implementations of a Physics-Informed Attention Unet (PIA-UNET) model, in accordance with the present technology, to reconstruct the entire iAP waveform from the eAP signal.

[0020] FIGS. 5A-5G show data plots from example implementations of example embodiments of the Al-based system and method, in accordance with the present technology, demonstrating high-throughput pharmacology by reconstructing iAPs from multi-channel eAP recordings and predicting the drug dose response of a cell type (e.g., cardiac myocytes).

[0021] FIG. 6 shows data plots depicting a comparison between patch clamp recorded intracellular electrical signals from extracellular signal recordings on example micro-electrode arrays (MEAs) and reconstructed extracellular signals created by implementation of the disclosed Al-based technique, in accordance with the disclosed technology.

[0022] FIG. 7 shows data plots of a violin plot distribution of eAP and iAP features for theexample data utilized in the machine learning and deep learning models as depicted in FIG. 4B.

[0023] FIG. 8A-8I show data plots depicting the iAPs in response to drugs in the exemplary implementations.

[0024] FIG. 9A-9F show a schematic of an example embodiment of a Quantile-Physics-Informed Attention Unet (QPIA-UNET), in accordance with the present technology, and data plots depicting example results from implementations of the exemplary QPIA-UNET, demonstrating the capability of reconstructing entire iAP waveforms from eAP signals with confidence interval.

[0025] FIGS. 10A-10F show diagrams and data plots for example implementations of simultaneous iAP recordings from cells using patch clamp and nanoelectrode arrays.

[0026] FIG. 11 shows data plots depicting an experiment-level comparison of normalized iAP from patch clamp and nanoelectrode array recording techniques.

[0027] FIGS. 12A-12D show diagrams and data plots for example implementations of simultaneous iAP recordings from cells using patch clamp and nanoelectrode arrays with introduction of an example drug (dofetilide).

[0028] FIG. 13 shows data plots depicting an experiment- level comparison of normalized iAPs from neighboring NEA channels.

[0029] FIGS. 14A-14D show data plots depicting a comparison of eAP and iAP baseline recordings across different experiments and a comparison of the drugs used in the example implementations, their dosages, and their impact on normalized iAP shapes.

[0030] FIGS. 15A and 15B show heatmap data plots depicting a correlation between temporal-related eAP features and between measurable eAP features, respectively, for eAP / iAP data utilized in the machine learning and deep learning models.

[0031] FIGS. 16A-16C show data plots depicting predicted iAPs using PIA-UNET, actual iAPs, iAPs generated by the Aliev-Panfilov model based on PIA-UNET physics parameter output values, and actual iAPs, for three exemplary test sets.

[0032] FIG. 17 shows a diagram depicting an example NEA layout showing an example of the distance between two neighboring channels.

[0033] FIGS. 18A and 18B show data plots of an example NEA-recorded eAP Activation map vs. its reconstructed action potential durations (e.g., APD50) map and a diagram of an exemplary star- shaped pattern of the NEA that indicates the channels arrangement.DETAILED DESCRIPTION

[0034] 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) recordings and extracellular action potential (eAP) recordings with high throughput. However, accessing intracellular potentials with NEAs remains a challenge.

[0035] The disclosed technology introduces a technique for characterizing intracellular electrophysiology supported by artificial intelligence (Al). The disclosed techniques are able to capitalize on thousands of synchronous eAP and iAP pairs collected from cells, such as stemcell-derived cardiomyocytes on microelectrode arrays (MEAs) and / or nanoelectrode arrays (NEAs). The disclosed Al-based platform and techniques include a physics-informed deep learning model trained on these datasets that successfully reconstructs iAP waveforms from extracellular recordings from NEAs and / or MEAs. Exemplary implementations of the disclosed Al-based platform and techniques demonstrated analysis on exemplary unique datasets that uncovered strong correlations between specific features of eAP and iAP waveform features, such as amplitude and spiking velocity, which were not detected before, indicating that extracellular signals can be reliable indicators of intracellular signal activity. Further, the example implementations of the disclosed Al-based platform and techniques demonstrated the exemplary deep learning model’s capability for non-invasive. long-term, and high-throughput assessments of drug cardiotoxicity.

[0036] The exemplary machine learning model can predict intracellular electrical signals from extracellular recordings without the need to penetrate or puncture the cell. The disclosed machine learning model incorporates bespoke training algorithms on multiple types of cellular signal datasets that apply relevant physics parameters to produce cellular function predictions with high confidence. Implementations of exemplary systems utilizing the disclosed machine learning model can provide a variety of advantages that reduce data processing resources as compared to conventional algorithms used to predict cellular function.

[0037] The disclosed Al-based platform and techniques can be used on any extracellularrecording obtained from electrode arrays, e.g., MEAs and NEAs. The MEAs and / or NEAs are generally used in biotechnology applications (e.g., commercially and in research laboratories) for drug discovery and to study the affect of drugs on electrical signals of anatomical and physiological aspects of living things, including but not limited to heart cells (e.g., cardiomyocytes). In some example implementations of the disclosed Al-based platform and techniques described below, an exemplary method, model, and computing instructions (e.g., code) can be used to obtain new information about how a drug or drugs impact the anatomical or physical aspect of the living being, such as the heart, using new and / or existing MEAs and / or NEAs.

[0038] Introduction

[0039] The drug development process is costly and inefficient, taking 10-15 years and approximately $5 billion per drug, with 63% of costs in preclinical development and 32% in clinical studies. A key challenge in drug development is the limited predictive power of preclinical screening, which relies on animal models and cell lines that may not accurately represent human physiology due to interspecies differences. Only 12% of drugs entering clinical trials are approved, with cardiotoxicity and hepatotoxicity as major reasons for drug failure, primarily due to drug-induced cardiotoxicity stemming from adverse effects on ion channels critical for action potential generation, which alter cardiomyocyte electrophysiology and elevate arrhythmic risks.

[0040] Electrophysiology, which investigates the electrical characteristics of biological cells and tissues, is crucial for understanding drug mechanisms, developing cardiac and neurological therapies, and evaluating cardiotoxicity across various drugs. Preclinical proarrhythmic assessments currently rely on in vitro hERG (human Ether-a-go-go Related Gene) potassium channel assays or in vivo electrocardiogram measurements in large animals. However, these conventional methods are costly and labor-intensive, so they are usually employed late in development, limiting the feasibility of making chemical modifications. Furthermore, evaluating a drug’s effect solely on the hERG channel can be insufficient, as drugs can affect multiple ion channels. For example, verapamil, a strong hERG blocker used to treat high blood pressure, abnormal heart rhythms, and chest pain (angina), is clinically safe due to its calcium channel blocking effect, which offsets its impact on hERG channels. Intracellular action potentials (iAPs) reflect the activity of multiple cardiac ion channels, and subtle changes in iAPs canindicate cardiotoxicity or serve as biomarkers. Advances in cardiomyocyte culture and the use of human induced pluripotent stem cell-derived cardiomyocytes (hiPSC-CMs) have greatly enhanced the availability of human cardiomyocytes for in vitro studies. Consequently, the Comprehensive in vitro Proarrhythmia Assay (CiPA) initiative, launched by the FDA and other international agencies, proposes in vitro screening of cardiac iAPs for assessing a drug’s risk of causing heart rhythm problems (proarrhythmia). This approach may offer a more accurate assessment of cardiac toxicity than the hERG assay and facilitate toxicity evaluation of early drug candidates.

[0041] The current gold standard for intracellular electrophysiology technique currently used in all areas of medical sciences is the patch clamp technique. The patch clamp method is very powerful and can measure intracellular potentials with very high precision. However, this technique is low in throughput, manual, and remains invasive to the recorded cell. Also, automated patch clamp systems, while improving throughput, require enzymatic dissociation of cells, which can alter cardiomyocyte electrophysiological properties. Alternative conventional techniques for assessing intracellular electrophysiology include optical recording of iAPs using voltage-sensitive dyes or proteins, but optical iAP recordings, while versatile and scalable, are significantly limited by low sampling rates, reduced sensitivity, and potential cytotoxicity from photobleaching, which may affect cellular behavior.

[0042] Extracellular electrophysiology techniques such as micro-electrode arrays (MEAs) overcome the invasiveness and throughput limitations of patch clamp, however, as the recording electrode remains outside the cell, this technique provides limited information on the shape of the electrical signals and cannot resolve the subtle changes in the intracellular potentials required for cardiotoxicity assessment.

[0043] Recently. nanoelectrode arrays, which can include free-standing nano- scale electrodes, e.g., up to 200 times smaller than the size of the cell, have emerged as a promising platform that combine the advantages of intra and extracellular electrophysiology techniques and have the ability to record high throughput extracellular signals, and on demand intracellular signals, in parallel, from single cells in various cell types such as neurons and cardiomyocytes. Intracellular signals include intracellular action potentials (iAPs). Extracellular signals include extracellular action potentials (eAPs). also referred to as field potentials. A variety of NEAs have recently been developed, each differentiated by their shape, throughput, and accessmechanisms to intracellular potential. Yet, a key challenge with NEAs lies in accessing the intracellular potential of cells. Certain NEAs can spontaneously access the intracellular space, but this generally suffers from limited probability and control. An alternative is the use of transient and controlled electroporation. This method, though more complex, uses a short electric pulse to temporarily create localized pores in the cell membrane, allowing the electrode to gain intracellular access and record iAPs at specific intervals. While these pores are highly localized, they can still disrupt cellular physiology at the nanoscale, and they eventually reseal, limiting the duration of intracellular recordings. A recording method which has the advantages of being high-throughput, long-term and non-invasive, like extracellular measurements, yet as accurate as patch clamp or intracellular recordings, would be optimal.

[0044] Example Embodiments

[0045] Disclosed are methods, systems, and devices for characterizing intracellular electrophysiology supported by artificial intelligence (Al). The disclosed techniques, systems, and devices utilize large amounts of iAP data and eAP data, including iAP and eAP recordings acquired from cells with high throughput using ME As and / or NEAs, to design and train a machine learning model that created an exemplary Al system in accordance with the disclosed technology. The disclosed Al system is able to predictively determine intracellular signals (e.g., iAPs) based on extracellular signal data for cell types that the machine learning model was trained on, including determining the intracellular response of a cell type to a variety of stimuli and conditions on, such as substances including but not limited to molecules (e.g., double or single stranded nucleic acids, ribonucleic acids, proteins, peptides, or gene editing components, such as viral vectors), nanoparticles, and / or chemical compounds, such as toxins or drugs, including known drugs and experimental drug designs.

[0046] For instance, some substances can act as a drug or a toxin, depending various factors including the cell type, uptake mechanism, concentration or dose, exposure duration, or other. As an illustrative example, the substances quinidine, flecainide, and dofetilide can be considered drugs and toxins (e.g., discussed later in section III of this disclosure), and other substances used in the example implementations described herein include cardiac cells ion channel blockers, which can be cardiotoxic at certain doses.

[0047] Example embodiments disclosed herein include a new approach for a machine learning model to reconstruct iAP waveforms using eAP recordings. Existing models use circuitelements relating the physical parameters such as gap size, double layer capacitance, and electrode properties to model the relationship between intracellular and extracellular membrane potentials. Models such as the Bidomain, Extracellular-Membrane-Intracellular, and Kirchhoff Network Model have attempted to describe physical relations between extracellular field potentials to intracellular waveforms in cardiac cells. However, a fundamental challenge common to these existing models is their heavy reliance on empirical data for parameter estimation and validation. The data, crucial for the accuracy of these models, is often sparse and not readily available, and might vary from cell to cell, restricting their broad applicability.

[0048] The disclosed machine learning model is created with a neural network architecture that integrates a U-Net architecture, attention gate mechanisms, and physics-informed learning, referred to herein as the Physics-Informed Attention-UNET (PIA-UNET). For example, in general, U-Net is a convolutional neural network configured for image segmentation, and attention gates refer to gated modules in machine learning model architecture that learn a function to produce a single-channel feature map. which is elementwise multiplied with the input to control the flow of information. Example embodiments of the PIA-UNET model, in accordance with the present technology, provides a deep learning approach for reconstructing iAP waveforms from eAP waveforms. Unlike traditional models that rely on extensive parameter estimation, the disclosed PIA-UNET intuitively translates the relationship between eAPs and iAPs by focusing on intrinsic patterns, thus, for example, bypassing the complex parameter estimation step For instance, the exemplary PIA-UNET deep learning model is substantially modified for optimization through hyperparameter tuning on a validation set derived from the training set and was subsequently evaluated on the training set and multiple test sets to assess the model’s performance. The disclosed technology advances NEA sensor technology and utilizes NEA device advancements to simultaneously record large amounts (e.g., thousands to millions) of eAP / iAP pairs.

[0049] In contrast with the disclosed embodiments of the machine learning model, for example, conventional approaches, such as classical forward / inverse models of a cell-electrode interface and tissue, require estimating many latent, cell-specific and / or device-specific parameters from sparse data. For instance, physical models such as bidomain / EMI / Kirchhoff-network and equivalent-circuit approaches heavily rely on empirical data for parameter estimation and validation, with parameters that vary from cell to cell, limitinggeneralizability. Examples of complex parameter estimations include cell-electrode interface inversion estimations, tissue / bidomain conductivity fitting, electrode geometry and coupling deconvolution, and ion-channel kinetic parameter fitting. Cell-electrode interface inversion estimations involve estimating, for each recording site, one or more of: seal resistance, cleft gap distance, access resistance, electrode double-layer capacitance / impedance spectrum, membrane capacitance and leak conductance, and the frequency-dependent transfer impedance between intracellular membrane currents and measured eAP. Tissue / bidomain conductivity fitting involves estimating intracellular vs extracellular conductivities and anisotropy ratios to relate field potentials to transmembrane potentials (bidomain / EMI / Kirchhoff models), where these parameters depend on cell morphology, coupling, and bath conductance and typically require separate calibration datasets or invasive measurements. Electrode geometry and coupling deconvolution involves inferring the unknown impulse response of each electrode-cell pair (a convolution of geometric spread, cleft impedance, and membrane pore state after electroporation) so that an eAP can be deconvolved into an iAP, which, notably, is typically underdetermined in practice without dense priors. Ion-channel kinetic parameter fitting is a complex parameter estimation that reproduces whole-cell iAP shapes from extracellular signals (multi-current Hodgkin-Huxley-type models), which requires non-unique fits and drug-specific parameter re-identification. Each of these estimation approaches are heavily complex, which the disclosed machine learning model (e.g., PIA-UNET deep model) is able to avoid for inference.

[0050] Through quantitative analysis of correlations between eAP and iAP pairs implemented using the disclosed Al-based platform and techniques, information in eAPs can be used to reconstruct iAPs for the given cells (e.g., cardiomyocytes). As demonstrated by the example implementations described below, the exemplary physics-informed deep learning model of the disclosed Al-based platform and techniques fully and accurately reconstruct and predict the iAPs from the eAP recordings.

[0051] In some embodiments in accordance with the disclosed technology, a method for characterizing intracellular electrophysiology includes providing eAP data of cardiomyocyte cells to an exemplary trained machine learning model; and reconstructing iAP waveforms by the trained machine learning model, where the reconstructing the iAP waveforms includes interpreting intrinsic, imperceptible patterns in the relationship between the eAP data and the iAP data. Implementations of the method provide numerous technical advantages over conventionaltechniques, e.g., particularly the advantage of bypassing complex parameter estimation techniques required in conventional approaches. In some implementations of the method, the reconstructing the iAP waveforms further includes analyzing the reconstructed iAP waveforms by correlating between at least some eAP and iAP pairs.

[0052] For example, an extracellular signal (e.g., eAP) is a spatially filtered, differentiated measure of transmembrane currents convolved with an unknown, site-specific transfer function that depends on cleft geometry, seal resistance, double-layer capacitance, and membrane pore state. This yields a highly nonlinear, context-dependent mapping from iAP to eAP. Notably, hand-crafted, low-dimensional algorithms built on a few features or on fixed-parameter circuit models are not expressive enough to invert that mapping across devices and preparations without extensive parameter fitting per site. The exemplary implementations of some embodiments in accordance with the disclosed technology, described later in this disclosure, show that even for arrhythmic signals, feature correlations exist but are subtle and multi-factorial, and that an example machine learning model, e.g., with attention and a physics-informed loss, is able to reconstruct full iAP shapes (e.g., with r~0.99) across multiple test sets, including different device classes (e.g., agnostic to NEAs or MEAs), without per-site parameter estimation. As such, the relevant patterns are intrinsic to eAP waveforms but imperceptible to conventional parametric techniques.

[0053] In some implementations of the method, for example, the machine learning model is configured as a deep learning model, and the reconstructing the iAP waveforms from eAP recordings uses time- synchronized pairs of eAP and iAP recordings with nanoelectrodes to train the deep learning model. Exemplary implementations of the method addressed a threshold question of whether enough information is contained in the eAP data (e.g., eAP waveforms recorded from NEAs or MEAs) for the exemplary deep learning model to be able to reconstruct the iAP waveforms. The results of such exemplary implementations showed that, based on the design of the deep learning model and the data signal processing technique to condition the training data for the model, unique but otherwise imperceptible patterns exist in the iAP and eAP waveforms to conclude that eAP waveforms contain sufficient information to serve as a stand-in for iAPs in biomedical applications of cardiomyocytes, such as drug screening or toxicity screening applications.

[0054] In some embodiments in accordance with the disclosed technology, a method forpredictively determining a cellular response, e.g., to a stimulus or stimuli, includes training a machine learning (ML) model using a cell electrophysiology data set comprising at least a thousand pairs of synchronous eAP waveforms and iAP waveforms of cardiomyocyte cells recorded simultaneously, the training the ML model comprising interpreting intrinsic, imperceptible patterns in a relationship between the eAP waveforms and the iAP waveforms; and generating, using the ML model, a predictive result of a chemical stimulus with respect to a viability or toxicity state of cardiomyocyte cells of a subject or a plurality of subjects.

[0055] Implementations of the disclosed Al-based system and method for predictively determining cellular response have shown that the disclosed technology can determine intracellular signals (e.g., iAPs) of a cell type (e.g., cardiomyocyte cells) on MEA sensor devices and NEA sensor devices alike. This technological breakthrough, for example, enables the characterization of a cell type’s intracellular electrophysiology while its extracellular signals are measured on ubiquitous biosensor platforms like MEAs, thereby enabling accurate predictive determinations of the cells’ intracellular signal response to various stimuli that can be used in drug screening applications and personalized medicine. Moreover, based on this high-throughput, the exemplary machine learning (ML) model can be further trained with the additional data from the MEAs (and / or NEAs) to continually refine its pattern recognition and signal determination capabilities, thereby adapting the exemplary ML model.

