Fig. 1: Framework overview.

a By collecting 86 data comprising chemical structures and their bioactivity of β-amino acid polymers, we develop an AI-guided few-shot inverse design framework to find promising polymers with broad-spectrum antibacterial efficacy and low cytotoxicity. In addition, we conduct a refined classification according to the different position of side chain substituents and cyclic or non-cyclic substitution pattern, which defines a scaffold set for the following polymer generation. x and y are defined as the percentages of a positively charged subunit and a hydrophobic subunit in β-amino acid polymers, respectively. “R₁, R₂, R₃, R₄” means that more than one substitution point should be decorated. Note that all subunits are achiral. b We conduct a multi-modal polymer representation method, including text sequence, graph with additional polymer settings and descriptors embedded with 2D- and 3D-properties of subunits to expand for multi-scale polymer information to realize few-shot polymer prediction. c We develop a graph grammar distillation method in which we utilize β-amino acids and natural α-amino acids to learn the split graph grammar fragments. These fragments are reconstructed as distilled molecule dataset to pre-train the generative model to restrict the huge chemical space for exploration.