Fig. 1: Overview of SMRTnet.
From: Predicting small molecule–RNA interactions without RNA tertiary structures
a, Schematic overview of SMRTnet, a deep learning method designed to predict SRIs. The method accepts an RNA sequence along with its secondary structure and a small molecule represented by its SMILES notation and outputs the binding score and potential binding site. b, Architecture of SMRTnet. The training dataset was derived from the PDB. SMRTnet comprises three main components: an RNA encoder, a small-molecule encoder, and an MDF module. The RNA encoder processes RNA sequences using an RNA language model (RNASwan-seq) and RNA structure using CNNs with ResNets. The small-molecule encoder processes small-molecule SMILES using a chemical language model (MoLFormer) and the chemical structure of small molecules using GATs. The MDF module uses attention-based neural networks to progressively integrate pairwise binding information, which is decoded by the fully connected neural network to predict the binding score for the input small molecule and RNA pair. Finally, an ensemble scoring strategy is applied to generate the final binding score.