[0056] One of the example objectives of the Al-based system and method is to create an ML model that can provide high quality intracellular action potentials, e.g., intracellular data of the accuracy level of a patch clamp-acquired action potential data, from electrophysiological data provided on electrode arrays, e.g., NEAs and MEAs. First is to analyze the quality of the intracellular signals acquired using NEAs and MEAs (e.g., as compared with intracellular signals acquired using the patch clamp technique), for the same type of cells. Once verified, extracellular and intracellular pair action potential data was acquired on the NEAs (and MEAs) for a diverse set of conditions, including but not limited to different drugs or chemical compounds to simultaneously measure the cells’ action potentials intracellularly and extracellularly on the selected electrode sensor platform, i.e., NEAs and MEAs, respectively. This created a large, robust, and diverse data set that could be used to initially and adaptively train the ML model. The disclosed ML model was architected with physics parameters to enable the Al system to accurately predict the shape of the measured action potentials to their true shape(as compared with the best standard measurements: patch clamp), e.g., thereby reducing error of the output shape predictions. As a technical effect of the ML model’s design and configuration, for example, the ML model is generalizable, i.e.. able to predict the intracellular response to unseen chemical compounds, such as untested drugs. Moreover, as another exemplary technical effect of the ML model’s design and configuration, it is applicable across multiple platforms, i.e., agnostic to the data source — namely nanoelectrode array sensor devices and (more widespread) microelectrode array sensor devices, e.g., without having to perform any additional data processing to adjust from one platform to the other. For example, the ML model can predict the intracellular signal response to a drug, toxin, or chemical compound based on the sample data from a NEA device or a MEA device, without having to pre-process the data to indicate that the source data is NEA-based or MEA-based.

[0057] FIG. 1A shows a diagram of an example embodiment of an Al-based system 100 for determining intracellular signal responses, in accordance with the present technology. In some embodiments, as shown in FIG. 1A, the Al-based system 100 includes a data processing module 120, a machine learning module 130 for training data and data integration, and an Al signal analysis module 140 for reconstructing intracellular signal response results of cells (e.g., cardiomyocyte cells). In various embodiments of the Al-based system 100, the machine learning module 130 and the Al signal analysis module 140 are embodied on a computer system comprising one or more computing devices, e.g., including but not limited to one or more servers, one or more databases, and / or one or more desktop, laptop, or mobile computing devices (e.g., smartphones or smartwearable devices), which can be in data communication with each other. In some embodiments of the system 100, the machine learning module 130 and the Al signal analysis module 140 are configured as the same module on the computer system (e.g., referred to as an Al module 105), for example, such as in embodiments where the Al-based system 100 includes an adaptive trained ML model that is configured to provide reconstructed intracellular signal data and may still undergo additional data training. For some example embodiments of the Al module 105 (e.g., where the machine learning module 130 and the Al signal analysis module 140 are configured as the same module), the machine learning module 130 can include an embodiment of the previously-trained ML model that receives and further trains on additional data sets (e.g.. additional iAP and eAP pairs and / or other data), and the Al signal analysis module 140 can include an embodiment of the adaptively-trained ML model thatis configured to perform new iAP reconstructions after additional adaptive training.

[0058] In some embodiments, for example, the Al-based system 100 may (optionally) include a cell sensor platform 110. The (optional) cell sensor platform 110 can include, but is not limited to, one or more nanoelectrode array sensor device(s), one or more microelectrode array sensor device(s), and / or one or more patch clamp devices or systems. In some example embodiments of the (optional) cell sensor platform 110 including the NEA sensor device(s) and / or MEA sensor device(s) for intracellular signal measurements, the (optional) cell sensor platform 110 can further include a signal generator (e.g., signal pulse generator) to create electroporation signal stimuli to electroporate cells on the NEA(s) and / or MEA(s). For example, when training a machine learning model of the machine learning module 130, the (optional) cell sensor platform 110 can be configured to have the exemplary NEA(s) and / or MEA(s) simultaneously acquire pairs of intracellular action potential (iAP) signals and extracellular action potential (eAP) signals from cells on one or more electrodes of the NEA(s) and / or MEA(s).

[0059] In some embodiments, the data processing module 120 is configured to receive a data set, e.g., from a data storage device and / or from the (optional) cell sensor platform 110, which the data set includes electrophysiological signals acquired from one or more cells (e.g., cardiomyocyte cells) on a cell sensor device (e.g., the (optional) cell sensor platform 110). In some implementations, for example, the data processing module 120 is configured to process the electrophysiological signals based on a threshold and select certain electrophysiological signal data that meets the threshold. For example, the threshold can include a signal-to-noise (SNR) threshold, e.g., indicative of a giga-ohm (GO) seal of the one or more cells attached to the cell sensor device. The data processing module 120 is configured to provide the machine learning module 130 and / or the Al signal analysis module 140 with the electrophysiological signal data, e.g.. such as the electrophysiological signal data as received in the data set or processed electrophysiological signal data (e.g., threshold- screened and selected electrophysiological signal data).

[0060] In some implementations for training of an exemplary model of the machine learning module 130, for example, the machine learning module 130 is configured to undergo training of the ML model to reconstruct an iAP data set from the selected electrophysiological signal data by recognizing patterns indicative of relationships between eAP waveforms and iAP waveformsof the selected electrophysiological signal data. In some embodiments of the machine learning module 130, for example, the training data includes simultaneously acquired pairs of iAP and eAP signals that have signal characteristics related to an intracellular electrophysiological response of the cardiomyocyte cells due to exposure to a drug or chemical compound under a set of conditions.

[0061] In some implementations for predictively determining or characterizing an intracellular electrophysiological response of cells, the data processing module 120 provides the machine learning module 130 and / or the Al signal analysis module 140 with the electrophysiological signal data. In some implementations, the data processing module 120 provides the machine learning module 130 and / or the Al signal analysis module 140 with instructions, which can include a condition parameter (e.g., identification or information about a substance, such as drug(s) or toxin(s)). The machine learning module 130 and / or the Al signal analysis module 140 is / are configured to reconstruct an iAP data set from the electrophysiological signal data that predictively determines iAP waveforms of the one or more cardiomyocytes, which can be in response to a condition parameter. In some embodiments, for example, the machine learning module 130 and / or the Al signal analysis module 140 is configured to reconstruct the iAP data set by translating ML-trained iAP waveform characteristics into the iAP waveforms based on determined patterns indicative of relationships between eAP waveforms and iAP waveforms; and in some embodiments, the reconstruction of the iAP data set can include information about the condition parameter.

[0062] FIG. IB shows a block diagram illustrating an example embodiment of the data processing module 120 that can be included or implemented with the example embodiments of the disclosed Al systems and methods. The data process unit can include one or more processors (referred to as the processor) that can be in communication with one or more memory units (referred to as the memory or memory unit), one or more input / output units (referred to as the I / O unit), and / or one or more (optional) output unit(s). The processor is configured to process data, and the memory unit is in communication with the processor to store and / or buffer the data. To support various functions of the data processing module 120, the processor can be included to interface with and control operations of other components of the disclosed systems or other data processing devices, such as via the I / O unit and / or the (optional) output unit. The processor can include one or more processors, e.g., including but not limited to microprocessors such as acentral processing unit (CPU), microcontrollers, graphics processing unit (GPU), or the like. The memory unit can include and store processor-executable code, which when executed by the processor, configures the data processing module 120 to perform various operations, e.g.. such as receiving information, commands, and / or data, processing information and data, and transmitting or providing information / data to another device. The memory unit can store other information and data, such as instructions, software, values, images, and other data processed or referenced by processor. For example, various types of Random Access Memory (RAM) devices, Read Only Memory (ROM) devices. Flash Memory devices, and other suitable storage media can be used to implement storage functions of memory unit. The memory unit can store data and information, which can include sensor data, device and / or system parameters, data processing parameters, and processed parameters and data that can be used in the implementation of data processing techniques, e.g., techniques in accordance with the disclosed technology. The memory unit can store data and information that can be used to implement a data acquisition and / or processing method, e.g., including one or more algorithms for implementing a data acquisition or signal processing protocol in accordance with the disclosed technology.

[0063] In some implementations, the data processing module 120 includes an input / output unit (I / O) to interface the processor and / or memory unit to other modules, units or devices associated with the disclosed system, and / or external devices. The I / O unit can connect to an external interface, source of data storage, or display device. Various types of wired or wireless interfaces compatible with typical data communication standards, such as Universal Serial Bus (USB), IEEE 1394 (FireWire), Bluetooth, Bluetooth low energy (BLE), ZigBee, IEEE 802.11, Wireless Local Area Network (WLAN), Wireless Personal Area Network (WPAN), Wireless Wide Area Network (WWAN), WiMAX, IEEE 802.16 (Worldwide Interoperability for Microwave Access (WiMAX)), 3G / 4G / LTE / 5G / 6G cellular communication methods, and parallel interfaces, can be used to implement I / O unit. In some implementations, for example, the data processing module 120 includes a wireless communications unit (not shown), e.g., such as a transmitter (Tx) or a transmitter / receiver (Tx / Rx) unit, which can be in data communication with the processor and / or the memory. The I / O unit can interface the processor and memory unit with the wireless communications unit to utilize various types of wireless interfaces, such as the examples described above. The I / O unit can interface with other external interfaces, sources of data storage, and / or visual or audio display devices, etc. to retrieve and transfer data andinformation that can be processed by the processor, stored in the memory unit, or exhibited on an output unit of a user device (e.g., display screen of a computing device) or an external device.

[0064] To support various functions of the data processing module 120, the data processing module 120 may optionally include one or more output units that can be used to exhibit data implemented by the example system disclosed herein. The (optional) output unit(s) can include various types of display, speaker, or printing interfaces to implement output functionalities the system. In some embodiments, for example, the (optional) output unit(s) can include cathode ray tube (CRT), light emitting diode (LED), or liquid crystal display (LCD) monitor or screen as a visual display. In some examples, the (optional) output unit can include toner, liquid inkjet, solid ink, dye sublimation, inkless (such as thermal or UV) printing apparatuses to implement some output modalities of the (optional) output unit(s). In some examples, the (optional) output unit(s) can include various types of audio signal transducer apparatuses.

[0065] FIG. 1C shows a diagram of an example embodiment of a method 160 for characterizing an intracellular electrophysiological response of cells supported by artificial intelligence, in accordance with the present technology. The method 160 includes a process 162 to receive, at a data processing module (e.g., data processing module 120) operating on a computing device, a data set comprising electrophysiological signals obtained from one or more cardiomyocyte cells on a sensor device (e.g., cell sensor platform 110). The method 160 includes a process 164 to provide the data set to an artificial intelligence (Al) module (e.g., one or both of machine learning module 130 or Al signal analysis module 140, or the Al module 105) that includes a trained machine learning (ML) model operating on a computer system. The method 160 includes a process 166 to reconstruct, by the trained ML model, an intracellular action potential (iAP) data set from the data set that predictively determines iAP waveforms or iAP characteristics of the one or more cardiomyocytes.

[0066] Example embodiments and implementations of the method 160 can include the following features. For example, in some implementations of the method 160, the electrophysiological signals include extracellular action potential (eAP). For example, in some embodiments of the method 160, the process 164 to reconstruct the iAP data set predictively determines the iAP waveforms or iAP characteristics of the one or more cardiomyocytes in response to a substance or a condition on the cardiomyocytes. For example, in some embodiments, the method 160 further includes a process to provide to the Al system aninstruction specifying the substance or condition to predictively determine the response. For example, in some embodiments of the method 160, the process 164 to reconstruct the iAP data set includes a process to translate ML-trained iAP waveform characteristics into the iAP waveforms based on determined patterns indicative of relationships between extracellular action potential (eAP) waveforms and iAP waveforms. For example, in some implementations of the process 164, the process to translate the ML-trained iAP waveform characteristics into the iAP waveforms can be (further) based on information about the substance or condition. For example, in some implementations of the process 164, the process to translate the ML-trained iAP waveform characteristics into the iAP waveforms can be without any information about the substance or condition. For example, in some embodiments, the method 160 further includes a process to data-process, e.g., at the data processing module, the electrophysiological signals based on a signal-to-noise (SNR) threshold indicative of a giga-ohm (GΩ) seal of the one or more cardiomyocyte cells to an electrode of the sensor device, the processing including selecting electrophysiological signal data that meets the SNR threshold. For example, in some embodiments, the method 160 further includes a process to data-process, e.g., at the data processing module, the electrophysiological signals based on a signal threshold comprising an amplitude level of the one or more cardiomyocyte cells recorded by the sensor device, the processing including selecting electrophysiological signal data that meets the signal threshold.

[0067] FIG. ID shows a diagram of an example embodiment of a method 170 for training a machine learning (ML) model for predicting intracellular response signals, in accordance with the present technology. The method 170 includes a process 172 to simultaneously acquire, at a sensor device (e.g., cell sensor platform 110). pairs of intracellular action potential (iAP) signals and extracellular action potential (eAP) signals from cardiomyocyte cells on one or more electrodes of the sensor device. The method 170 includes a process 174 to data-process, at a data processing module (e.g., data processing module 120) operating on a computing device, the simultaneously acquired pairs of iAP and eAP signals based on a signal-to-noise (SNR) threshold indicative of a giga-ohm (GΩ) seal of the one or more cardiomyocyte cells to the electrode of the sensor device. In some implementations of the process 174, the data-processing can include selecting electrophysiological signal data that meets the SNR threshold. The method 170 includes a process 176 to provide the selected electrophysiological signal data to a machine learning (ML) model (e.g., of the machine learning module 130) operating on a computersystem. The method 170 includes a process 178 to train the ML model to reconstruct an iAP data set from the selected electrophysiological signal data by recognizing patterns indicative of relationships between eAP waveforms and iAP waveforms of the selected electrophysiological signal data.

[0068] Example embodiments and implementations of the method 170 can include the following features. For example, in some implementations of the method 170, the simultaneously acquired pairs of iAP and eAP signals can include signal characteristics related to an intracellular electrophysiological response of the cardiomyocyte cells due to exposure to a drug or chemical compound under a set of conditions. For example, in some embodiments of the method 170, the process 174 to data-process the simultaneously acquired pairs of iAP and eAP signals can include: normalizing eAP waveforms and iAP waveforms; identifying action potential peaks of the normalized eAP waveforms and the normalized iAP waveforms; and segmenting the normalized eAP waveforms and the normalized iAP waveforms around the identified action potential peaks. Example embodiments and implementations of the method 170 can further include a process to validate the training of the ML model by calculating action potential durations (APDs) from the reconstructed iAP waveforms, wherein the APDs comprise measurements that measure timeframes for an action potential to decrease to respective percentages of peak value and recover. For example, in some implementations of the method 170, the process 178 to train the ML model can further include estimating confidence intervals for iAP values.

[0069] Exemplary Results and Discussion of Example Implementations

[0070] Example implementations of some embodiments of the disclosed Al-based systems and methods for training a machine learning model on eAP and iAP data and for predictively determining intracellular signals, e.g., in response to known and unknow substances, are described. The example implementations were performed on multiple types of cardiomyocyte cells using a variety of substances, including drugs and toxins to the cardiomyocyte cells. The example implementations were also performed using NEA, MEA, and patch clamp sensor devices.

[0071] Recording of time-synchronized pairs of eAP and iAP.

[0072] FIGS. 2A-2H show images, diagrams, and data plots depicting example embodiments of nanoelectrode-based eAP and iAP data collection and pre-processing techniques, inaccordance with the present technology.

[0073] FIG. 2A shows an illustrative diagram (right) and a scanning electron microscopy (SEM) image (left) of an exemplary NEA channel of an example embodiment of a NEA device (not shown), where the NEA channel shown in the diagram and images includes a plurality of (e.g., nine) nanocrown electrodes per cell-electrode channel. The illustrative diagram shows a side view and top view of the nine nanopillar structures for the example NEA channel; and the SEM image depicts the example NEA channel (scale bar 10 pm) with a zoomed- in SEM image view showing three of the conductive material-coated nanopillars and another zoomed-in SEM image view showing a single nanopillar structure of the array (scale bars 2 pm). Example multichannel NEA device layouts are shown later in FIGS. 17 and 18B. Also, some example embodiments of NEA devices that can be implemented for the disclosed methods and systems are described in U. S. Patent Application Publication No. US2021 / 0008363 Al, which is incorporated by reference as part of this disclosure for all purposes.

[0074] FIG. 2B shows a diagram depicting a methodological comparison for capturing cardiac action potentials, presenting NEA-based eAPs and NEA-based iAPs alongside the reference patch clamp technique (e.g., ordered by their degree of invasiveness from left to right). Notably, the exemplary NEA’s functionality to convert eAPs into iAPs through precise biphasic electric pulses (electroporation) is demonstrated.

[0075] FIG. 2C shows a diagram depicting simultaneous iAP recordings from neighboring channels via non-invasive extracellular action potential (NEA) electroporation.

[0076] FIG. 2D shows a data plot depicting the exemplary NEA intracellular and extracellular recordings for an example multi-step addition process of an example drug, dofetilide (an antiarrhythmic drug used to treat and maintain normal sinus rhythm in patients with atrial fibrillation and atrial flutter), which was administered in concentrations of 0.3, 1, 3, and finally 10 nM. The process was conducted in three stages at approximately 400. 800. and 1,200 seconds during the recording session. This method was employed to collect a diverse range of iAP shapes, facilitating a more comprehensive analysis. The expanded diagram of FIG. 2D depicts the process similar to PC vs. NEA iAP recording, e.g., a comparison of iAP recordings from neighboring channels by normalizing and segmenting them into arrays of length 800 indices or 1.6 s, where the scaling is between 0 to 1.

[0077] FIG. 2E shows box plots depicting the distribution of the difference in cycle time(dCT), correlation coefficient (r), mean absolute error (MAE), and APD50% and APD90% errors between neighboring and NEA normalized iAP pairs with NEA S / N > S / N* (= 90), respectively. The exemplary data in the box plots is for n=2,661 samples, comparing 22 pairs of neighboring iAP channels from two independent cultures. The box plots show the median (center line), the interquartile range (IQR) as box bounds, the whiskers (1.5xIQR), and the outliers (points beyond whiskers).

[0078] FIG. 2F shows data plots depicting examples of iAP pairs from neighboring channels with the highest MAE, APD50% errors, and APD90% errors, as indicated in the box plots.

[0079] FIG. 2G shows a diagram depicting simultaneous iAP and eAP recordings from neighboring channels, including NEA electroporation of cardiac myocyte cells.

[0080] FIG. 2H shows a data plot depicting a collection of diverse iAP waveforms and corresponding eAPs from neighboring channels on an exemplary NEA in hiPSC-CM cells, following the addition of drugs. This process involves applying electroporation to obtain iAPs and non-electroporation techniques for eAPs. Similar to the patch clamp vs. NEA iAP recording method, the iAPs are scaled between 0 to 1, but this scaling is different from what was used in the training data. For eAPs, a specific segment of the signal, particularly values between indices 1150 and 1350, is selected to characterize signal noise. The function calculates the standard deviation (std) of this segment. The eAP signal is then processed by subtracting the mean of the entire eAP signal from each value, followed by normalization relative to its noise level. This is achieved by dividing the mean-adjusted values by 60 times the calculated standard deviation. The signals are then segmented into windows of 800 indices or 1.6 seconds. The figure presents a wide range of collected eAPs and iAPs, including an overlay of the non-normalized eAP and iAP pairs.

