Protein folding path prediction and energy optimization driven by deep reinforcement learning

Protein folding is the key process to determine its biological activity. Traditional experimental methods are time-consuming and laborious, while the method based on deep reinforcement learning (DRL) can effectively improve the prediction efficiency and accuracy. In this study, a DRL model based on actor critical framework was constructed. The amino acid sequence was used as the input, and the features were extracted by convolutional neural network (CNN) and bidirectional long-term and short-term memory network (BiLSTM), and the intermediate conformation and energy value in the process of protein folding were output. The model encourages the generation of low-energy stable folding paths by designing reasonable reward function and state transition function, and optimizes conformational stability by using energy constraint reward function. The experiment uses 1200 single-domain protein of CATH 4.2 database, combined with the prediction structure of AlphaFold DB as the reference. The results show that the topological similarity (TSI) of this model is 0.83, which is 10.7% higher than that of AlphaFold2, the path convergence rate (PCR) is 0.91, and the trajectory energy fluctuation is reduced by 42%, and the calculation efficiency is significantly better than the traditional method. The ablation experiment verifies the important role of each component of the reward function on the model performance. This study proves the effectiveness of DRL in protein folding path prediction and energy optimization.