3dDNAi: An Integrated Approach for 3D Structure Prediction of Single-Stranded DNAs

Machine-learning approaches are now widely used to predict the 3D structures of proteins and RNA molecules. However, prediction accuracy for RNAs is much lower than for proteins due to the scarcity of experimental structural data and homologous sequences. This situation is even worse for single-stranded DNA (ssDNA) molecules, such as DNA aptamers, which have diverse clinical and biotechnological applications. To address this problem, we integrated our physics-based method 3dDNA with an RNA language model and a deep-learning RNA structure prediction model to build 3D structures of DNA aptamers. This integrated approach largely overcomes the data scarcity problem for ssDNA molecules. The resulting method, 3dDNAi, improves global and backbone-level structural agreement for ssDNAs in the evaluated benchmarks, while the comparison with AlphaFold3 is metric-dependent. The framework established by 3dDNAi may also help address structure prediction challenges in other biomolecular systems where experimental data is limited.