[0081] The example results depicted in FIGS. 2A-2H demonstrate the successful reconstruction of iAP waveforms from eAP recordings by using time- synchronized pairs of eAP and iAP recordings with nanoelectrodes to train a deep learning model. The example data shown in these figures and throughout this disclosure demonstrated that the eAP data contains sufficient information for the disclosed deep learning model to accurately reconstruct iAP data. The disclosed deep learning model achieved this finding through its design and by the quality of the training dataset.

[0082] While previous studies have demonstrated some similarity between normalized iAPwaveforms recorded by NEAs and the gold standard patch clamp technique, such existing approaches are ineffective at identifying the otherwise undetectable patterns within the iAP and eAP waveforms to distinguish the relationship that can be relied upon for determining cellular electrophysiological behavior (of cardiomyocytes) to various stimuli, whether physical, biological, or chemical. The disclosed machine learning approach conducted a more comprehensive analysis (as detailed in the section I {'Assessing the Accuracy of NEA Recordings When Compared to Patch Clamp”) of this disclosure) that informed the design of the deep learning model and the data processing techniques for training the model. For example, the disclosed methods and systems include a data processing technique that establishes a signal-to-noise ratio (S / N) cutoff, e.g., of 90 dB (the eAP magnitude relative to the recording noise), to ensure the fidelity of the NEA recordings relative to standard patch clamp recordings. This S / N cutoff is enforced when selecting data in all subsequent analyses.

[0083] Furthermore, in the example implementations described herein, for example, to obtain precise time- synchronized pairs of eAP and iAP waveforms, recordings from neighboring nanoelectrode channels were used in a confluent monolayer of human stem-cell-derived cardiomyocytes that are in close physical proximity, where one nanoelectrode channel measured intracellularly while the other extracellularly. Moreover, for example, in a separate experimental implementation, it was demonstrated that two neighboring nanoelectrode channels exhibit highly similar iAP waveforms when comparing them with various metrics including differences in their cycle times (dCT) and action potential durations (APD), as well as their correlation (r) and Mean Absolute Error (MAE) over an extended period of time, and under various drug conditions, as shown in FIGS. 2C-2F, and detailed in section II (“Tw Neighboring Channels on NEA Show Similar iAPs") of this disclosure). Therefore, in the example implementations, it was assumed that the iAP collected from a nanoelectrode channel can act as an accurate training target to reconstruct from the eAP measured in a neighboring nanoelectrode channel, as the iAP waveforms are extremely similar over nanoelectrode channels in close proximity. To collect a diverse spectrum of synchronized eAP and iAP pairs for the exemplary training dataset, various ion-channel blockers were introduced in a dose-dependent manner to hiPSC-derived cardiomyocytes (hiPSC-CMs) to achieve different action potential shapes. For example, the drugs used in these example implementations included dofetilide (primarily blocks hERG (IKr) potassium channels); quinidine (blocks Na-i- channels (INa), K+ channels (IKr and IKs), andCa2+ channels (ICa)); nifedipine (selectively blocks L-type Ca2+ channels (ICa-L)); flecainide (blocks Na+ channels (INa) and, to a lesser extent, K+ channels (IKr)); lidocaine (primarily blocks Na+ channels (INa)); and propranolol, a beta-blocker that also blocks Na-i- channels (INa). For instance, a total of 2364 synchronized eAP / iAP pairs were collected from independent recording sets (see Exemplary Methods section). Data were obtained from ten independent experiments using different cell cultures on different NEAs from different batches and / or wafers.

[0084] Correlations between eAP and iAP waveform features.

[0085] The distinctive capability of the disclosed Al-based platform to record time-synchronized eAP and iAP pairs enables the direct comparison of these two waveforms and computing correlations between their shape features. This is important to quantify whether there is any plausibility to the claim that the eAP holds enough information to reconstruct the iAP. For instance, eAP and iAP pairs recorded from arrhythmic cells revealed a strong correlation (r=0.903) between the eAP amplitude and the iAP spiking velocity, as shown in FIG. 3A. To quantitatively describe the correlations between iAP and eAP pairs, the waveforms were characterized by defining several features on each waveform (FIG. 3B) and calculating their cross-correlations (FIG. 3C). The distribution of these eAP and iAP features for the dataset is depicted in FIG. 7.

[0086] FIG. 3A-3I shows data plots and data tables depicting quantitative relationships between eAPs and iAP waveform features.

[0087] FIG. 3A show data plots of a representative recording of simultaneous eAP and iAP from arrhythmic cells (top), and an overlay of eAP amplitude (representing its maximum voltage during the spiking phase [mv]) with iAP spike velocity (represented as the percentage change in iAP voltage during its spiking phase over time [% change in iAP / s]), showing a strong correlation (r=0.903). This association is further evident as oscillations in the extracellular recordings appear to reflect the action potential's repolarization phase.

[0088] FIG. 3B shows data plots illustrating key points that were identified on the eAP (left) and iAP (right) recordings to describe their waveforms. Key points on the eAP include: break point 1 (bpi) - just before the spike where its derivative notably rises; and break point 2 (bpz) -immediately post its minimum, as the derivative starts decreasing. Other defining points are: ymax (the peak positive value), ymin (negative spike minimum following the positive peak), andbP3 (following the first minimum, marking where the eAP derivative becomes positive). The vertical distance from bpi to ymaxis termed AVi; AV2 defines the vertical distance from bp2 to ymin; ATi defines the width of the positive spike from bpi to its rightmost pre-minima point; AT2 signifies the width of the negative spike from bp2 to its leftmost pre-minima point; ATd represents the horizontal distance between bp2 and bp3. The decay rate (DR, v / s) is the average slope between bp2 and bps, reflecting the voltage change over time, while the increase rate defines the slope just after ymin. In the context of iAP signals, APD10 to APD100 describe the duration it takes for the action potential to return to 10% and 100% of its amplitude, respectively. The increase rate (IR, v / s) is determined by the slope from iAP bpi to its peak, while the decay rate gauges the slope from the peak to bp2. AVi is the vertical span between bpi and the peak, and ATScharts the vertical distance between bpi and bp2. Additionally, examples of eAPs with distorted spikes are also presented.

[0089] FIG. 3C shows a data table demonstrating a correlation analysis of eAP features with iAP APD values on undistorted eAP / iAP pairs (n=1049), e.g., showcasing the relationship between various features of eAPs and iAP APD values ranging from APD 10 to APD 100. This section provides insights into how eAP characteristics correlate with the corresponding iAP APD metrics.

[0090] FIG. 3D shows a violin plot distribution of normalized eAP and iAP waveform features. This plot illustrates the distribution of specific eAP features (ATS, AVa, DR, and AVi) and iAP features (APD30, APD50, APD70, and APD90), where ntraining=1512, nTesti=272, nTest2=171 and, nTest3=91).

[0091] FIG. 3E shows a data table demonstrating a comparison of predicted and actual APD, e.g., which showcases reconstructed APD lines for the shortest and longest action potential durations from the test set, illustrating model accuracy across varying AP durations. An example XGBoost model underwent optimization through hyperparameter tuning on a validation set derived from the training set and was evaluated on the training set and three test sets to assess the model’s performance. Test 1 involved hiPSC-CMs exposed to dofetilide, a drug included in the training set, which were tested on a different NEA from a separate device to obtain eAP / iAP pairs. Test 2 involved propranolol, a drug not used in the training set, tested on a different NEA device to obtain eAP / iAP pairs. Test 3 involved eAP / iAP pairs recorded from a different laboratory using separate hiPSC-CMs (e.g., commercial hiPSC-CMs) on a MEA platform (e.g.,commercial MEA device), where simultaneous patch clamp recordings of the same cells were used to obtain iAPs.

[0092] FIG. 3F shows a data plot depicting a comparative analysis of predicted APD error values for test and training data. For example, this section evaluates the average prediction errors for APD-30, APD-50, APD-70, and APD-90 across different test sets. For Test Set 1. the errors were 0.010 ± 0.006 s, 0.016 ± 0.012 s, 0.019 ± 0.011 s, and 0.033 ± 0.018 s. respectively. In Test Set 2, the errors were 0.061 ± 0.014 s, 0.054 ± 0.011 s, 0.025 ± 0.007 s, and 0.028 ± 0.011 s. For Test Set 3, the errors were 0.011 ± 0.007 s, 0.055 ± 0.045 s, 0.059 ± 0.070 s, and 0.058 ± 0.063 s. When expressed as percentages, the prediction errors for the APD benchmarks are as follows. For Test Set 1, the errors were 2.390 ± 1.358% for APD-30, 2.460 ± 1.706% for APD-50, 2.373 ± 1.393% for APD-70. and 3.584 ± 2.059% for APD-90. In Test Set 2, the errors were 13.320 ± 2.978% for APD-30, 8.154 ± 1.732% for APD-50, 3.474 ± 0.908% for APD-70, and 3.655 ± 1.368% for APD-90. For Test Set 3, the errors were 2.695 ± 1.671% for APD-30, 11.093 ± 9.079% for APD-50, 10.598 ± 12.540% for APD-70. and 9.326 ± 10.222% for APD-90. These figures demonstrate the robustness of the XGBoost model in accurately predicting APD values ranging from APD 10 to APD 100.

[0093] FIG. 3G shows a data plot of a violin plot of average APD error on test and training sets. The mean absolute APD error from APD10 to APD100 for the test sets were as follows: Test 1 had a mean of 0.020 ± 0.007 s, Test 2 had a mean of 0.040 ± 0.006 s, and Test 3 had a mean of 0.047 ± 0.038 s. For the training set, the mean error was 0.002 ± 0.001 s.

[0094] FIG. 3H shows a data plot showing feature significance based on their SHapley Additive exPlanations (SHAP) values indicates respectively, Td, VI / AV2. ATS2, AVi, and IRi as the most important eAP features to predict iAP features. SHAP values illustrate how features alter the predicted value from the model’s average output. Feature importance is defined as the average of absolute changes imposed on the predicted values by varying features within their range.

[0095] FIG. 31 shows data plots depicting the ranked significance of eAP signal features in predicting APD30, APD50, APD70, and APD90 is illustrated, along with their local partial dependency. Each dot signifies a single predicted APD value, its color indicates the feature's value, and its position on the X-axis represents its SHAP value reflects the expected deviation in APD prediction. For example, as the part APD90 shows, increasing ATS(going from blue to redcolor), results in increasing the predicted APD90 value, while ATSand APD30 exhibit the opposite direction. The observed SHAP values could be influenced by local minima, potentially limiting their representation of the global relationship between features. This underscores the importance of considering broader contextual factors when interpreting these dependencies. The box plots show the median (center line), interquartile range (IQR; box bounds), whiskers (1.5xIQR), and outliers (points beyond whiskers).

[0096] From the example data, it is noted that the amplitude-related characteristics, but not the shape, of iAPs are significantly influenced by the NEA recording technique (e.g., the quality of electroporation and size of membrane pores that electrically connect NEAs to the intracellular domain). Conversely, the amplitude of eAPs can be impacted by variations in iAPs waveforms. This is further evidenced by the gradual decrease in iAP amplitude over time, due to gradual resealing of induced membrane pores, an effect not observed with eAPs (e.g., depicted in FIG.8D-8G). Hence, the focus was on attributes less sensitive to recording technique for iAPs while detailing the iAP waveform traits. The exemplary APD metrics, e.g.. specifically APD10 to APD100, measure the timeframe for an action potential to decrease to 10% and recover to 100% of its peak value, providing insight into the waveform’s recovery phase independent of amplitude variations. These temporal aspects offer a more stable basis for comparing iAP and eAP dynamics.

[0097] Analysis of several eAP features, e.g., features described in FIG. 3B and the section IV (“ P Features Determination Methods”) of this disclosure, showed strong correlations with iAP durations (APDs), with correlation coefficients greater than 0.60, e.g., shown in FIG. 3C. To identify which specific features of the eAP signal can precisely predict iAP features and to quantitatively explore the relationships between these eAP and iAP features, an efficient, fast and scalable tree boosting machine learning method, XGBoost, was first used. XGBoost was used as a feature-engineering baseline to quantify which eAP features predict APD metrics and to demonstrate generalizability. Next, the data was split to assess its accuracy and generalizability.

[0098] The training dataset included eAP / iAP recordings from NEA using hiPSC-CMs exposed to multiple drugs, for example: dofetilide, quinidine, nifedipine, flecainide, and lidocaine. To validate the generalizability of the example study, three different test sets were used:1. Test 1: hiPSC-CMs exposed to dofetilide, a drug included in the training set, were tested on a different NEA device that was not used for training recordings, to obtain eAP / iAP pairs.2. Test 2: hiPSC-CMs exposed to Propranolol (a drug not included in the training set) were tested on a different NEA device to obtain eAP / iAP pairs.3. Test 3: eAP / iAP pairs were recorded from a different laboratory using commercial hiPSC-CMs on an example MEA. For this test, simultaneous patch clamp recordings of the same cells were used to obtain iAPs.

[0099] FIG. 3D shows the distributions of the training and test sets for some of the eAP features and iAP APD values, and FIGS. 10A-10F shows the distributions of all features. The example results indicate that XGBoost can accurately predict iAP features from eAP features across a wide range of waveforms, as shown in FIG. 3E, which presents reconstructed APD lines comparing predicted values with actual ones for all test set scenarios. The example XGBoost model’s prediction accuracy for APD benchmarks (APD30, APD50, APD70, and APD90) was assessed across training and test sets (FIG. 3F and the Additional Exemplary Results section of this disclosure). The mean absolute APD error across all benchmarks was 0.002 ± 0.002 s for the training set, 0.027 ± 0.009 s for Test Set 1, 0.046 ± 0.012 s for Test Set 2, and 0.046 ± 0.036 s for Test Set 3 (FIG. 3G). When expressed as percentages, the mean errors were 0.48 ± 0.37% for the training set, 4.45 ± 1.32% for Test Set 1, 10.13 ± 2.70% for Test Set 2, and 11.00 ± 6.39% for Test Set 3.

[0100] To understand which eAP features are most important in predicting iAP features, SHapley Additive exPlanations (SHAP) summary plots were utilized (FIGS. 3H and 31). These plots rank features by their influence on APD predictions, quantified as the average absolute shift in predicted APDs caused by variations in each feature. The exemplary analysis indicates that ATd, V1 / AV2, ATS. and increase rate (IR) (FIG. 3B) are the most important predictors of APD values (FIG. 3H). These exemplary results are in agreement with previous qualitative relationships shown between eAP features and iAP features. Furthermore, as shown in FIG. 31, the top three eAP features locally affecting APD30, APD50, APD70, and APD90 predictions (reflecting 30, 50, 70 and 90% cell repolarization levels) were identified (FIG. 31). For APD70 and 90, ATd, increase rate (IR). and ATSwere found to be crucial predictors. Higher ATd and, to a lesser extent, ATSpredict increased APD values, whereas a lower IR suggests lower APDvalues. Td plays the same role concerning APD30 and APD50. Furthermore, for APD30 and APD50, ATSand the ratio AV1 / AV2 are the other most significant influencers. Interestingly, for both APD30 and APD50, higher ATS, as opposed to APD70 and APD90, is associated with lower APD values. These new relationships revealed between eAP and iAP features were identifiable only because of the availability of the unique dataset produced by the disclosed technology, which included several eAP and iAP pairs. Having identified eAP features as accurate predictors of iAP features, it was then investigated whether the entire iAP waveform could be reconstructed from eAP using deep learning.

[0101] Deep Learning for reconstructing iAPs from eAP.

[0102] Deep learning algorithms excel at processing and analyzing high-dimensional data (i.e., data with many features or variables), as is the case with the eAP and iAP signals. In efforts to reconstruct the entire iAP waveform from the corresponding eAP, the disclosed technology produced a modified Attention-Residual-Block UNET model, enhanced with a pseudo-physics loss function.

[0103] FIG. 4A-4G shows a diagram and data plots associated with an example embodiment and exemplary implementations of a Physics-Informed Attention Unet (PIA-UNET) model, in accordance with the present technology, to reconstruct the entire iAP waveform from the eAP signal.

[0104] FIG. 4A shows a diagram of an example embodiment of a PIA-UNET Model, in accordance with the present technology. The PIA-UNET model was utilized in the study. The exemplary Attention Physics informed UNET (PIA-UNET) model includes an encoder and decoder, both constructed with Residual Block (ResBlock) functions. The encoder compresses eAP features while the decoder reconstructs the iAP. Each ResBlock contains convolutional, batch normalization, and ReLU (CBR) sequences with squeeze-and-excitation (SE) blocks for enhanced feature representation. The architecture also incorporates skip connections and attention mechanisms, enhancing gradient flow and focus on significant data features for precise iAP reconstruction from eAPs. The Aliev-Panfilov is Incorporated into the PIA-UNET as the physics part of the hybrid loss function.

[0105] FIG. 4B shows a set of data plots depicting the Comparison of reconstructed iAPs using PIA-UNET with (orange dashed line) and without (blue dashed line) physics-informed loss functions, alongside the actual iAPs derived from corresponding eAPs (black line), wasperformed on three test sets. Test 1 involved hiPSC-CMs exposed to dofetilide, a drug included in the training set, tested on a different NEA from a separate device. Test 2 used propranolol, an unseen drug, tested on a different NEA from a separate batch and wafer. Test 3 comprised eAP / iAP pairs recorded from a different laboratory using commercial hiPSC-CMs on an example MEA platform (e.g., commercially available MEA device), where simultaneous patch clamp recordings provided iAPs. Incorporating physics corrected the brief undershoot at the end of repolarization in some NEA iAP recordings (Test 2), and resulted in more accurate predictions on eAPs recorded with the MEA (Test 3).

[0106] FIG. 4C shows a data plot depicting the exemplary model’s predicted potential values closely aligning with the actual values, with a coefficient of determination (r = 0.99) calculated across all 8000 points, based on 421, 187, and 149 samples for the three respective test sets.

[0107] FIG. 4D shows a data plot depicting the mean absolute error (MAE) exhibited by the exemplary model. The data plot shows MAE of 0.032 ± 0.041 on Test 1, 0.051 ± 0.050 on Test 2, and 0.038 ± 0.045 on Test 3. The MAE on the training set is 0.021 ± 0.031, indicating lower error during training. The higher MAE on Test 2 is attributed to the correction of NEA iAP through the incorporation of physics into the loss function.

