Generative Bayesian Optimization for Structured Design

Black-box optimization is a general framework for problems where our goal is to find an input that maximizes some real-valued objective function. Typically, no mathematical form is available for this objective function, and it may be expensive to evaluate, making it a black-box. Many practical scientific discovery problems can be framed as black-box optimization. For example, in drug discovery one might seek to design a small molecule that maximizes a real-valued objective function, such as predicted therapeutic activity against a target pathogen. Such objectives lack closed-form expressions and may require costly physics-based simulations or laboratory experiments to evaluate, motivating sample-efficient optimization methods. Bayesian optimization (BO) is a widely used machine learning method for sample-efficient black-box optimization. However, standard BO methods assume a continuous, numerical input space and are not directly applicable to structured, discrete domains such as molecules or proteins. Latent space Bayesian optimization (LS-BO) is a newly emerging approach for this setting. LS-BO employs generative models to embed discrete, structured inputs into continuous latent spaces where BO techniques can be applied. This dissertation presents recent work that advances LS-BO, enabling more effective optimization of challenging real-world objectives in scientific discovery. In addition, it extends the LS-BO framework to new problem settings that broaden its applicability and better align with the practical needs of scientific practitioners.

Author dc:creator

• Maus, Natalie

Advisor dc:contributor.advisor

• Gardner, Jacob, R

• Computer Sciences

Language dc:language.iso

en

Repository record dc:identifier.uri

https://repository.upenn.edu/handle/20.500.14332/62344

OAI identifier oai:identifier

oai:repository.upenn.edu:20.500.14332/62344