Chemical-agnostic hazard prediction: statistical inference of toxicity pathways from proteomics responses to chemical mixtures.

Toxicity pathways have been defined as normal cellular pathways that, when sufficiently perturbed as a consequence of chemical exposure, lead to an adverse outcome. If an exposure alters one or more normal biological pathways to an extent that leads to an adverse toxicity outcome, a significant correlation must exist between the exposure, the extent of pathway alteration, and the degree of adverse outcome. Biological pathways are regulated at multiple levels, including transcriptional, post-transcriptional, post-translational, and targeted degradation, each of which can affect the levels and extents of modification of proteins involved in the pathways. Significant alterations of toxicity pathways resulting from changes in regulation at any of these levels therefore are likely to be detectable as alterations in the proteome. We hypothesize that significant correlations between exposures, adverse outcomes, and changes in the proteome have the potential to identify putative toxicity pathways, facilitating selection of candidate targets for high throughput screening, even in the absence of knowledge of either the specific pathways involved or the specific agents inducing the pathway alterations. We explored this hypothesis in BEAS-2B human airway epithelial cells exposed to different concentrations of Ni, Cd, and Cr, alone and in defined mixtures. Levels and phosphorylation status of a variety of signaling pathway proteins and cytokines were measured after 48 hours exposure, together with cytotoxicity. Least Absolute Shrinkage and Selection Operator (LASSO) multiple regression was used to identify a subset of these proteins that constitute a putative toxicity pathway capable of predicting cytotoxicity. The putative toxicity pathway for cytotoxicity of these metals and metal mixtures identified by LASSO is composed of phospho-RPS6KB1, phospho-p53, cleaved CASP3, phospho-MAPK8, IL-10, and Hif-1α. As this approach does not depend on knowledge of the chemical composition of the mixtures, it may be generally useful for identifying sets of proteins predictive of adverse effects for a variety of mixtures, including complex environmental mixtures of unknown composition.