Four CiPA compounds, read end to end.

Dofetilide is the reference positive control for TdP. It is a near-selective hERG blocker with a free Cmax of 2 nM and a hERG IC50 of 4.87 nM, so it sits at roughly one IC50 of its own target at therapeutic exposure. If a mechanism read cannot get dofetilide right, nothing else it says is worth reading. This is the easy case, and it is here to set the ceiling.

Free Cmax 2 nM. Concentrations examined: 2, 4, 6 and 8 nM.

Distinguishable. The true block is 30.4%, so the true conductance scale is 0.696. The search brackets it with the two adjacent grid points and puts the next current three times further away. The gap between rank 2 and rank 3 is a factor of three in fit error, which is the signature of an identifiable single-current mechanism.

Ambiguous. Allowing two currents instead of one produces a better number: hERG/IKr 0.75 paired with IKs 0.65 fits at 0.044. IKs is not part of dofetilide's mechanism. A second free parameter always improves a fit. We only promote a pair over a single when the single grid fails outright, which here it does not.

None is needed to identify the mechanism, and that is the finding. The useful experiment here is quantitative rather than discriminating: a concentration series at 0.25x to 2x free Cmax on iPSC-CM optical action potentials, to test the predicted +15.3% APD90 at 2 nM.

Do not run it at 4x. At 4x the model stops repolarising, and every hypothesis that also stops repolarising scores identically. Collapse is a saturated signal: it tells you the compound is dangerous and nothing about which current did it. Three of the top hypotheses at 4x are tied at the same fit value for exactly this reason, including one, IK1 raised twofold, that has nothing to do with dofetilide.

Also correct, at a fraction of the effort. A free hERG margin of -0.386 log units puts dofetilide over the High boundary. On this compound the mechanism read buys you the dose-response curve and the recommended concentration window, not a better class call.

Verapamil is the compound that broke hERG-only screening. It is a potent hERG blocker that never causes TdP in the clinic, because it blocks L-type calcium at the same time and the two effects cancel. It is the canonical argument for reading mechanism instead of a single channel. It is also the drug most often quoted in our favour, and the quote does not survive contact with the result: we get verapamil wrong too, just less wrong.

Free Cmax 81 nM. Concentrations examined: 81, 162, 243 and 324 nM. The two IC50s are within a factor of 1.4 of each other, which is what makes this case hard and what makes it interesting.

Every single-current answer is wrong. Verapamil's phenotype is a mild APD prolongation with the calcium contribution partly cancelling it, and no single conductance change in this model reproduces that shape. Publishing this failure is the point: a single-current search on a multi-channel compound returns a confident, wrong answer.

The pair grid recovers it. True scales are IKr 0.772 and ICaL 0.732; the search returns 0.75 and 0.75. At 4x the true scales are 0.472 and 0.373 and the search returns 0.50 and 0.35. Both currents, both magnitudes, from the action-potential shape alone.

Two hypotheses survive. Rank 1 and rank 2 are within a factor of 1.5 of each other, and they say different things: "the drug blocks calcium" against "the drug does not touch calcium and the cell is more polarised". Five voltage biomarkers cannot separate them. That is the ambiguity the confirming experiment exists to kill.

Compare with dofetilide: at 4x, dofetilide has stopped repolarising while verapamil has prolonged by 19%. That separation is produced entirely by the calcium block, and it is invisible to any hERG-only method.

Record the cytosolic calcium transient on the same iPSC-CM well as the action potential, drug against paired vehicle, and report transient amplitude as a percentage of the matched control.

The two surviving hypotheses move calcium in opposite directions, about 14 percentage points apart. That is a sign test, not a magnitude test, which is what makes it usable on a plate. A calcium readout adds one dye and no new instrument to an assay the customer is already running.

Caveat, because it matters: 14 points is the model's predicted separation, not a measured one. Well-to-well variability in iPSC-CM calcium amplitude is real, so this needs a paired vehicle design and enough wells to resolve 14 points. We have not run it on cells.

Worse, and this is the one place where that is true. The hERG margin is -0.551 against a frozen High boundary of -0.589, so it calls verapamil High and misses by two classes. It is the only compound in the 28 on which any method makes a two-class error. It is not, however, a compound we get right, and it is not the only one where we beat the margin. Head to head across the 28, our call is correct on two drugs the margin misses (astemizole and risperidone) and wrong on three it gets (loratadine, nifedipine, nitrendipine). Net, we are one compound behind. Verapamil shows what a multichannel read is for. It does not show that ours works.

