Computational prediction of aquatic toxicity provides a rapid and cost-efficient alternative to resource-intensive in vivo ecotoxicological testing. However, most existing models adopt a fragmented modeling strategy, developing separate models for each organism group and toxicity end point. Such task-specific models restrict knowledge transfer across related biological contexts and fail to account for the biological ordering between multiple effect concentrations (EC50/EC10). Here, we propose a unified multitask framework to simultaneously predict effect concentrations for three aquatic organism groups (fish, crustaceans, algae) and two end points (EC50/EC10) within a single model. This framework integrates a molecular encoder with a Multigate Mixture-of-Experts (MMoE) network, a soft end point-ordering term that encourages the biological relation EC50 ≥ EC10, and a heteroscedastic regression output that provides per-prediction uncertainty estimates. We evaluated a multirepresentation fusion encoder combining physicochemical descriptors, molecular fingerprints, and a 2D molecular graph neural network (GNN), together with 3D-pretrained Uni-Mol and TRIDENT baselines. Under a common global scaffold/acyclic-cluster 5-fold pooled out-of-fold (OOF) protocol, the multirepresentation fusion model demonstrated robust predictive performance across all six organism-end point tasks, with lower prediction errors than TRIDENT across all six tasks and Uni-Mol in four tasks. Attention-based interpretation further highlighted chemically meaningful atoms and substructures that were consistent with plausible toxicity-relevant structural motifs. Together, these findings demonstrate that integrating multitask learning, biologically informed end point constraints, and complementary molecular representations provides a promising foundation for broadly applicable and uncertainty-aware models of aquatic toxicity.