📄 medicine

Virtual Phase I trials predict drug induced QT prolongation from human ventricular tissue

This paper presents a validated in silico Phase I trial framework that integrates human-calibrated cardiomyocyte populations, ion-channel blocking, tissue-level propagation, and physiological fibrosis to accurately predict drug-induced QT prolongation and inter-individual variability, demonstrating that including low-degree tissue heterogeneity is essential for achieving high concordance with clinical safety data.

Original authors: Matteo Costi, Jose Felix Rodriguez Matas

Published 2026-06-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Matteo Costi, Jose Felix Rodriguez Matas

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a chef trying to create a new dish (a new medicine). Before you serve it to the public, you need to make sure it doesn't give people heart palpitations. One specific warning sign is a condition called QT prolongation, which is like a "stutter" in the heart's electrical rhythm that can be dangerous.

Traditionally, to check for this, scientists have to test the drug on real people in "Phase I" clinical trials. This is expensive, time-consuming, and sometimes ethically tricky (you can't always test toxic cancer drugs on healthy volunteers).

This paper introduces a Virtual Phase I Trial. Think of this as a high-tech "flight simulator" for the human heart. Instead of testing the drug on real people, the researchers built a digital world where they can simulate how a drug affects a virtual heart.

Here is how their "digital heart simulator" works, broken down into simple steps:

1. The Digital Actors (Virtual Patients)

The researchers didn't just build one perfect, average heart. They created 10 unique "virtual patients."

The Analogy: Imagine a casting call for a play. Instead of finding 10 actors who all look and act exactly the same, they found 10 actors with slightly different voices, heights, and personalities.
In the paper: They created 10 different digital heart cells, each with slightly different electrical properties (some conduct electricity faster, some slower). This mimics the real-world fact that every human heart is slightly different.

2. The Stage (The Tissue with "Scars")

They placed these 10 different heart cells onto a digital stage shaped like a piece of heart muscle.

The Analogy: Imagine a wooden floor. If the floor is perfectly smooth, sound travels one way. But real floors have knots, cracks, and uneven spots.
In the paper: The researchers added 4% "physiological fibrosis" (tiny, natural "scars" or gaps) to the digital tissue. This represents the tiny, natural imperfections found in healthy adult hearts.
The Big Discovery: When they removed these natural "scars" and made the floor perfectly smooth, their predictions were way off (they overestimated the danger). When they added the natural 4% "scars," the predictions became incredibly accurate. It turns out, the "imperfections" in the heart are actually necessary to get the math right.

3. The Test (The Drug)

They introduced a virtual drug into this digital heart.

The Analogy: They poured a specific amount of "virtual poison" (the drug) into the system to see how the actors (the heart cells) reacted.
In the paper: They used real-world data on how the drug blocks specific electrical channels in the heart, combined with the exact amount of drug a patient would have in their blood during a real trial.

4. The Result (The Pseudo-ECG)

The simulator ran the experiment and produced a "Pseudo-ECG."

The Analogy: This is like a microphone recording the heart's electrical song. The researchers measured the length of the song (the QT interval) before and after the drug was added.
In the paper: They calculated exactly how many milliseconds the heart rhythm slowed down.

The Verdict: Did it Work?

The researchers tested this simulator against 61 real drugs where the actual results from human trials were already known.

The Score: The virtual trial was incredibly accurate. On average, it was off by only 2.35 milliseconds (a tiny fraction of a second).
The Safety Net: They used a strict rule: "Is there any chance the worst-case scenario for a patient is off by more than 10 milliseconds?"
- The Result: For 60 out of 61 drugs, the virtual trial said "Yes, we are safe within the 10ms limit."
- Without the "natural scars" (fibrosis), the simulator would have failed on 15 of those drugs, falsely flagging them as dangerous.

The One Glitch

There was one drug, Ranolazine, where the virtual trial was slightly less accurate (off by about 11.5 ms in the worst-case scenario). The authors suggest this is because this specific drug is very complex and interacts with the heart in a tricky way that is hard to predict without knowing every tiny detail of how the drug behaves in different people.

Summary

This paper claims that by building a digital heart that includes real human variety and natural tiny imperfections (fibrosis), scientists can predict how a drug will affect a human heart's rhythm with near-perfect accuracy.

They didn't just guess; they proved it by running 61 "virtual tests" and showing that the results matched real human history almost exactly. This suggests that in the future, we might be able to use these "flight simulators" to screen drugs for heart safety before ever testing them on a single human being.

Technical Summary: Virtual Phase I Trials for Predicting Drug-Induced QT Prolongation

Problem Statement
Drug-induced QT prolongation is a critical endpoint in early cardiac safety assessment, yet conventional Phase I clinical studies face significant limitations. These include high costs, resource intensity, and ethical constraints that prevent the systematic exploration of inter-individual variability or the testing of intrinsically toxic compounds (e.g., in oncology) in healthy volunteers. Furthermore, animal models often fail to translate due to interspecies differences in ion channel expression and repolarization reserves. While computational approaches like the Comprehensive in Vitro Proarrhythmia Assay (CiPA) have advanced the mechanistic understanding of proarrhythmic risk, a gap remains in directly predicting clinically measurable Phase I QT/QTc outcomes using a human-based, digital workflow that incorporates tissue-level propagation and patient variability.

