DiffDose: Differentiable Programming for Personalized Dose-Regimen Optimal Control

Dose-regimen design involves determining the quantity of drug administered, the timing of administration, and how treatment should fluctuate across different patients. Mechanistic pharmacokinetic-pharmacodynamic (PK/PD) and quantitative systems pharmacology (QSP) models can forecast treatment outcomes; however, optimizing dosing inputs relies on model-specific sensitivity calculations or derivative-free search techniques.

Researchers have now developed DiffDose, a differentiable programming framework for mechanistic open-loop dose-regimen optimization. This innovative approach employs automatic differentiation (AD) to treat clinically interpretable dose amounts and administration times as differentiable controls.

Three distinct scenarios were examined to evaluate the effectiveness of DiffDose. First, the method was applied to fixed-schedule dose-amplitude optimization within OptiDose PK/PD benchmarks. Second, DiffDose was utilized for dose-timing optimization in a chemotherapy-induced neutropenia model featuring state-dependent delay. Lastly, the framework was employed to create individualized dosing schedules for mosunetuzumab within a QSP virtual population.

The results across these diverse examples revealed that AD produced gradients that were consistent with references, minimizing the need for model-specific derivative calculations. Furthermore, DiffDose significantly reduced the time required to find the optimal solution compared to traditional methods. By transforming mechanistic PK/PD and QSP models into gradient-based engines for personalized regimen design, DiffDose has the potential to revolutionize the field of dose-regimen optimization.