Developing Deep Learning Models for RNA Inverse Folding

Develop a deep learning model for RNA inverse folding that (i) captures the geometric features of flexible RNA conformations, (ii) handles the non-unique mappings between RNA structures and sequences, and (iii) provides alternative sequence design options for different preferences, thereby addressing the stated requirements that make this task remain open.

Background

RNA inverse folding seeks sequences that fold into predefined structures, and most prior computational approaches have focused on secondary structures, which provide limited information for achieving functional tertiary constraints. Designing sequences directly from 3D structures is challenging due to data scarcity, non-unique structure-to-sequence mappings, and RNA conformational flexibility.

The paper highlights that despite deep learning’s potential to narrow the vast sequence space, an appropriate deep learning model for RNA inverse folding remains to be developed, specifically one that can capture complex geometric features, manage multiple valid sequence solutions for the same structure, and support design preferences with diverse candidates.