Building predictive models for mechanism-of-action classification from phenotypic assay data sets

Area of Science:

• Pharmacology
• Toxicology
• Computational Biology

Background:

• Mechanism of action (MoA) is crucial for drug development but challenging to determine in phenotypic drug discovery.
• Phenotypic screening may identify compounds with known, undesirable MoAs, posing a risk.
• Early identification of compound MoA is essential for efficient drug development and safety assessment.

Purpose of the Study:

• To develop and validate a predictive modeling approach for assigning mechanism classes to bioactive compounds.
• To enable the triage of phenotypic screening hits by predicting their MoA.
• To support drug discovery by identifying potential off-target toxicities and facilitating target deconvolution.

Main Methods:

• Utilized an 84-feature cell-based signature from BioMAP systems to characterize compounds.
• Developed predictive models using support vector machines (SVMs) trained on a reference dataset of well-characterized compounds.
• Classified compounds into 28 distinct mechanism classes, including safety and efficacy-related pathways.

Main Results:

• Successfully developed predictive models capable of assigning compounds to 28 mechanism classes.
• Demonstrated the application of these models in a decision scheme for analyzing phenotypic screening hits.
• Validated the approach using a dataset of 309 environmental chemicals from the EPA's ToxCast program.

Conclusions:

• The described method provides a robust system for predicting compound mechanism classes from cell-based signatures.
• This approach aids in identifying off-target toxicity mechanisms and deconvoluting targets for phenotypic drug discovery hits.
• The predictive models offer a valuable tool for informed decision-making in drug development pipelines.