EVENT DOSSIER
The research team proposed a discrete denoising diffusion model named Juniper, aimed at solving the problem of reconstructing molecules from coarse-grained beads. This model uses the syngas-water partition free energy (ΔG_W→O) as the generation condition, and the training data covers chemical spaces with up to 9 heavy atoms mapped to one or two beads. Experiments show that for the dual-bead target, Juniper generates molecules with an effective rate of 93% and uniqueness of 92%, and its predicted free energy distribution shows a high linear correlation with the target values (r²≥0.96). Even when only individual scalar free energy information is provided, the model can systematically adjust the functional group structures, successfully achieving transformations from non-polar branched hydrocarbons to polar amides, isocyanates, and other structures. This achievement enables the conversion of bead combinations based on coarse-grained screening markers into candidate molecules for atomic-level research or synthesis.
研究团队提出 juniper 模型，通过自由能条件生成实现从粗粒化珠子到分子的重建。该离散去噪扩散模型以辛醇 - 水分配自由能 $\Delta G_{\mathrm{W} \mapsto \mathrm{O}}$ 为条件，训练于最多 9 个重原子映射至一或两个珠子的数据上。实验显示，针对双珠子目标，juniper 生成的分子有效率为 93%，唯一性为 92%，其自由能分布与目标值呈线性关系（$r^{2} \geq 0.96$）。尽管仅接收单个标量信息，模型仍能系统调节官能团，从非极性支链烃过渡到极性酰胺、异氰酸酯等结构。该成果使粗粒化筛选标记的珠子组合可转化为原子级研究或合成的候选分子。