Dual-atom catalysts (DACs) offer a powerful platform for modulating oxygen electrocatalysis, yet rational design is limited by lack of transferable, mechanism-level principles. While machine learning (ML) is increasingly used for catalyst screening, its potential for mechanistic discovery is underexploited. Here, we elevate interpretable ML from an accelerator to a mechanism-discovery engine exposing the electronic rules underlying DACs. Using extended phthalocyanines (M1M2-ePc), we establish an integrated DFT-ML-experiment framework that maps catalytic performance onto a low-dimensional, interpretable electronic landscape. Screening 81 DFT-computed and 360 ML-predicted metal pairs identifies FeM-ePc as a promising bifunctional catalyst family. Crucially, SHapley Additive exPlanations (SHAP) analysis reveals an intrinsic role asymmetry: Fe serves as a stable adsorption anchor, while the secondary metal (M2) tunes the activity window via d-electron filling and charge affinity. First-principles analysis further uncovers a spin-dependent dual-descriptor mechanism, in which d-band center and charge transfer act as co-descriptors governing bifunctional activity. Guided by this principle, atomically dispersed FeCo-ePc-L is synthesized with a Fe-Co separation consistent with theory and delivers a bifunctional gap (ΔE = 0.60 V), outperforming Pt/C + RuO2 and enabling stable cycling in quasi-solid-state Zn-air batteries. This work establishes a transferable, ML-derived paradigm for programming electronic cooperation in multi-metal catalysts.