Aim 1

Expand antibiotic design into unexplored and underexploited reaches of “chemical space”

Historically, drug discovery pipelines have been limited to screening compound libraries of a few million compounds at most. However, the searchable “chemical space” for small molecule drugs consists of ~10³³ compounds. While this enormous search space is intractable for screening using experimental approaches, AI-based approaches open up the possibility of virtual screening. In this project, we will build AI-based generative drug design models to expand antibiotic drug design into novel areas of chemical space.

The active-learning based generative drug design models will be trained with data from chemical libraries and drug screens, to provide information about the chemical structures, pharmacokinetic properties, toxicity profiles, ease of synthesis, binding properties, etc., for millions of compounds. After successive iterations of training, we expect the models to predict both existing and never-before-seen drug designs that optimize the multiple objectives for an antibiotic. These include target specificity (against a given bacterial target), efficacy, safety, pharmacokinetic properties, known mechanisms of action, ease of synthesis, etc.

In silico drug design predictions from the model will then be tested using the screening approaches described in Aims 2 and 3, with data from the screens fed back into the generative drug design models for continued training.