Аннотация:In biology, it remains challenging to predict interactions between proteins and DNA or RNA. When it comes to nucleic acids, existing methods of binding site identification or interaction prediction are inefficient, especially in minor cases, such as aptamer binding. In order to predict NA-protein interactions, we use a deep-learning framework called dMaSIF. Therefore, we modified the atom encoding module to reflect atom positions and relationships more precisely and used parallel calculation to optimize training process. The framework showed effectiveness on two tasks: identifying NA binding sites and predicting NA-protein interactions. This approach can thereby be used to find potential NA binding sites, to perform NA-protein docking and virtual screening, etc.