Optimizing organic reaction conditions to obtain a high-yielding target product is crucial for overcoming technical issues in synthetic chemistry. Owing to the extremely expensive and time-consuming with chemical experiments, extensive efforts have been devoted toward accelerating and facilitating the process. With advancements in artificial intelligence, varying data-driven approaches have been applied to predict suitable chemical reaction conditions. But, in many case of novel synthetic experiments, it is inevitable to optimize the reaction reagents by repeating an experiments to getting high-yield products. Recently, Bayesian optimization (BO), an iterative global optimization algorithm, has demonstrated exceptional performance in finding suitable reagents compared to synthesis experts. However, Bayesian optimization needs randomly selected several initial experiments results (yields) to train yield prediction surrogate models (about 10 trials). This process is inefficient parts, such as cold-start problem in recommender systems. In this paper, we present an efficient Bayesian reaction optimization algorithm with Message Passing Neural Network (MPNN) trained by database extracted from vast experimental studies. We demonstrate the effectiveness of the proposed algorithm on benchmark datasets and in-house datasets.