Generative adversarial networks for construction of virtual populations of mechanistic models: simulations to study Omecamtiv Mecarbil action

Jaimit Parikh, Timothy Rumbell, Xenia Butova, Tatiana Myachina, Jorge Corral Acero, Svyatoslav Khamzin, Olga Solovyova, James Kozloski, Anastasia Khokhlova, Viatcheslav GurevJournal of Pharmacokinetics and Pharmacodynamics 2021 · DOI

Abstract

Biophysical models are increasingly used to gain mechanistic insights by fitting and reproducing experimental and clinical data. However, inherent variability in recorded datasets presents a significant challenge for parameter inference. This study introduces a novel approach integrating mechanistic modeling and machine learning to analyze in vitro cardiac mechanics data. We designed a specialized generative adversarial network (GAN) to construct virtual populations of cardiac ventricular myocyte models, specifically to study the action of Omecamtiv Mecarbil (OM), a positive cardiac inotrope. The GAN effectively inferred model parameters while incorporating prior information about OM's targets. Our results suggest a novel mechanism for OM involving modified interactions between myosin and tropomyosin. Validation against independent experimental protocols confirmed the model's predictive power, demonstrating its utility in exploring variability in complex biological systems.

Addressing the Inverse Problem in Cardiac Modeling

In biological research, cell-to-cell variability makes "one-size-fits-all" mechanistic models inadequate. While these models are excellent for understanding physical and chemical processes, fitting them to variable experimental datasets—the "inverse problem"—is computationally demanding. Determining the underlying physical parameters (e.g., protein stiffness or reaction rates) for each cell in a population typically requires expensive methods like Markov Chain Monte Carlo (MCMC).

Furthermore, traditional methods struggle with complex constraints, such as ensuring a drug affects only specific target parameters while leaving others constant across experimental groups. This study addresses these challenges by using machine learning to automate the creation of "virtual populations" that statistically match real-world data distributions.

Methodology: Mechanistically Grounded GANs

The researchers developed a specialized GAN architecture to solve this stochastic inverse problem. Unlike standard GANs used for image generation, this architecture is grounded in biophysical reality through three key components:

  1. Multi-Generator Architecture: The system employs three generators ($G_1, G_2, G_3$). $G_1$ learns parameters that remain constant across all cells (e.g., baseline stiffness), while $G_2$ and $G_3$ learn how specific drug-target parameters vary between control and treated groups.
  2. Surrogate Modeling: Mechanistic models often rely on slow-to-solve Ordinary Differential Equations (ODEs). To enable efficient training, the authors trained a neural network "surrogate" to mimic the ODE model, providing the speed and differentiability required for gradient-based optimization.
  3. Discriminator Feedback: Multiple discriminators ($D_1$ through $D_4$) ensure that both the inferred parameters and the resulting model outputs match the distributions observed in experimental data.
The GAN architecture uses multiple generators and discriminators to incorporate prior biophysical knowledge into the parameter inference process.

Data and Feature Extraction

The study utilized experimental recordings of unloaded sarcomere shortening from isolated rat ventricular myocytes. The dataset included a control group (24 cells) and an OM-treated group (17 cells treated with $1 \mu M$ of Omecamtiv Mecarbil). From each contraction trace, three key biomarkers were extracted: diastolic length (dSL), systolic length (sSL), and time to peak shortening (TTP).

Feature extraction from experimental traces: dSL (resting length), sSL (peak contraction length), and TTP (time to peak).

Insights into Omecamtiv Mecarbil Action

By splitting the generators, the model was forced to respect biological constraints—specifically, that a drug should only alter protein kinetics rather than a cell's fundamental physical structure. This "factorized learning" allowed the GAN to map experimental variability back to the parameter space, revealing new insights into OM's mechanism.

The inverse mapping suggested that OM does more than just increase calcium sensitivity; it likely modifies the cooperative interactions between myosin and tropomyosin proteins. A significant validation occurred when parameters inferred from simple shortening experiments were used to simulate "skinned" cell force-calcium tests. The model successfully predicted known OM effects—increased calcium sensitivity and a decreased Hill coefficient—without having been trained on that specific data.

Key Results

  • Population Matching: The GAN-generated virtual populations accurately reproduced experimental distributions. OM was shown to increase the mean TTP by approximately 50 ms and decrease the resting diastolic length by 0.2 $\mu m$.
  • Parameter Shifts: The model identified that OM significantly decreases the Hill coefficient ($h$) of the force-calcium curve (from 8.15 to 4.99) and increases calcium sensitivity (decreasing $Ca_{50}$ from 2.06 $\mu M$ to 1.25 $\mu M$).
  • Mechanistic Prediction: Simulations indicated that OM prolongs the "strong-bound" state of myosin cross-bridges, explaining the prolonged contraction (increased TTP) observed under the drug's influence.

Conclusion and Future Directions

This approach demonstrates the power of combining mechanistic models with generative machine learning. While successful, the method currently relies on the availability of a differentiable surrogate model. Future work could involve moving toward more complex, stochastic models of thin-filament activation to provide even more granular insights into drug actions. Additionally, larger experimental datasets would allow for the exploration of higher-dimensional parameter spaces, further refining our understanding of biological variability.

Summary updated

Abstract

Biophysical models are increasingly used to gain mechanistic insights by fitting and reproducing experimental and clinical data. The inherent variability in the recorded datasets, however, presents a key challenge. In this study, we present a novel approach, which integrates mechanistic modeling and machine learning to analyze in vitro cardiac mechanics data and solve the inverse problem of model parameter inference. We designed a novel generative adversarial network (GAN) and employed it to construct virtual populations of cardiac ventricular myocyte models in order to study the action of Omecamtiv Mecarbil (OM), a positive cardiac inotrope. Populations of models were calibrated from mechanically unloaded myocyte shortening recordings obtained in experiments on rat myocytes in the presence and absence of OM. The GAN was able to infer model parameters while incorporating prior information about which model parameters OM targets. The generated populations of models reproduced variations in myocyte contraction recorded during in vitro experiments and provided improved understanding of OM's mechanism of action. Inverse mapping of the experimental data using our approach suggests a novel action of OM, whereby it modifies interactions between myosin and tropomyosin proteins. To validate our approach, the inferred model parameters were used to replicate other in vitro experimental protocols, such as skinned preparations demonstrating an increase in calcium sensitivity and a decrease in the Hill coefficient of the force-calcium (F-Ca) curve under OM action. Our approach thereby facilitated the identification of the mechanistic underpinnings of experimental observations and the exploration of different hypotheses regarding variability in this complex biological system.
BibTeX
@article{parikh2021generative,
  author = {Jaimit Parikh and Timothy Rumbell and Xenia Butova and Tatiana Myachina and Jorge Corral Acero and Svyatoslav Khamzin and Olga Solovyova and James Kozloski and Anastasia Khokhlova and Viatcheslav Gurev},
  title = {{Generative adversarial networks for construction of virtual populations of mechanistic models: simulations to study Omecamtiv Mecarbil action}},
  journal = {Journal of Pharmacokinetics and Pharmacodynamics},
  year = {2021},
  doi = {10.1007/s10928-021-09787-4},
}
Figure