(320a) Automated, Multi-Task Bayesian Optimization of Pharmaceutical Processes

In recent years, there has been an increased interest in the use of automated, self-optimizing continuous flow platforms for the optimization of chemical processes.^[1] This has arisen from the desire for faster reaction optimization, improved process sustainability and cheaper overall process development. The use of these platforms aims to enhance the capabilities of the researcher by removing the need for labour-intensive experimentation, allowing them to focus on more challenging tasks. These platforms use optimization (machine-learning) algorithms to use only minimal reaction material, but ultimately gain the most process information possible, making their deployment in fine chemical and pharmaceutical industries very attractive.^[2]

Recent work has shown that Bayesian optimization in particular is a powerful tool for these applications.^[3] However, in all previous studies, each optimization has begun using no a priori information about the chemical landscape for the reaction of interest. This protocol therefore requires many initial experimental iterations whereby the algorithm is learning about the experimental design space, without any information on where the optimal reaction conditions may be. This can be wasteful in terms of both cost and time, particularly when there may already be a wealth of information on the chemical transformation of interest from previous optimization campaigns. For medicinal chemistry applications, such as developing chemistries for the elaboration of small polar fragments into nanomolar leads (i.e. Fragment-Based Drug Discovery)^[4,5], the use of efficient optimization techniques is paramount due to the minimal quantity and increased preciousness of starting materials.

This work shows the first example of leveraging previous reaction optimization data for unseen chemical transformations, using multi-task Bayesian optimization (MTBO). The framework of MTBO, first introduced by Swerksy et al.,^[6] replaces the standard probabilistic model in Bayesian optimization with a multi-task model. As these multi-task models can be trained on data from related tasks, we can therefore utilise data from similar reactions - both from the laboratory and from the literature. In this work, we sequentially optimize the reaction conditions for five similar palladium-catalyzed C-H activation reactions using our automated reactor platform. Each reaction yields a pharmaceutically relevant oxindole product, as the optimization efficiency increases with further experimental data - this is true even with different chemical substrates that host different functionalities. One of these five processes is shown in the corresponding scheme, highlighting the reactivity that we targeted in this study.

This work represents a step-change in self-optimization, enabling ultra-efficient reaction optimization through the use of previous experimental data.

References: [1] React. Chem. Eng., 2019, 4, 1545-1554. [2] Chem. Eng. J., 2020, 384, 123340. [3] Nature, 2021, 590, 89-96. [4] Nat. Chem., 2009, 1, 187-192. [5] Chem. Sci., 2021, 12, 11976-11985. [6] NIPS Proc., 2013, 2004-2012.