[0108] FIG. 4E shows a data plot depicting APD errors across the training set and the three test sets. Reconstructed iAPs are compared with actual ones in terms of APD. For Test 1, the APD errors are APD30: 0.029 ± 0.020 s, APD50: 0.028 ± 0.022 s, APD70: 0.017 ± 0.010 s, and APD90: 0.045 ± 0.021 s. For Test 2, the APD errors are APD30: 0.036 ± 0.026 s, APD50: 0.024 ± 0.028 s, APD70: 0.015 ± 0.025 s, and APD90: 0.027 ± 0.013 s. For Test 3, the APD errors are APD30: 0.012 ± 0.009 s, APD50: 0.016 ± 0.017 s, APD70: 0.021 ± 0.022 s, and APD90: 0.029 ± 0.018 s. For the training set, the APD errors are APD30: 0.017 ± 0.015 s, APD50: 0.018 ± 0.018 s, APD70: 0.017 ± 0.021 s, and APD90: 0.017 ± 0.022 s. In terms of percentage error, for Test 1, the model shows errors of 7.33 ± 5.05% at APD30, 4.27 ± 3.31% at APD50. 2.15 ± 1.17% at APD70, and 4.90 ± 2.00% at APD90. For Test 2, the percentage errors are 7.62 ± 5.47% at APD30, 3.49 ± 4.02% at APD50, 1.97 ± 3.21% at APD70, and 3.47 ± 1.70% at APD90. For Test 3, the percentage errors are 3.15 ± 2.50% at APD30, 3.27 ± 3.45% at APD50, 3.66 ± 3.91% at APD70, and 4.66 ± 2.95% at APD90. On the training set, the percentage errors are 4.26 ± 4.26% at APD30, 3.07 ± 3.07% at APD50, 2.50 ± 2.50% at APD70. and 2.41 ± 2.41% at APD90.

[0109] FIG. 4F shows a data plot depicting mean APD errors. The mean absolute APD error from APD10 to APD100 was 0.030 ± 0.024 s for Test 1, 0.027 ± 0.023 s for Test 2, and 0.020 ± 0.019 s for Test 3, with the training set showing a lower error of 0.017 ± 0.019 s. In terms of percentage error, these values correspond to 5.63 ± 4.50% for Test 1, 5.56 ± 5.91% for Test 2, 4.33 ± 4.03% for Test 3, and 3.61 ± 4.26% for the training set.

[0110] FIG. 4G shows a data plot depicting a comparison of the MAE and eAP amplitude / noise. The MAE comparison on the test set versus eAP shows that the eAP waveform amplitude / noise ratio is the primary factor influencing prediction error. Signals with lower noise levels are expected to be reconstructed more accurately. This ratio is calculated by dividing the eAP maximum value by the noise level.

[0111] As shown in FIG. 4A, in some example embodiments, the exemplary Attention Physics informed UNET (PIA-UNET) model can include an encoder and decoder, where the encoder compresses eAP features, and the decoder reconstructs the iAP. Furthermore, the PIA-UNET model incorporates a modified version of the Aliev-Panfilov model as the foundational physics component (see the Exemplary Methods section). The modified Aliev-Panfilov model, both simple and efficient, describes cardiac electrophysiology by depicting it as a traveling excitation wave followed by a non excitable (refractory) region. The Aliev-Panfilov model component of the disclosed PIA-UNET model was considered as a “pseudo physics” function, owing to its phenomenological approach to mimic action potentials. The goal with the integration of physics-informed loss functions was to ensure predictions that are not only data-consistent but also physics-plausible. This approach helps prevent unrealistic predictions when working with new eAP recordings from different devices, maintaining robustness across varying data sources. The model captures iAP waveforms and prevents super-repolarization (e.g., a brief undershoot at the end of repolarization) in the system.

[0112] The successful reconstruction of normalized iAP waveforms by the exemplary PIA-UNET is demonstrated through representative overlays between reconstructed and actual iAPs for four distinct waveforms in the test sets (FIG. 4B). This comparative visualization highlights the model’s proficiency in accurately capturing the nuanced shapes of iAPs across a diverse spectrum of eAP shapes. Incorporating a pseudo-physics function into the model’s loss function corrected the slight undershoot observed in NEA iAPs (FIG. 4B for Test 2 data) and enhanced generalizability, particularly in Test 3 (MEA data) (FIG. 4B for Test 1 and Table SI-1). Thisapproach facilitated effective adaptation across various scenarios, as shown in the reconstruction results.

[0113] The model’s predicted values for membrane potentials across the three test sets were shown by the exemplary data to align closely with the actual values, demonstrating a high correlation (r = 0.99) (FIG. 4C). Additionally, the model exhibits an MAE of 0.032 ± 0.041 on Test 1, 0.051 ± 0.050 on Test 2, and 0.038 ± 0.045 on Test 3. while the MAE on the training set remains lower at 0.017 ± 0.027 (FIG. 4D). Furthermore, the average APD errors were less than 4.3% for APD70 and APD90, and less than 7.6% for APD30 and APD50 (FIG. 4E). The mean absolute APD error from APD10 to APD100 was low at 0.012 ± 0.016 s for the training set, and 0.030 ± 0.024 s for Test 1, 0.027 ± 0.023 s for Test 2, and 0.020 ± 0.019 s for Test 3 (FIG. 4F). Altogether, these various error metrics demonstrate the exemplary PIA-UNET model’s ability to accurately reconstruct iAP waveforms from eAP recordings.

[0114] While it has been shown that eAP recordings are good predictors of iAP, for example, a certain level of fidelity in the eAP recording is essential for accurately encoding the information needed to reconstruct the iAP waveform. For instance, if the eAP signal lacks the detailed features depicted in FIG. 3B, the precision of the exemplary PIA-UNET model in predicting the corresponding iAPs is expected to decrease due to the absence of necessary information within the eAP signal. Consistent with this, the PIA-UNET model demonstrated reduced accuracy for eAP signals with low amplitudes of the positive eAP spike (AVi) and high noise, as illustrated in FIG. 4G. As previously indicated (FIG. 3A), lower eAP spike amplitude may result from a biological phenomenon related to reduced spiking velocity of the iAP or poor cell / electrode coupling. In any scenario, an eAP signal deficient in essential information compromises the model’s ability to accurately reconstruct the iAP signal. The increased MAEs associated with these findings, along with the rigorous model training design, strengthen confidence that the model is effectively learning the relationship between eAP and iAP. See also FIGS. 9A-9F.

[0115] The exemplary PIA-UNET was shown to outperform the feature-based XGBoost in action potential analysis. While XGBoost is limited to predicting APD values and requires extensive feature engineering, the exemplary PIA-UNET reconstructs entire iAP shapes directly from raw eAP data, providing a more comprehensive understanding of cardiac electrophysiology. PIA-UNET is also configured to handle noisy or distorted data moreeffectively and demonstrates better generalization across diverse test sets, minimizing the risk of overfitting.

[0116] Application of the exemplary Al-based system for high-throughput multi-channel assessment of drug induced cardiotoxicity.

[0117] The distinctive ability of the model to accurately reconstruct iAP from non-invasive eAP recordings can hold significant potential for applications in electrophysiology. One such application is within the context of the CIPA initiative, aimed at developing improved in vitro models for more accurate evaluation of cardiotoxicity using stem-cell-derived cardiomyocytes. The disclosed Al-based non-invasive electrophysiology approach can detect subtle changes in the iAP waveforms, such as the repolarization prolongation of stem-cell-derived cardiomyocytes in the presence of a cardiotoxic drug (dofetilide).

[0118] FIGS. 5A-5G show data plots from example implementations of the disclosed AI-based system and method demonstrating high-throughput pharmacology by reconstructing iAPs from multi-channel eAP recordings and predicting the drug dose response of a cell type (e.g., cardiac myocytes).

[0119] FIG. 5A shows a data plot depicting an example of a one-channel recording for an extended duration, with 10 nM of Dofetilide added to the dish around t ~ 300 s.

[0120] FIG. 5B shows a series of data plots showing reconstructed iAPs by the exemplary PIA-UNET model from eAP recording samples at 200. 350. 500, and 850 seconds, and an overlay (right). The solid line represents the q=0.5 prediction; the shaded band between the q=0.05 and q=0.95 quantiles indicates the model’s prediction uncertainty.

[0121] FIG. 5C shows a data plot demonstrating changes in APD50. APD70. and APD90 values over time, calculated from the reconstructed iAP signals at the corresponding sample times. Central points represent APD values from the q =.5 predictor; vertical error bars indicate uncertainty spanning the range between the q=0.05 and q=0.95 quantiles.

[0122] FIG. 5D shows a diagram depicting an example NEA platform with a snapshot of simultaneous eAP recordings from 50 channels on the NEA at t ~ 445 s (7.42 mins), along with iAP recordings from the neighboring channel next to the second channel on NEA using electroporation (*).

[0123] FIG. 5E shows a series of data plots of reconstructed iAPs from simultaneous eAP recordings across 49 channels on nanoelectrode arrays, captured at t = 445 s, with actual iAPrecording (*) from the neighboring channel next to the second channel. The solid line shows the iAP predicted by QPIA-UNET with q =.5; the shaded band between the q= =0.05 and q=0.95 quantiles indicates the model’s prediction uncertainty.

[0124] FIG. 5F shows a series of data plots demonstrating changes in APD values (APD50, APD70, and APD90) calculated from reconstructed iAPs. showing variation over the recording period (~19 minutes: 0 to 19 minutes), with 10 nM of Dofetilide added to the dish around the 5 minute time pint (t « 300 s). The data plots show the APD values (s), where the x-axis is labeled at 5 minutes and 15 minutes. Central points represent APD values from the q=0.5 predictor; vertical error bars indicate uncertainty spanning the range between the q=0.05 and q=0.95 quantiles.

[0125] FIG. 5G shows box plot distributions showing maximum drug-induced APD changes at 40%, 50%, 70%, and 90% repolarization levels upon exposure to 10 nM dofetilide, expressed in milliseconds and as a percentage (n-50). A one-sided t test was conducted to assess if the changes in APD are greater than zero. *** indicates p-value < 0.001 compared to no change. P-values were calculated using a one-sided t-test, assessing whether the drug-induced APD changes were significantly greater than zero. Observed p-values: 2.02xl0-13, 1.04xl0-17, 4.09xl0-18, and 2.81X10-18. The box plots show the median (center line), interquartile range (IQR; box bounds), whiskers (1.5 x IQR), and outliers (points beyond whiskers).

[0126] As shown in FIG. 5A, the drug dosage can be added during a long-term non-invasive eAP recording, and the iAP can be reconstructed at any given time point before or after drug administration, as shown in FIG. 5B. Furthermore, the exemplary PIA-UNET model can report on changes in any desired iAP feature including APD values throughout the recording, as shown in FIG. 5C. Acknowledging the importance of addressing uncertainties in the example iAP predictions, which is critical for drug screening applications, a confidence interval was also incorporated that accompanies the example iAP predictions (FIGS. 5B-5C). The 90% confidence interval was evaluated by reconfiguring the final layer of the exemplary model to yield three outputs: V0.05. V0.5, and V0.9, which represent the 0.05, 0.5, and 0.95 quantiles (see Exemplary Methods section).

[0127] Building on the exemplary Al-based system’s ability to detect drug-induced variations in iAP waveforms at the single-channel level, the exemplary implementations further included long-term, multi-channel parallel recordings of eAPs. This was followed by thereconstruction of iAPs across multiple cells in a network of stem-cell-derived cardiomyocytes (FIG. 5E). Using this approach, the drug-induced effects as indicated by changes in APD values, can be non-invasively monitored over extended periods for several cells simultaneously (FIG. 5F). Using this high-throughput approach, population-level analyses were conducted to identify APD changes (FIG. 5G), e.g., a key parameter for cardiotoxicity assessment and drug screening. The disclosed example method enables detailed observation of variations within the cardiomyocyte population, facilitating in-depth cardiac electrophysiology studies at both individual cell and broader population levels.

[0128] FIG. 6 shows data plots depicting a comparison between patch clamp recorded intracellular electrical signals from extracellular signal recordings on example micro-electrode arrays (MEAs) and reconstructed extracellular signals created by implementation of the disclosed Al-based technique, in accordance with the disclosed technology.

[0129] The example data demonstrates accurate reconstruction of intracellular electrical signals from extracellular signals recorded on the microelectrode arrays. The plots show overlays of the exemplary Al-based system’s reconstructed signal using the example model (red), e.g., compared with ground truth signals recorded with the standard, conventional patch clamp technology (black). The mean absolute error shown in the plots of FIG. 6 is 0.07.

[0130] Summary of Example Implementations and Results

[0131] The example implementations of the disclosed Al-based method and system present a non-invasive, intelligent electrophysiology technique that utilized (1) nanoelectrode arrays, which can simultaneously record intracellular and extracellular signals from thousands of interconnected cells, and (2) the exemplary PIA-UNET, which enables fast and precise reconstruction of iAP signals. Using the exemplary nanoelectrode arrays, a unique dataset of thousands of diverse iAP and eAP pairs were obtained from monolayers of human stem-cell-derived cardiomyocytes. Through this dataset, new relationships between eAP and iAP features were uncovered, and a physics-informed deep learning model was created to accurately reconstruct iAP waveforms from eAP signals. The exemplary implementations evaluated the performance and generalizability of the example model, referred to as the PIA-UNET model, trained on NEA iAP / eAP pairs, using eAPs recorded from NEAs exposed to a drug from the training set, a different unseen drug, and eAPs from commercial MEA recordings. Additionally, the example implementations demonstrated the disclosed technique’s utility for high-throughput,long-term monitoring of proarrhythmic drug effects at both single-cell and population levels.

[0132] Exemplary Methods

[0133] Error definition

[0134] Fabrication of Nano Electrode Arrays (NEAs)

[0135] The fabrication process can include, for example, using maskless photolithography followed by deep reactive ion etching to develop vertical SiCh nanopillars, which were then coated with Pt metal to achieve conductivity. The metal was etched from the tip of the pillars using a directional dry etch to achieve the nanocrown shape.

[0136] hiPSC-CM differentiation and characterization. The non-commercial hiPSC-CMs used in the example implementations were differentiated using highly standardized protocols. Briefly, hiPSCs (line SCVI-273) were treated with 6 p. M CHIR99021 (Selleck Chemical) in RPMI supplemented with B27 without insulin for 2 days, followed by recovery in RPMI supplemented with B27 without insulin for 1 day, and then by treatment with 5 LIM IWR-1 (Selleck Chemical) for 2 days. After recovery in fresh RPMI plus B27 without insulin medium for 2 days, cells were switched to RPMI plus B27 with insulin for 2 days. The hiPSC-CMs were purified with glucose free RPMI plus B27with insulin medium for 2-4 days and maintained in RPMI plus B27with insulin medium for subsequent experiments. Using patch clamp techniques, these cells were demonstrated to be heterogeneous with ventricular-like cells being the predominant (57%) along with atrial-like and nodal-like cells. The chemically defined differentiation method provides a reproducible and scalable method for deriving cardiomyocytes from hiPSCs. hiPSC-CMs derived from iPSC lines using this protocol can be consistently recovered after cryopreservation and demonstrate comparable and functional sarcoplasmic reticulum calcium handling. The commercial hiPSC cardiomyocytes were purchased from Celogics (Celo. Cardiomyocytes, Celogics, CAT # C50) and used as an additional test set due totheir low variability and high sensitivity to electrophysiological changes from drug treatment as shown by the vendors). Previous characterization of these cells by patch-clamp and fluorescent microscopy demonstrated their human cardiomyocyte-like electrophysiology, as well as ventricular cardiomyocyte markers and structural characteristics.

[0137] hiPSC-CMs culture on NEA and MEA devices. Prior to plating cells on NEA devices, the device was coated with 1 mg / mL poly-L- Lysine at room temperature for 15 min, then treated with 0.5% Glutaraldehyde in PBS at room temperature for 10 min, followed by 1:200 Matrigel in DMEM / F12 at 37 °C for 3 hours before seeding cells. The cultured hiPSC-CM were disassociated from the plate with TryPLE select lOx at 37°C for 5 min after 25-60 days of differentiation. Cells were resuspended in culture medium supplemented with 10% KnockOut Serum Replacement (KSR), then seeded at ~1.2xl05cells / device. Measurements were taken for over a month from 5 days post cell attachment. For MEA recordings, commercial hiPSC-ventricular cardiomyocytes (Celo. Cardiomyocytes, Celogics, CAT # C50) were used. The cells were thawed and prepared according to the manufacturer’s protocol. These cardiomyocytes were derived from a proprietary human iPSC line (fibroblast, Caucasian male donor) and exhibited spontaneous beating starting on day 2 post-thaw, with stabilization at 45-60 beats per minute (bpm) by day 7. Celo. Cardiomyocytes formed synchronous monolayers and expressed ventricular cardiomyocyte- specific markers, such as connexin 43, indicating high interconnectivity. These cells also demonstrated physiologically relevant electrophysiological properties, as evidenced by their response to various ion channel modulators.

[0138] NEA and MEA recordings. For both NEA and MEA recordings, measurements were taken at 32°C in RPMI with B27 and 10 mM HEPES. This temperature was chosen because it keeps the membrane pores, which form during electroporation, open for longer iAP recordings compared to 37°C. HEPES was used to stabilize the pH at room atmosphere. Recordings were taken using a 60-channel voltage amplifier (MEA1060-Inv-BC, Multi-Channel Systems, Reutlingen, Germany), e.g. with a sampling rate of 5 kHz for NEA recordings and 10 kHz for MEA recordings. For NEA recordings, electroporation was achieved by delivering a ±4V biphasic square wave with a 200 ps duration at each phase at each electrode. For MEA recordings, commercial MEA devices (60MEA100 / 10iR-Ti) from Multichannel Systems, powered by Harvard Bioscience. Inc., were sterilized and prepared for cell culture. The MEAs first underwent UVO treatment for 10 minutes using a UVO-CLEANER, Model 42 (JelightCompany, Inc.). Following this, the MEAs were washed three times with phosphate-buffered saline (PBS) and then washed twice with 70% ethanol inside a biosafety cabinet (BSC). The devices were air-dried for 30 minutes to 1 hour and subsequently sterilized under ultraviolet (UV) light in the BSC for 1 hour. After sterilization, the MEAs were coated with fibronectin (Millipore Sigma, F1141-1MG) at a concentration of 50 pg / mL (1:20 dilution). A droplet of 8 pL of the diluted fibronectin was applied to the center of the MEA, covering the electrodes, and incubated for at least 1 hour at 37°C. Cells were seeded onto the fibronectin-coated MEAs at a density of 50,000 cells in 8 pL. The devices were then incubated at 37°C for 60 to 90 minutes, followed by the slow addition of 400 pL of plating medium (Celogics, CAT # C50-PM, supplemented with C50-PS) while the MEAs were positioned at a 30° angle. The MEAs were then returned to a flat position and incubated for 24 hours before replacing 100% of the plating medium with advanced medium (Celogics, CAT # CM200, supplemented with C50-MS). Media changes of 50% were performed every 48 hours thereafter. By day 5, the human iPSC-derived ventricular cardiomyocytes on the MEAs were ready for electrophysiology experiments.

[0139] Drug experiment. Nifedipine (Sigma), Quinidine (Sigma, Q3625), Propranolol (Sigma, P0884), Lidocaine (Sigma, L7757) and Flecainide (Sigma, F6777) were dissolved in DMSO to make 100 mM stock solution; Dofetilide (Sigma) was dissolved in DMSO to make 10 mM stock solution. For each dose, the stock drug was diluted in measurement medium into 2x of the targeted dose and wanned up to 32 °C. Upon drug administration, 500 pL of the 2x dose drug was added to the 500 pL existing measurement medium to help homogeneous diffusion. The cells were electroporated 1 min after the start of measurements; the drug was administered 400 s after electroporation: and the recording lasted for ~30 min for each repeat. The drug’s effect on action potential and the corresponding iAP waveforms are described in section III (“Drug Test Experiments”) of this disclosure.