Ranolazine is an anti-anginal whose therapeutic action is late sodium current block. It also blocks hERG. Clinically it is Low risk, because the late sodium block shortens the action potential by about as much as the hERG block lengthens it. We call it Intermediate. This case is on the site because the reason we get it wrong is structural and nameable, and a customer deserves to know the failure mode before they buy.

Free Cmax 1950 nM. Concentrations examined: 1950, 3900, 5850 and 7800 nM. Note the top two rows: the INaL and IKr IC50s are within 5% of each other. Ranolazine blocks both to almost exactly the same degree at every concentration tested, 21.2% against 21.7% at 1x and 49.8% against 48.7% at 4x. The Kernik-Clancy formulation carries only one of them.

The inverse read is self-consistent and useless here. True IKr scale is 0.783 and the search returns 0.75, correct to within one grid step at both 1x and 4x. It recovers exactly what it was given. It cannot recover the current it was never told about, and no amount of ranking fixes that.

The +31.6% at 4x is the number that pushes ranolazine over our Low boundary. A model carrying late sodium would shorten every row in that last column. We cannot tell you by how much, because we have not run one.

Record iPSC-CM action potentials with and without a selective late-sodium blocker in the background, then repeat the ranolazine concentration series in both conditions. If the compound's prolongation shrinks when late sodium is already blocked, the missing current is carrying the difference, and the size of the shrinkage is the correction our model owes you.

This experiment tells you how wrong the model is. An experiment that only confirms it tells you less.

Equally wrong. The hERG margin is -0.627 and also lands on Intermediate. Both methods miss because neither one sees the net effect of late-sodium block cancelling hERG block. We tested the fix directly: adding a late-sodium current to the model recovers mexiletine but still does not recover ranolazine. Simplicity is not the problem here, and complexity is not the fix.

Astemizole was withdrawn for QT prolongation and TdP. It is a very potent hERG blocker with a free Cmax of 0.26 nM, so the therapeutic margin against hERG looks enormous: an IC50 of 33.3 nM is 128-fold above exposure. A margin-based method reads that number and calls it safe. It was not safe. Astemizole is in the held-out validation half of the benchmark, so its class was never used to set our thresholds.

Free Cmax 0.26 nM. Concentrations examined: 0.26, 0.52, 0.78 and 1.04 nM. Only one current moves at all, and it moves by less than 8% even at the top concentration.

At 1x free Cmax no current is blocked by as much as 5% and the action potential is 1.25% off drug-free baseline. The search still returns a ranked list, because it always does. Its top answer is IKs reduced to 0.75, a current astemizole has no measured value for and does not meaningfully block. The three top hypotheses are separated by 0.016 in fit error.

Nothing here is identifiable, and the report refuses to name a winner. The true mechanism sits at rank 2, behind a wrong answer by 0.0053. Compare dofetilide, where the gap between the right current and the next one was 0.28. Fifty times smaller phenotype, fifty times less resolving power. A ranked list without a stated gap is a list that pretends to know things.

We also checked whether the calcium readout that rescued verapamil helps here. It does not: the two surviving hypotheses predict 99.6% and 99.3% of control calcium amplitude. That is 0.3 percentage points, well under any plate's noise floor. Reporting a discriminating assay that cannot discriminate would be worse than reporting none.

Measure IKs block directly by voltage clamp, KCNQ1 with KCNE1, at 1x to 10x free Cmax.

The rationale is not that we suspect IKs. It is that the public panel has no IKs value at all for astemizole, so the search treats IKs as an unconstrained free parameter and it wins on that freedom rather than on evidence. One measurement removes the degeneracy by removing the freedom. This is what a confirming experiment is for: it is chosen to kill an ambiguity, not to confirm a belief.

Second, cheaper option: run the concentration series out to 10x or 30x free Cmax. IKr block rises to a level the biomarkers can resolve, and the phenotype stops being degenerate. Only do this if a supratherapeutic exposure is meaningful for your programme.

Wrong. The margin is -2.107, comfortably under the frozen Low boundary of -2.034, so the baseline calls astemizole Low. A drug withdrawn from the market for TdP is read as safe by the 128-fold hERG margin, and read as Intermediate by a model that simulates what a 3% APD change does to a cell. That is the case for mechanism, stated on the single compound where it holds.