Predictive Models for Mechanism-of-Action Classification of Chemicals Using Phenotypic Data

Predictive Models for Mechanism-of-Action Classification of Chemicals Using Phenotypic Data
Version:
2014

File Name/Number:
Society of Toxicology

Year:
2014

Introduction. There is increasing interest in applying in vitro methods to characterize the potential toxicity of drugs and chemicals. Methods that would allow automatic classification of in vitro activity profiles into mechanism classes could be helpful in prioritizing novel agents.

Method. We have developed primary human cell-based assay panels in which chemical effects on the expression of protein biomarkers are measured.  Eighty-eight selective, well-characterized tool compounds representing 28 distinct mechanism classes relevant to key biological and toxicity mechanisms were profiled in a standardized panel of 8 assay systems that had been previously used for testing >1000 chemicals and materials for the Environmental Protection Agency’s ToxCast program (Houck, 2009). The resulting dataset of 83 endpoint measurements was then used to build predictive models for each of the 28 mechanism classes using machine learning. A support vector machine (SVM) approach gave the best performance, and so was applied to build a series of two-class SVM models for each mechanism class.

Results. These models were then applied to evaluate several libraries and collections of bioactive materials including diversity and kinase focused libraries, biologics and natural products. Compounds from all collections types were highly active in phenotypic assays (hit rates of 7-60% at 1-5 μM). Analysis of a kinase-focused library and a set of environmental bio-actives revealed 48-57% of actives could be classified into one of the 28 tested mechanism classes. The most common mechanisms were mitochondrial inhibitors (14-27%) and microtubule inhibitors (4-10%). Proteasome inhibitors and cAMP elevators were more frequent in the environmental bioactive collection.

Conclusions. These results suggest that compounds with potentially undesirable mechanisms are surprisingly common in most compound collections including libraries used for high throughput screening. This method may be useful for library characterization, determining compound mechanisms of action, and chemical prioritization.