Physically-Informed Machine Learning Enhances Predictive Design of Fluorescent DNA-Stabilized Silver Clusters

DNA-stabilized silver clusters (Ag_N-DNAs) are promising nanomaterials for applications in sensing, nanophononics, and bioimaging, due to their diverse sequence-encoded fluorescence color, high quantum yields, and unique rod-like cluster structures. However, our poor understanding of how DNA templates Ag_N growth, combined with the huge sequence space of possible DNA oligomers, challenges the fundamental understanding and rational design of Ag_N-DNAs. To address this problem, we are employing machine learning to discern how DNA sequence selects the fluorescence colors of Ag_N-DNAs. Building from our previous work using machine learning for Ag_N-DNA design, and enabled by well-controlled high-throughput experimental methods for Ag_N-DNA synthesis and characterization, we introduce a new physically informed machine learning model inspired by the recently resolved crystal structures of a few Ag_N-DNAs. We begin with feature vectors that quantify the prevalence of DNA base patterns, specifically, relative arrangement of pairs of nucleic acids within DNA template sequences. We then train an ensemble of support vector machines to assign the most probable fluorescence color class to input DNA sequence. Using this model, we screen all 4¹⁰ possible 10-base DNA sequences to select the most promising DNA templates for Ag_N-DNAs of select colors. In addition, feature selection tools allow us to gain insights into the most important DNA ligand motifs that are discriminative for the atomic sizes and fluorescence colors of Ag_N-DNAs. This work provides a case study for combining machine learning with known physics and experimental training data to improve both the design process and fundamental understanding of nanomaterials systems.