Prediction of Optimal Drug Schedules for Controlling Autophagy

Afroza Shirin, Isaac Klickstein, Song Feng, Yen Ting Lin, William S. Hlavacek, Francesco Sorrentino

TL;DR

This paper applies control engineering principles to biology by developing a mathematical framework to determine the most efficient drug dosing schedules for regulating autophagy. By treating drug delivery as a "minimum fuel" optimal control problem, the researchers identified pulsatile and sequential dosing strategies that achieve therapeutic targets with significantly lower total drug amounts than standard approaches.

What Problem Does This Paper Address?

Autophagy—the process by which cells recycle their own components—is a critical survival mechanism that plays a "double-edged sword" role in cancer: it can either protect tumor cells from stress or be harnessed to induce cell death. However, the regulatory network governing autophagy (involving kinases like MTORC1, AMPK, and ULK1) is highly nonlinear and complex, making it nearly impossible to predict the effects of drug combinations using intuition alone.

Current therapeutic strategies often rely on constant dosing or trial-and-error combinations. The authors address the lack of a formal, model-driven approach to therapy design, specifically seeking to minimize drug toxicity (off-target effects) while maintaining precise control over cellular behavior.

Methods

The authors combined biological modeling with advanced numerical optimization:

  • Mathematical Modeling: They constructed a minimalist system of nonlinear Ordinary Differential Equations (ODEs) representing the fractions of active/inactive kinases (MTORC1, ULK1, AMPK, VPS34) and the count of autophagic vesicles (AVs).
  • Optimal Control Theory: They framed the dosing problem as a "minimum fuel problem," where the objective is to minimize the total integrated drug dose ($\int u(t) dt$) over a specific time window while keeping the AV count within a target range.
  • Pseudo-Spectral Optimal Control (PSOC): To solve the complex nonlinear equations, they used PSOC to discretize the continuous-time problem into a Nonlinear Programming (NLP) problem.
  • Software Stack: The optimization was performed using the open-source C++ toolbox PSOPT, with Ipopt serving as the interior-point solver.

Datasets

As a computational modeling study, the paper does not use a traditional "big data" training set. Instead, it utilizes:

  • Kinetic Parameters: Derived from established theoretical models of the AMPK-MTORC1-ULK1 pathway (e.g., Szymanska et al.).
  • Experimental Validation Data: The model's predictions were compared against measured autophagic vesicle dynamics from literature (Martin et al.) to ensure the ODEs accurately reflected biological reality.
  • Pharmacokinetic Data: Drug clearance rates ($\delta$) were set based on measured half-lives of representative small-molecule inhibitors (e.g., Rapamycin, Buparlisib) found in clinical and pharmacological databases.

Key Insights and Ideas

  • The Pulsatile Nature of Optimality: The most efficient way to control the network is not a constant infusion, but rather intermittent "pulses" of drug. This allows the drug concentration to decay naturally while the system remains within the target threshold, minimizing the total "fuel" consumed.
  • Synergy through Nonlinearity: Synergistic effects between drugs (like a VPS34 inhibitor paired with a PI3K inhibitor) emerge naturally from the network's nonlinear dynamics. In one case, a combination reduced the required dose of a primary drug by more than five-fold.
  • Sequential Logic: For certain drug pairs, the order of operations matters. The model suggests that in oscillatory systems (moderate stress), one drug must be applied first to "break" the oscillation before the second drug can effectively drive the system to the target state.
  • Target Control vs. Attractor Control: Unlike traditional control goals that drive a system to a natural "rest state" (attractor), this approach allows researchers to maintain the system at an arbitrary, non-attracting operating point indefinitely.

Key Results

  • Monotherapy: Identified optimal schedules for six classes of drugs, showing that drugs 1 and 5 upregulate autophagy, while 2, 3, 4, and 6 downregulate it.
  • Combination Efficiency: The combination of Drug 2 (MTORC1/VPS34 inhibitor) and Drug 6 (VPS34 inhibitor) was particularly potent. Under severe stress, Drug 6 alone could not reach the target, but when paired with Drug 2, it achieved the goal with a massive reduction in the total amount of Drug 2 needed.
  • Stabilization: The framework successfully designed schedules to stabilize systems that naturally exhibit oscillatory behavior (common in intermediate nutrient stress conditions), providing a steady "low-autophagy" state that single drugs often fail to maintain.

Limitations and Future Work

  • Non-Interacting Drug Assumption: The current model assumes drugs do not interact directly with each other (e.g., metabolic interference), only through their shared signaling network.
  • Cellular Heterogeneity: The model represents an "average" cell. Future work is needed to account for stochasticity and cell-to-cell variability within a tumor population.
  • Experimental Gap: While the model is grounded in literature data, the specific "optimal schedules" predicted (the exact timing and dosage of pulses) still require validation in live-cell experiments.
  • Resistance: The authors note that future iterations could incorporate the emergence of drug resistance as a constraint in the optimization process.

Summary updated

Abstract

The effects of molecularly targeted drug perturbations on cellular activities and fates are difficult to predict using intuition alone because of the complex behaviors of cellular regulatory networks. An approach to overcoming this problem is to develop mathematical models for predicting drug effects. Such an approach beckons for co-development of computational methods for extracting insights useful for guiding therapy selection and optimizing drug scheduling. Here, we present and evaluate a generalizable strategy for identifying drug dosing schedules that minimize the amount of drug needed to achieve sustained suppression or elevation of an important cellular activity/process, the recycling of cytoplasmic contents through (macro)autophagy. Therapeutic targeting of autophagy is currently being evaluated in diverse clinical trials but without the benefit of a control engineering perspective. Using a nonlinear ordinary differential equation (ODE) model that accounts for activating and inhibiting influences among protein and lipid kinases that regulate autophagy (MTORC1, ULK1, AMPK and VPS34) and methods guaranteed to find locally optimal control strategies, we find optimal drug dosing schedules (open-loop controllers) for each of six classes of drugs and drug pairs. Our approach is generalizable to designing monotherapy and multi therapy drug schedules that affect different cell signaling networks of interest.
BibTeX
@inproceedings{shirin2019prediction,
  author = {Afroza Shirin and Isaac Klickstein and Song Feng and Yen Ting Lin and William S. Hlavacek and Francesco Sorrentino},
  title = {{Prediction of Optimal Drug Schedules for Controlling Autophagy}},
  booktitle = {Scientific},
  year = {2019},
  doi = {10.1038/s41598-019-38763-9},
}
Figure