Identification of p53 Activators in a Human Microarray Compendium

Abstract

Biomarkers predictive of molecular and toxicological effects are needed to interpret emerging high-throughput transcriptomic data streams. The previously characterized 63 gene TGx-DDI biomarker that includes 20 genes known to be regulated by p53 was previously shown to accurately predict DNA damage in chemically-treated cells. We comprehensively evaluated whether the molecular basis of the DDI predictions was based on a p53-dependent response. The biomarker was compared to microarray data in a compendium derived from human cells using the Running Fisher test, a nonparametric correlation test. Using the biomarker, we identified conditions that led to p53 activation, including exposure to the chemical nutlin-3 which disrupts interactions between p53 and the negative regulator MDM2 or by knockdown of MDM2. The expression of most of the genes in the biomarker (75%) were found to depend on p53 activation status based on gene behavior after TP53 overexpression or knockdown. The biomarker identified DDI chemicals that were strong inducers of p53 in wild-type cells; these p53 responses were decreased or abolished in cells after p53 knockdown by siRNAs. Using the biomarker, we screened ∼1950 chemicals in ∼9800 human cell line chemical vs. control comparisons and identified ∼100 chemicals that caused p53 activation. Among the positive chemicals were many that are known to activate p53 through direct and indirect DNA damaging mechanisms. These results contribute to the evidence that the TGx-DDI biomarker is useful for identifying chemicals that cause DDI and activate p53.

FOR TOC ONLY

INTRODUCTION

High-throughput transcriptomic (HTTr) technologies are increasingly being used to screen environmental chemicals in in vitro cell systems. The EPA ToxCast effort is now utilizing HTTr technologies such as targeted RNA-Seq ¹ to substantially expand coverage of the biological pathways examined in the ToxCast screening program ². As part of this screening effort, full-genome gene expression profiles have been generated in human cells for thousands of chemicals in concentration response format ³. Predictions from these “Tier 0” screens (defined as chemical screens carried out before more conventional assays) will be followed by targeted HT Tier 1 assays to uncover or confirm the underlying mechanisms of action. A critical challenge in HTTr profiling as well as other large genomic studies including the Connectivity Map (CMAP) project ⁴ and the TG-GATES study ⁵ continues to be linking changes in gene expression to alterations in specific molecular targets including molecular initiating events (MIEs) and downstream key events (KEs) that lead to adverse outcomes in adverse outcome pathways (AOPs) ⁶. Gene expression signatures (or biomarkers) are increasingly being used to make predictions of MIEs ^7–13 or organ toxicities ^14–20. Most biomarkers developed thus far have been used to examine chemical effects in rodent livers. However, strategies for building, testing, and validating human gene expression biomarkers have been described ^21,22 that are useful in interpreting Tier 0 HTTr chemical screening data.

A panel of genes useful for identifying chemicals that are DNA damage inducing (DDI) in human cells has been characterized and is called the TGx-DDI biomarker (originally called the TGx-28.65 biomarker). The biomarker was developed by exposing TK6 cells to 28 prototype agents that are either DDI or non-DDI alongside solvent controls. Expression profiles were then produced using DNA microarrays. The data set was analyzed using a machine learning algorithm (nearest shrunken centroid model) that identified 63 genes consistently induced or repressed by DNA damage ²³. The DDI compounds used to create the biomarker covered a broad range of mechanisms including DNA alkylating agents, DNA strand breaking agents, topoisomerase inhibitors, nucleotide antimetabolites, heavy metals, and inhibitors of energy metabolism, histone deacetylase, and microtubule formation. Initial studies indicated that the TGx-DDI biomarker differentiates DDI and non-DDI (NDDI) compounds with 100% accuracy in human lymphoblast TK6 cells ²³. In follow up studies, the efficiency of the biomarker for accurately classifying compounds that require metabolic activation was demonstrated using different metabolic activating systems ²⁴. The biomarker was also 100% accurate in identifying DDI compounds evaluated by another microarray platform (Affymetrix) in the metabolically competent human liver cell line, HepaRG ²⁵. The biomarker accurately classified 45 test agents across a broad set of chemical classes using the nCounter high throughput cell based testing platform coupled with probability analysis, principle components analysis, and two-dimensional clustering ²⁶. In a recent study, our group confirmed that the biomarker could identify DDI chemicals using a different computational method, the Running Fisher test. Using the same chemicals that were tested in earlier studies ^25,26, the method was very accurate (up to 97%) in distinguishing between chemicals that are DDI and non-DDI in two human cell lines (TK6 and HepaRG cells) ²⁷. To increase the use of these classification methods among the regulatory community, an online tool (the TGx-DDI Biomarker for DNA Damage Classification Tool; https://manticore.niehs.nih.gov/tgclassifier/) was developed that evaluates user microarray data predicting the DDI potential of tested chemicals ²⁸. Overall, these studies indicate that the TGx-DDI biomarker used in conjunction with either the nearest shrunken centroid model or the Running Fisher test will be useful in evaluating data from HTTr studies that could complement or be used for prioritization of regulatory genotoxicity testing.

