Data-driven optimization of diet formulation to enhance survival and growth in Japanese Eel (Anguilla japonica) larvae

Abstract

Developing an optimal larval diet for Japanese eel ( Anguilla japonica ) remains a major bottleneck in artificial seedling production due to the limited efficacy of conventional heuristic formulation methods. We implemented a data-driven, human-in-the-loop optimization framework based on Bayesian optimization (BO) and Gaussian process regression (GPR) to systematically refine shark egg yolk-free diets. Across eight sequential feeding trials, the ingredient ratios of seven key components were iteratively adjusted, guided by empirical expert insight. Survival rate and total length (TL) were measured from 6 to 100 days post-hatch (dph). Multi-objective optimization targeted simultaneous improvement in survival and growth during the early (6–40 dph) and late (41–80 dph) periods. The optimized diet formulations achieved survival rates up to 74.9% versus 30.7% in controls and mean TL of 40.1 mm versus 35.2 mm ( p  < 0.05) at 100 dph. These results were consistently reproduced in independent validation trials. Mechanistic modeling via Random Forest and SHapley Additive exPlanations (SHAP) revealed that stage‑specific balancing of yeast extract-derived nutrients and other macronutrient components were critical drivers of improved survival and growth. To our knowledge, this is the first reported application of BO to larval feed optimization in Japanese eel. Our framework integrates probabilistic modeling with expert input to reduce experimental burden and accelerate development of superior larval diet formulations. This approach offers a practical, broadly applicable template for feed innovation across aquaculture species.