📄 中文摘要
设计用于玻璃化的冷冻保护剂（CPA）混合物面临挑战，配方必须足够浓缩以抑制冰的形成，同时又不能具有毒性以保持细胞活力。这种权衡导致了一个庞大的多目标设计空间，传统的发现过程往往依赖专家直觉或耗时的实验。提出了一种数据高效的框架，通过结合高通量筛选与基于多目标贝叶斯优化的主动学习循环，加速CPA混合物的设计。从初始测量的混合物出发，训练概率代理模型以预测浓度和活力，并量化候选配方的的不确定性。然后进行迭代优化，以寻找最佳的CPA配方。该方法显著提高了发现新冷冻保护剂混合物的效率。
通过多目标贝叶斯优化加速冷冻保护剂混合物的发现
📄 English Summary
Accelerated Discovery of Cryoprotectant Cocktails via Multi-Objective Bayesian Optimization
Designing cryoprotectant agent (CPA) cocktails for vitrification is a complex challenge, as formulations must be concentrated enough to suppress ice formation while remaining non-toxic to preserve cell viability. This trade-off results in a large multi-objective design space, where traditional discovery methods are often slow and rely heavily on expert intuition or exhaustive experimentation. A data-efficient framework is presented that accelerates CPA cocktail design by integrating high-throughput screening with an active-learning loop based on multi-objective Bayesian optimization. Starting from an initial set of measured cocktails, probabilistic surrogate models are trained to predict concentration and viability, while quantifying uncertainty across candidate formulations. Iterative optimization is then performed to identify the optimal CPA formulations, significantly enhancing the efficiency of discovering new cryoprotectant cocktails.