RNADiffFold: generative RNA secondary structure prediction using discrete diffusion models.

Area of Science:

• Molecular Biology
• Computational Biology
• Bioinformatics

Background:

• Ribonucleic acid (RNA) molecules are crucial macromolecules with diverse biological functions.
• Accurate prediction of RNA secondary structures is vital for understanding their 3D architecture and function.
• Existing methods often provide static predictions and require strict prior constraints.

Purpose of the Study:

• To introduce RNADiffFold, a novel generative approach for RNA secondary structure prediction.
• To leverage multinomial diffusion models for refining RNA contact maps.
• To develop a conditioning mechanism for accurate structure generation using sequence features.

Main Methods:

• Reconceptualizing contact map prediction as pixel-wise segmentation.
• Training a denoising model to progressively refine contact maps from a noisy state.
• Employing a conditioning mechanism using one-hot encoded sequences, probabilistic maps, and RNA foundation model embeddings.

Main Results:

• RNADiffFold demonstrates competitive performance against state-of-the-art methods on within- and cross-family datasets.
• The model shows proficiency in capturing dynamic aspects of RNA structures.
• Effective performance was observed on datasets containing multiple RNA conformations.

Conclusions:

• RNADiffFold offers an innovative generative approach for RNA secondary structure prediction.
• The method successfully integrates sequence information to guide structure generation.
• RNADiffFold advances the field by providing insights into the dynamic nature of RNA structures.