A cross-scale causal model of a human iPSC-cardiomyocyte. Knock down a gene with CRISPRi, or switch to drug-block mode and apply a percent block to an ion current. Either way the action potential recomputes in mV and ms on Kernik-Clancy 2019, published and validated. Each channel case moves the action potential in the literature direction.
A mechanistic prior finds every active perturbation in far fewer experiments than random, because the model scores the next experiment. A DEG-count proxy does worse than random. This is expected: the loudest perturbations are ion channels, network leaves that move no other gene, so a DEG-count proxy ranks them last. Fewer experiments is better.
The ranker scores every conductance-change hypothesis on its grid from five biomarkers: APD90, APD50, MDP, dV/dt max and beat rate. The fit score is a normalized RMSE over the five biomarkers, so lower is closer. The case below is a two-current effect: hERG and Cav1.2 block together. The single-current read fails here; the pair read recovers both.
| rank | hypothesis | fit score |
|---|
These targets come from the model's own forward simulation. This page shows internal consistency. It is not evidence of skill on real wet-lab traces.
Three-current mechanisms are not identifiable from these five biomarkers. On the chlorpromazine-like case the pair grid returns true and wrong pairs at near-equal score.
Small-magnitude causes are noise-fragile. An IKs gain shifts APD90 by 9 ms; at 5% noise the top single hypothesis is right one third of the time. Large-magnitude causes, like this one, survive 5% noise about 80% of the time. Small ones do not. Five biomarkers cannot separate six single currents; stronger single-current causes separate cleanly.
What it is: the cross-scale causal architecture running on a published, validated electrophysiology model (Kernik-Clancy 2019, via Myokit/CVODE). Its baseline reproduces the published biomarkers within ~1% (APD90 413.4 vs 414 ms), and each channel knockdown moves the action potential in the literature direction. A gene knockdown flows through a gene-regulatory network to ion-channel conductances to the action potential, in real units.
What it is not: the calibrated model. The gene-regulatory network here is a hand-seeded placeholder, and the gene-to-conductance step is an uncalibrated first-order map, not a per-gene fit. The functional layer matches the literature; the cross-scale composition is shown here and testable against the program's CRISPRi data. The calibrated network and per-gene gains come from that data.
On drug-block mode: it scales one channel's conductance and shows how the single-cell action potential reshapes, the way an IKr or ICaL blocker would in this model. It is not a cardiotoxicity prediction, nor a clinical ECG or QRS reading. This is a single iPSC-cardiomyocyte action potential, not tissue or a body-surface signal. It shows direction and rough magnitude of the AP change from a current block in a validated model.