M-DeepAssembly2: A Web Server for Predicting Multiple Conformations of Multi-domain Proteins Using Deep Learning

The biological functions of proteins often depend on dynamic transitions between multiple conformational states, and for multi-domain proteins, inter-domain dynamics play a critical role in their biological functions. Despite significant advances in static structure prediction, accurately modeling these multiple conformations remains a major challenge. Here, we present M-DeepAssembly2, a web server designed for modeling multiple conformations of multi-domain proteins using deep learning-predicted inter-domain and intra-domain constraints. The server operates in three stages. Firstly, evolutionary, physicochemical, and geometric features are extracted from the input and fed into AlphaFlex, a flexible residue prediction network. This process generates multiple distance maps that effectively capture both inter-domain and intra-domain interactions. Secondly, supplementary distance maps that characterize inter-domain interactions are generated using the DeepAssembly method. Finally, all distance maps are integrated as constraints in a multi-objective protein optimization to generate multiple conformational states of the protein. Compared to its predecessor, M-DeepAssembly2 leverages a flexible residue prediction network that directionally decouples coevolutionary information to generate heterogeneous distance maps. Benchmark results demonstrate that this strategy enables accurate prediction of protein multiple conformations and significantly improves the precision of static multi-domain protein structure prediction.