Why every read runs on a human iPSC-CM model.

We score every compound on the Kernik-Clancy 2019 human induced-pluripotent-stem-cell cardiomyocyte model. It trades the top of the classifier leaderboard for being the cell your confirming experiment runs in. This note sets out the trade, with the numbers on both sides of it.

request a free first read see four worked reads
01 · the question

What question is the model being asked?

There are two different questions you can put to a cardiac cell model, and they have different right answers.

The first is “how dangerous is this compound?” That is a classification problem. It wants the model that best separates torsadogenic compounds from safe ones on a reference set, whatever cell type that model happens to describe. The CiPA initiative answered it with an adult ventricular model and a metric called qNet, and on the published 28-drug panel that pairing is better than ours; the benchmark page shows the comparison.

The second is “what would this ion-channel profile do to the cell I am about to put it on?” That is a mechanism problem. The answer is only useful if the simulated cell and the experimental cell are the same cell. If the model predicts a 12 % action-potential prolongation in an adult ventricular myocyte and you go and measure iPSC-CMs on a multi-electrode array, a disagreement tells you nothing: a wrong mechanism and a wrong cell type look alike.

We sell the second answer. So we run the cell that the follow-up assay runs in.

Model choice here is a claim that a prediction and its test should be made in the same preparation. It is the claim that fits the way we sell the read.
02 · the model

What Kernik-Clancy 2019 is

A published single-cell model of the human iPSC-derived cardiomyocyte action potential, from Kernik, Morotti and colleagues with Colleen Clancy, in The Journal of Physiology 2019, 597(17):4533–4564 (doi:10.1113/JP277724, PMID 31278749). We run the CellML release through Myokit with a CVODE solver, as published, with its equations unchanged.

It carries 27 components. The membrane currents are the fast sodium current, L-type and T-type calcium, the rapid and slow delayed rectifiers, the inward rectifier, the transient outward current, the hyperpolarization-activated funny current, background sodium and calcium, sodium-calcium exchange, the sodium-potassium pump and the sarcolemmal calcium pump. Intracellular handling covers cytosolic and sarcoplasmic-reticulum calcium, release, uptake and leak.

Two of them drive the behaviour. The funny current and the T-type calcium current are both associated with the immature, spontaneously beating phenotype. The model beats on its own, without a stimulus, which is exactly what an iPSC-CM monolayer does in a plate.

drug-free baseline, our run against the published values
quantityour runpublished Kernik-2019 baseline
APD90413.31 ms414 ms
maximum diastolic potential−75.60 mV−76 mV
spontaneous rate61.04 bpm61 bpm
APD50356.94 msnot quoted in that table
peak upstroke velocity33.09 V/snot quoted in that table
amplitude109.89 mVnot quoted in that table

The published comparators are the Kernik-2019 baseline row of a 2021 Frontiers in Pharmacology comparison of iPSC-CM models. This is a reproduction check on our solver settings, not a validation of the model against experiment. Agreement to 0.2 % tells you we are running the published equations correctly. It tells you nothing about whether the published equations are right.

03 · the comparison

How it differs from the adult ventricular model

The reference point is the O’Hara-Rudy dynamic model, ORd, published in PLOS Computational Biology 2011 (doi:10.1371/journal.pcbi.1002061). It is built from measurements in more than 100 undiseased human hearts, and a variant of it is what CiPA’s regulatory-facing pipeline uses. It is a serious piece of work and it describes a different cell.