[0140] Obtaining a diverse dataset of eAP and iAP pairs. Two approaches were tried for data collection: incremental drug addition and single high-dose drug application. In the first approach, the drugs were added in multiple steps to capture a spectrum of iAP shapes and their corresponding eAPs. However, a relatively high concentration of the drugs was also applied in a single step during each recording session (FIG. 2G), which enabled achievement of a wider range of iAP durations (shown in FIG. 8A). The single high-dose approach minimized the noise typically associated with incremental drug addition. Given that both amplitude and S / N ratiotend to decrease over time — resulting in significantly lower S / N ratios and amplitudes in the final traces compared to the initial ones (FIGS. 8B-8E, with paired t-test p-values of 8.6xl0-15and I. lx 10x. respectively) — this strategy allowed for collection of high S / N ratio data at higher drug dosages. Moreover, the APD values changed gradually until stabilization (FIG. 8H), providing a spectrum of iAPs with APD values ranging from normal to fully impacted by the drug. Consequently, this approach, similar to multi-step drug addition, resulted in a diverse range of eAP and iAP pairs, and as shown in FIGS. 8H and 81, it facilitated the acquisition of longer APDs in a shorter time with improved S / N ratios while allowing the APD values to change incrementally.

[0141] Bandpass filtering. The data underwent bandpass filtering to selectively isolate the frequencies of interest, utilizing an acausal third-order Butterworth filter renowned for its flat frequency response within the passband. With a sampling frequency set at 5000 Hz, the filter coefficients were designed for a low-cut frequency at 0.1 Hz and a high-cut frequency at 2499 Hz. Furthermore, to extract the signal’s noise, the signal was subjected to a filter between 2499 Hz and 4000 Hz and the standard deviation of this filtered signal was taken. However, for both training and validation purposes, raw eAP and iAP signals are used directly.

[0142] Filtering proper recordings. Identifying which neighboring eAP and iAP recordings to include was the only manual part of data analysis. Recordings of neighboring cells were plotted to identify quality pairs of extracellular and intracellular data. These pairs were chosen qualitatively based on a high signal to noise ratio, a stable baseline voltage, and having an intracellular amplitude of above ImV for a long duration.

[0143] Data filtering and data segmentation. The initial step in data processing involved identifying high-quality pairs of neighboring cells for inclusion in the dataset. To determine the action potential peaks in intracellular recordings, an exemplary peak-finding algorithm was developed (e.g.. from the scipy.signal library). The example peak-finding algorithm was configured to identify peaks with a minimum height of twelve times the noise level and a separation distance of at least 60% of the average action potential period. Once the peaks were identified, the data were segmented around these points. For example, 8000 points (equivalent to 1.6 seconds), starting 1000 points (or 0.2 seconds) before the identified peak. Following segmentation, stringent selection criteria were applied for inclusion in the analysis. For example, only segments exhibiting a high signal to noise for intracellular channels (>90 dB) wereconsidered. Additionally, segments were required to have an action potential amplitude of at least 0.5 mV in the intracellular channel. This process ensured the quality and relevance of the data for subsequent analyses.

[0144] An example embodiment of a data processing algorithm includes an (optional) initial process to perform QA filtering and noise estimation. The (optional) initial process includes applying an example acausal 3rd-order Butterworth bandpass at 0.1-2499 Hz (e.g., 5 kHz sampling) to obtain a QA view, and estimating noise by filtering 2499-4000 Hz and computing c (used below). It is noted that training / validation can use the raw signals; and the (optional) filtering here supports S / N and artifact checks. The data processing algorithm can include a process to perform peak detection (e.g., on iAPs) by estimating the mean cycle period from the local trace; and detecting peaks with (i) minimum height > 12 x noise level, and / or (ii) minimum distance > 0.60 x mean cycle period. For example, this can yield robust, non-overlapping beats even in mildly arrhythmic traces. The data processing algorithm can include a segmentation process, which can include, for each detected peak, extracting a window of samples (e.g., 8,000 samples (-1.6 s), beginning 1,000 samples (0.2 s) before the peak); and extracting the synchronous eAP window from the neighboring electrode. The data processing algorithm can include a quality gating process, e.g., including checking whether eAP / iAP pairs spike with same frequency, computing iAP S / N in dB (e.g., using RMSsignai / RMSnoise, e.g., require > 90 dB; and / or require iAP amplitude > 0.5 mV). Segments that fail any criterion are excluded from training / analysis. The data processing algorithm can include an (optional) normalization process, according the following. For an (optional) feature extraction analysis, the process can include identifying bpl, Vmax, Vmin, bp2, bp3 and computing AVI, AV2, AVd, ATI, AT2, ATd, DR, IR. The data processing algorithm can include assembling the data set, e.g., by collecting the synchronized, normalized eAP / iAP windows that passed the quality gating process; such that, training uses these; and where independent test sets are composed (an example being: Test 1 NEA same drug different device; Test 2 NEA unseen drug; Test 3 MEA + simultaneous patch clamp). In some embodiments, the data processing algorithm can include an APD-based validation process (e.g., post-reconstruction of the iAP data). For example, the APD-based validation process can include smoothing the reconstructed iAPs with moving average (e.g., window = 20 samples); using peak_widths to compute APD10, 20,.... 100 across the beat Step H — APD-based validation (post-reconstruction); and reporting absolute andpercentage APD errors versus ground truth for training and reporting test sets.

[0145] Characterizing the eAP and iAP waveforms. To characterize eAP waveforms, five critical points where the waveform gradient changes significantly were identified: just before the spike (bpi), maximum point (Ymax), minimum point (Ynun), right after the minimum point (bp ), and where it starts to rise again (bps). The vertical span from bpi to Ymax is labeled AVi, and from Ymin to bp2 is AV2. Similarly, AVd is the vertical distance from bp2 to bps. To describe the horizontal dimensions within the spike, ATi and AT2 describe the span of the positive and negative spike phases, respectively, and ATd denotes the distance between bp2 and bps. Decay Rate (DR, v / s) and Increase Rate (IR, v / s), represent the slope from bps to bps and just after bps, respectively. The action potential duration (APD) metrics, e.g., specifically APD10 to APD100, measure the timeframe for an action potential to decrease to 10% and recover to 100% of its peak value, providing insight into the waveform's recovery phase independent of amplitude variations. Further details are provided in section IV (“eAP Features Determination Methods'') of this disclosure.

[0146] Signal to Noise ratio (S / N) calculation. To calculate the S / N for iAPs, filtering techniques were initially employed on the raw recording. A low-pass filter is used to isolate the signal component, while a high-pass filter helps in extracting the noise component. After filtering, these filtered components were segmented into windows of length d. This results in arrays for the signal, V = [v1,v2...,vd], and for the noise. N = [n1,n2...,nd]. Following the segmentation, the Root Mean Square of the Signal (RMSsignal) and the Root Mean Square of the Noise (RMSnoise) were calculated. RMSsignalis computed as:

[0147] Similarly, the RMSnoiseis calculated as:

[0148] Finally, the Signal to Noise Ratio in decibels (S / N [dB]) is determined using the formula:

[0149] For eAP signals, the S / N was calculated as the ratio between the maximumamplitude in signal array over the standard deviation of the noise array.

[0150] eAP and iAP normalization. Normalization is important to bring all data to a comparable scale and to emphasize relative changes in signal features over absolute values. The windows of iAPs are normalized to a range between 0 and 1 starting from 0.1. For eAPs, a distinct approach is used for normalization. A specific segment of the eAP signal, e.g., particularly the values between indices 1150 and 1350, is selected. This example segment is important as it characterizes the noise within the eAP signal, based on the assumption that it accurately represents the noise characteristics of the entire signal. The peak index in each window is typically around index 1000, just before this segment. To quantify the noise, the function calculates the standard deviation (o) of this selected segment, which serves as a measure of variation or dispersion within the values. Next, the eAP signal undergoes normalization by subtracting the mean of the entire eAP signal from each value. To ensure the normalization starts from 0, the initial value of the eAP signal is subtracted from all subsequent values. The final step scales the signal relative to its noise level, achieved by dividing the mean-adjusted values by 60 times the calculated standard deviation. The normalization methods applied to eAP and iAP arrays given that eAP or iAP are arrays of values (xl, x2,..., x8000), are as follows:

[0151] PIA-UNET architecture. UNET is an autoencoder architecture often used in biomedical applications for image segmentation and data reconstruction. Here, the UNET was modified to engineer a new structure to produce a new architecture for an attention UNET. For example, the exemplary PIA-UNET machine learning architecture has an integrated U-Net architecture with attention gate mechanisms and physics and biophysics parameters for physics-informed learning, e.g., where the physics-informed learning incorporates a modified Aliev-Panfilov model that can account for single cell electrophysiology.

[0152] Input. A vector with dimensions (batch size x 8000 xl)

[0153] Convolution-BatchNormalization-ReLU (CBR) Block'. The CBR block forms the basic building block of the exemplary model architecture. It incorporates a ID convolution layerfollowed by a batch normalization layer and a ReLU activation function. The “He” normal initialization method is employed for the convolution layers.

[0154] Squeeze-and-Excitation (SE) Block'. Within residual blocks (Resblocks), SE blocks are to recalibrate channel-wise feature responses adaptively. The SE mechanism acts to recalibrate the preliminary features by adaptively reweighting the channel-wise feature responses. It accomplishes this by performing global average pooling, dimensionality reduction (by a factor of 8), and subsequent scaling using a two-layer fully connected network with ReLU and sigmoid activations, respectively on Resblock input. These scaling factors are then used to reweight the original Resblock output.

[0155] Residual Block (Resblock)'. Nested within the architecture, each resblock serves as a mechanism for efficient feature extraction and transformation. Each Resblock starts with two consecutive Convolution-BatchNormalization-ReLU (CBR) blocks, first applying a ReLU activation, then a Sigmoid activation. An SE block then refines this output by providing adaptive weights, helping the model focus on important features while ignoring the less relevant ones. The function concludes its operations by integrating the original input to this recalibrated output, forming a residual connection. This approach provides a balanced and refined feature representation and allows for the construction of deeper architectures without information loss.

[0156] Attention Mechanism'. The attention mechanism is to enhance the feature representations in the decoder by considering features at the corresponding level in the encoder. First the input from lower decoder level is upsampled by a factor of 5, then given the upsampled input from a lower decoder level and a shortcut connection from the encoder, both are transformed via ID convolutions and summed. A ReLU activation is applied to this sum, which is then processed through another ID convolution with a sigmoid activation. This forms the attention mask, which is multiplied with the encoder feature map.

[0157] Encoder. The encoder initiates with a CBR block and progresses through a sequence of eight Resblocks. With each stage, feature maps are condensed by a factor of five using strided convolutions, enhancing the receptive field while aggregating spatial information. Parallelly, the depth of these feature maps escalates, beginning at 32 channels beginning at 32 and adding 32 more at each step, reaching 32*4 by the last step. This simultaneous contraction of spatial resolution and channel expansion ensures a rich representation of features. Moreover, attention mechanisms at every encoder level capture crucial spatial cues, preparing for the subsequentdecoding phase.

[0158] Decoder. In the decoder phase, spatial resolution is progressively restored across five decoder levels. At each level feature maps are upsampled by a factor of five through the attention mechanism. These upsampled feature maps are combined with attention-refined feature maps from their corresponding encoder counterparts. Following this fusion, the concatenated feature maps pass through a CBR block, which mirrors the structure of the encoder. Simultaneously, the number of channels decreases progressively, transitioning from 32*4 down to 32*3, then 32*2, and finally 32*1 channel.

[0159] Quantile PIA-UNET with confidence interval (Q-PIA-UNET). Q-PIA-UNET is a modified version of the disclosed PIA-UNET architecture designed to estimate the 90% confidence interval for iAP values using a quantile loss function. The quantile loss function asymmetrically penalizes overestimation and underestimation, enabling the model to accurately predict specific quantiles (e.g., 0.05, 0.5, and 0.95) by minimizing errors for each. To address the issue of quantile crossing, the modified architecture simultaneously predicts the 0.05, 0.5, and 0.95 quantiles through three parallel output layers (e.g., V0.05, Vo.5, and V0.95), along with corresponding physical parameter predictors. Quantile crossing occurs when the predicted lower, median, and upper quantiles are out of order, which makes the predictions unreliable and inconsistent. By using separate output layers corresponding to each quantile in one model, the model ensures that the predicted values follow the correct order, improving both the accuracy and reliability of the results. The shared hidden layers in the Q-PIA-UNET architecture learn common underlying features from the input data, providing a consistent foundation for all quantile predictions. The separate output layers then specialize in predicting their respective quantiles, fine-tuning the shared representations to capture the unique aspects of each quantile level. Training these quantile predictions simultaneously within the same model allows for joint optimization, where the errors of all quantiles are considered together. This joint learning process implicitly enforces the natural ordering of quantiles, as the model adjusts its weights to minimize discrepancies between quantile levels. As a result, the predictions for the lower quantile (0.05) remain less than or equal to the median (0.5), which in turn remains less than or equal to the upper quantile (0.95), thus preventing quantile crossing. This allows for simultaneous prediction of iAP values and their associated physical parameters across all three quantiles. The incorporation of quantile loss into the proposed PIA-UNET is detailed in the nextsection. This approach aims to provide a more reliable prediction of iAP under various conditions.

[0160] Physics informed layer (a, x and k estimation)'. In the exemplary model, the parameters k, x, and a are estimated from the bottleneck layer or the layer just before the final layer. This process begins with the convolution of features, where the data is passed through a convolutional layer to extract relevant information. The resulting features are then flattened and processed through a dense layer, which effectively maps the high-level features to the desired parameters. This design allows the model to adapt and learn different sets of parameters for each quantile, accommodating the unique characteristics of each quantile distribution. By following this approach, the model is better equipped to capture the intricate relationships between the input data and the physics-informed parameters.

[0161] Incorporating the modified Aliev-Panfilov model into the PIA-UNET hybrid loss function. The modified Aliev-Panfilov model for a single cell is as follows:

[0162] In these equations, v and w represent the normalized iAP and recovery variables, respectively. To ensure the differentiability is loss function, the step function s(, a,x) is approximated using a sigmoid activation function, which smoothly transitions between values and preserves the exemplary model’s ability to backpropagate gradients effectively.s(v, a, x) ~ X. < J (n a — vf) + (1 — x). <j (n ( — a)) (14) in which n is a constant determining the sharpness of the stem function. In the example approach, n is chosen to be set to 1000. The parameter a is the excitation threshold, while k controls the magnitude of the transmembrane current. Both space units [s.u.] and time units [t.u.] are dimensionless. The parameter x controls the balance between excitation and recovery dynamics by modulating the smoothness of transitions in the recovery mechanism based on the membrane potential v and the threshold a. This flexibility allows the action potential model to adapt to a wider range of iAP shapes. This is further discussed in sections I and II of this disclosure. The loss function typically used in fully connected neural networks aims to reconstruct simulated iAPs from time and coordinates as inputs, leveraging calculable derivatives during propagation to ensure adherence to governing physical equations. However, in this case, the input is eAP, and the pseudo-physics loss function is employed to maintain the iAP shape.Derivatives are computed numerically due to the temporal nature of the data arrays. Furthermore, the term w is initially unknown but can be simplified with certain assumptions in the Aliev-Panfilov equations.

[0163] Incorporating the value of w from Equation (12) into Equation (13) and deriving it with respect to time (t), a single equation is obtained:

[0164] The detailed derivation of Equation (15) is provided in the section X of thisdisclosure. A small noise (eps = 10-3) was added the indices where v or = 0. Note thatthe exemplary model processes temporal arrays of eAP potentials of length 8000. outputting dv d2v temporal arrays of the same length for iAP potentials. The derivatives — and — for arrayV=[v1, v2...,vd] are calculated using discrete numerical methods:dv _ Vt+l- t-l / | Z, dt 2t ' d2v _ vt+l+ Vt-i-2Vt z,— dt2t2'

[0165] The parameters v, a and k and x are estimated by compressing the bottleneck or the final layer of the example PIA-UNET model, just before the iAP layer, into a single channel using a convolution operation. This is followed by flattening the output and applying a two-layer dense neural network to generate the parameter estimates. The hybrid loss function incorporates terms that account for both alignment with experimental measurements (LD) and adherence to physical laws Lp, as described by the pseudo-physics expression in Equation (15). This loss is computed for N predicted iAP arrays, each with d dimensions'^!?!, v2..., vd]), as follows: as follows:

[0166] Here, FAPrepresents the function describing the pseudo-physics relationship, with a and k as parameters. The overall hybrid loss function, L, combines these two components, integrating a logarithmic transformation on the physical law adherence termL = a LD+ / 3log(LP') (20)

[0167] The logarithm function is applied to the physics loss functiondue to its monotonic nature. This transformation serves to dampen significant deviations in Lp. thereby harmonizing the scales of the physics-based and data-based components of the loss function. This approach ensures a more balanced optimization, taking into account both experimental data and physical law adherence. The respective weights for the data-based and physics-based components of the loss function, denoted as a andare determined through a grid search hyperparameter tuning technique detailed in the Hyperparameter Tuning and Model Training section.

[0168] Q-PIA-UNET loss function. The quantile loss function for N samples and a given quantile q (0.05, 0.5, or 0.95) is defined as:

[0169] For predictions below the actual value, the error is weighted by q, and for predictions above, by 1-q. This asymmetry is beneficial for quantile regression. Additionally, the physics term and the total loss functions are:

[0171] Grid search, in conjunction with k-fold cross-validation, was employed for multi-step hyperparameter tuning. Distinctively, in each fold, one pair of neighboring channels was set aside. Five distinct pairs of neighboring channels — sourced from two different recording sets — were utilized for the training and hyperparameter tuning process. The tuning was conducted in a sequential three-step approach: initially, the kernel size. Resblock depth, and the number of channels in Convolution-Batch Normalization-Re LU (CBR) were optimized. Subsequently, the learning rate, number of epochs and batch size were optimized. The final step involved tuning of the weights assigned to the loss functions. The example model, trained with these optimal hyperparameters, underwent a two-fold cross-validation. In each fold, the data from one recording set were considered as the validation set. The optimal hyperparameters for the model are set as follows: Initial learning rate at 0.001. number of epochs at 100, number of channels in CBR at 32, kernel size at 11, Resblock depth at 8, batch size at 32, and the weighting factors lossfunctions, a and 0 at 10 and 0.01, respectively.

