전산의학연구실

Why Organoid AI?

Animal tests and single-channel assays often fail to predict how the human heart responds to a drug. Blocking the hERG potassium channel is the classic warning sign for drug-induced Torsades de Pointes (TdP), a potentially lethal arrhythmia. Current guidelines (ICH S7B and E14) still rely mainly on hERG and QT assays, which are sensitive but not very specific.

Human stem cell-derived models offer a more realistic alternative. Cardiomyocytes grown from human induced pluripotent stem cells (hiPSC-CMs) beat, carry the full set of human cardiac ion channels, and can be made from patients with a specific genetic background. The CiPA initiative (Comprehensive in vitro Proarrhythmia Assay) places these cells at the centre of next-generation cardiac safety testing.

These models produce rich, multi-parameter data, and machine learning is how we read it. Our work so far uses 2D hiPSC-CM monolayers recorded on multielectrode arrays (MEA). This is the foundation for more complex human cell-based models.

Two problems drive our work:

Interpretation: each drug and concentration yields several MEA biomarkers, and a threshold on a single parameter misses nonlinear patterns.
Biology: most models use healthy donor cells, while patients with inherited arrhythmias such as long QT syndrome (LQTS) and Brugada syndrome (BrS) are the most vulnerable.

Research Journey

Our hiPSC-CM work followed three questions: how far can a minimal feature set go, does the patient's genetic background change the answer, and can a high-fidelity cell model catch cardiotoxicity that viability assays miss? In each study, data were split by drug so that test compounds were never seen during training, and predictions were evaluated over 10,000 resampling runs.

Study 1 (2026): genotype-specific risk with patient-derived iPSC-CMs

Our first question was whether a model trained on healthy cells can judge risk for patients with inherited arrhythmias.

Cells: iPSC lines from 8 patients in the Korean Inherited Arrhythmia Registry (4 LQTS with KCNH2 or KCNQ1 mutations, 4 BrS with SCN5A, ANK2, KCNE1 or CACNB2b mutations) plus healthy controls. Only lines with both a confirmed mutation and a disease phenotype in patch-clamp recordings were kept: 3 LQTS lines (prolonged APD80) and 2 BrS lines (loss of the phase-2 dome). Three mutation carriers showed no disease phenotype, which shows that genotyping alone is not enough.
Data: MEA responses of 9 cell lines (4 normal, 3 LQTS, 2 BrS) to 26 CiPA drugs at four concentrations, using FPD, FPDc, beat period and amplitude. 11 drugs for training, 15 for testing, with binary high vs low TdP risk.
Model: neural network, random forest and XGBoost, tuned by grid search and tested in eight scenarios, including training on normal cells and testing on disease cells.
Results: the neural network trained and tested on LQTS or BrS cells reached AUC up to 0.94. Models trained only on normal cells and applied to disease lines dropped to accuracy 0.67–0.69 (AUC 0.74–0.78).
Biology: BrS cells were hypersensitive to Ca²⁺ channel blockers (nifedipine, nitrendipine, diltiazem). Quinidine changed FPDc by more than 40% at 3 µM in LQTS and BrS lines, against less than 30% in normal lines.
Application: two drugs without an official risk class, amiodarone and remdesivir, scored as high risk in all cell groups, highest in LQTS cells (score above 0.8).

Paper, Journal of Advanced Research 2026

Study 2 (2026): high-fidelity hiPSC-CMs and hidden cardiotoxicity

Patient cells showed that the cell source matters. We then asked whether a well-characterized cell model combined with AI can detect functional cardiotoxicity that standard viability assays miss.

Cells: commercial ventricular-enriched hiPSC-CMs (Cardiosight®-S, NEXEL), with cTnT+ above 95% and MLC-2V+ at 78–84% across three batches.
Data: 28 CiPA reference drugs at 0.1, 1, 10 and 100× Cmax, using FPDcF, beat period, spike amplitude and beating irregularity. 12 drugs for training, 16 for testing.
Model: logistic regression, neural network, random forest and XGBoost. The neural network performed best, with median AUC 0.982 on the held-out drugs. FPD and FPDcF were strongly correlated (r = 0.95), while the other biomarkers correlated weakly with them (r < 0.30), so each adds information.
Application: 12 anticancer drugs (anthracyclines, tyrosine kinase inhibitors, chemotherapies) were monitored by impedance for up to 120 h. Idarubicin, erlotinib, sunitinib and cyclophosphamide caused little loss of cell signal, yet MEA showed FPDcF prolongation above 20%. The model classified sunitinib (early time points, high concentrations) and erlotinib (from 72 h) as high to intermediate TdP risk, while idarubicin and cyclophosphamide stayed low risk.
Takeaway: functional cardiotoxicity can appear without structural damage, and long-term MEA monitoring combined with AI can detect it.

Paper, Archives of Toxicology 2026

Study 3 (2026): stacking ensemble on minimal MEA features

Rich multi-biomarker data is not always available. Our third question was whether two simple biomarkers are enough if the models are combined well.

Data: the public CiPA multisite MEA dataset (Blinova et al. 2018), 28 CiPA reference drugs with clinical TdP risk labels (high, intermediate, low). 12 drugs were used for training and 16 were held out for testing.
Features: only two predictors derived from ΔFPDc: the maximum change across tested concentrations and the interpolated change at clinical Cmax.
Model: a two-level stacking ensemble. Random forest, XGBoost and a shallow neural network form the first level, and their class probabilities feed a meta-classifier (random forest, XGBoost or support vector machine). Hyperparameters were tuned with drug-level cross-validation.
Results: the best single model reached AUC 0.817 on the test drugs. The XGBoost meta-classifier reached AUC 1.000 for both high+intermediate vs low and high vs intermediate+low, with pairwise ranking accuracy 1.000 and classification error of at most 0.125 across resampling runs. Training took about 1.5 s and inference 0.0004 ms per sample.
Scope: the test set contains 16 drugs, and resampling measures within-drug variability rather than generalization to new compounds. Validation on independent MEA datasets is the next step.

Paper, Annals of Biomedical Engineering 2026

Page updated

Google Sites

Report abuse