Development of omics biomarkers for estrogen exposure using mRNA, miRNA and piRNAs

Abstract

The number of chemicals requiring risk evaluation exceeds our capacity to provide the underlying data using traditional methodology. This has led to an increased focus on the development of novel approach methodologies. This work aimed to expand the panel of gene expression-based biomarkers to include responses to estrogens, to identify training strategies that maximize the range of applicable concentrations, and to evaluate the potential for two classes of small non-coding RNAs (sncRNAs), microRNA (miRNA) and piwi-interacting RNA (piRNA), as biomarkers. To this end larval Pimephales promelas (96 hpf +/- 1h) were exposed to 5 concentrations of 17α- ethinylestradiol (0.12, 1.25, 2.5, 5.0, 10.0 ng/L) for 48 h. For mRNA-based biomarker development, RNA-seq was conducted across all concentrations. For sncRNA biomarkers, small RNA libraries were prepared only for the control and 10.0 ng/L EE2 treatment. In order to develop an mRNA classifier that remained accurate over the range of exposure concentrations, three different training strategies were employed that focused on 10 ng/L, 2.5 ng/L or a combination of both. Classifiers were tested against an independent test set of individuals exposed to the same concentrations used in training and subsequently against concentrations not included in model training. Both random forest (RF) and logistic regression with elastic net regularizations (glmnet) models trained on 10 ng/L EE2 performed poorly when applied to lower concentrations. RF models trained with either the 2.5 ng/L or combination (2.5 + 10 ng/L) treatments were able to accurately discriminate exposed vs. non-exposed across all but the lowest concentrations. glmnet models were unable to accurately classify below 5 ng/L. With the exception of the 10 ng/L treatment, few mRNA differentially expressed genes (DEG) were observed, however, there was marked overlap of DEGs across treatments. Overlapping DEGs have well established linkages to estrogen and several of the 81 DEGs identified in the 10 ng/L treatment have been previously utilized as estrogenic biomarkers (vitellogenin, estrogen receptor-β). Following multiple test correction, no sncRNAs were found to be differentially expressed, however, both miRNA and piRNA classifiers were able to accurately discriminate control and 10 ng/L exposed organisms with AUCs of 0.83 and 1.0 respectively. We have developed a highly discriminative estrogenic mRNA biomarker that is accurate over a range of concentrations likely to be found in real-world exposures. We have demonstrated that both miRNA and piRNA are responsive to estrogenic exposure, suggesting the need to further investigate their regulatory roles in the estrogenic response.

Keywords: Fathead minnow; RNA-seq; epigenetics; estrogens; miRNA; piwiRNA.

Fig. 1.

Differentially expressed genes across all treatments. Larval fathead minnow were exposed to one of five concentrations of EE2 for 48 h and RNA-seq was conducted to identify differentially expressed genes (DEGs). DEG analysis was performed using limma (FDR < 0.05)

Fig. 2.

Prediction probabilities of random forest classifiers trained on either 10, 2.5, or a combination of 10 and 2.5 ng/L EE2 fathead minnow larvae. (A) Prediction probabilities of all three training scenarios, (B) 10 ng/L EE2 scenario with error bars, (C) 2.5 ng/L EE2 scenario with error bars or (D) combination training set with error bars.