Methodology
The authors developed a fully computational "Virtual Phase I" framework designed to emulate the logic of early clinical cardiac safety trials. The workflow integrates six sequential steps:

Virtual Patient Generation: A population of 168,000 candidate human ventricular cardiomyocytes was generated using the O'Hara-Rudy model. Nine ionic conductances ( $I_{CaL}, I_{to}, I_{NaCa}, I_{K1}, I_{Kr}, I_{Ks}, I_{Na,f}, I_{Na,L}, I_{NaK}$ ) were varied via Latin Hypercube Sampling. Cells were filtered against experimental biomarkers (RMP, APA, APD30/50/90) to retain 35,000 physiologically calibrated cells. From this pool, 10 non-overlapping virtual patients were constructed, each assigned a unique subset of 2,184 cells mapped to a ventricular tissue mesh.
Tissue Substrate Construction: Patients were embedded in a cuboid ventricular wedge (1.2 × 8 × 10 mm³) with transmural layering (endocardial, mid-myocardial, epicardial). Crucially, the model incorporated physiological interstitial fibrosis by randomly assigning 4% of tissue nodes to passive fibroblasts (MacCannell model), electrically coupled to myocytes with reduced local conductivity.
Pharmacological Exposure: Drug effects were modeled using a static Hill-type pore-block formulation. Multichannel ion-channel block parameters ( $IC_{50}$ , Hill coefficient) were derived from the Crumb and Kramer datasets. Simulations were conducted at the clinically matched free maximum effective therapeutic plasma concentration ( $EFTPC_{max}$ ).
Electrical Propagation: Tissue-level signal propagation was simulated using the monodomain equation with anisotropic conductivity. A pacing protocol of 75 beats/min (800 ms BCL) was applied.
Pseudo-ECG Reconstruction: A Geselowitz-based formulation reconstructed a digital pseudo-ECG from a fixed virtual electrode position. QT intervals were measured automatically using a tangent-based method for both control and drug-treated conditions.
Endpoint Comparison: Simulated $\Delta QT$ $Δ QT$ values were compared against clinical reference data. Two metrics were used:
- Mean-level accuracy: Bias, Mean Absolute Error (MAE), and Root Mean Square Error (RMSE).
- Regulatory concordance: A conservative Upper Bound of the 95% confidence interval (UB95) of the absolute prediction error across the 10 virtual patients, assessed against an ICH-aligned 10-ms threshold.

Key Results
The framework was validated against a cohort of 61 compounds with available clinical Phase I QT/QTc data.

Prediction Accuracy: In the configuration with 4% physiological fibrosis, the model achieved near-zero global bias (-0.001 ms), a mean absolute error (MAE) of 2.35 ms, and a root mean square error (RMSE) of 3.10 ms.
Regulatory Concordance: Using the conservative UB95 endpoint, 60 out of 61 compounds (98.4%) remained within the 10-ms agreement criterion. Only ranolazine exceeded the threshold (UB95 = 11.5 ms).
Impact of Fibrosis (Ablation Analysis): Removing physiological fibrosis (0% configuration) significantly degraded performance. The non-fibrotic model exhibited a positive bias of 8.09 ms (systematic overestimation), an MAE of 9.97 ms, and an RMSE of 19.74 ms. Consequently, the concordance rate dropped from 98.4% to 75.4%, with 15 compounds exceeding the 10-ms threshold compared to only 1 in the fibrotic model.
Robustness: The high concordance was maintained across different ion-channel pharmacology sources (Crumb vs. Kramer) and was not driven by any single therapeutic class, as confirmed by leave-one-class-out sensitivity analyses.

Significance and Claims
The paper claims that virtual Phase I trials can serve as a human-based digital approach for early QT-focused cardiac safety assessment. The primary significance lies in the framework's ability to reproduce clinical QT outcomes with sufficient quantitative accuracy and robustness to support translational decision-making without relying on animal models or clinical testing in healthy volunteers for every compound.

The authors emphasize that the inclusion of physiological interstitial fibrosis is a key determinant of predictive performance, correcting systematic overestimation of QT prolongation seen in homogeneous tissue models. The framework is positioned not as a replacement for clinical evidence but as a complementary digital layer that integrates human-calibrated electrophysiological variability, tissue-level propagation, and patient-variability-aware endpoints. This allows for the interpretation of drug-induced QT changes in the same conceptual space as regulatory guidelines (ICH E14), potentially aiding in compound prioritization and the safety assessment of toxic compounds where traditional Phase I trials are ethically constrained.

The study acknowledges limitations, including the use of a simplified ventricular wedge rather than a full heart, a static Hill-type block model, and the lack of explicit propagation of pharmacological uncertainty (e.g., in $IC_{50}$ values). However, the results suggest that a physiologically motivated, tissue-level approach can effectively bridge the gap between mechanistic cellular modeling and clinical safety endpoints.