[0172] Action Potential Durations (APDs) calculation. To accurately determine the APDs, a smoothing technique was first applied to the segmented intracellular traces using a moving average filter (window size = 20) to reduce noise and enhance the detectability of the action potential features. Next, the standard deviation of the smoothed data was utilized to identify the initial upward spike of the action potential. For the quantification of the APDs, the 'peak_widths' method was employed from the SciPy Python package. This method is particularly effective in measuring the widths of action potentials at various levels of repolarization. Ten distinct APD measurements were obtained to comprehensively describe the shape of each action potential. For instance, APD 10 and APD20 represent the widths of the intracellular action potential at 10% and 20% of repolarization, respectively.

[0173] Additional Exemplary Results of Example Implementations

[0174] FIG. 7 shows data plots showing the violin plot distribution of eAP and iAP features for the data utilized in the machine learning and deep learning models, as depicted in FIG. 4B, shown in green. The blue and red plots represent the distribution of the training set and test set for eAP and iAP, respectively. The blue and red plots represent the distribution of the training set and test set for eAP and iAP, respectively (ntraining-vai=I209, UTesti=272, nTest2=171 and, UTest3=91).

[0175] FIG. 8A-8H show data plots depicting the iAPs in response to drugs in the exemplary implementations. FIG. 8A shows a data plot depicting a comparison of iAP APD values upon drug addition, based on experiments comparing neighboring NEA channels’ iAPs, from 88 unique iAP channels across two sets of recordings. The dashed line indicates the time of drug addition. FIG. 8B shows a data plot depicting the variation in iAP signal-to-noise ratio (S / N) over time from the same recordings as above. FIG. 8C shows a data plot comparing the last iAP in the recording with the first one in terms of S / N, demonstrating a significant drop in value (P value = [***p-value: 8.63 x 10-15]) as the experiment progresses, due to the resealing of cell membrane pores. FIG. 8D shows a data plot presenting the variation in iAP window's maximum amplitudes (spike amplitude) over time from the same recordings as above. FIG. 8E shows a data plot comparing the last iAP in the recording with the first one in terms of amplitude, showing a significant drop in value (***p- value = [1.1 x 10-08]) as the experiment progresses, attributed to the resealing of cell membrane pores. FIG. 8F shows a data plot presenting thevariation in eAP maximum amplitude (spike amplitude) over time for 49 distinct eAP channels. FIG. 8G shows a data plot comparing the last eAP in the recordings with the first one in terms of amplitude, indicating no significant change (p- value = [0.262]). FIG. 8H shows a data plot depicting iAPs collected from distinct sets of recordings by the addition of dofetilide, used for the machine learning and deep learning aspects of this study. The test set here corresponds to Testi. FIG. 81 shows a data plot comparing max induced APD100 between single-step vs. multi-step drug addition.

[0176] FIG. 9A-9F show a schematic of an example embodiment of a Quantile-Physics-Informed Attention Unet (QPIA-UNET), in accordance with the present technology, and data plots depicting example results from implementations of the exemplary QPIA-UNET, demonstrating the capability of reconstructing entire iAP waveforms from eAP signals with confidence interval. FIG. 9A shows a schematic diagram of an example embodiment of a QPIA-UNET Model, in accordance with the present technology. This provides a visual representation of the QPIA-UNET model used in the example implementations. As shown in the diagram, the exemplary model is able to simultaneously predict three potentials (VI, V2, V3) represented in green (VI, q = 0.05), red (V2, q = 0.5), and blue (V3, q = 0.095), along with their corresponding physics parameters (kl, xl and al in green, k2, x2 and a2 in red, and k3, x3 and a3 in blue) from the bottleneck. FIG. 9B shows data plots that illustrates how the model’s predicted potential values closely align with actual values, in three different tests sets, achieving a correlation of r = 1.00, 1.00 and 0.98, respectively on testl, test2 and test3. FIG. 9C shows a data plot depicting reconstructed iAPs from the median predictor (q = 0.5) are compared with actual ones in terms of APD. FIG. 9D shows a data plot depicting examples of predicted iAP with their CI (confidence interval) vs the actual ones. FIG. 9E shows a data plot depicting how the model exhibits a MAE of 0.04 ± 0.01 on Test 1, 0.04 ± 0.01 on Test 2, and 0.05 ± 0.01 on Test 3. The MAE on the training set is 0.02 ± 0.01, indicating lower error during training. FIG. 9F shows a data plot depicting the mean absolute APD error from APD 10 to APD 100 was 0.05 + 0.04 s for Test 1, 0.02 ± 0.02 s for Test 2, and 0.03 ± 0.03 s for Test 3, with the training set showing a lower error of 0.02 + 0.02 s. In terms of percentage error, these values correspond to 13.07+13.45% for Test 1, 4.89 + 5.29% for Test 2, 6.95 + 4.36% for Test 3, and 4.49 + 4.99% for the training set.

[0177] I. Assessing the Accuracy of NEA Recordings When Compared to Patch Clamp

[0178] Ensuring the accuracy of the training data is vital for the effectiveness of any deep learning model. To determine the accuracy of iAP waveforms obtained from NEAs, simultaneous NEA and patch clamp recordings were performed from the same cell and compared various features of their waveforms. Nanocrown- shaped NEAs (referred to simply as NEAs herein) were fabricated. iPSC-CM cells were then seeded onto the NEAs employing a differentiation and culture protocol.

[0179] FIGS. 10A-10F show diagrams and data plots for example implementations of simultaneous iAP recordings from cells using patch clamp and nanoelectrode arrays. FIG. 10A shows a diagram of an example platform for simultaneous iAP recording from cells (e.g., iPSC-CMs) using patch clamp (PC) and nanoelectrode array (NEA) via electroporation. FIG. 10B shows a data plot depicting the comparison of iAP recordings from PC and NEA. The scaling was between 0 to 1 and segmenting into arrays was of length 8000 indices or 1.6s. The data plot includes an exploded / expanded data plot for a few second portion that shows important features such as cycle time, APD50 (action potential duration at 50% repolarization), and APD90. FIG. 10C shows a data plot comparison between windows of iAP from PC and NEA by cycle time, APD50, and APD90. FIG. 10D shows a data plot illustrating mean absolute error (MAE), correlation coefficient (r), and NEA signal to noise ratio (S / N) changes over time during one set of the experiment. FIG. 10E shows a data plot depicting a comparison between MAE, APD50% error, and APD90% error vs NEA S / N. This exemplary data highlights the changes in three critical errors describing the similarity of NEA-recorded and PC-recorded normalized iAPs (MAE, APD50 percentage error, and APD90 percentage error), with thresholds set at 0.05. 10%, and 10% respectively to ensure reasonable similarity between iAP pairs. Also included is the comparison between iAP pairs with the highest MAE with NEA S / N > S / N threshold (S / N*), shown in blue, as well as examples of iAP pairs with error exceeding the threshold from the region with NEA S / N < 90. FIG. 10F shows a box plot distribution of cycle time difference (dCT(s)), r, MAE, APD50%, and APD90% errors between PC and NEA normalized iAP pairs with NEA S / N > S / N* (= 90).

[0180] For example, 15 sets of simultaneous iAP recordings from single cells were performed, utilizing both patch clamp and NEA methods, with a comparative analysis of their waveforms shown in FIG. 10 A. Each recording set, as depicted in FIG. 10B, demonstrated thatalthough NEA iAP traces exhibit a lower amplitude, they closely align with the patch-clamp recordings when normalized. This normalization process allowed for a more accurate comparison between the two techniques. The total duration of these simultaneous recordings was approximately 46 minutes. To quantify the similarity between waveforms, high-pass and low-pass filtering was initially applied, as detailed in the Exemplary Methods section, to decompose the waveform into isolated noise and refined action potential waveform. To account for various iAP shapes collected during recording, the signals were then segregated using a window length of 8000 timepoints or 1.6 seconds.

[0181] This exemplary process yielded a total of 3363 pairs of iAP recordings from both NEA and patch clamp methods. The S / N for both NEA and patch recordings was calculated by utilizing the filtered signal and noise vectors, assigning an S / N value to each window of the respective recording methods. Each window, containing pairs of action potential signals, was then normalized to a scale ranging from 0 to 1 to facilitate a more precise comparison. Furthermore, the APD values and their differences within each window were quantified, along with the cycle time for each recording (FIG. 10C). The discrepancy in cycle times between the NEA and patch clamp methods was also calculated. Additionally, the Mean Absolute Error (MAE) was determined and the correlation (r) within each window was analyzed (FIG. 10D). This comprehensive analysis, which included comparisons of NEA and patch clamp recordings as depicted in the FIG. 10 A, provided an in-depth evaluation of the compatibility between these two recording techniques.

[0182] The measured low MAE of 0.046 ±0.028 and a high average correlation (r) of 0.989 + 0.012 pointed to near-perfect agreement between the NEA iAP and patch-clamp recordings. Further, key parameters linked to drug-induced heart rhythm abnormalities were investigated by assessing iAP measurements from NEAs against patch clamping, focusing on APD50, and APD90 (cell repolarization markers) and cycle time. The average errors for APD50, and APD90 between NEA iAPs and patch clamps during the experiment were 0.032+0.034(s), and 0.01 l±0.016(s), respectively. Expressed as percentage errors, these values equate to 11.516+10.010%, and 3.875+3.425%, respectively. For cycle time, the mean difference between two consecutive spikes was 0.011 + 0.024 (s). FIG. IDE illustrates how the MAE, APD50, and APD90 errors vary across a range of values depending on the S / N in NEA recording traces.

[0183] The figure highlights that as the NEA S / N ratio increases, the maximum values ofthese errors decrease. This trend suggests that low S / N ratios may cause distortions in iAP shape, potentially due to an imperfect cell-to-NEA seal. These distortions seem to follow a probability distribution dependent on the S / N value, with deviations diminishing at higher S / N ratios. Based on these findings, to ensure waveform accuracy, maximum acceptable thresholds were set at 0.05 for MAE, and 10% for both APD50 and APD90 percentage errors. The exemplary analysis revealed that applying a stringent threshold of 90 dB (S / N* =90 dB) and filtering signals with S / N ratios above S / N* ensured that the APD50 percentage error and MAE remained below 10% and 0.1, respectively, as demonstrated in FIG. 10E. The data plot also presents examples of eAP and iAP pairs that illustrate the maximum errors observed at or above the S / N*.

[0184] Upon comparing NEA iAPs that satisfied the S / N* with corresponding patch clamp iAP recordings, a significant reduction in errors was observed. MAE decreased to 0.024+0.006, and the average correlation coefficient (r) increased to 0.996 ± 0.003. The average errors for APD50 and APD90 between NEA iAPs and patch clamps improved to 0.017+0.020 seconds and 0.004+0.007 seconds, respectively. When expressed as percentage errors, these values correspond to 4.815±2.342% for APD50 and 1.343±0.993% for APD90. Additionally, the mean cycle time difference between two consecutive spikes was recorded at 0.007 + 0.006 seconds. The distributions of aforementioned errors are shown in FIG. 10F. Furthermore, the experimentlevel comparison of normalized iAPs from NEA recording and patch clamp is provided in FIG. 11. These findings align with a comprehensive comparison of NEA and patch clamp iAPs, particularly when various drugs were introduced during recordings. The example findings revealed a near-perfect match between iAPs from the two recording methods when the S / N* was exceeded. This S / N* threshold was used for processing of data used to train the exemplary model in subsequent steps.

[0185] FIG. 11 shows data plots depicting an experiment-level comparison of normalized iAP from patch clamp (PC) and nanoelectrode array (NEA) recording techniques. Fifteen unique sets of successfully simultaneous iAP recordings using NEA and PC are presented, showcasing various errors (r, APD50, APD90, MAE, cycle time difference, APD 50%, and APD 90%). The left panel displays comparisons for all recordings, while the right panel focuses on recordings with NEA iAP signal-to-noise ratio (S / N) greater than 90 and comprising more than 10 samples.

[0186] II. Two Neighboring Channels on NFA Show Similar iAPs

[0187] To obtain synchronized eAP and iAP recordings from NEAs, it was assumed that neighboring cells in a confluent monolayer of iPSC-CMs that are in close physical proximity on the NEAs exhibit similar AP waveform. This assumption was tested by thoroughly assessing the similarity between neighboring channels over an extended period of time, and under various conditions (FIGS. 12A-12D). To this end, dofetilide (dissolved in DMSO) - a compound known to prolong iAP duration - was incrementally introduced into the cell culture while recording iAP signals from many channels. The drag’s concentration was progressively increased from 0.3 to 1, then to 3, and finally to 10 nM through a multi step addition process. This process was done in four stages, occurring approximately at 400, 800, 1200, and 1700 seconds during the recording session, as illustrated in FIG. 12A.

[0188] FIGS. 12A-12D show diagrams and data plots for example implementations of simultaneous iAP recordings from cells using patch clamp and nanoelectrode arrays with introduction of an example drag (dofetilide). FIG. 12A shows a diagram illustrating simultaneous iAP recording from neighboring channels via non-invasive extracellular action potential (NEA) electroporation. This section details the multi-step addition process of dofetilide, administered in concentrations of 0.3, 1, 3, and finally 10 nM. The process was conducted in four stages at approximately 400, 800, 1200, and 1700 seconds during the recording session. This method was employed to collect a diverse range of iAP shapes, facilitating a more comprehensive analysis. FIG. 12B shows a data plot showing a comparison process similar to PC vs NEA iAP recording, where the scaling is between 0 to 1 and segmenting is into arrays of length 8000 indices or 1.6s. FIG. 12C shows a data plot depicting a box plot distribution of dCT(s), correlation coefficient (r), MEA, and APD50% and APD90% errors between neighboring and NEA normalized iAP pairs with NEA S / N > S / N* (= 90). FIG. 12D shows a data plot depicting examples of iAP pairs from neighboring channels with the highest MEA, APD50, and APD90% errors, as indicated in the box plots.

[0189] For example, dofetilide addition enables studying the similarity between two neighboring channels for a wider range of iAP durations. In total, 2661 pairs for iAP spikes from 22 pairs of simultaneous recordings from neighboring channels were collected with S / N > 90 (dB). The same similarity analysis was applied for NEA versus patch clamp iAPs. Paired NEA neighboring channels iAPs were scaled to a range of 0 to 1 (FIG. 12B), then the MAE wascalculated, and correlation between them along with comparing their APD50, APD90 and cycle time was conducted (FIG. 12B). The average difference (MAE) across two iAP traces was 0.011 ± 0.006. Moreover, the correlation coefficients (r) between the iAPs two neighboring channels were quantified as 0.999 + 0.001 respectively. The average errors for APD50, and APD90 between two neighboring NEA channels iAPs were 0.015±0.009(s), and 0.006±0.005(s), respectively. Expressed as percentage errors, these values equate to 3.943+2.233%, and 3.943+2.233 %, respectively. For cycle time, the mean difference between two consecutive spikes was 0.002±0.002 (s).

[0190] FIG. 12D exhibits pairs of iAPs from neighboring channels, highlighting the maximum MAE, APD 50 percentage error, and APD 90 percentage error observed. This evidence underscores the consistency in iAP traces from neighboring NEA channels when the S / N ratio is greater than the S / N*. This noteworthy outcome underscores the potential to use two neighboring channels to record synchronously from two adjacent cells in a confluent monolayer effectively obtaining both eAP from one channel and iAP from the other, as if acquiring both from a single cell. Leveraging this understanding, in the next step, a cell was electroporated through one channel while recording eAP from the adjacent cell, enabling the simultaneous capture of both eAP and iAP data. This approach was implemented to record the input (eAP) and corresponding output (iAP) data of the data analysis, machine learning, and deep learning models.

[0191] FIG. 13 shows data plots depicting an experiment-level comparison of normalized iAPs from neighboring NEA channels, e.g., with a minimum S / N > 90, involving 17 pairs of channels recorded in two distinct experiments. This comparison highlights various errors, including correlation coefficient (r), APD50, APD90, mean absolute error (MAE), cycle time difference, and percentage errors for APD 50% and APD 90% (e.g., for n = 2661 samples, comparing 22 pairs of neighboring iAP channels from two independent cell cultures). The red lines show the median value.

[0192] It should be noted, for example, that recording from the same electrode during repetitive stimulation to obtain two consecutive identical iAPs, rather than using neighboring channels for iAPS, presents two main challenges.

[0193] A first challenge includes a gradual change in iAP shapes (e.g., electroporation-induced state changes). For instance, when cells are exposed to drugs, the iAPshapes can gradually change over time. As a result, two subsequent iAPs may not necessarily be similar, especially if the drug effects are still evolving. This can make it difficult to ensure that the recorded iAPs are truly identical or representative of a steady- state condition. For example, repetitive stimulation at the same site alters membrane pore size / seal over time can cause amplitude drift and subtle morphology changes, so consecutive iAPs are not identical. In the example implementations, the method can use neighboring channels in a confluent monolayer to acquire a high-fidelity iAP concurrently with the eAP on an adjacent channel, avoiding the need to re-porate the same site. Example results described herein shows quantified neighboring-channel iAPs are extremely similar over time and drug perturbations (e.g., small dCT / APD / MAE differences), as shown in FIG. 2C-2F, e.g., thereby validating this training surrogate.

[0194] A second challenge includes an electroporation recovery time, e.g., baseline drift and undershoot artifacts in NEA iAPs. For instance, after a cell is electroporated, it can typically take around 30 to 60 minutes for the membrane to reseal and recover fully. If an eAP were recorded, for example, followed by the next iAP after electroporation, the 30-minute waiting period would significantly reduce the throughput of data recording. This process would hinder the ability to collect the large number of synchronized eAP / iAP pairs necessary for training deep learning models, which require substantial amounts of data. For example, in the example implementations, the example embodiment of the model normalized iAPs to [0.1, 1] to suppress amplitude drift, segmented peaks for temporal alignment, and added a physics-informed loss that penalizes non-physiological super-repolarization, improving shape fidelity and cross-device generalization (e.g., MEA Test 3).

[0195] III. Drug Test Experiments

[0196] The following provides a detailed overview of the drugs used in the experiments, their interactions with cardiac ion channels, and their impact on the shape of action potentials.

[0197] Dofetilide: Class III antiarrhythmic drug, dofetilide primarily blocks the hERG (IKr) potassium channels. By inhibiting this channel, it prolongs the repolarization phase, resulting in an increased action potential duration (APD) and a longer QT interval.

[0198] Quinidine: Class IA antiarrhythmic, quinidine blocks sodium (Na+) channels (INa), effectively depressing the rapid initial depolarization phase of the action potential. This mechanism reduces the sodium current and affects the overall electrical activity of the heart.

[0199] Nifedipine: Calcium channel blocker in the dihydropyridine class, nifedipine selectively blocks L-type calcium channels (ICa-L). This results in a shortened plateau phase of the action potential, leading to a reduced action potential duration.

[0200] Flecainide: Class IC antiarrhythmic drug, flecainide blocks both Na+channels (INa) and, to a lesser extent, potassium (IKr) channels. Its overall effect is to prolong the duration of the action potential and the effective refractory period in ventricular fibers, while both are shortened in the Purkinje system, which is likely consistent with its sodium channel blockade.

[0201] Lidocaine: Class IB antiarrhythmic, lidocaine primarily blocks Na+channels (INa). It may shorten the APD, though the effect observed in the experiments varied based on the sample conditions.

