Leveraging Chemical Hidden-Space Representations Effectively in Bayesian Optimization for Experiment Design through...

Area of Science:

• Computational Chemistry
• Machine Learning
• Chemical Informatics

Background:

• Hidden-space molecular representations from AI models offer potential for chemical Bayesian optimization (BO).
• Current BO workflows often fail to leverage these advanced representations effectively.
• Performance discrepancies are poorly understood and difficult to assess in practice.

Purpose of the Study:

• Investigate the reasons behind the underperformance of hidden-space molecular representations in BO.
• Identify the interaction between representation dimensionality and Gaussian process kernel hyperpriors.
• Develop a novel approach to enhance the performance of these representations in chemical BO.

Main Methods:

• Analyzed the impact of representation dimensionality on BO surrogate learning and acquisition optimization.
• Introduced a dimension-aware adaptive hyperprior for Gaussian process kernels.
• Benchmarked the adaptive hyperprior against fixed hyperpriors across diverse reaction optimization datasets and molecular representations.

Main Results:

• Demonstrated that representation dimensionality significantly affects BO performance through interaction with hyperpriors.
• Showed that fixed or mismatched hyperpriors lead to degraded surrogate learning and optimization.
• The proposed adaptive hyperprior consistently restored and amplified the advantages of hidden-space featurizations.

Conclusions:

• Hyperprior calibration is critical for evaluating and deploying modern molecular representations in chemical BO.
• The dimension-aware adaptive hyperprior effectively addresses performance limitations.
• This work enables fairer assessment and improved utilization of advanced molecular featurizations in chemical design.