The molecule at the center of the study is CXCL9, a chemokine that acts as a beacon for effector CD8-positive T cells. Higher CXCL9 expression has been linked to better responses to immunotherapy across multiple cancer types, making it an attractive lever for converting cold tumors into inflamed ones. The problem is that the signaling circuitry governing CXCL9 production inside tumor cells is poorly understood. The two best-known inducers, the inflammatory cytokines interferon-gamma and TNF-alpha, activate the JAK-STAT and NF-kappaB pathways respectively, but these pathways crosstalk extensively with PI3K-AKT, MAPK and p53 signaling, all of which are frequently rewired in cancer. Untangling which of these interactions actually matter in a given tumor cell is a combinatorial nightmare.

The Eindhoven team’s solution was to build interpretable mechanistic models of the signaling network and then let an active learning algorithm decide which experiments to run next. They started by curating a prior knowledge network specific to CXCL9 regulation, drawing on literature and the DoRothEA database of transcription factors. The network spans five major pathways: JAK-STAT, NF-kappaB, PI3K-AKT, MAPK and p53, connected to upstream cytokines including IFN-gamma, TNF-alpha, IFN-alpha and EGF. This scaffold was converted into a set of logic-based ordinary differential equations, a formalism that turns qualitative wiring diagrams into continuous dynamical systems without requiring the detailed kinetic parameters that are usually unknown in cancer signaling. Each edge in the network carries an adjustable strength parameter, which makes the fitted models biologically interpretable rather than black boxes.

To train these models, the researchers worked with two pancreatic cancer cell lines, AsPC1 and BxPC3, chosen because reanalysis of the Genomics of Drug Sensitivity in Cancer database showed they respond very differently to drugs. They exposed the cells to the two cytokines alone and in combination, alongside five clinically relevant inhibitors targeting JAK, IKK, PI3K, MEK and RAS, and measured secreted CXCL9 protein using a bead-based immunoassay with flow cytometry readout. The results confirmed the central role of JAK-STAT signaling: the JAK inhibitor momelotinib strongly suppressed CXCL9 in both lines, while IFN-gamma drove robust induction. More intriguingly, the PI3K inhibitor taselisib and the MEK inhibitor trametinib boosted CXCL9 expression, especially when combined with dual cytokine stimulation, pointing to previously underappreciated regulatory routes.

The fitted models, ensembles of ten optimizations per cell line, reproduced the experimental data with striking accuracy, achieving Pearson correlations of 0.998 for AsPC1 and 0.995 for BxPC3. In silico knockout experiments, in which individual regulatory edges were systematically removed from the models, then revealed context-specific control points. Deleting the ERK-AR interaction reduced CXCL9 in BxPC3 but not AsPC1, while JAK-STAT1 and STAT1-CXCL9 knockouts affected only AsPC1. The NF-kappaB pathway emerged as the key mediator of synergy between IFN-gamma and TNF-alpha in both cell lines, whereas JAK-STAT interactions contributed to synergy specifically in BxPC3. Bootstrapped parameter comparisons quantified these differences, showing that eight pathway parameters were significantly stronger in BxPC3 while two, including IFNGR-JAK, were stronger in AsPC1, providing a mechanistic explanation for the cell lines’ divergent drug responses.

The truly novel element, however, was the active learning pipeline coupled directly to these mechanistic models. Active learning is well established in drug discovery, where it helps algorithms pick the most informative compounds to test next, but it had never before been integrated with mechanistic biological models of this kind. The workflow is elegantly cyclical: the model ensemble predicts CXCL9 responses for all untested perturbation conditions, an acquisition function selects a small batch of the most valuable candidates, those are measured in the wet lab, and the models are retrained on the expanded dataset. The researchers benchmarked four acquisition strategies on synthetic data: greedy sampling, which chases conditions predicted to produce the highest CXCL9; uncertainty sampling, which targets conditions where the model ensemble disagrees most; a hybrid of the two; and random selection as a baseline.

The benchmarking produced a clear and practically useful picture. Greedy and the hybrid strategy discovered 1.4 to 1.9 times more CXCL9-inducing conditions than random sampling after five rounds, but they also generated more false positives when too many conditions were added per round. Uncertainty sampling was less aggressive at finding hits but delivered the best model generalization, reaching a mean R-squared of 0.93 across all conditions, including unseen ones, significantly outperforming every other strategy. The choice of initial training set also mattered: a carefully hand-picked set of ten conditions yielded seventeen final hits on average compared with nine for the worst random set, though the pipeline proved capable of recovering from suboptimal starts. These findings offer concrete design guidance for anyone attempting similar iterative experiments under real resource constraints.

Crucially, the team then took the pipeline back into the laboratory, running two rounds of active learning with real measurements in both cell lines. The qualitative differences between strategies seen in silico reproduced experimentally. Greedy and hybrid selections produced the strongest CXCL9 induction, while uncertainty-guided choices explored a broader response range and most consistently shrank the model’s prediction uncertainty. One complication surfaced: some greedy-selected drug combinations, such as PI3K plus ERK inhibition, yielded lower CXCL9 than expected because the high cumulative drug concentration triggered apoptosis. A Caspase-3 assay confirmed strong negative correlations between cell death and chemokine secretion, and after correcting for apoptosis, the expected hierarchy of acquisition strategies re-emerged. This observation may also help explain conflicting reports in the literature linking CXCL9 to both favorable and unfavorable prognosis in pancreatic cancer, since cytotoxicity can mask genuine immunostimulatory effects.

The study’s broader significance lies in demonstrating that mechanistic modeling and active learning, usually pursued on separate tracks, can be fused into a data-efficient engine for biological discovery. The interpretable logic-ODE framework kept the experimental design grounded in prior biological knowledge, while the learning loop squeezed maximum information from minimal measurements. Among the most tantalizing findings were the frequent selections of AKT and p53 inhibitors by the hybrid strategy, both validated as CXCL9 inducers despite sitting outside the canonical JAK-STAT and NF-kappaB regulatory axes, suggesting that less-characterized signaling mechanisms may hold untapped potential for immunomodulation. The authors caution that predictions remain constrained by the structure of the prior knowledge network, and that future work could expand it with transcriptomic data, add multiplexed readouts such as PD-L1 or TGF-beta, and employ Bayesian parameter inference for better-calibrated uncertainty. But the proof of principle stands: rational, mechanism-driven design of combination therapies aimed at warming up cold tumors is no longer a distant aspiration, but an iterative workflow that a small lab can start running today.

Subject of Research: Active learning-guided mechanistic modeling of CXCL9 chemokine regulation in pancreatic cancer cells

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

Article References: Wang, B.-R., Yelleswarapu, M., Descamps, L., & Eduati, F. (2026). Active learning-guided mechanistic modeling reveals context-specific regulators of CXCL9 expression in pancreatic cancer cells. Molecular Systems Biology, 22(8), 1360-1375. https://doi.org/10.1038/s44320-026-00221-w

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00221-w

Keywords: pancreatic cancer, CXCL9, active learning, mechanistic modeling, logic-ODE, immunotherapy, JAK-STAT, NF-kappaB, chemokines, cold tumors, drug combinations, systems biology