Multi-objective reinforcement learning for optimizing JAK/STAT pathway interventions: a quantitative systems pharmacology study for preclinical drug development

Treatment optimization targeting JAK/STAT inhibition remains challenging due to pathway complexity and competing clinical objectives. We developed a computational framework integrating mechanistic modeling with multi-objective reinforcement learning to identify potential intervention strategies for preclinical development. Our quantitative systems pharmacology model incorporated signaling dynamics, tumor-immune interactions, resistance evolution, and drug pharmacokinetics/pharmacodynamics. Five reinforcement learning agents with different objective weight configurations (tumor reduction, immune preservation, resistance prevention, and toxicity minimization) revealed pronounced, model-predicted trade-offs under the present assumptions and parameterization, suggesting that simultaneous optimization across all four objectives may be difficult to achieve in the modeled setting. Balanced configurations appeared to fall short of a true compromise, with mean scores of [0.299, 0.227, 0.201, and 0.007], respectively. The resistance-focused configuration achieved strong performance in resistance prevention (mean score 0.949), but with limited tumor reduction (0.097), moderate immune preservation (0.259), and reasonable toxicity minimization (0.254). We also explored a composite configuration aimed at maximizing tumor reduction and immune preservation while minimizing resistance development and toxicity, resulting in scores of [0.282, 0.244, 0.214, 0.024]. Sensitivity analysis using three complementary approaches identified SOCS3 inhibition, STAT phosphorylation, and resistance parameters as potentially important determinants of treatment outcomes. This framework illustrates distinct treatment paradigms for targeting JAK/STAT and offers valuable guidance before resource-intensive animal studies in preclinical development.