A preliminary pathway analysis of the genes in the TGx-DDI biomarker found that a subset of 20 genes was associated with activation of the transcription factor p53 ²³, not surprising given that most compounds that are DDI activate p53. The p53 protein is a critical tumor suppressor in human cancers, as evidenced by the high frequency of p53 mutations in tumors, with ∼95% of the mutations found within the DNA binding domain ²⁹. Under basal conditions, the activity of p53 is held in check by the negative regulator and E3 ubiquitin ligase, MDM2, which binds and targets p53 for proteasomal degradation ³⁰. In response to multiple types of stress stimuli, this inhibition is relieved, leading to the release of p53, increased binding of the p53 tetramer to p53 response elements, and trans-activation of genes involved in different cellular functions including cell cycle arrest, apoptosis, senescence, and autophagy ^31,32. Many mechanisms exist within the cell to fine-tune the p53 transcriptional program including posttranslational modifications of p53, interactions with covalent and noncovalent p53 binding partners, and differences in DNA binding affinities to p53 response elements ³¹.

Given the importance of p53 as a potential therapeutic target in treating cancers as well as an indicator of DNA damage, a number of high throughput screens have been carried out to identify compounds that activate p53 ²⁹. One class of small molecules identified by a cell free high throughput screen targeted the interaction between p53 and MDM2, allowing bypass of stress signaling to trigger p53 activation. The first-in-class compound, nutlin-3, acts as a competitive inhibitor by binding to the hydrophobic pocket in MDM2 required for interaction with p53 ³³. Peltonen et al. ³⁴ screened chemicals in A375 melanoma cells using a p53 trans-activation assay and identified six novel activators that likely activate p53 by intercalating with DNA, which for some of the compounds occurred in the absence of DNA damage. Witt et al. ³⁵ recently screened 7849 compounds in a p53 trans-activation assay carried out in the human colon carcinoma HCT-116 cell line and identified 365 chemicals that activated p53, including many that are known DDI. To our knowledge, screens for p53 activating chemicals have not been carried out using a gene expression-based strategy. Such an approach would be useful given that public databases contain large numbers of microarray studies of chemical-induced gene expression which will eventually include those that are part of HTTr profiling in chemical screening programs.

In the present study, we hypothesized that the TGx-DDI biomarker could be used to identify DDI chemicals that activate p53, whether or not such activity is due to prior DNA damage. A number of strategies were employed to demonstrate that the biomarker coupled with the Running Fisher test could identify bone fide p53 activators. These strategies included assessment of biomarker predictions after exposure to known p53 activators, determination of p53-dependent regulation of genes in the biomarker, and comparison of chemical effects in cells with normal expression vs. cells with decreased expression of p53. We utilized the biomarker to perform a virtual screen of ∼1950 chemicals in ∼9800 microarray comparisons. A majority of the ∼100 chemicals identified are known to damage DNA through a number of direct and indirect mechanisms.

EXPERIMENTAL PROCEDURES

Overarching strategy for identification of DDI chemicals that modulate p53 using the TGx-DDI biomarker.

Detailed experimental procedures are provided in subsequent sections. Here we describe our overall approach which is similar to that described earlier for identification of genotoxic chemicals ²⁷. Briefly, we used publicly available data to identify chemicals or genetic perturbations inducing a pattern of gene expression that is positively correlated to the TGx-DDI biomarker (Figure 1). To do this, the expression patterns of the DDI agents versus solvent controls were used to calculate average fold-change values for each gene across 13 genotoxic agents from the original TGx-DDI training set ²³. We used this expression profile to identify correlated data sets in a compendium of gene expression comparisons from a commercially available gene expression database called BaseSpace Correlation Engine (BSCE) (https://www.illumina.com/products/by-type/informatics-products/basespace-correlation-engine.html; formally NextBio) that contains over 22,443 highly curated, publicly available, omic-scale studies from 15 species including ∼137,000 lists of statistically filtered genes (as of November, 2018). Statistically filtered gene lists are referred to as biosets. We extracted available information about each bioset from BSCE and used the information to populate a database of experimental details focusing on human cells. Each bioset was annotated for information that facilitates interpretation of TGx-DDI biomarker activity. To assess correlation, the biomarker was compared in a pair-wise fashion to each bioset in the compendium using the Running Fisher algorithm ³⁶. The method allows an assessment of whether the overlapping genes in each comparison are regulated in a similar or opposite manner. Biosets that exhibit significant positive correlation with the biomarker are predicted to reflect p53 activation in those cells. Results of the comparisons were exported and used to populate the annotated compendium with a Running Fisher test p-value and direction of correlation for each bioset. We have used this analysis strategy to identify chemicals that activate or suppress transcription factors in the mouse and rat liver (AhR, CAR, Nrf2, PPARα, SREBP, STAT5b) ^7–11,13,37 as well as the estrogen receptor (ER) in MCF-7 cells ²² and the androgen receptor (AR) in human prostate cancer cell lines ²¹.

Figure 1. Overarching strategy for using the TGx‐DDI biomarker to identify chemicals that damage DNA.

Left, generation of gene lists to produce biosets. Filtered gene lists were derived from treatment vs. control experiments in human cells. Specifically, after microarray analysis of RNA, statistically filtered differentially expressed gene lists (DEGs) were derived using standard data analysis pipelines in BaseSpace Correlation Engine (BSCE). (Right) The DEGs from each of these experiments were then compared to the pattern of gene expression changes for the 63 genes in the TGx-DDI biomarker. To do this, the DEG lists were uploaded to the BSCE environment in which internal protocols rank‐ordered the genes based on their absolute average fold‐change. Comparison of the biomarker to each gene list was carried out using a pair‐wise rank‐based algorithm (the Running Fisher test). The results of the test, including the direction of correlation and the p-values for the evaluated biosets, were exported and used to populate a master table containing bioset experimental detail enabling the identification of chemical exposure conditions in which there was evidence of genotoxicity. The figure was adapted from Corton et al. ²⁷.

