Active learning-guided mechanistic modeling reveals context-specific regulators of CXCL9 expression in pancreatic cancer cells.

Cold tumors like pancreatic cancer suffer from poor immune infiltration, limiting effective anti-tumor responses. The chemokine CXCL9 promotes immune cell recruitment, but the signaling mechanisms regulating its expression in tumor cells remain poorly understood and underexplored as targets for modulation. We present a framework that integrates active learning with mechanistic logic-ODE models to guide perturbation screenings and uncover regulators of CXCL9 in pancreatic cancer cells. Using perturbation-response data and curated prior knowledge, we trained interpretable models to identify signaling mechanisms that enhance CXCL9 expression and prioritize drug combinations. Active learning enabled data-efficient model refinement and guided informative experiments under resource constraints. Benchmarking on synthetic data and experimental validation confirmed the performance of different acquisition strategies and its applicability to feasible iterative wet lab experiments. Our results demonstrate how combining active learning with mechanistic modeling supports rational, targeted experimental design.