[0202] Propranolol: Known as a beta-blocker 45 and classified as a Class II antiarrhythmic, propranolol also exhibits sodium (INa) channel blocking properties, further influencing cardiac electrical activity. It can result in shortening APD90, by increasing the repolarization slope.

[0203] This comprehensive understanding of the drug-channel interactions and their resultant impact on action potentials aids in evaluating the electrophysiological responses during experiments.

[0204] FIGS. 14A-14D show data plots depicting a comparison of eAP and iAP baseline recordings across different experiments, as well as a comparison of the drugs used in the example implementations, their dosages, and their impact on normalized iAP shapes. Additionally, the wafer (W) and device (D) numbers of the unique NEAs used in the study are included. A color gradient from red to blue indicates the timeline of the experiments, with the bluer tones representing data from more recent experiments. FIG. 14A shows the example iAP intracellular baseline recordings; FIG. 14B shows the example eAP extracellular baseline recordings; FIG. 14C shows the iAP recordings after drug administration; and FIG. 14D shows the eAP recordings after drug administration.

[0205] IV. eAP Features Determination Methods

[0206] Determination of Starting Point bpl: bpl, is identified through a series of signal processing steps aimed at detecting significant changes in the signal’s behavior.

[0207] 1. Signal Smoothing. The initial portion of the extracellular action potential (eAP) signal, from the beginning up to the point of maximum voltage (xl), is extracted for analysis. To reduce noise and highlight the essential features of the signal, a Savitzky-Golay filter is applied.This filtering technique smooths the signal while preserving its overall shape, which is critical for accurately detecting transitions in the signal.

[0208] 2. Gradient Calculation. After smoothing, the gradient of the signal is computed to accentuate the areas where rapid changes occur, which are often indicative of important transitions such as the start of the signal. This gradient is further smoothed to reduce noise, which ensures that only the most significant changes are considered in the next step.

[0209] 3. Change Point Detection. The processed gradient signal is analyzed using a method based on Jenks natural breaks optimization, which is designed to find natural divisions in data by minimizing the variance within groups and maximizing the variance between groups. In this context, the method is used to identify distinct changes in the signal that likely correspond to the initiation of the eAP. Specifically, the algorithm segments the gradient into three distinct classes, and the most relevant change point for bpl is selected as the last point where the gradient signal shifts from one class to another.

[0210] Determination of bp2: bp2, is determined using a similar methodology, tailored to identify the conclusion of the signal.

[0211] 1. Signal Segmentation. A specific segment of the eAP signal, starting just beyond the point of maximum voltage and extending a short distance thereafter, is isolated for this analysis. This region is chosen because it is expected to contain the signal's descent back to the baseline, marking the end of the action potential.

[0212] 2. Signal Smoothing and Gradient Calculation. As with the determination of bpl, the signal segment is smoothed using the Savitzky-Golay filter, and its gradient is calculated and smoothed. This step is crucial for emphasizing the downward trend as the signal returns to baseline levels.

[0213] 3. Change Point Detection. The processed gradient of this segment is then analyzed using a slightly modified version of the Jenks natural breaks optimization, this time dividing the data into two classes. This adjustment reflects the simpler nature of the signal at this stage, which primarily involves a transition back to the baseline. The algorithm identifies the most significant change point as the first point where the gradient signal transitions between classes, corresponding to the point where the signal's descent begins to level off, marking the end of the eAP.

[0214] V. eAP Features Screening

[0215] Distorted spikes in eAP signals, as shown in FIG. 3B, and the strong correlation between ATi and AT2 with ATS(r = 0.93 and r = 0.87 respectively), led us to opt ATSthe representative of the eAP spike temporal features.

[0216] FIG. 15A shows a heatmap data plot depicting a correlation between temporal-related eAP features (ATi, AT2, ATS, and Td) for the undistorted portion of eAP / iAP data collected to be utilized in the machine learning and deep learning models.

[0217] FIG. 15B shows a heatmap data plot depicting a correlation between measurable eAP features (excluding ATi and AT2) for all of the eAP / iAP data (with distorted and undistorted eAP spikes) utilized in the machine learning and deep learning models (e.g., n=1049 samples).

[0218] VI. Physics Loss in PIA- UNET

[0219] FIG. 16A shows data plots showing an overlay of predicted iAPs using PIA-UNET, actual iAPs, and iAPs generated by the Aliev-Panfilov model based on PIA-UNET physics parameter output values, for all three test sets.

[0220] FIG. 16B shows data plots showing an overlay of the physics-informed loss values from the PIA-UNET for all three test set predictions. The parameter “a” was predicted to be 0.057, “k” was fixed at 100.00 for all samples, and the distribution of “x” is shown in FIG. 16C. The initial membrane potential uo was set to 0.1, with du / dt at to initialized to 0.4 and the equation was solved for t = 1.6 sec.

[0221] FIG. 16C shows data plots showing a distribution of the x value associated with the physics-informed loss values from the PIA-UNET for all three test set predictions. The box plots show the median (center line), interquartile range (IQR: box bounds), whiskers (1.5xIQR), and outliers (points beyond whiskers).

[0222] VII. Comparison of Model Performance With and Without Physics-based Loss Incorporation

[0223] Table SI-1, below, shows a comparison of Model performance with and without physics-based loss incorporation. The results are reported as mean ± standard deviation.Table SI-1

[0224] The results show that incorporating physics into the loss function led to a notable improvement in APD error, enhancing the waveform similarity. However, this adjustment caused a higher mean absolute error due to the correction of super-repolarization — a brief undershoot at the end of repolarization observed in some NEA recordings. In terms of generalizability, the physics-based approach performed significantly better in Test 3, which involved eAPs on MEA, when compared to patch clamp iAPS recordings. This suggests improved robustness across different test conditions despite the increase in absolute error.

[0225] VIII. Nanoelectrode Arrays Layout and Neighboring Channels Distance

[0226] The following figure shows an example distance between two neighboring channels that was used to record pairs of iAP / eAP signals. The average distance between these channels is approximately 120 pm.

[0227] FIG. 17 shows a diagram depicting an example NEA layout showing an example of the distance between two neighboring channels.

[0228] IX. Activation Map and APD Map

[0229] To study the application of PIA-UNET on NEA multi-channel recordings, the APD map and Activation map are calculated. It is noted that the channels are arranged in a starshaped nanoelectrode dish, which was considered when comparing the Activation map and APD map.

[0230] FIG. 18A shows data plots of an example NEA-recorded eAP Activation map (top) vs. its reconstructed APD 50 map (bottom).

[0231] FIG. 18B shows a diagram of an exemplary star-shaped pattern of the NEA that indicates the channels’ arrangement.

[0232] X. Simplification of the Aliev-Panfili Model

[0233] The following are governing equations:= fcl?(l — l’)(v“ «) “ wdt(SI)

[0236] For v fi 0, v is divided and simplified.

[0237] 3. Simplification of the left side of Equation (S6)

[0238] Each term is differentiated.(S9)

[0239] 4. Put together

[0241] Putting the left and right sides together: Equation (S12)

[0242] XL Example XGBoost Model Performance on the Training and Test Sets

[0243] Table SI-2, below, shows XGBoost Model performance on the training and test sets. The results are reported as mean ± standard deviation.Table SI-2Examples

[0244] In some embodiments in accordance with the present technology (example Al), a method for characterizing intracellular electrophysiology supported by artificial intelligence (Al) includes reconstructing intracellular action potential (iAP) waveforms from extracellular action potential (eAP) waveforms, wherein the reconstructing the iAP waveforms is based on employing a machine learning model that intuitively translates the relationship between eAPs and iAPs by focusing on intrinsic patterns, thus bypassing the complex parameter estimation; and analyzing the reconstructed iAP waveforms by correlating between at least some eAP and iAP pairs.

[0245] Example A2 includes the method of example Al, wherein the method further includes acquiring iAP recordings and eAP recordings simultaneously from cells on the one orboth of the MEA and the NEA; and providing the simultaneously acquired iAP recordings and eAP recordings to the machine learning model.

[0246] In some embodiments in accordance with the present technology (example A3), a system for characterizing intracellular electrophysiology supported by artificial intelligence (Al) includes a sensor device comprising one or both of a microelectrode array (MEA) and a nanoelectrode array (NEA); and a computing device in communication with the sensor device, the computing device comprising a processor, and a memory coupled to the processor and storing instructions that, when executed by the processor, cause the system to perform operations comprising: reconstructing intracellular action potential (iAP) waveforms from extracellular action potential (eAP) waveforms, wherein the reconstructing the iAP waveforms is based on employing a machine learning model that intuitively translates the relationship between eAPs and iAPs by focusing on intrinsic patterns, thus bypassing the complex parameter estimation; and analyzing the reconstructed iAP waveforms by correlating between at least some eAP and iAP pairs.

[0247] Example A4 includes the system of example A3, wherein the instructions are further configured to cause, when executed by the processor of the computing device, the system to control the sensor device to simultaneously acquire intracellular action potential (iAP) recordings and extracellular action potential (eAP) recordings from cells on the one or both of the MEA and the NEA.

[0248] In some embodiments in accordance with the present technology (example Bl), a method for characterizing an intracellular electrophysiological response of cells supported by artificial intelligence includes receiving, at a data processing module operating on a computing device, a data set comprising electrophysiological signals obtained from one or more cardiomyocyte cells on a sensor device; providing, to an artificial intelligence (Al) system comprising a trained machine learning (ML) model operating on a computer system, the data set and an instruction comprising a condition parameter; and reconstructing, by the trained ML model, an intracellular action potential (iAP) data set from the electrophysiological signals that predictively determines iAP waveforms of the one or more cardiomyocytes in response to the condition parameter.

[0249] Example B2 includes the method of example Bl or any of examples B1-B23, wherein iAP waveform response to the condition parameter for the one or more cardiomyocytes ispredictively determined without performing complex parameter estimations. For instance, the complex parameter estimations can include any of: electrode-cell seal resistance, electrode double-layer capacitance, cleft gap impedance, intracellular / extracellular bidomain conductivities or anisotropy ratios, membrane access resistance, or a site-specific electrode-cell impulse response.

[0250] Example B3 includes the method of example Bl or any of examples B1-B23, wherein the reconstructing the iAP data set comprises: translating ML- trained iAP waveform characteristics into the iAP waveforms based on determined patterns indicative of relationships between extracellular action potential (eAP) waveforms and iAP waveforms.

[0251] Example B4 includes the method of example B3 or any of examples B1-B23, wherein the translating the ML-trained iAP waveform characteristics into the iAP waveforms is further based on information about the condition parameter.

[0252] Example B5 includes the method of example B3 or any of examples B1-B23, wherein the translating the ML-trained iAP waveform characteristics into the iAP waveforms is without any information about the condition parameter.

[0253] Example B6 includes the method of example B3 or any of examples B1-B23, wherein the determined patterns are derived by the trained ML model from implementation of a training method where a ML model interprets patterns in the relationship between simultaneously-measured pairs of eAP signals and iAP signals measured from cardiomyocyte cells.

[0254] Example B7 includes the method of example B6 or any of examples B1-B23, wherein a quantity of the simultaneously-measured pairs of the eAP signals and the iAP signals measured from the cardiomyocyte cells is at least a thousand.

[0255] Example B8 includes the method of example B6 or any of examples B1-B23, wherein the simultaneously-measured pairs of eAP signals and iAP signals are measured at a same time and on a same cell or neighboring cells of a plurality of cells of the cardiomyocyte cells.

[0256] Example B9 includes the method of example Bl or any of examples B1-B23, wherein the trained ML model comprises a deep learning model.

[0257] Example B10 includes the method of example B9 or any of examples B1-B23, wherein the deep learning model includes an autoencoder U-Net or transformer architecture.

[0258] Example Bll includes the method of example Bl or any of examples B1-B23, wherein the trained ML model comprises a Physics-Informed Attention-Gated U-Net (PIA-UNET) machine learning architecture having an integrated U-Net architecture with attention gate mechanisms and physics and / or biophysics parameters for physics-informed learning.

[0259] Example B12 includes the method of example Bll or any of examples B1-B23, wherein the PIA-UNET machine learning architecture includes: an encoder configured to initiate with a Convolution-BatchNormalization-ReLU (CBR) block and progress through a sequence of residual blocks (Resblocks); a decoder configured to progressively restore spatial resolution across multiple decoder levels; and an attention mechanism configured to enhance feature representations in the decoder by considering features at corresponding levels in the encoder.

[0260] Example B13 includes the method of example B12 or any of examples B1-B23, wherein each Resblock of the Resblocks comprises: two consecutive CBR blocks; a squeeze-and-excitation (SE) block configured to recalibrate channel-wise feature responses adaptively; and a residual connection configured to integrate an original input to a recalibrated output.

[0261] Example B14 includes the method of example B13 or any of examples B1-B23, wherein the SE block is configured to implement global average pooling, dimensionality reduction, and subsequent scaling using a two-layer fully connected network with ReLU and sigmoid activations.

[0262] Example B15 includes the method of example Bl or any of examples B1-B23, further comprising: pre-processing the electrophysiological signals in the received data set by processing, at the data processing module, the electrophysiological signals based on a threshold indicative of a signal quality condition of the one or more cardiomyocyte cells recorded by the sensor device; selecting electrophysiological signal data that meets the threshold; and providing the selected electrophysiological signal data to the Al system for implementing the reconstructing the iAP data set from the electrophysiological signals.

[0263] Example B16 includes the method of example B15 or any of examples B1-B23, wherein the threshold includes signal-to-noise (SNR) value or signal amplitude of the one or more cardiomyocyte cells.

[0264] Example B17 includes the method of example B16 or any of examples B1-B23, wherein the signal quality condition includes a giga-ohm (GΩ) seal associated with a patch clamp measurement of the one or more cardiomyocyte cells.

[0265] Example B18 includes the method of example Bl or any of examples B1-B23, wherein the sensor device includes a nanoelectrode array (NEA) sensor device.

[0266] Example B19 includes the method of example Bl or any of examples B1-B23, wherein the sensor device includes a microelectrode array (MEA) sensor device.

[0267] Example B20 includes the method of example Bl or any of examples B1-B23, wherein the method is agnostic to the sensor device being an NEA sensor device or a MEA sensor device.

[0268] Example B21 includes the method of example Bl or any of examples B1-B23, wherein the sensor device includes a patch clamp sensor device.

[0269] Example B22 includes the method of example Bl or any of examples B1-B23, wherein the condition parameter includes one or more drugs and / or one or more toxins.

[0270] Example B23 includes the method of example B22 or any of examples B1-B21, wherein the one or more drugs and / or one or more toxins includes dofetilide. quinidine, nifedipine, flecainide, lidocaine, and propranolol.

[0271] In some embodiments in accordance with the present technology (example B24), a method for training a machine learning (ML) model for predicting intracellular response signals includes simultaneously acquiring, at a sensor device, pairs of intracellular action potential (iAP) signals and extracellular action potential (eAP) signals from cardiomyocyte cells on one or more electrodes of the sensor device; processing, at a data processing module operating on a computing device, the simultaneously acquired pairs of iAP and eAP signals based on a signal-to-noise (SNR) threshold indicative of a giga-ohm (GΩ) seal of the one or more cardiomyocyte cells to the electrode of the sensor device, the processing including selecting electrophysiological signal data that meets the SNR threshold; providing the selected electrophysiological signal data to a machine learning (ML) model operating on a computer system; and training the ML model to reconstruct an iAP data set from the selected electrophysiological signal data by recognizing patterns indicative of relationships between eAP waveforms and iAP waveforms of the selected electrophysiological signal data.

[0272] Example B25 includes the method of example B24 or any of examples B24-B33, wherein the simultaneously acquired pairs of iAP and eAP signals include signal characteristics related to an intracellular electrophysiological response of the cardiomyocyte cells due to exposure to a drug or chemical compound under a set of conditions.

[0273] Example B26 includes the method of example B24 or any of examples B24-B33, wherein the processing the simultaneously acquired pairs of iAP and eAP signals includes:normalizing eAP waveforms and iAP waveforms; identifying action potential peaks of the normalized eAP waveforms and the normalized iAP waveforms; and segmenting the normalized eAP waveforms and the normalized iAP waveforms around the identified action potential peaks.

[0274] Example B27 includes the method of example B26 or any of examples B24-B33, wherein the normalizing the iAP waveforms comprises scaling the iAP waveforms to a range between 0.1 and 1.

[0275] Example B28 includes the method of example B26 or any of examples B24-B33, wherein the normalizing the eAP waveforms comprises subtracting a mean of the eAP signals from each value and scaling the eAP signals relative to a noise level.

[0276] Example B29 includes the method of example B26 or any of examples B24-B33, wherein the processing the simultaneously acquired pairs of iAP and eAP signals further includes: filtering the eAP signals and iAP signals using a bandpass filter at a sampling rate of at least 5 kHz with a low-cut frequency at a LCF value in a range of 0.05 to 0.2 Hz and a high-cut frequency at a HCF value in a range of 2000 to 2500 Hz.

[0277] Example B30 includes the method of example B26 or any of examples B24-B33, wherein the segmenting comprises extracting 8000 data points equivalent to 1.6 seconds, starting 1000 points before an identified peak.

[0278] Example B31 includes the method of example B24 or any of examples B24-B33, further comprising: validating the training of the ML model by calculating action potential durations (APDs) from the reconstructed iAP waveforms, wherein the APDs comprise measurements that measure timeframes for an action potential to decrease to respective percentages of peak value and recover.

[0279] Example B32 includes the method of example B24 or any of examples B24-B33, wherein the training the ML model further includes estimating confidence intervals for iAP values.

[0280] Example B33 includes the method of example B32 or any of examples B24-B31, wherein the ML model simultaneously predicts 0.05. 0.5, and 0.95 quantiles through parallel output layers to prevent quantile crossing.

[0281] In some embodiments in accordance with the present technology (example B34), a system for characterizing intracellular electrophysiology supported by artificial intelligence includes a sensor device to acquire extracellular action potential (eAP) signals from one or morecardiomyocyte cells; and a computing system in communication with the sensor device, the computing system comprising one or more computers each including a processor, and a memory coupled to the processor and storing instructions that, when executed by the processor, cause the computing system to perform operations comprising: receiving a data set including the eAP signals, and reconstructing, by a machine learning (ML) model operating on the computing system, intracellular action potential (iAP) signals from the eAP signals, wherein the reconstructing the iAP signals predictively determines iAP waveforms or iAP characteristics of the one or more cardiomyocytes.

[0282] Example B35 includes the system of example B34 or any of examples B34-B46, wherein the reconstructed iAP signals include at least one of iAP waveforms or iAP characteristics of the one or more cardiomyocyte cells.

[0283] Example B36 includes the system of example B34 or any of examples B34-B46, wherein the reconstructed iAP signals are indicative of a response by the one or more cardiomyocytes to a substance or a condition on the cardiomyocytes.

