Phenotypic screening tests the effect of compounds on living systems. In many cases, the assay readout is at the end of a biological pathway, such that disruption of proteins along the pathway generates a positive signal in the assay. Therefore, the screening hits can target different proteins involved in the pathway. However, it is not possible to know a priori which targets are disrupted from the screening results. Target deconvolution studies are needed to understand the activity profile of phenotypic screening hits. Prediction of activity profiles provides hypotheses that can guide target deconvolution studies. Multitask machine learning methods are able to predict several outputs with a single model, making them ideal to predict activity profiles. They have become widely used in chemoinformatics, thanks to the popularity of deep neural networks. They have previously been used to perform large-scale activity prediction over public databases such as ChEMBL and PubChem, as well as other chemical endpoint of interest. Here, we use multitask prediction methods to aid phenotypic screens. We propose a model that is able to predict if a compound is active in a phenotypic screen and to which target it might bind. We perform focused prediction over targets known to be involved in the pathway studied in the screen. We test our model on a proof-of-concept data set that incorporates bioactivity, screening, and pathway information obtained from ChEMBL, PubChem, WikiPathways, and literature sources. We expect our strategy will provide useful assistance in phenotypic screening campaigns.

Phenotypic screening aided by multitask prediction methods

256th National Meeting (2018)

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