In the machine learning literature, accuracy is often assessed on held-out data or through cross-validation.  Unfortunately artifacts in chemical data, such as redundancy and analogue bias, make a naive partitioning of data inappropriate for performance assessment.  We demonstrate that even widely-used techniques for redundancy elimination, such as scaffold clustering, can leave significant bias in the data.  Furthermore, we demonstrate that these biases may explain much of the performance of popular machine learning techniques for ligand-based virtual screening.  We conclude with best practices to adopt when seeking predictive techniques that extrapolate to new biochemical space, rather than merely interpolate among known answers.

Pitfalls in the assessment of ligand-based virtual screening accuracy

251st National Meeting (2016)

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