Kernik-Clancy 2019 · what we run O’Hara-Rudy 2011 · ORd
preparation human iPSC-derived cardiomyocyte undiseased adult human ventricular myocyte
activity spontaneous; 61.0 bpm with no stimulus applied paced; results reported at a chosen cycle length
peak upstroke velocity 33.1 V/s in our run 254 V/s simulated at 1 Hz, against 234 ± 28 V/s measured
late sodium current no separate late-sodium component present, fitted to nonfailing human ventricular data
cell subtypes one cell endocardial, epicardial and mid-myocardial variants
CaMKII regulation not represented represented across several current and calcium targets

Read the upstroke row twice. The two models disagree about the fast sodium current by nearly a factor of eight. Both numbers are correct. An iPSC-CM sits at a more depolarized maximum diastolic potential than an adult myocyte, so a larger fraction of its sodium channels are inactivated at rest, and it upstrokes slowly. The published experimental window for iPSC-CM upstroke velocity spans roughly 5 to 86 V/s. Our 33 V/s sits inside it. ORd’s 254 V/s sits inside the adult window. Both are right about their own cell.

The consequence for you is direct. Any prediction that depends on sodium-channel availability (conduction slowing, use-dependent block, a Brugada-type signal) transfers badly between these two models. A prediction that depends on repolarizing potassium current transfers better. Each read names which side of that line it sits on.

04 · the argument

Why we chose the less accurate classifier

what we give up

Aggregate risk-classification accuracy. On the public 28-drug CiPA panel the published ORd/qNet numbers are better than ours, and a free hERG margin computed on a spreadsheet beats our three-class call as well. The benchmark page carries every one of those numbers, including the ones that lose.

what we get

A prediction stated in the same units, in the same cell type, as the experiment you would run next to check it. APD90 on an iPSC-CM monolayer. A calcium transient on the same well. A voltage clamp on the channel the read is unsure about. No cell-type translation step sits between the claim and its test.

That is the entire argument. It only holds if you are going to run the confirming experiment. If your next step is a regulatory filing rather than a bench assay, the CiPA pipeline is the right tool and we will tell you so.

One thing we do not claim: regulatory acceptance. CiPA’s own documented pipeline is built on an adult ventricular model. Nothing on this site is qualified for a submission, and no output of ours has been reviewed by any agency.

05 · the cost

What this choice costs, itemised

limitations we accept by running this model
  1. No late sodium current. There is no late-sodium component in the model at all. Any compound whose safety story runs through INaL will be read wrong. Ranolazine is the clearest case: its published INaL and hERG IC50 values sit within 5 % of each other, so at every concentration it blocks both by almost the same fraction, and we simulate only one of them. We call it Intermediate risk. CiPA calls it Low. That miss is fully explained by this gap, and it is written up in full on the worked-examples page.
  2. Single cell, one myocyte. We can tell you a repolarization signal is present. Conduction, a wavefront, QRS and torsade in a whole heart live in the tissue and whole-organ models that sit downstream of the read.
  3. An immature phenotype. The spontaneous beating, the depolarized diastolic potential and the slow upstroke are all real features of iPSC-CMs and all departures from adult ventricular myocardium. This is a faithful model of an imperfect cell.
  4. Only three currents enter the risk score. The scored panel is IKr, ICaL and peak INa. Blocks you supply on other channels are reported and reasoned about; they leave the number unchanged.
  5. Mechanism ranking is scored against this model’s own output. When we rank candidate causes, the target phenotype was generated by the same equations that generate the candidates. That measures resolving power and degeneracy, no more. Accuracy against a real recording is a separate, slower test, and we state it as such.
  6. A model bounds a cause; the confirming experiment proves it. Every read we ship ends by naming the experiment that would settle it.
06 · the measured check

The drug response against a measured panel

the comparison

Blinova et al. 2018 measured field-potential duration across the 28 CiPA compounds in human iPSC-cardiomyocytes, in a blinded multisite study. We ran the same 28 compounds at the same 1×, 2×, 3× and 4× free Cmax through this model and compared its APD90 response to the measured change at each point. 112 paired points.

what came back

The direction and the ordering track the measurement. Rank correlation is 0.67 across all points, and 0.62 when repolarization collapse is left out of the comparison. Direction agrees on 89 percent of the points that move. All six compounds the model sends into repolarization collapse carry a measured early-afterdepolarization; no collapse fires without one. The collapse signal is a continuous measure, and it tracks the measured early-afterdepolarization rate across the panel: rank correlation 0.72, and area under the curve 0.82 against the 1,887 measured wells.

