Abstract
<title>Abstract</title> <p>Marine food webs structure the flow of energy, nutrients, and biomass through the ocean, yet our understanding of their global organization is severely limited by the scarcity and geographically bias of documented species interactions. To address this gap, we constructed a global probabilistic metaweb for 2,454 marine fish species using a positive-unlabeled machine-learning framework. By integrating traits, taxonomic hierarchies, and environmental niches with 10,549 verified interactions, we identified 50,725 novel, high-probability feeding links, raising global connectance from an empirical baseline of 0.18% to 0.98%. We demonstrate that these missing interactions are highly structured, with the largest gaps occurring in hyper-diverse tropical realms, including the Central Indo-Pacific, the High Seas, and upper pelagic zones between 50 and 500 m. Structurally, the undocumented component of the global web is overwhelmingly wider rather than taller. Predicted local connectance and diet breadth increase sharply, whereas mean trophic level and food-chain length remain largely unchanged. This lateral expansion shifts our view of marine trophic architecture from sparse and fragmented to densely connected and trophically redundant, implying empirical records substantially underestimate ecosystem resilience. Ultimately, this baseline provides a crucial trophic prior for modelling extinction risks and forecasting food-web rewiring under climate-driven range shifts.</p>