A compendium of gene expression experiments carried out in human cells.

Information in the BSCE database was used to build an annotated compendium of gene expression biosets derived from experiments carried out in human tissues. The BSCE database contains over approximately 21,600 highly curated, publically-available, omic-scale studies across 15 species including approximately 134,000 lists of statistically filtered genes (as of October 2017). Annotated information from BSCE about human-derived biosets was used to populate a master file with information about each bioset including Biodesign, Biosource, Chemical Name, Gene, Gene Mode, Phenotype, Tissue, and Study ID (last update, August 17, 2017). Each bioset was annotated for category and name of the perturbant examined based on the name of the bioset. For example, the bioset called “Hepatocellular carcinoma cell line HepG2 – 7uM cisplatin treated 12hr _vs_ PBS control” is in the category “Chemical” and the specific perturbant is “cisplatin”. Biosets that examined more than 2 perturbants at 1 time (e.g., exposure to 3 chemicals vs control) or that could not be interpreted were not used in any further analyses. The final compendium contained ∼37,000 biosets, of which ∼9800 were derived from chemical vs. control comparisons. Information about the biosets examined in the study that exhibited p53 activation is found in Supplemental File 1.

Identification of differentially expressed genes in BSCE microarray datasets.

All differentially regulated genes were identified using the criteria in the BSCE analysis pipeline and are described in detail in Kupershmidt et al. ³⁶. Briefly, following platform-appropriate processing and normalization, statistical analyses to identify differentially-expressed genes involved Welch or standard t tests with a p-value cutoff of 0.05 (without multiple test correction) and a minimum absolute fold-change cutoff of 1.2. Justification for implementing these cutoffs can be found in ³⁶. The CMAP database was downloaded as CMAP 2.0 build 01 into BSCE. Even though there was only 1 biological sample per chemical exposure (i.e., one Affymetrix .cel file per treatment) within this dataset, statistically significant genes were identified by comparing each treatment with a group of control samples using a t-test to calculate the p-value with an assumption of equal variance between case and controls ³⁶.

Comparison of the TGx-DDI biomarker to biosets in the compendium.

The strategy for comparison of a biomarker to collections of biosets has been described in our previous studies ^7–9. The Running Fisher algorithm was used to compare the biomarker to each bioset in BSCE. The analysis output includes the number of overlapping genes, p-value of the Running Fisher correlation test, and direction of the correlation. P-values were converted to –Log(p-value)s, and those with negative correlations were converted to negative numbers. The final list of –Log(p-value)s was used to populate the table containing information about each bioset. Prior studies with gene expression biomarkers for xenobiotic receptors showed that a cutoff p-value ≤ 10⁻⁴ after a Benjamini Hochberg correction of α = 0.001 resulted in a balanced accuracy for activation of 95%, 97%, and 98% for AhR, CAR, and PPARα, respectively ^7–9. This cutoff resulted in a balanced accuracy of DDI of 90–97% in TK6 and HepaRG cells ²⁷. We provide information about all biosets with -Log(p-value) ≥ 4 (Supplemental File 1).

Expression of the biomarker genes across the biosets.

The statistically-significant fold-change values for each gene in the biomarker across all human biosets were extracted from BSCE. Fold-change values were used to populate the table containing biomarker correlation p-values and information about the biosets. Heatmaps were visualized using TreeView (http://rana.lbl.gov/EisenSoftware.htm).

Identification of genes in the biomarker regulated by p53.

Genes in the biomarker were examined after overexpression or knockdown of p53. Seven biosets were used in the comparisons, as they came from studies in which the p53 gene was overexpressed or knocked down. The biosets used in the comparisons were the following, taken verbatim from BSCE (GEO accession numbers are in parentheses):

1. Small cell lung cancer NCI-H1299 cells w/o p53 - overexpressing WTp53 _vs_ vector controls (GSE37271);

2. Small cell lung cancer NCI-H1299 cells w/o p53 - overexpressing a deltaNp53:WTp53 _vs_ vector controls (GSE37271);

3. Malignant pleural mesothelioma MES1-MM cells - P53 overexpression _vs_ control vector (GSE64738);

4. ME16C cell line treated with vehicle - p53-shRNA transduced _vs_ not transduced (GSE3178);

5. HME-CC cell line treated with vehicle - p53-shRNA transduced _vs_ not transduced (GSE3178);

6. Fibroblast IMR90 cells RAS[G12V] (oncogene) induced senescence - TP53 shRNA 5d_vs_vector (GSE53379);

7. Umbilical vein endothelial cells after 48hr TP53 siRNA _vs_ scrambled (GSE84877).

The “meta-analysis” function in BSCE was used to identify the top 5000 genes common to the largest number of the biosets described above. All data were exported and used to classify each gene in the biomarker for p53 regulation. Biomarker genes were classified as p53 dependent if: 1) the up-regulated genes were increased in expression or down-regulated genes were decreased in expression after exogenous overexpression of p53, 2) the up-regulated genes were decreased in expression or down-regulated genes were increased in expression after knockdown of p53 expression, or 3) both. To be called a p53-dependent gene, the gene had to be altered in the expected direction in at least one of the 7 biosets. However, most of the genes classified as p53-dependent had more than one incidence of expected change among the 7 biosets.

