Molecular property predictors form the core of any AI-enabled drug discovery strategy. In recent years, there has been significant research in this area, resulting in the development of powerful predictors and representations. However, these diverse predictors have different software interfaces, dependencies, and levels of documentation. Due to lack of a unified API for molecular property prediction, an AI-enabled drug discovery endeavor often necessitates a tangled web of scripts, notebooks, and configuration. This makes it is needlessly difficult to share, distribute, and manage predictors, to ensemble predictors together, and to provide universal AI explainability tools. To this end, we present Oloren ChemEngine (OCE), an open-source Python library with a unified API for molecular property predictors with simplified model management and reproducibility. Using OCE, we create models which achieve superior performance on ADME/Tox prediction tasks by ensembling and integrating many different molecular property prediction methods. We include model-agnostic uncertainty quantification using calibrated confidence intervals and probabilities as well as interpretability using counterfactual methods.
Benchmark Model Parameters
Supporting Information. Model parameter strings for models named in Table 1.