Active learning in drug discovery

Active learning and Bayesian optimization have become important/prevalent tools in drug discovery and design. They provide a systematic way to select samples with a high amount of information. This property has made them an ideal choice in early-stage drug discovery, e.g., in hit identification through virtual screening and also lead optimization. Also, they can be used to search through usually very large drug combination space to find drug pairs with the highest synergy score. In this workshop, first, we read and review the theoretical aspects of active learning and bayesian optimization.

For example, we will talk about:

• Surrogate models
• Uncertainty Quantification
• Acquisition functions
• Structure-based input space such as strings and graphs

In the next step, we will discuss some areas/steps in the drug discovery pipeline in which active learning can play an important role. In particular, we will talk about the following steps:

• Virtual screening of huge (in the order of billion) compound library
• Lead optimization through a sequential optimization approach
• Finding synergistic drug combinations
• QSAR with limited available data