Comparison of biomarker predictions to an HTS assay for p53 activation.

We compared the results of the biomarker approach with a recent HTS for p53-activating chemicals that was carried out in human colon carcinoma HCT-116 cells ³⁵. This assay examined 7849 chemicals after 16 h of exposure in dose-response format by measuring the ability of p53 to activate a reporter gene under control of p53 response elements. There were 629 chemicals examined by both methods. The chemicals examined in the microarray compendium were eliminated if they were inactive at a dose that was ≤ the effective concentration for a 50% response (EC50) in the HTS assay due to the fact that a negative response could be due to insufficient concentration tested and differences in the cell lines used for testing. Biosets derived from cell lines that were classified as p53-null or p53 mutant were not used in the analysis (http://p53.iarc.fr/). If there was more than one bioset for a chemical, the bioset with the highest -Log(p-value) was used in the analysis. For calls of p53 activity in the HTS assay, we used the supplemental data for the assay results which accounted for cytotoxicity; positive calls for p53 activation had to occur at doses lower than those that induced cytotoxicity ³⁵.

Results and Discussion

The TGx-DDI biomarker identifies pharmacological and genetic perturbations that activate p53.

The TGx-DDI biomarker was previously shown to accurately identify chemicals that are DDI in TK6 and HepaRG cell lines ^23–27. Previous pathway analysis by Li et al. ²³ found that 20 biomarker genes are regulated by p53. We compared the biomarker genes to the canonical pathways in the Broad MSigDB collection and found that the genes in the biomarker were significantly enriched for p53-linked pathways (Supplemental Figure 1). We hypothesized that the TGx-DDI biomarker could be used to identify chemicals and other factors that activate p53 independent of DNA damage. To test this hypothesis, we first examined 8 biosets derived from four cell types (HCT-116, MCF-7, primary sarcoma cells, U-2 OS) treated with the MDM2 inhibitor nutlin-3. When these biosets were compared to the biomarker, all 8 exhibited significant (-Log(p-value ≥ 4) positive correlations to the biomarker (Figure 2A). Examination of the expression of the biomarker genes across the comparisons showed that nutlin-3 exposures generally led to increases in the positively-regulated genes and to lesser extents, decreases in the negatively-regulated genes.

Figure 2. The TGx-DDI biomarker identifies pharmacological and genetic perturbations that activate p53.

The biomarker was compared to biosets derived from cells treated with (A) nutlin-3 or (B) a set of compounds that activate p53. All of the treatments were for 24 or 48 hrs at the indicated concentrations. (C) Effects of gene activation or suppression on p53 activity. (Left) Suppression of MDM2 expression in A549 or TOV21G cells leads to increased expression of biomarker genes and positive correlation to the biomarker. (Right) Effect of increased or decreased expression of genes known to affect p53 activity.

The top panels represent the -Log(p-value)s of the Running Fisher correlation test between the biomarker and the indicated bioset. The bottom panels are heatmaps of the expression of the biomarker genes in each of the comparisons. The expression of the 63 genes in the biomarker are shown to the left. The study from which the microarray data was derived is indicated as a Gene Expression Omnibus (GEO) accession number.

Next, we examined 6 chemicals (BMH-7, −9, −15, −21, −22, −23) identified as p53 activators in a high-throughput screen for potential anti-cancer agents ³⁴. When gene expression was examined in MCF-7 cells after exposure for 6 hrs, 5 of the compounds induced a gene expression profile that was positively correlated to the biomarker (Figure 2B). Four of the compounds (BMH-9, −21, −22, −23) achieved statistical significance (-Log(p-value) ≥ 4), BMH-7 approached significance (-Log(p-value) = 3.7), and BMH-15 did not appear to be active (only one gene in the biomarker was modulated). Peltonen et al. ³⁴ found that the induction of target genes by BMH-9, −21, −22, and −23 was clearly dependent on p53, whereas BMH-7 and BMH-15 had more restricted capacity for p53-dependent gene regulation based on comparisons between responses in wild-type and p53-null HCT-116 cells. In their analysis, BMH-15 was the only compound out of the 6 that lacked significant enrichment of p53 signaling genes. Furthermore, BMH-7 and BMH-15 but not the other compounds activated a DNA damage response. Peltonen et al. ³⁴ speculated that BMH-9, −21, −22, and −23 interfere with DNA topology in the absence of direct DNA damage and may act as topoisomerase inhibitors as the predominant means of pharmacological activation of p53.

