Predicting trophic links across freshwater fish food webs from species traits

Body size is the principal trait governing trophic interactions in freshwater fish communities, yet the body size data available for most species are a species-level maximum. Whether this global proxy can predict trophic links in systems the model has never seen is untested. We evaluated four models on 37 freshwater fish food webs under leave-one-study-out cross-validation (LOSO-CV), predicting each held-out network from species traits with no observed interactions. Locally measured body mass ratios predicted links accurately (median ROC-AUC 0.973). FishBase traits predicted fish-associated links nearly as well (0.887 against 0.912 on the same networks). Whole-network scores were lower (0.607); the shortfall was confined to pairs between non-fish taxa, which FishBase does not describe. A habitat overlap score from FishBase depth-stratum categories added no signal, and a graph attention network did not outperform the random forest. For held-out systems within the range of the training corpus, fish-associated links are recovered from globally available traits alone. This did not extend to the three out-of-region networks, where prediction dropped to near or below chance. Cross-region transfer will need a larger and more geographically varied set of fish food webs.