[0284] Example B37 includes the system of example B34 or any of examples B34-B46, wherein the instructions, when executed, cause the computing system to perform further operations comprising: pre-processing the eAP signals based on a threshold to select eAP data that meets the threshold.

[0285] Example B38 includes the system of example B34 or any of examples B34-B46, wherein the instructions, when executed, cause the computing system to perform further operations comprising: determining electrophysiological metrics from the reconstructed iAP signals, including action potential durations (APD) at one or more particular repolarization percentages.

[0286] Example B39 includes the system of example B34 or any of examples B34-B46, wherein the instructions, when executed, cause the computing system to perform operations, prior to performing the receiving and the reconstructing operations, comprising: initially training the ML model on data sets that include time-synchronized iAP-eAP data pairs that are processed by one or more of normalization, peak detection, or segmentation into at least one fixed-duration window.

[0287] Example B40 includes the system of example B34 or any of examples B34-B46, wherein the instructions, when executed, cause the computing system to perform furtheroperations comprising: adaptively training the ML model on additional data sets that include time-synchronized iAP-eAP data pairs to augment the ML model.

[0288] Example B41 includes the system of example B34 or any of examples B34-B46, wherein the sensor device includes at least one of a nanoelectrode array (NEA) sensor device, a microelectrode array (MEA) sensor device, or a patch clamp sensor device.

[0289] Example B42 includes the system of example B34 or any of examples B34-B46, wherein the ML model comprises a deep learning model.

[0290] Example B43 includes the system of example B42 or any of examples B34-B46, wherein the deep learning model includes an autoencoder U-Net or transformer architecture.

[0291] Example B44 includes the system of example B34 or any of examples B34-B46, wherein the ML model comprises a Physics-Informed Attention-Gated U-Net (PIA-UNET) machine learning architecture having an integrated U-Net architecture with attention gate mechanisms and physics and / or biophysics parameters for physics-informed learning.

[0292] Example B45 includes the system of example B34 or any of examples B34-B46, wherein the computing system includes a data processing module configured to receive the data set including the eAP signals and an artificial intelligence (Al) system module that includes the ML model and is configured to reconstruct the iAP signals from the eAP signals.

[0293] Example B46 includes the system of example B45 or any of examples B34-B44, wherein the Al system module comprises one or both of a machine learning module and an Al signal analysis module.

[0294] In some embodiments in accordance with the present technology (example B47), a method for characterizing an intracellular electrophysiological response of cells supported by artificial intelligence includes receiving, at a data processing module operating on a computing device, a data set comprising electrophysiological signals obtained from one or more cardiomyocyte cells on a sensor device; providing, to an artificial intelligence (Al) system comprising a trained machine learning (ML) model operating on a computer system, the data set; and reconstructing, by the trained ML model, an intracellular action potential (iAP) data set from the data set that predictively determines iAP waveforms or iAP characteristics of the one or more cardiomyocytes.

[0295] Example B48 includes the method of example B47 or any of examples B47-B55, the electrophysiological signals include extracellular action potential (eAP).

[0296] Example B49 includes the method of example B47 or any of examples B47-B55, wherein the reconstructing the iAP data set predictively determines the iAP waveforms or iAP characteristics of the one or more cardiomyocytes in response to a substance or a condition on the cardiomyocytes.

[0297] Example B50 includes the method of example B49 or any of examples B47-B55, wherein the providing further includes providing to the Al system an instruction specifying the substance or condition to predictively determine the response.

[0298] Example B51 includes the method of example B49 or any of examples B47-B55, wherein the reconstructing the iAP data set comprises: translating ML-trained iAP waveform characteristics into the iAP waveforms based on determined patterns indicative of relationships between extracellular action potential (eAP) waveforms and iAP waveforms.

[0299] Example B52 includes the method of example B51 or any of examples B47-B55, wherein the translating the ML-trained iAP waveform characteristics into the iAP waveforms is further based on information about the substance or condition.

[0300] Example B53 includes the method of example B51 or any of examples B47-B55, wherein the translating the ML-trained iAP waveform characteristics into the iAP waveforms is without any information about the substance or condition.

[0301] Example B54 includes the method of example B47 or any of examples B47-B55, further comprising: processing, at the data processing module, the electrophysiological signals based on a signal-to-noise (SNR) threshold indicative of a giga-ohm (GΩ) seal of the one or more cardiomyocyte cells to an electrode of the sensor device, the processing including selecting electrophysiological signal data that meets the SNR threshold.

[0302] Example B55 includes the method of example B47 or any of examples B47-B54, further comprising: processing, at the data processing module, the electrophysiological signals based on a signal threshold comprising an amplitude level of the one or more cardiomyocyte cells recorded by the sensor device, the processing including selecting electrophysiological signal data that meets the signal threshold.Conclusion

[0303] Implementations of the subject matter and the functional operations described in this patent document can be implemented in various systems, digital electronic circuitry, or in computer software, firmware, or hardware, including the structures disclosed in this specificationand their structural equivalents, or in combinations of one or more of them. Implementations of the subject matter described in this specification can be implemented as one or more computer program products, i.e., one or more modules of computer program instructions encoded on a tangible and non-transitory computer readable medium for execution by, or to control the operation of, data processing apparatus. The computer readable medium can be a machine-readable storage device, a machine-readable storage substrate, a memory device, a composition of matter effecting a machine-readable propagated signal, or a combination of one or more of them. The term “data processing unit” or “data processing apparatus” encompasses all apparatus, devices, and machines for processing data, including by way of example a programmable processor, a computer, or multiple processors or computers. The apparatus can include, in addition to hardware, code that creates an execution environment for the computer program in question, e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, or a combination of one or more of them.

[0304] A computer program (also known as a program, software, software application, script, or code) can be written in any form of programming language, including compiled or interpreted languages, and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. A computer program does not necessarily correspond to a file in a file system. A program can be stored in a portion of a file that holds other programs or data (e.g., one or more scripts stored in a markup language document), in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files that store one or more modules, sub programs, or portions of code). A computer program can be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a communication network.

[0305] The processes and logic flows described in this specification can be performed by one or more programmable processors executing one or more computer programs to perform functions by operating on input data and generating output. The processes and logic flows can also be performed by, and apparatus can also be implemented as, special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application specific integrated circuit).

[0306] Processors suitable for the execution of a computer program include, by way ofexample, both general and special purpose microprocessors, and any one or more processors of any kind of digital computer. Generally, a processor will receive instructions and data from a read only memory or a random access memory or both. The essential elements of a computer are a processor for performing instructions and one or more memory devices for storing instructions and data. Generally, a computer will also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data, e.g., magnetic, magneto optical disks, or optical disks. However, a computer need not have such devices. Computer readable media suitable for storing computer program instructions and data include all forms of nonvolatile memory, media and memory devices, including by way of example semiconductor memory devices, e.g., EPROM, EEPROM, and flash memory devices. The processor and the memory can be supplemented by, or incorporated in, special purpose logic circuitry.

[0307] While this patent document contains many specifics, these should not be construed as limitations on the scope of any invention or of what may be claimed, but rather as descriptions of features that may be specific to particular embodiments of particular inventions. Certain features that are described in this patent document in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.

[0308] Similarly, while operations are depicted in the drawings in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. Moreover, the separation of various system components in the embodiments described in this patent document should not be understood as requiring such separation in all embodiments.

[0309] Only a few implementations and examples are described and other implementations, enhancements and variations can be made based on what is described and illustrated in this patent document.

Claims

CLAIMSWhat is claimed is:

1. A method for characterizing an intracellular electrophysiological response of cells supported by artificial intelligence, comprising:receiving, at a data processing module operating on a computing device, a data set comprising electrophysiological signals obtained from one or more cardiomyocyte cells on a sensor device;providing, to an artificial intelligence (Al) system comprising a trained machine learning (ML) model operating on a computer system, the data set and an instruction comprising a condition parameter; andreconstructing, by the trained ML model, an intracellular action potential (iAP) data set from the electrophysiological signals that predictively determines iAP waveforms of the one or more cardiomyocytes in response to the condition parameter.

2. The method of claim 1, wherein iAP waveform response to the condition parameter for the one or more cardiomyocytes is predictively determined without performing complex parameter estimations.

3. The method of claim 1, wherein the reconstructing the iAP data set comprises:translating ML-trained iAP waveform characteristics into the iAP waveforms based on determined patterns indicative of relationships between extracellular action potential (eAP) waveforms and iAP waveforms.

4. The method of claim 3, wherein the translating the ML-trained iAP waveform characteristics into the iAP waveforms is further based on information about the condition parameter.

5. The method of claim 3, wherein the translating the ML-trained iAP waveform characteristics into the iAP waveforms is without any information about the condition parameter.

6. The method of claim 3, wherein the determined patterns are derived by the trained ML model from implementation of a training method where a ML model interprets patterns in therelationship between simultaneously-measured pairs of eAP signals and iAP signals measured from cardiomyocyte cells.

7. The method of claim 6, wherein a quantity of the simultaneously-measured pairs of the eAP signals and the iAP signals measured from the cardiomyocyte cells is at least a thousand.

8. The method of claim 6, wherein the simultaneously-measured pairs of eAP signals and iAP signals are measured at a same time and on a same cell or neighboring cells of a plurality of cells of the cardiomyocyte cells.

9. The method of claim 1, wherein the trained ML model comprises a deep learning model.

10. The method of claim 9, wherein the deep learning model includes an autoencoder U-Net or transformer architecture.

11. The method of claim 1, wherein the trained ML model comprises a Physics-Informed Attention-Gated U-Net (PIA-UNET) machine learning architecture having an integrated U-Net architecture with attention gate mechanisms and physics and / or biophysics parameters for physics-informed learning.

12. The method of claim 11, wherein the PIA-UNET machine learning architecture includes:an encoder configured to initiate with a Convolution-BatchNormalization-ReLU (CBR) block and progress through a sequence of residual blocks (Resblocks);a decoder configured to progressively restore spatial resolution across multiple decoder levels; andan attention mechanism configured to enhance feature representations in the decoder by considering features at corresponding levels in the encoder.

13. The method of claim 12, wherein each Resblock of the Resblocks comprises:two consecutive CBR blocks:a squeeze-and-excitation (SE) block configured to recalibrate channel- wise feature responses adaptively; anda residual connection configured to integrate an original input to a recalibrated output.

14. The method of claim 13, wherein the SE block is configured to implement global average pooling, dimensionality reduction, and subsequent scaling using a two-layer fully connected network with ReLU and sigmoid activations.

15. The method of claim 1, further comprising:pre-processing the electrophysiological signals in the received data set by processing, at the data processing module, the electrophysiological signals based on a threshold indicative of a signal quality condition of the one or more cardiomyocyte cells recorded by the sensor device;selecting electrophysiological signal data that meets the threshold; andproviding the selected electrophysiological signal data to the Al system for implementing the reconstructing the iAP data set from the electrophysiological signals.

16. The method of claim 15, wherein the threshold includes signal-to-noise (SNR) value or signal amplitude of the one or more cardiomyocyte cells.

17. The method of claim 16, wherein the signal quality condition includes a giga-ohm (GΩ) seal associated with a patch clamp measurement of the one or more cardiomyocyte cells.

18. The method of claim 1, wherein the sensor device includes a nanoelectrode array (NEA) sensor device.

19. The method of claim 1, wherein the sensor device includes a microelectrode array (MEA) sensor device.

20. The method of claim 1, wherein the method is agnostic to the sensor device being an NEA sensor device or a MEA sensor device.

21. The method of claim 1, wherein the sensor device includes a patch clamp sensor device.

22. The method of claim 1, wherein the condition parameter includes one or more drugs and / or one or more toxins.

23. The method of claim 22, wherein the one or more drugs and / or one or more toxins includes dofetilide, quinidine, nifedipine, flecainide, lidocaine, and propranolol.

24. A method for training a machine learning (ML) model for predicting intracellular response signals, comprising:simultaneously acquiring, at a sensor device, pairs of intracellular action potential (iAP) signals and extracellular action potential (eAP) signals from cardiomyocyte cells on one or more electrodes of the sensor device;processing, at a data processing module operating on a computing device, the simultaneously acquired pairs of iAP and eAP signals based on a signal-to-noise (SNR) threshold indicative of a giga-ohm (GΩ) seal of the one or more cardiomyocyte cells to the electrode of the sensor device, the processing including selecting electrophysiological signal data that meets the SNR threshold;providing the selected electrophysiological signal data to a machine learning (ML) model operating on a computer system; andtraining the ML model to reconstruct an iAP data set from the selected electrophysiological signal data by recognizing patterns indicative of relationships between eAP waveforms and iAP waveforms of the selected electrophysiological signal data.

25. The method of claim 24, wherein the simultaneously acquired pairs of iAP and eAP signals include signal characteristics related to an intracellular electrophysiological response of the cardiomyocyte cells due to exposure to a drug or chemical compound under a set of conditions.

26. The method of claim 24, wherein the processing the simultaneously acquired pairs of iAP and eAP signals includes:normalizing eAP waveforms and iAP waveforms;identifying action potential peaks of the normalized eAP waveforms and the normalized iAP waveforms; andsegmenting the normalized eAP waveforms and the normalized iAP waveforms around the identified action potential peaks.

27. The method of claim 26, wherein the normalizing the iAP waveforms comprises scaling the iAP waveforms to a range between 0.1 and 1.

28. The method of claim 26, wherein the normalizing the eAP waveforms comprises subtracting a mean of the eAP signals from each value and scaling the eAP signals relative to a noise level.

29. The method of claim 26, wherein the processing the simultaneously acquired pairs of iAP and eAP signals further includes:filtering the eAP signals and iAP signals using a bandpass filter at a sampling rate of at least 5 kHz with a low-cut frequency at a LCF value in a range of 0.05 to 0.2 Hz and a high-cut frequency at a HCF value in a range of 2000 to 2500 Hz.

30. The method of claim 26, wherein the segmenting comprises extracting 8000 data points equivalent to 1.6 seconds, starting 1000 points before an identified peak.

31. The method of claim 24, further comprising:validating the training of the ML model by calculating action potential durations (APDs) from the reconstructed iAP waveforms, wherein the APDs comprise measurements that measure timeframes for an action potential to decrease to respective percentages of peak value and recover.

32. The method of claim 24, wherein the training the ML model further includes estimating confidence intervals for iAP values.

33. The method of claim 32, wherein the ML model simultaneously predicts 0.05, 0.5, and 0.95 quantiles through parallel output layers to prevent quantile crossing.

34. A system for characterizing intracellular electrophysiology supported by artificial intelligence, comprising:a sensor device to acquire extracellular action potential (eAP) signals from one or more cardiomyocyte cells; anda computing system in communication with the sensor device, the computing system comprising one or more computers each including a processor, and a memory coupled to the processor and storing instructions that, when executed by the processor, cause the computing system to perform operations comprising:receiving a data set including the eAP signals, andreconstructing, by a machine learning (ML) model operating on the computing system, intracellular action potential (iAP) signals from the eAP signals, wherein the reconstructing the iAP signals predictively determines iAP waveforms or iAP characteristics of the one or more cardiomyocytes.

35. The system of claim 34, wherein the reconstructed iAP signals include at least one of iAP waveforms or iAP characteristics of the one or more cardiomyocyte cells.

36. The system of claim 34, wherein the reconstructed iAP signals are indicative of a response by the one or more cardiomyocytes to a substance or a condition on the cardiomyocytes.

37. The system of claim 34, wherein the instructions, when executed, cause the computing system to perform further operations comprising: pre-processing the eAP signals based on a threshold to select eAP data that meets the threshold.

38. The system of claim 34, wherein the instructions, when executed, cause the computing system to perform further operations comprising: determining electrophysiological metrics from the reconstructed iAP signals, including action potential durations (APD) at one or more particular repolarization percentages.

39. The system of claim 34, wherein the instructions, when executed, cause the computing system to perform operations, prior to performing the receiving and the reconstructing operations, comprising: initially training the ML model on data sets that includetime-synchronized iAP-eAP data pairs that are processed by one or more of normalization, peak detection, or segmentation into at least one fixed-duration window.

40. The system of claim 34, wherein the instructions, when executed, cause the computing system to perform further operations comprising: adaptively training the ML model on additional data sets that include time-synchronized iAP-eAP data pairs to augment the ML model.

41. The system of claim 34, wherein the sensor device includes at least one of a nanoelectrode array (NEA) sensor device, a microelectrode array (MEA) sensor device, or a patch clamp sensor device.

42. The system of claim 34, wherein the ML model comprises a deep learning model.

43. The system of claim 42, wherein the deep learning model includes an autoencoder U-Net or transformer architecture.

44. The system of claim 34, wherein the ML model comprises a Physics-Informed Attention-Gated U-Net (PIA-UNET) machine learning architecture having an integrated U-Net architecture with attention gate mechanisms and physics and / or biophysics parameters for physics-informed learning.

45. The system of claim 34, wherein the computing system includes a data processing module configured to receive the data set including the eAP signals and an artificial intelligence (Al) system module that includes the ML model and is configured to reconstruct the iAP signals from the eAP signals.

46. The system of claim 45, wherein the Al system module comprises one or both of a machine learning module and an Al signal analysis module.

47. A method for characterizing an intracellular electrophysiological response of cells supported by artificial intelligence, comprising:receiving, at a data processing module operating on a computing device, a data set comprising electrophysiological signals obtained from one or more cardiomyocyte cells on a sensor device;providing, to an artificial intelligence (Al) system comprising a trained machine learning (ML) model operating on a computer system, the data set; andreconstructing, by the trained ML model, an intracellular action potential (iAP) data set from the data set that predictively determines iAP waveforms or iAP characteristics of the one or more cardiomyocytes.

48. The method of claim 47, the electrophysiological signals include extracellular action potential (eAP).

49. The method of claim 47, wherein the reconstructing the i AP data set predictively determines the iAP waveforms or iAP characteristics of the one or more cardiomyocytes in response to a substance or a condition on the cardiomyocytes.

50. The method of claim 49, wherein the providing further includes providing to the Al system an instruction specifying the substance or condition to predictively determine the response.

51. The method of claim 49, wherein the reconstructing the iAP data set comprises:translating ML-trained iAP waveform characteristics into the iAP waveforms based on determined patterns indicative of relationships between extracellular action potential (eAP) waveforms and iAP waveforms.

52. The method of claim 51, wherein the translating the ML-trained iAP waveform characteristics into the iAP waveforms is further based on information about the substance or condition.

53. The method of claim 51, wherein the translating the ML-trained iAP waveform characteristics into the iAP waveforms is without any information about the substance or condition.

54. The method of claim 47, further comprising:processing, at the data processing module, the electrophysiological signals based on a signal-to-noise (SNR) threshold indicative of a giga-ohm (GΩ) seal of the one or more cardiomyocyte cells to an electrode of the sensor device, the processing including selecting electrophysiological signal data that meets the SNR threshold.

55. The method of claim 47, further comprising:processing, at the data processing module, the electrophysiological signals based on a signal threshold comprising an amplitude level of the one or more cardiomyocyte cells recorded by the sensor device, the processing including selecting electrophysiological signal data that meets the signal threshold.

Images