We determined if modulation of the expression of genes that encode p53-regulating proteins would result in TGx-DDI biomarker effects. Knockdown of the negative regulator MDM2 by siRNA in A549 or TOV21G cells resulted in increased expression of a subset of the up-regulated biomarker genes that was strongest at 12 h after siRNA transfection compared to 24 or 48 h, with some of the biosets reaching significance for positive correlation (Figure 2C, left). Additional gene modulations were tested that are known to increase p53 expression or stability (Figure 2C, right). When overexpressed these included the microRNAs MIR542 and MIR30A, as well as PPARG ^{38 39 40 41} and when knocked down, the genes included CDK2, SCYL1, and SNCA ^{42 42 43}. Two other genes when overexpressed (FAM123B/WTX) or knocked down (CTBP1) activate p53. FAM123B/WTX binds to the DNA-binding domain of p53, enhances p53 binding to CBP/p300, and increases CBP/p300-mediated acetylation of p53 at Lys373 and Lys382 ⁴⁴. CTBP1 is a known transcriptional repressor that is recruited to the MDM2-p53 complex by MDM2; inhibition of CTBP1 expression led to p53 activation ⁴⁵. None of the studies in which expression of the genes was altered noted DNA damage. These results indicate that the TGx-DDI biomarker can identify chemical or genetic perturbations that activate p53 independent of overt DNA damage.

The majority of the TGx-DDI biomarker genes are regulated by p53.

To determine if there are p53-regulated genes in the biomarker other than the 20 identified by Li et al. ²³, biomarker genes were examined in microarray comparisons in which the TP53 gene was either overexpressed or knocked down. Used in the analysis were three biosets in which TP53 was overexpressed and four biosets in which TP53 was knocked down by siRNA technologies. One of the biosets with overexpression was a comparison between control cells and cells expressing a p53 molecule comprised of a p53 lacking the N-terminal 40 amino acids (deltaNp53) expressed as a fusion protein with a wild-type p53 ⁴⁶. The deltaNp53:WTp53 tetramer activated transcription equally well compared with WTp53 tetramers in an in vitro reconstituted transcription system ⁴⁶. All biosets gave the expected biomarker response, i.e., overexpression of TP53 led to positive correlation to the biomarker, while knocking down TP53 expression led to negative correlation to the biomarker (Figure 3, top).

Figure 3. The majority of the TGx-DDI biomarker genes are regulated by p53.

(Top) Correlation -Log(p-value)s of comparisons between the biomarker and biosets in which p53 was either overexpressed or knocked down. (Bottom) The 63 genes in the TGx-DDI biomarker were compared to their expression in the biosets. The genes were divided into those that exhibited p53-dependence and those which were p53-independent. The biosets used in the comparisons and their GEO accession numbers were the following (bioset names from BSCE):

1. Small cell lung cancer NCI-H1299 cells w/o p53 - overexpressing WTp53 _vs_ vector controls (GSE37271);

2. Small cell lung cancer NCI-H1299 cells w/o p53 - overexpressing a deltaNp53:WTp53 _vs_ vector controls (GSE37271);

3. Malignant pleural mesothelioma MES1-MM cells - P53 overexpression _vs_ control vector (GSE64738);

4. ME16C cell line treated with vehicle - p53-shRNA transduced _vs_ not transduced (GSE3178);

5. HME-CC cell line treated with vehicle - p53-shRNA transduced _vs_ not transduced (GSE3178);

6. Fibroblast IMR90 cells RAS[G12V] (oncogene) induced senescence - TP53 shRNA 5d_vs_vector (GSE53379);

7. Umbilical vein endothelial cells after 48hr TP53 siRNA _vs_ scrambled (GSE84877).

The expression of biomarker genes was examined to identify those that were regulated by p53. Out of the 63 genes, 47 showed a pattern of expression that was consistent with p53 regulation (Figure 3, bottom). The genes up-regulated in the biomarker were increased by TP53 overexpression or decreased by TP53 knockdown or both. To a lesser extent the genes down-regulated in the biomarker were decreased by TP53 overexpression or increased by TP53 knockdown or both. Seven of the biomarker genes had expression patterns that were not consistent with p53 regulation (ARRDC4, CBLB, CCP110, E2F8, HIST1H2BG, ID2, SEL1L). Nine of the genes did not exhibit any expression changes after TP53 genetic manipulation (B3GNT2, COIL, DAAM1, HIST1H1E, LCE1E, LRRFIP2, RAPGEF2, RBM12B, SEMG2) (not shown). To summarize, the majority of the TGx-DDI biomarker genes (75%) exhibited behavior consistent with p53 regulation.

p53-dependence of biomarker responses.

As the majority of genes in the biomarker are regulated by p53, we predicted that biomarker responses to p53-activating chemicals would be altered in a predictable fashion by concurrent genetic manipulation of TP53 expression. To test this, biomarker responses were compared in 20 pairs of biosets in which gene expression was examined after exposure to p53-activating chemicals in cells treated with either scrambled siRNAs (“wild-type” cells) or siRNAs against TP53 (“p53 knockdown” cells). In 15 of the cases involving exposure to 8 chemicals, there were significant positive correlations to the biomarker in wild-type cells (Figure 4A-D). The significance of the correlations was diminished in all of these studies when p53 expression was knocked down, and in 8 of the cases, the significance of the correlations was abolished (-Log(p-value) < 4). One of the compounds that showed p53-dependence, RITA (reactivation of p53 and induction of tumor cell apoptosis), is a small molecule inhibitor of p53-MDM2 interaction (Figure 4D) and is thought to be the only known compound that binds directly to p53 ^{47 29}. Overall, these results indicate biomarker responses are p53-dependent.

Figure 4. p53-dependence of biomarker responses.

The biomarker was compared to pairs of biosets in which a chemical was examined for gene expression in cells treated with either scrambled siRNAs (“Wild-type” cells) or siRNAs against p53 (“p53 knockdown” cells). Chemical, time or concentration, and GEO number of each study are indicated. In GSE3178, the cell lines used in the comparisons are also indicated.

