Bayesian optimisation and graph-based rheology enable sequence-dependent modelling of DNA materials

Bridging molecular and emergent properties is essential for designing soft matter. Synthetic DNA materials are attractive in this context because their sequence design space supports a wide range of material properties. Targeted design of DNA materials is hindered by scale differences and manual exploration of vast design spaces. We address this challenge with a computational workflow that links sequence-level design to rheological material properties. Concretely, we use machine learning to parametrise scalable, DNA-sequence-aware simulations, which we then evaluate using graph-based rheology. In our example, we study materials composed of self-interacting, multivalent DNA nanostars assembled from single strands. Structure and flexibility of nanostars are quantified with nucleotide-level oxDNA simulations, enabling Bayesian optimisation of a more coarse-grained bead-spring model. The bead-spring model allows efficient simulation of network formation between nanostars, governed by hybridisation free energies, which are computed with oxDNA and NUPACK. Nanostar valency and network connectivity are translated into rheological material properties with a graph-based method that we extend to include hydrodynamic interactions, yielding good agreement with experimental reference data. We generalise our findings by analysing theoretical graph representations of DNA materials and show how machine learning can optimise sequence affinities to produce desired rheological responses. Our work illustrates how machine learning can bridge scales and automate coarse-graining to facilitate targeted design of DNA materials through sequence-property relationships.