Narayanan Harini, Hinckley Joshua A, Barry Rachel, Dang Brendan, Wolffe Lenna A, Atari Adel, Tseng Yuen-Yi, Love J Christopher
Koch Institute for Integrative Cancer Research at MIT, 500 Main Street, Cambridge, MA, 02139, USA.
Broad Institute of MIT and Harvard, 415 Main Street, Cambridge, MA, 02142, USA.
Nat Commun. 2025 Jul 1;16(1):6055. doi: 10.1038/s41467-025-61113-5.
Optimizing operational conditions for complex biological systems used in life sciences research and biotechnology is an arduous task. Here, we apply a Bayesian Optimization-based iterative framework for experimental design to accelerate cell culture media development for two applications. First, we show that this approach yields new compositions of media with cytokine supplementation to maintain the viability and distribution of human peripheral blood mononuclear cells in the culture. Second, we apply this framework to optimize the production of three recombinant proteins in cultivations of K.phaffii. We identified conditions with improved outcomes for both applications compared to the initial standard media using 3-30 times fewer experiments than that estimated for other methods such as the standard Design of Experiments. Subsequently, we also demonstrated the extensibility of our approach to efficiently account for additional design factors through transfer learning. These examples demonstrate how coupling data collection, modeling, and optimization in this iterative paradigm, while using an exploration-exploitation trade-off in each iteration, can reduce the time and resources for complex optimization tasks such as the one demonstrated here.
优化生命科学研究和生物技术中使用的复杂生物系统的操作条件是一项艰巨的任务。在此，我们应用基于贝叶斯优化的迭代框架进行实验设计，以加速两种应用的细胞培养基开发。首先，我们表明这种方法产生了添加细胞因子的新培养基成分，以维持培养物中人类外周血单核细胞的活力和分布。其次，我们应用此框架优化法夫酵母培养物中三种重组蛋白的生产。与初始标准培养基相比，我们通过比其他方法（如标准实验设计）估计的实验次数少3至30倍的实验，确定了两种应用中具有更好结果的条件。随后，我们还展示了我们的方法通过迁移学习有效考虑其他设计因素的可扩展性。这些例子表明，在这种迭代范式中耦合数据收集、建模和优化，同时在每次迭代中使用探索-利用权衡，可以减少诸如本文所示的复杂优化任务的时间和资源。
