Machine learning (ML) including deep learning (DL) is increasingly applied in pharmaceutical research for a variety of applications including predictive modeling of high-throughput screening (HTS) data. However, HTS data display a large imbalance between active and inactive compounds and contain experimental noise, which complicates ML-based predictions.<br/>Compound profiling matrices represent a special form of HTS data. Such matrices are obtained by screening a compound collection against multiple targets and present challenging prediction tasks for ML (Figure 1).<br/>We have compared a variety of ML and DL methods for the prediction of large compound profiling matrices. Different molecular representations and prediction strategies were explored. Although predictive performance varied across assays comprising matrices, promising results were obtained for many targets. Perhaps surprisingly, DL did not further increase the predictive performance of standard ML methods such as random forests (RF). RF models correctly prioritized active compounds for most assays of test matrices indicating the ability of standard ML techniques to identify novel hits under rather challenging experimental conditions.

Machine learning-based prediction of compound profiling matrices

2019 Spring National Meeting

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