Fig. 4: Results of the generative model and visualized analysis.

a–c Average reward curves show opposite model performance when fine-tuning model pre-trained by graph grammar distillation (red) and ChEMBL dataset (blue) with reinforcement learning. Total reward consists of the constraints about the values of the minimum inhibitory concentration for S. aureus (MIC_S.aureus) and number of carbon numbers (less than 11). A higher reward means that the model generates more desired structures as expected. d Overview of the Topological Data Analysis Mapper (TMAP) for 2114 generated hydrophobic subunits colored by the corresponding scaffolds. Subunits with the same scaffolds are generally clustered together. Note that all subunits are achiral. Cluster A–E include the representative six styles of β-amino acid polymers which are mostly appeared in our data. e–j Property prediction distributions on three bioactivities, including the values of the minimum inhibitory concentration for S. aureus (MIC_S.aureus), E. coli (MIC_E.coli) and the value of the minimum concentration to cause 10% hemolysis (HC₁₀), for the generated 19,026 β-amino acid polymers with fixed dimethyl (DM, e–g) or monomethyl (MM, h–j) positively charged subunits. Ratios of polymers in different carbon range which reach the threshold of specific bioactivity are calculated (Source data are provided as a Source Data file).