A. Effect of actinomycin D in HCT116 cells (GSE12459).

B. Effect of hydrogen peroxide (H2O2), hydroxyurea, or nitric oxide (NO) in HCT116 cells (GSE3176).

C. Effect of doxorubicin in four breast cancer cell lines (GSE3178).

D. Effect of RITA in HCT116 cells, MLN4924 in A375 cells, nutlin-3 in A375 cells, or doxorubicin in HCT116 cells.

Identification of p53-modulating chemicals in a human microarray compendium.

We performed a screen of our human microarray compendium to identify p53-activating chemicals. There were ∼9800 human cell line chemical vs. control comparisons encompassing ∼1950 chemicals. There were 98 chemicals in 274 biosets that induced a gene expression pattern with significant positive correlation to the biomarker (Table 1; Figure 5; Supplemental File 1). Regulation of the genes in the biomarker by the chemicals is shown in Figure 5, bottom, showing the similarity of the positive biosets to the biomarker in both direction and level of regulation. There were multiple biosets from exposures to 8 well known chemotherapeutic or environmental chemicals that resulted in a positive response (Supplemental Figure 4). Below we discuss the effects of specific chemical exposures on p53 activation. The p53-modulating chemicals identified in the CMAP and TG-GATES studies are discussed and shown in Supplemental File 2 including Supplemental Figures 2 and 3.

Table 1.

Chemicals identified in the study that activate p53.

Class	Chemicals
DNA damage	2-(Chloromethyl) pyridine hydrochloride, 2-amino-1-methyl-6-phenylimidazo(4,5-b)pyridine, 4-(acetoxymethylnitrosamino)-1-(3-pyridyl)-1-butanone, 4-Nitro-o-phenylenediamine, Aflatoxin B1, Benzo(a)pyrene, Chlorambucil, Cisplatin, Cyclophosphamide monohydrate, Ethyl Methanesulfonate, Ethylnitrosourea, formalin, fotemustine, Hycanthone, Lomustine, Methyl Methanesulfonate, Mitomycin C, Oxaliplatin, resorcinol, Tobacco smoke, Zalypsis
Topoisomerase inhibitors	Actinomycin D, Alpha-Amanitin, BMH-21, BMH-22, BMH-23, BMH-9, Camptothecin, Ciprofloxacin, clinafloxacin, Daunorubicin, Dexazosane, Doxohycanrubicin, Doxycycline, Etoposide, gatifloxacin, irinotecan, Tacrine, trovafloxacin
Antimetabolite	5109870, 4-amino-6-hydrazino-7-beta-D-ribofuranosyl-7H-pyrrolo(2,3-d)-pyrimidine-5-carboxamide, 5-Aza-CdR, 5-Fluorouracil, Azacitidine, ciclopirox, decitabine, deferasirox, Deferoxamine mesylate, Hydroxyurea, Methotrexate, Mycophenolic Acid, sangivamycin, Trifluridine
Oxidative stress	Acetaminophen, Bleomycin, Cbdive_003943, Hydrogen Peroxide, Nitric Oxide, Sodium Arsenite, Vitamin K 3
Microtubule disruptor	Colcemid, Docetaxel, Vincristine
Disruption of p53 signaling	benzyloxycarbonylleucyl-leucyl-leucine aldehyde, bortezomib, CG-1520, cidofovir, 7,8-diacetoxy-4-methylcoumarin, Enterobactin, geldanamycin, glucose, Mepacrine, MLN4924, nutlin-3, PHA 739358, Quercetin, resveratrol, Ribavirin, RITA, Thiostrepton, Y15
Miscellaneous	0175029–0000, Harmine, R547, Tetra(4-N-Methylpyridyl)Porphine
Unknown/Novel	Apigenin, Ethacrynic Acid, Metyrapone, monobenzone, OICR623, Ornithogalum glycoside, Oxyphenisatin Acetate, Phenoxybenzamine, Pyridoxine, pyrvinium, RPI-1, si-wu-tang, Sparstolonin B

Information about chemical exposure conditions are found in Supplemental File 1. Miscellaneous chemicals include those that block the cell cycle, act as kinase inhibitors, or inhibit telomerase.

Figure 5. Identification of p53-activating chemicals in a human microarray compendium.

There were∼1950 chemicals screened in ∼9800 human cell line chemical vs. control comparisons. (Top) The biosets were rank-ordered based on the -Log(p-value)s of the pair-wise comparisons with the biomarker. The chemicals that were predicted to activate p53 are indicated as red dots. Fifteen of the top-ranked chemicals are indicated. (Bottom) Heatmap of the expression of the biomarker genes in each of the comparisons. The expression of the 63 genes in the biomarker are shown to the left.

We determined the effects of the 16 genes in the biomarker that were not regulated in a p53-dependent manner. These genes were removed from the biomarker and then the resulting biomarker was compared to the original 274 biosets with positive correlations. Despite the lack of these genes, we found that only 19 biosets fell under our threshold of -Log(p-value =4), and those biosets were never below a value of 2.4. Thus, 93% of the bioset positives were also called positive without the 16 genes (Supplemental Figure 5). These results support the fact that these genes do not appreciably affect the biomarker results consistent with their apparent p53-independent regulation.

