(184ab) Efficient Multi-Objective Bayesian Optimization of Chemical Reactions

Chemical Reaction Optimization in Early-Stage Drug Development typically involves two stages. In the first stage, the objective(s) of the optimization campaign are clearly formulated and the design parameters that have a measurable impact on the reaction are identified. Stage 2 involves running experiments within the identified design space to discover the optimal conditions that maximize the objectives. Even with the adoption of automated reaction platforms^[1] and parallel experimentation, a thorough exploration of the design space is impractical for most real-world problems and the complexity scales with the dimensionality of the design space. Machine Learning algorithms are useful in this context since they can guide the sequence of experiments to be run by making informed data-driven decisions. Bayesian Optimization (BO) has been the algorithm of choice in recent data-driven reaction optimization literature due to its ability to find the optimal reaction conditions while simultaneously satisfying time and material constraints. In the simplest case where a single objective has to be maximized, BO works by fitting a probabilistic surrogate model that maps the design parameters to the objective. Based on the predictions of the learned surrogate model and the associated uncertainties, the experiment that is likely to improve upon the objective is carried out.

Often, there is more than one objective of interest to the chemist and these objectives can compete with each other (For example, Yield, Space-Time-Yield, E-factor, cost-of-materials etc.). Here, a multi-objective BO (MOBO) algorithm is preferred where independent surrogate models are built for each objective. The goal in this scenario is not to maximize any one objective but to discover the set of pareto-optimal reaction conditions. In this work, we identify two inefficiencies associated with how MOBO algorithms are set up and propose solutions that can improve the efficiency with which they discover the optimal reaction conditions.

Current MOBO algorithms treat each objective as independent unknown black-box functions. In practice, this is often not the case. As a trivial example, if cost-of-materials is one of the objectives that needs to be optimized, a surrogate model of cost-of-materials does not need to be learnt from data since the mapping between the design parameters and the cost-of-materials can be calculated beforehand. While most objectives are not this trivial, many are often not completely black box in nature either. For example, Space-Time-Yield, which is a common objective of interest in MOBO, can be expressed as the composition of an unknown black-box function (yield) and known design parameters. By treating objectives as grey-box instead of black-box in this way, Raul Astudillo et al.^[2] have shown that single-objective BO algorithms can significantly improve their query-efficiency. Inspired by this approach, we extend this idea to the multi-objective scenario and show improvements in time and material requirements.

Another limitation with current MOBO algorithms is that all objectives are weighted equally, and the algorithm is incentivized to find the trade-off between these objectives (pareto front). In reality, some objectives are more important to the chemist than others (e.g. Yield over E-factor) resulting in only a portion of the pareto front being relevant. This leads to the algorithm expending its resources wastefully on discovering uninteresting conditions. Lin et al.^[3] have recently shown that this problem can be alleviated by having a surrogate model learn the preferences of the expert decision maker through pairwise comparisons. After learning a preference surrogate model, the BO algorithm is tasked with optimizing the expert’s preferences. It is not clear at the outset if this approach can transfer over to the reaction optimization problem since a chemist’s preferences have never been learnt through pairwise comparisons. In the second part of this work, we model the preferences of a chemist using a preference based Gaussian Process model and compare this approach of preference-based BO to conventional MOBO.

1. Sagmeister, P., Ort, F. F., Jusner, C. E., Hebrault, D., Tampone, T., Buono, F. G., Williams, J. D., & Kappe, C. O. (2022). Autonomous Multi‐Step and Multi‐Objective Optimization Facilitated by Real‐Time Process Analytics. Advanced Science, 9(10), 2105547. https://doi.org/10.1002/advs.202105547
2. Astudillo, R., & Frazier, P. I. (2019). Bayesian Optimization of Composite Functions. ArXiv. https://doi.org/10.48550/arxiv.1906.01537
3. Lin, Z. J., Astudillo, R., Frazier, P. I., & Bakshy, E. (2022). Preference Exploration for Efficient Bayesian Optimization with Multiple Outcomes. ArXiv. https://doi.org/10.48550/arxiv.2203.11382