Abstract
AbstractDNA folding thermodynamics are central to many biological processes and biotechnological applications involving base-pairing. Current methods for predicting stability from DNA sequence use nearest-neighbor (NN) models that struggle to accurately capture the diverse sequence-dependency of elements other than Watson-Crick base pairs, likely due to insufficient experimental data. We introduce a massively parallel method–Array Melt–that uses fluorescence-based quenching signals to measure equilibrium stability of millions of DNA hairpins simultaneously on a repurposed Illumina sequencing flow cell. By leveraging this dataset of 27,732 sequences with two-state melting behavior, we derived a refined NUPACK-compatible NN model, a richer parameterization NN model that exhibits higher accuracy, and a graph neural network (GNN) model that identifies relevant interactions within DNA beyond nearest neighbors. All models provide improved accuracy in predicting DNA folding thermodynamics, providing improvements relevant forin silicodesign of qPCR primers, oligo hybridization probes, and DNA origami.
Publisher
Cold Spring Harbor Laboratory