Constrained Bayesian optimization for automatic chemical design using variational autoencoders.

Area of Science:

• Computational Chemistry
• Machine Learning
• Drug Discovery

Background:

• Automatic Chemical Design (ACD) aims to generate novel molecules with desired properties.
• Existing ACD frameworks using variational autoencoders (VAEs) and Bayesian optimization (BO) often produce invalid molecular structures.
• This invalidity stems from querying latent space regions distant from the VAE's training data.

Purpose of the Study:

• To empirically demonstrate the cause of invalid molecular generation in ACD.
• To mitigate the pathology of invalid structures in VAE-based ACD.
• To improve the validity of generated molecules in chemical design.

Main Methods:

• Empirical analysis of VAE-BO pathology.
• Reformulation of the search procedure as a constrained Bayesian optimization (CBO) problem.
• Evaluation of CBO for molecular validity in ACD.

Main Results:

• Confirmed that invalid molecular structures arise from BO querying VAE latent space points far from training data.
• Demonstrated that CBO significantly mitigates the pathology, improving molecular validity.
• Showcased marked improvements in the validity of generated molecules using CBO.

Conclusions:

• The pathology in VAE-BO for ACD is linked to training set mismatch.
• Constrained Bayesian optimization effectively addresses this training set mismatch.
• CBO offers a promising approach for robust molecular generation in ACD and other VAE-based generative tasks.