Bioinformatic analysis of xenobiotic reactive metabolite target proteins and their interacting partners

Abstract

Background: Protein covalent binding by reactive metabolites of drugs, chemicals and natural products can lead to acute cytotoxicity. Recent rapid progress in reactive metabolite target protein identification has shown that adduction is surprisingly selective and inspired the hope that analysis of target proteins might reveal protein factors that differentiate target- vs. non-target proteins and illuminate mechanisms connecting covalent binding to cytotoxicity.

Results: Sorting 171 known reactive metabolite target proteins revealed a number of GO categories and KEGG pathways to be significantly enriched in targets, but in most cases the classes were too large, and the "percent coverage" too small, to allow meaningful conclusions about mechanisms of toxicity. However, a similar analysis of the directlyinteracting partners of 28 common targets of multiple reactive metabolites revealed highly significant enrichments in terms likely to be highly relevant to cytotoxicity (e.g., MAP kinase pathways, apoptosis, response to unfolded protein). Machine learning was used to rank the contribution of 211 computed protein features to determining protein susceptibility to adduction. Protein lysine (but not cysteine) content and protein instability index (i.e., rate of turnover in vivo) were among the features most important to determining susceptibility.

Conclusion: As yet there is no good explanation for why some low-abundance proteins become heavily adducted while some abundant proteins become only lightly adducted in vivo. Analyzing the directly interacting partners of target proteins appears to yield greater insight into mechanisms of toxicity than analyzing target proteins per se. The insights provided can readily be formulated as hypotheses to test in future experimental studies.

Figure 1

ROC curve using the false-positive rate and true positive rate to evaluate the performance of the model to predict target proteins. The area under the curve is 0.857.

Figure 2

Relative importance of the top 15 features as ranked by the random forest algorithm.

Figure 3

Cumulative distribution of lysine (panel A) and cysteine (panel B) for 62 thiobenzamide target proteins (open circles), 45 bromobenzene target proteins (filled triangles) and 11482 non-target proteins (solid line).