the limits of this check

Field-potential duration and APD90 are different measures, so this check covers direction and rank. Waveform and absolute magnitude, which the model overstates, sit outside it. The hard collapse flag is specific but conservative: it names six of the twenty-seven compounds that showed a measured early-afterdepolarization, so a collapse call is reliable while the detector misses most cases. At high exposure the model can over-call collapse, and the ranking signal degrades with it. Verapamil is wrong in the measured data too, prolonging in the model and shortening in the cells, the calcium-block case now confirmed against a recording. The mechanism ranking still awaits a matched wet dataset, and this page says where that test stands.

07 · in practice

How the model is used in a read

forward · from your panel to a phenotype

Each measured IC50 and Hill coefficient becomes a static fractional block, 1 / (1 + (IC50 / C)h), at 1×, 2×, 3× and 4× free Cmax. Those fractions scale the corresponding conductances, the model is pre-paced to steady state and run free, and we read APD90 off the resulting beats. The risk metric is the largest fractional APD90 prolongation across that concentration range, with a repolarization failure scored at a fixed ceiling. Two ordinal cut points, fitted once on 12 training drugs and then frozen, turn that number into a class.

inverse · from a phenotype back to candidate causes

Given a set of action-potential biomarkers, we sweep conductance scalings across six candidate currents, IKr, IKs, ICaL, INa, IK1 and Ito, on a fixed ten-point grid, singly and in pairs, and rank each hypothesis by normalized error against the observed biomarkers. The output is a ranked list with its scores, each hypothesis carrying a stated error.

identifiability · which of those ranks you can trust

Five biomarkers separate six currents only when one cause dominates; a single strong cause is recoverable. A two-current mechanism needs the pair grid and is then recoverable to within one grid step. A three-current mechanism sits outside what these biomarkers identify, and we do not pretend they reach it. Small-signal cases return noise: when a compound shifts APD90 by about one percent, the ranking is noise, and we report it as noise. The confirming experiment we name is chosen to break exactly the ambiguity that survives the ranking.

The useful output of this model is a short list of causes with the distance between them measured, and the confirming experiment that separates them.
see the model do the work

Four public CiPA compounds are read end to end on this model, including the two it gets wrong and the reason it gets them wrong. The interactive page lets you move conductances yourself and watch the action potential change.

four worked reads run the model live

Sources. Model: Kernik DC, Morotti S, Wu H, Garg P, Duff HJ, Kurokawa J, Jalife J, Wu JC, Grandi E, Clancy CE. A computational model of induced pluripotent stem-cell derived cardiomyocytes incorporating experimental variability from multiple data sources. J Physiol 2019;597(17):4533–4564. doi:10.1113/JP277724. Comparison model: O’Hara T, Virág L, Varró A, Rudy Y. PLoS Comput Biol 2011;7(5):e1002061. doi:10.1371/journal.pcbi.1002061; the upstroke-velocity and late-sodium rows in the comparison table are taken from that paper’s own text. Published Kernik-2019 baseline biomarkers are the corresponding row of a 2021 Frontiers in Pharmacology comparison of iPSC-CM models; the iPSC-CM experimental windows quoted for upstroke velocity derive from Doss et al. 2012 as tabulated in a 2021 Frontiers in Physiology review. Ion-channel block data and TdP classes are the CiPA reference set, from the FDA/CiPA repository and Colatsky et al., J Pharmacol Toxicol Methods 2016. Measured field-potential-duration data are the CiPA myocyte validation study, Blinova K, et al., Cell Reports 2018;24(13):3582–3592. doi:10.1016/j.celrep.2018.08.079. Simulations run in Myokit 1.39.2 with CVODE. Our own figures on this page were regenerated on 2026-09-06.

get in touch
[email protected]