E2EDA: Protein domain assembly based on end-to-end deep learning

Abstract

With the development of deep learning, almost all single-domain proteins can be predicted at experimental resolution. However, the structure prediction of multi-domain proteins remains a challenge. Achieving end-to-end protein domain assembly and further improving the accuracy of the full-chain modelling by accurately predicting inter-domain orientation while improving the assembly efficiency will provide significant insights into structure-based drug discovery. In addition, the available GPU memory limits the size of a full-chain protein which can be predicted. Based on the divide-and-conquer strategy, the single-domain structure is predicted by the state-of-the-art prediction method, such as AlphaFold2, and then assembled into a full-chain model through the domain assembly method, which can effectively reduce the demand for hardware resources. In this work, we propose an End-To-End Domain Assembly method based on deep learning, named E2EDA. We first develop an EffificientNetV2-based deep learning model (RMNet), which is specialised for predicting inter-domain orientations. The RMNet uses an attention mechanism to predict inter-domain rigid motion by fusing sequence features, multiple template features and single-domain features. Then, the predicted rigid motions are converted into inter-domain spatial transformations to assemble full-chain models of multi-domain proteins directly without time-consuming simulation processes. Finally, a scoring strategy, RMscore, is designed to select the best model from multiple assembled models to improve assembly accuracy. The experimental results show that the average TM-score of the model assembled by E2EDA on the benchmark set (356) is 0.84, which is better than other domain assembly methods SADA (0.80), DEMO (0.74) and AIDA (0.63). Meanwhile, on our constructed human protein dataset from AlphaFold DB, the model reassembled by E2EDA is 6.8% higher than the full-chain model predicted by AlphaFold2, indicating that E2EDA can capture more accurate inter-domain orientations to improve the quality of the model predicted by AlphaFold2. Furthermore, the average running time of E2EDA on the benchmark is reduced by 74.6% compared with the domain assembly simulation method SADA, which indicates that E2EDA can effectively improve assembly efficiency through an end-to-end manner.The online server is at http://zhanglab-bioinf.com/E2EDA/ .