Information about the known or possible mechanisms of DNA damage or p53 activation by the 98 chemicals are summarized in Table 1 and in greater detail in Supplemental File 1. For most of the chemicals, there is evidence for direct DNA damage. Other chemicals may act as DNA topoisomerase inhibitors, DNA synthesis blockers by acting as antimetabolites, inducers of indirect DNA damage through increases in oxidative stress, and microtubule disruptors. A number of miscellaneous mechanisms were also found for 4 chemicals including those that block the cell cycle, act as kinase inhibitors, or that inhibit telomerase.

For many of the chemicals, there is evidence for alteration in p53 expression, stability, and/or post-translational modifications (Table 1). The proteasome inhibitor bortezomib stabilizes p53 protein and increases p53 activity through an unknown mechanism ⁴⁸; cidofovir stabilizes p53 by suppression of the expression of human papilloma virus E6 and E7 proteins that bind and neutralize the activity of p53 ⁴⁹; the histone deacetylase inhibitor CG-1521 prevents the deacetylation at p53 Lys373 and increases p53 activity ⁵⁰; Y15 inhibits the ability of the negative regulator tyrosine kinase Fak to interact with p53 ⁵¹; DAMTC (7,8-diacetoxy-4-methylcoumarin), geldanamycin, mepacrine, and thiostrepton increase the expression of p53 ^{52 53 54 55}; PHA 739358 blocks p53 degradation by blocking interaction with MDM2 ⁵⁶; benzyloxycarbonylleucyl-leucyl-leucine aldehyde stabilizes p53 protein through proteasome inhibition ⁵⁷; enterobactin may induce p53 by inhibition of the cell cycle ⁵⁸; in at least one condition, glucose can upregulate p53 transcriptional activity ⁵⁹; MLN4924 may induce p53 by increasing nucleolar stress ^60,61; quercetin stabilizes the p53 protein ⁶²; resveratrol upregulates the expression and phosphorylation of p53 ⁶³. RITA and nutlin-3 were discussed above.

A number of chemicals or chemical mixtures were identified that appear to be novel p53 activators, as there is no existing evidence that they activate p53 or damage DNA. These included apigenin, ethacrynic acid, metyrapone, monobenzone, OICR623, ornithogalum glycoside, oxyphenisatin acetate, phenoxybenzamine, pyridoxine, pyrvinium, RPI-1, si-wu-tang, and sparstolonin B. Some of these compounds have been or are currently used as drugs. Acetaminophen is an analgesic, metyrapone is occasionally used in the treatment of Cushing’s syndrome, monobenzone is used for medical depigmentation, phenoxybenzamine is used to treat hypertension, pyrvinium is an anthelmintic effective for pinworms, oxyphenisatin acetate is a laxative, OICR623 is a Wnt signaling inhibitor ⁶⁴, and RPI-1 is a RET tyrosine kinase inhibitor ⁶⁵. Apigenin is a flavone found in many plants. Pyridoxine is a form of vitamin B6 and used as a dietary supplement. Ornithogalum glycoside, si-wu-tang, and sparstolonin are herbal mixtures. Further studies are needed to determine how these chemicals/mixtures cause p53 activation. These compounds may be DDI or initiate TGx-DDI through non-genotoxic mechanisms. It is important to note that false positive DDI calls may result from overtly cytotoxic exposure conditions, and it is important to consider cytotoxicity in evaluating the results. Overall, our analysis reveals how the TGx-DDI biomarker may be used in chemical screening.

Comparison of biomarker predictions to those identified by a p53 high throughput screen.

We compared the results of the biomarker approach with a recent screen for p53-activating chemicals carried out in human colon carcinoma HCT-116 cells ³⁵. There were 629 chemicals that were evaluated by both methods. Microarray data came from the CMAP study and 59 additional studies. Fourteen chemicals were positive in both assays, 16 chemicals were positive using the biomarker approach only, and 19 chemicals were positive using the p53 screening assay only (Figure 6). There were 580 chemicals that were negative in both assays.

Figure 6. Comparison of the results of the TGx-DDI biomarker approach to an HTS p53 activation assay.

Overlapping among active chemicals in our biomarker approach from ∼60 microarray studies and an HTS screen for p53-activating chemicals ³⁵ were compared. Chemicals that were identified using either or both approaches are indicated. The 578 chemicals that were negative in both assays are not shown.

The 14 chemicals positive in both assays included a number which directly or indirectly damage DNA by alkylating DNA, acting as DNA crosslinkers, causing DNA strand breaks, inhibiting DNA methylation, or inhibiting topoisomerase (chlorambucil, daunorubicin, decitabine, etoposide, hycanthone, methotrexate, mitomycin C, oxaliplatin (Supplemental File 1)). Three of the compounds increase the expression or stabilization of p53 protein (bortezomib, ribavirin, thiostrepton). Four other compounds act as metabolic inhibitors or microtubule disruptors (5-fluorouracil, mycophenolic acid, trifluridine, vincristine). Apigenin, ethacrynic acid, and oxyphenisatin acetate have no known mechanism of action.

