RNADiffFold: Generative RNA Secondary Structure Prediction using Discrete Diffusion Models

Abstract

As a crucial class of macromolecules, RNA plays a vital role in various biological functions within living organisms. Accurately predicting the secondary structure of RNA contributes to a better understanding of its intricate three-dimensional structure and functionality. Previous energy-based and learning-based methods model RNA secondary structures in a static view and impose strong prior constraints. Inspired by the success of diffusion models, in this work, we propose a generative prediction method based on multinomial diffusion. We consider the prediction of contact maps as a pixel-level segmentation task and train the denoise model to iteratively refine contact maps from noise. Additionally, we design an effective condition to extract features from sequences, guiding the model to generate the corresponding secondary structure. These features include sequence one-hot encoding, probability maps from a pre-trained score network, as well as embeddings and attention maps from RNA-FM. Experimental results on both within- and cross-family datasets demonstrate RNADiffFold’s competitive performance compared with current state-of-the-art methods. Moreover, RNADiffFold moderately captures dynamic structural features of RNA, as validated on a multi-conformational dataset.