Huge variability captured in simulated iPSC-derived cardiomyocyte populations.

(A) To illustrate population variability, 20 action potentials (APs) were shown, each resulting from ±40% random variation applied to 52 parameters governing six key ionic currents (IKr, ICaL, INa, IKs, IK1, and If,) in the baseline Kernik iPSC-CM model50, within a simulated population of 200,000 spontaneously beating cells. (B) These perturbations yielded a wide spectrum of APs, with substantial variation in both waveform and frequency. (C) The corresponding total ionic current and its decomposition into individual current components are visualized in panels (IKr, ICaL, INa, IKs, IK1, and If, D–I, respectively).

Digital twins of human iPSC-CMs from one simple voltage-clamp recording.

Top row (pipeline overview): A large population of synthetic iPSC-CMs (top left panel) is generated by introducing variation to 52 biophysical parameters governing key ionic currents in the baseline model: IKr, ICaL, INa, IKs, IK1, and If (highlighted with red asterisks in the schematic on right). A computationally optimized voltage-clamp protocol (black trace, top left center panel) is applied to generate distinct whole cell current (ITotal, red trace), enabling cell-wide excitation of key ion channels. The simulated whole cell currents ITotal from 200,000 synthetic cells serve as inputs to a fully connected neural network, trained to map raw current responses to the underlying 52 model parameters with high accuracy, as demonstrated by the low MSE across training and test sets (top center right panel). The deep learning model is trained to predict optimized parameter values by inferring gating kinetics and maximal conductance for each ionic species. An example formulation for the fast sodium current (INa) is shown, with inferred parameters (x₁–x₅) contributing to gating and conductance (top right panel). In the bottom row, the resultant digital twin output is shown with the inferred parameters used to simulate the AP, whole cell currents ITotal and each key ion channel INa, ICaL, IK1, IKs, IKr, and If.

Deep learning guided optimization of an ideal voltage clamp protocol.

Each iteration began with a −100 mV holding potential for 250 ms, followed by sequential testing potentials from −120 mV to +50 mV in 10 mV increments. A total of 200,000 synthetic samples were generated per training cycle. The optimal testing potential, identified by the lowest MSE from the deep learning model, was then applied for 250 ms. This optimization cycle repeated every 7000 ms. At 6000 ms, the potential was transiently stepped to −120 mV for 250 ms before resuming the next testing potential sequence. The schematic illustrates the iterative loop of model evaluation, MSE-based selection, and protocol updating, leading to the optimized composite voltage waveform shown in the lower panel.

Massive synthetic data to train and test deep learning algorithm for ion channel parameter estimation.

(A) A deep learning derived optimized voltage-clamp protocol (left panel) was designed to activate a broad set of ionic currents in iPSC-CMs by applying dynamic membrane potential steps (black trace), producing distinctive whole-cell current responses (ITotal, red trace). The inset highlights fine-scale current kinetics captured during the protocol. These Vm-ITotal pairs serve as network inputs for parameter inference. (B) Prediction accuracy for individual model parameters was evaluated using training datasets of 1,000, 10,000, and 200,000 synthetic cells (left to right). For the smallest dataset, the test error exhibited an asymmetric U-shaped curve across training epochs, indicating limited generalizability. As dataset size increased, training and test errors converged, with median mean absolute errors (MAE) of 0.041, 0.038, and 0.020 for the three dataset sizes, respectively. Bottom panels show MAE distributions across parameters, demonstrating progressively narrower error ranges with larger training datasets. (C) APs (Vm), intracellular calcium transients (Cai), total ionic current (ITotal), and six major individual ionic currents (IKr, ICaL, IKs, IK1, INa, If) simulated from the predicted parameters of a single test cell using the deep learning network trained on 200,000 samples. Blue traces represent the original simulated data (input), and red traces show the model outputs. The close overlap confirms accurate parameter recovery and faithful reproduction of cell-specific electrophysiological behavior.

Digital twin generation and predictive modeling from real cells in iPSC-CM experimental recordings.

(A, B, top) Ten simulated APs from a synthetic population of 1,100,000 induced pluripotent stem cell-derived cardiomyocytes (iPSC-CMs) at room-temperature were generated by introducing ±40% random variation to 52 biophysical model parameters from baseline, governing six major ionic currents: (IKr, ICaL, IKs, IK1, INa, If). Colored traces represent exemplar simulated cells from training data; overlaid black traces are experimentally recorded APs from two representative iPSC-CMs from hiPSC cell line iPS-6-9-9T.B. Insets show the experimental whole-cell current responses to a deep learning optimized voltage-clamp protocol (as described in Figure 3), used for parameter inference. (C, D, bottom) Digital twin models of the same experimental iPSC-CMs shown in panels A and B. Digital twins were created by extracting all 52 model parameters from the experimental whole-cell current using the deep learning inference pipeline. These parameters were used to instantiate cell-specific computational models, and AP simulations (red traces) were generated. The close overlay between the experimental traces (black) and digital twin predictions (red) demonstrates the success of the framework to accurately derive cell-specific digital twins from real cells.

Cell-to-cell variability drives population-level responses to drug application in large synthetic iPSC-CM digital twins.

(A) iPSC-CM model Cell 1 (corresponding to Figure 5A). (B) iPSC-CM model Cell 2 (corresponding to Figure 5B). In both panels, the top traces show control simulations at physiological temperature (37 °C). The middle row show responses in the presence of E-4031 (50 nM). Early afterdepolarizations (EADs) occurred in Cell 1 (red) at 50 nM, but not in Cell 2 (blue). (C) We applied ±20% perturbations to all parameters governing six key ionic currents (IK1, IKr, IKs, ICaL, INa, If) in Cell 1 (orange) and Cell 2 (cyan) to generate a population of virtual cells (n = 4000), capturing the full spectrum of cell variability within a cell line. Overlaid membrane potential traces (black) show APs from Cell 1 and Cell 2. The population average action potential duration at 90% repolarization (APD90) was 403 ± 47 ms. (D) Incidence of EADs (%) in the virtual population as a function of E-4031 concentration. This population-based digital twin modeling framework enables statistical comparisons across cell lines or patient-specific digital twins.