Predictive models based on neural networks have seen increased adoption in lead optimization, such as in the prediction of ADMET properties or bioactivity optimization. However, such models can behave in an arbitrary and unknown manner in regions of chemical space far outside of the training set (i.e., out-of-distribution). Furthermore, the properties of interest often exhibit aleatoric noise owing in part to experimental error, which can challenge the robustness of deep neural networks with sufficient flexibility to fit noise arbitrarily well.<br/><br/>In this work, we develop robust and distribution-free approaches to uncertainty quantification for the construction of predictive models in ligand-based drug design. We specifically aim to develop models that achieve two goals. First, these models can detect out-of-distribution ligands with high probability and detect in-distribution ligands with a user-specified marginal coverage guarantee - for instance, rejecting a random, unseen in-distribution ligand as being out-of-distribution with probability not exceeding 5%. Second, they provide prediction intervals with guaranteed marginal coverage in-distribution; for example, failing to cover the true response with probability not exceeding 10% for a random, unseen in-distribution ligand. The ligands which fail to be rejected as in-distribution are then ranked according to the lower bound of their prediction interval for the target in question, which has connections to robust procedures in optimization (e.g., value-at-risk). We investigate how these methods compare to alternatives in their detection of out-of-distribution ligands, their ability to produce valid prediction intervals for in-distribution ligands, and for ranking.

Uncertainty quantification for ligand-based drug design

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.