Out of the 16 chemicals that were positive using the biomarker only, 6 are known to damage DNA directly (2-(chloromethyl) pyridine hydrochloride, cisplatin, cyclophosphamide monohydrate, ethyl methanesulfonate, formalin, methyl methanesulfonate). Among the others, hydrogen peroxide and acetaminophen causes oxidative stress, gatifloxicin is a topoisomerase inhibitor, and actinomycin D inhibits topoisomerase. Metyrapone has no clear linkage to induction of DNA damage or p53 activation. More than half of these 16 compounds were tested at concentrations >100 uM, the highest concentration tested in the p53 HTS assay. For example, acetaminophen was active at 1 mM, a concentration 10 times higher than the limit concentration in the HTS assay. While it is not completely clear why, aside from tested concentration levels, these compounds would be active using the biomarker but not active in the p53 assay, we note that cyclophosphamide monohydrate was only active in hepatocytes with the ability to metabolize the compound to the active genotoxicant ⁶⁶, which may not occur in HCT-116 cells.

Compounds that were positive in the p53 HTS assay but negative using the biomarker approach included two that may be directly genotoxic (azathioprine, chlormadinone acetate), three that increase p53 expression or acetylation (chlorpromazine, niclosamide, riluzole), three that may increase oxidative stress-induced DNA damage (albendazole, atovaquone, pentachlorophenol), and five that act as microtubule destabilizing agents (albendazole, mebendazole, nocodazole, fenbendazole, parbendazole) (Supplemental Table 2). There were 7 compounds with no clear mechanism of p53 activation or DNA damage (benzbromarone, bisacodyl, dienestrol, diethylstilbestrol, enzastaurin, nifuroxazide, rimexolone). Differences in identifying p53-activating chemicals could be due to differences in responsiveness between the HCT-116 cell line and the different cell lines evaluated using the biomarker. Differences may also be due to the characteristics of the assays: the biomarker approach often evaluated chemicals at one dose and time point in the absence of any information about cytotoxicity. The Tox21 assay examined p53 responses at multiple doses at 16 h and in parallel identified doses that were cytotoxic. Compounds that were negative in the HTS assay may exhibit peaks of DNA damage within hours of exposure initiation that would not be observed by 16 h. Most of the chemicals positive in the p53 HTS assay but negative in the biomarker approach were examined in the biomarker approach at 6 h only, a time point that may not be of sufficient duration to induce secondary responses such as oxidative stress, a contributing factor in DNA damage for a subset of the chemicals.

Summary

The development of accurate in vitro molecular biomarkers for regulatory toxicology is a priority to improve testing strategies for human health risk assessment. Previous work has extensively characterized the TGx-DDI biomarker used in conjunction with a number of computational methods as a useful strategy for identifying chemicals that damage DNA ^23–28 and activate p53. Here, we showed through a number of lines of investigation that the biomarker coupled with the Running Fisher test can identify p53-activating chemicals. 1) The biomarker predicted p53 activation after exposure to chemicals known to activate p53 in the absence of DNA damage, including nutlin-3, which disrupts p53-MDM2 interactions, or known to intercalate DNA. 2) By comparing the gene expression pattern of the biomarker genes to profiles from cells in which TP53 was either overexpressed or knocked down by siRNA, we found that most of the genes in the biomarker (75%) were dependent on p53. 3) The induction of p53 by a number of chemicals was decreased or abolished in TP53 knockdown cells. The biomarker was used to identify p53-activating compounds in a compendium of ∼9800 human cell line chemical vs. control comparisons encompassing ∼1950 chemicals. We identified 98chemicals that caused p53 activation. Activators included those that are known to damage DNA directly or indirectly or to modulate p53 expression or signaling. A small number of identified chemicals or mixtures appear to be novel DDI/p53 activators.

Overall, these results demonstrate the utility of the biomarker approach to identify chemicals that damage DNA and activate p53 through one or more of the large number of known mechanisms and further support a strategy using well characterized and validated biomarkers to identify chemicals that modulate key events in the network of AOPs. The findings clarify that both DDI and non-DDI (p53 targeting) mechanisms can lead to a positive call for TGx-DDI. Thus, evaluation of a positive TGx-DDI result in a chemical screen should consider both mechanisms through structure-activity analysis to determine what further testing is required.

Supplementary Material

Supplemental file 2

Supplemental File 2. Contains supplemental sections: 1) “Identification of p53-modulating chemicals in MCF-7 cells.”, and 2) “Identification of p53-modulating chemicals in human primary hepatocytes”. This file also contains supplemental figures: 1) Supplemental Figure 1. Canonical pathways that significantly overlap with the biomarker genes, 2) Supplemental Figure 2. Identification of p53-activating chemicals in MCF-7 microarray comparisons from the CMAP study, 3) Supplemental Figure 3. Identification of p53-modulating chemicals in human primary hepatocytes, 4) Supplemental Figure 4. Regulation of genes in the TGx-DDI biomarker by chemicals predicted to activate p53, and 5) Supplemental Figure 5. Effect of removal of 16 p53-independent genes on biomarker scores.

Supplemental File 1

Supplemental File 1. Contains information about the microarray comparisons which exhibited a positive response examined in this study and treatments of chemicals found to be negative by the biomarker but positive in the Tox21 p53 assay.

Abbreviations

AOPs

adverse outcome pathways

BSCE

BaseSpace Correlation Engine

CMAP

Connectivity Map

DDI

DNA damage inducing

HTTr

High-throughput transcriptomic

KEs

key events

MIEs

molecular initiating events

NDDI

non-DDI