Turbulent drag is a dominant contributor to energy losses in transportation and industrial flows, accounting for a substantial fraction of global energy consumption. Despite decades of progress, controlling turbulence remains a fundamental challenge due to its high dimensionality, nonlinearity, and the incomplete understanding of the mechanisms sustaining turbulent dynamics. Recent advances in deep reinforcement learning (DRL) have demonstrated the potential to discover effective flow-control strategies directly from data, yet most approaches remain black box in nature and provide limited physical insight. This work introduces a novel framework that integrates explainable deep learning (XDL) with deep reinforcement learning to develop physically informed, energy-efficient turbulence-control strategies. The central idea is to move beyond traditional reward formulations (such as direct drag minimization or suppression of classically defined coherent structures) and instead use explainability tools to identify the flow regions that are causally most influential for the future evolution of turbulence. These regions are identified using Shapley additive explanations (SHAP) computed from a deep neural network trained to predict future velocity fluctuations in wall-bounded turbulence. The proposed approach is demonstrated in turbulent open-channel flow at a friction Reynolds number of 180. A DRL agent controls the flow via wall-based blowing and suction, using only limited near-wall velocity measurements. Four reward strategies are systematically compared: direct drag reduction, suppression of Q-events, suppression of streaks, and suppression of SHAP-identified high-importance regions. While conventional rewards targeting Qevents and streaks effectively reduce their respective structures, they do not optimally reduce drag. In contrast, the SHAP-based reward leads to the highest drag reduction (exceeding even the policy trained explicitly for drag minimization), while requiring only about half of the control power input. These results reveal that the structures most responsible for sustaining turbulence and drag are not fully captured by classical coherent-structure definitions. The SHAP-based strategy disrupts the self-sustaining process of near-wall turbulence in a more direct and energy-efficient manner. Importantly, the control policies are trained in a minimal flow unit and successfully generalized to larger domains with multiple interacting turbulence cycles