Bastion3 is a two-layer approach and predictor for identifying type III secreted effectors (T3SEs) using ensemble learning. By employing a wide range of sequence-derived features, Bastion3 trained models using a powerful gradient boosting machine, namely, LightGBM, and further boosted the models' performances through a novel genetic algorithm (GA)-based two-step parameter optimization strategy. These models were finally integrated into a two-layer architecture based on ensemble learning, which enabled Bastion3 server to achieve a stable and accurate prediction performance for identifying T3SEs.