Computer-aided molecular design from virtual screening to de novo generation
The prototypical discovery workflow for molecules and materials involves iterating through design-build-test loops. One paradigm of chemical space exploration is virtual screening, where pre-enumerated libraries are used as a source of ideas for simulation or experimental testing. Another paradigm is generative molecular design, which may suggest novel structures beyond what is available in virtual libraries. I will describe our recent algorithmic approaches with both goals in mind, including the use of model-guided Bayesian optimization for virtual screening. I will discuss new solutions to address common failure modes of deep learning-based generative tools, emphasizing the consideration of synthetic accessibility. Finally, I will describe an optimization framework for the downselection of molecular structures from large candidate sets that strikes a balance between information gain and experimental cost. By providing effective and controllable navigation within synthesizable chemical space, we can provide actionable suggestions of new small organic molecules across a range of fields, including drug development and materials science.
Hosted by Professor Joel Freundlich