Go to ICLR 2026 Conference homepageSIDDesign: Sequence-Informed Distillation for Tertiary Structure-Based RNA Design
Keywords: Tertiary Structure-Based RNA Design, Cross-Modal Distillation, Stochastic Topology Regularization
TL;DR: We present SIDDesign, a distillation framework that transfers sequence-level priors from RNA foundation models to enhance tertiary structure-based RNA design through contextual refinement and stochastic topology regularization.
Abstract: Tertiary structure-based RNA design, which aims to design nucleotide sequences that fold into a given 3D structure, is a fundamental challenge in synthetic biology and structure-guided design. Although recent work has advanced geometric encoders and architectural innovations, progress remains constrained by the scarcity and bias of resolved RNA tertiary structures. Considering that RNA folding and design are approximate inverse tasks, and that RNA foundation models employed in folding models capture rich RNA priors from large-scale sequence data, we are inspired to exploit these representation-level priors to enhance the RNA design task. Motivated by this perspective, we propose the Sequence-Informed Distillation framework for structure-based RNA Design (SIDDesign), which aligns structure-derived embeddings with sequence-level representations obtained from RNA foundation models. To bridge modality and contextual gaps between sequence and structure representations, we design a Similarity-aware Contextual Refinement (SimACR) module based on cross-attention. To further mitigate edge noise introduced during graph construction, we introduce a Stochastic Topology Regularization (STR) strategy during training that improves the robustness of message passing. Extensive experiments on benchmark datasets demonstrate the effectiveness of SIDDesign, with consistent and significant improvements over existing approaches.
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
Submission Number: 1845
Loading