A PBN-RL-XAI Framework for Discovering a "Hit-and-Run" Therapeutic Strategy in Melanoma

Innate resistance to anti-PD-1 immunotherapy remains a major clinical challenge in metastatic melanoma, with the underlying molecular networks being poorly understood. To address this, we constructed a dynamic Probabilistic Boolean Network model using transcriptomic data from patient tumor biopsies to elucidate the regulatory logic governing therapy response. We then employed a reinforcement learning agent to systematically discover optimal, multi-step therapeutic interventions and used explainable artificial intelligence to mechanistically interpret the agent’s control policy. The analysis revealed that a precisely timed, 4-step temporary inhibition of the lysyl oxidase like 2 protein (LOXL2) was the most effective strategy. Our explainable analysis showed that this ‘‘hit-and-run" intervention is sufficient to erase the molecular signature driving resistance, allowing the network to self-correct without requiring sustained intervention. This study presents a novel, time-dependent therapeutic hypothesis for overcoming immunotherapy resistance and provides a powerful computational framework for identifying non-obvious intervention protocols in complex biological systems.

💡 Research Summary

This paper addresses the pressing clinical problem of innate resistance to anti‑PD‑1 checkpoint blockade in metastatic melanoma by integrating three cutting‑edge computational approaches: probabilistic Boolean network (PBN) modeling of patient‑derived transcriptomic data, reinforcement learning (RL) for dynamic therapy design, and explainable artificial intelligence (XAI) for mechanistic interpretation.

First, the authors re‑analyzed the publicly available RNA‑seq dataset (GSE78220) from Hugo et al., selecting 28 pretreatment tumor biopsies and dividing them into responders (n = 15) and non‑responders (n = 13). After variance‑stabilizing transformation with DESeq2, each gene’s expression distribution was modeled with a Gaussian mixture, and a data‑driven threshold was applied to binarize the data into OFF/ON states. A core network of 12 genes was constructed, comprising five IPRES signature genes (AXL, ROR2, WNT5A, LOXL2, TAGLN) and seven additional genes chosen by a composite score that combined differential‑expression significance and dynGENIE3‑derived network centrality.

For each cohort, a PBN was inferred. Candidate regulators for each target were limited to the top four dynGENIE3 rankings, and up to four Boolean predictor functions per gene were evaluated. Function probabilities were assigned proportionally to their mutual‑information scores, yielding two context‑specific PBNs with a state space of 2¹² = 4096. Attractor analysis showed that the responder network is highly plastic, featuring 22 distinct attractors with no single basin dominating, whereas the non‑responder network collapses into a rigid landscape dominated by one attractor that captures roughly 50 % of the probability mass.

Network‑level comparison revealed a dramatic rewiring: the transcription factor JUN becomes a master regulator in the resistant model, and LOXL2 emerges as its primary effector. In the resistant PBN, JUN’s update rule simplifies to an AND gate (JUN = MAP2K3 ∧ NRAS ∧ RELA ∧ LOXL2), while LOXL2’s rule collapses to LOXL2 = JUN ∧ MAP2K3. This hierarchical canalization creates a hard‑wired JUN/LOXL2 axis that locks the system into the resistant phenotype.

To discover therapeutic interventions capable of destabilizing this axis, the authors framed the control problem as a Markov decision process using the gym‑PBN environment. The action space included “Do Nothing” and single‑gene flips; the reward function granted +100 for reaching a sensitive attractor, –5 for remaining in a resistant attractor, and zero cost for actions, encouraging active but parsimonious interventions within a 15‑step horizon. Proximal Policy Optimization (PPO) was trained with standard hyperparameters (γ = 0.99, λ = 0.95, ε = 0.2).

The RL agent identified a striking “hit‑and‑run” protocol: a transient four‑step inhibition of LOXL2 (forcing LOXL2 = 0 for four consecutive time steps) followed by a policy that simply does nothing. This protocol achieved a 93.45 % success rate in silico, outperforming longer or shorter inhibition windows and also surpassing analogous JUN or MAPK3 priming strategies. The non‑monotonic relationship between inhibition duration and success underscores the importance of precise timing; sustained inhibition actually reduces efficacy, likely because the network adapts to continuous pressure.

To make the learned policy biologically interpretable, SHAP (Shapley Additive Explanations) values were computed for each decision step. SHAP analysis showed that the four‑step LOXL2 knock‑down dramatically reduces the contribution of the JUN/LOXL2 axis to the agent’s action scores, after which the agent’s highest‑value action becomes “Do Nothing”. Trajectory visualizations confirm that once the resistant attractor is destabilized, the network’s intrinsic dynamics naturally flow toward sensitive states without further external manipulation.

In summary, the study delivers three major contributions: (1) a patient‑specific PBN framework that quantifies how network rewiring creates a rigid, therapy‑resistant attractor; (2) a reinforcement‑learning pipeline that discovers optimal, time‑dependent, multi‑step interventions that would be infeasible to test experimentally at scale; and (3) an XAI layer that translates the black‑box policy into mechanistic insight, revealing that brief LOXL2 inhibition is sufficient to “prime” the system for self‑correction. The authors acknowledge that experimental validation is required, but the proposed PBN‑RL‑XAI pipeline represents a powerful, generalizable tool for uncovering non‑obvious therapeutic strategies in complex diseases.