Synthetic RNA Evolution Enables Accurate Alignment-Free 3D Structure Prediction

Accurate RNA 3D structure prediction remains a bottleneck in computational biology. Although protein structures can now be predicted with near-experimental fidelity, RNA 3D prediction still lags behind. One reason is that state-of-the-art tools such as AlphaFold 3 require deep multiple sequence alignments (MSAs) that are considerably more difficult to obtain for RNA than for proteins. Here, we remove this bottleneck by synthesizing homologous RNAs. Starting from a single RNA sequence, our novel deep learning model RNAformer predicts secondary structure with high fidelity, which we use to generate structurally consistent synthetic homologs through lightweight, evolution-inspired mutation rules. This process produces deep, MSA-like sequence ensembles in seconds, without reliance on natural sequence databases while bypassing alignment steps. When supplied to AlphaFold 3, synthetic homologs substantially improve local RNA structural accuracy and rescue predictions for orphan RNAs where no natural alignment exists. Secondary structure-driven synthetic evolution, therefore, unlocks deep alignment benefits for RNA 3D structure prediction.