The efficient and effective design of chemical processes and products heavily relies on the accurate prediction of essential properties. In this work, a deep-learning architecture integrating a bidirectional long short-term memory (Bi-LSTM) network, an attention mechanism, and a back-propagation neural network (BPNN) is developed for the prediction of energy conversion efficiency of organic solar cells. Inspired by the success of artificial intelligence in natural language processing, we first developed a novel strategy for molecular signature encoding and information embedding in order to depict the compositional structures of molecules. Then, an advanced recurrent neural network named Bi-LSTM is employed to process the molecular information, while the BPNN is applied to correlate energy conversion efficiency. During this procedure, the attention mechanism is used to identify molecular constituents that are important to the property of interest. To evaluate the performance of the proposed approach, the energy conversion efficiencies of more than 20,000 organic photovoltaics are used to train and test the model. Result comparisons with several other modeling approaches indicate that the proposed method is competitive in prediction accuracy and possesses good transferability to small data sets. Additionally, the proposed method is capable of identifying decisive molecular constituents, providing instructive information for the reverse design of organic solar cells.