Identifying Gene Predictors of Chemicals Linked With Breast Cancer: A Machine Learning Analysis of MCF7 Cellular Transcriptomic Screening Data

ABSTRACT

Breast cancer is the most prevalent cancer in women and has been linked to exposure to environmental chemicals. However, many chemicals have not been evaluated for relationships with this outcome. In this study, we analyzed RNA sequencing data from human breast cancer‐derived MCF7 cells exposed to hundreds of individual chemicals. These chemicals were binned into three categories: (1) chemicals with known associations to breast cancer (BCs); (2) chemicals with a lack of relationship to breast cancer (NBCs); and (3) chemicals that remain understudied for breast cancer risk (UCs). Machine learning models were trained to discriminate between BCs and NBCs based on transcriptomic and physicochemical property data. The best model yielded a balanced accuracy of 80% and was applied to the UCs. A total of 170 genes were found to contribute to model performance, including Claspin (CLSPN), Runt‐related Transcription Factor 2 (RUNX2), and Ubinuclein 2 (UBN2). These genes further informed enriched pathways relevant to inflammation, ferroptosis signaling, and cell proliferation. Additionally, 97 UCs were predicted to be more analogous to BCs, including select biocides and dyes. To ground results in human population data, expression profiles for the 170 genes were assessed in tumor samples from The Cancer Genome Atlas, yielding overlap in human cancer‐relevant alterations and in vitro chemical‐induced alterations. Collectively, this study addresses a gap related to understanding which chemicals may be of interest for further characterization of breast cancer risk by prioritizing chemicals and underlying mechanisms using high‐throughput transcriptomic screening data.

1. Introduction

Breast cancer is a highly prevalent disease that is currently increasing in incidence and associated mortality in some developed countries, including China and South Korea (Lei et al. 2021). Furthermore, breast cancer risk is now recognized as heavily impacted by an individual's environment. For example, certain environmental chemicals have been evaluated through epidemiological study designs and identified as having potential relationships to human breast cancer incidence (Eldakroory et al. 2017; Khanna et al. 2014; White et al. 2019). Still, the majority of chemicals in the environment have yet to be evaluated for their potential impact on breast cancer risk. To put this in perspective, approximately 100,000 substances have been inventoried for commercial use in the United States (Isaacs et al. 2022). Tens of thousands of these substances are thought to be used extensively, produced at over one ton per year (Egeghy et al. 2012; Isaacs et al. 2022; Muir and Howard 2006). Further complicating this issue is the fact that breast cancer‐related biological alterations resulting from environmental chemical exposures are not fully understood. In order to more efficiently investigate health outcomes, including breast cancer, associated with environmental chemical exposures, new approach methodologies (NAMs) such as in vitro high‐throughput screening (HTS), ‐omics technologies, and predictive modeling via machine learning reflect essential tools now poised to address these gaps (Schmeisser et al. 2023).

Breast cancer is recognized as a heterogenous disease, wherein patients commonly exhibit varying clinical presentations with underlying differences in cancer cellular morphologies and genomic, transcriptomic, epigenomic, and proteomic signatures (Guo et al. 2023). Still, common patterns of cell morphology, protein markers, and gene expression have yielded molecular subtypes including (1) luminal A, (2) luminal B, (3) HER2‐enriched, and (4) basal‐like (Johnson et al. 2021; Tsang and Tse 2020). Many published reviews explore these subtypes in depth, but generally, subtypes are determined by expression patterns of genes relevant to estrogen and progesterone receptor (ER and PR) signaling and the HER2 oncogene (Johnson et al. 2021; Tsang and Tse 2020). Understanding these subtypes can prove instrumental in the design of treatment strategies and investigations of how environmental exposures may initiate or progress disease (Johnson et al. 2021; Tsang and Tse 2020).

In a recent attempt to better understand and characterize carcinogens, the International Agency for Research on Cancer (IARC) working group identified 10 key characteristics that are common between established carcinogens (Guyton et al. 2018; Smith et al. 2016). Broadly, these key characteristics pertain to genotoxicity and genomic instability, oxidative stress and inflammation, epigenetic alterations, immunosuppression, receptor‐mediated effects, and immortalization and cell proliferation (Guyton et al. 2018; Smith et al. 2016). Given the well‐documented linkage between breast cancer development and endocrine function, carcinogens that operate through receptor‐mediated effects are of high relevance (Kay et al. 2024). Specifically, endocrine disrupting chemicals (EDCs) that affect ER and PR pathways are of high importance to study in relation to breast cancer risk, though other relevant mediators include the peroxisome proliferator‐activated receptor (PPAR) and the aryl hydrocarbon receptor (AhR) (Kay et al. 2024; Smith et al. 2016). Nonetheless, other key characteristics linked to breast cancer pertain to genotoxicity, epigenetic alterations, and cell proliferation, cell death, or nutrient supply, and these characteristics should also be considered when screening substances for breast cancer risk (Kay et al. 2024; Smith et al. 2016). Notably, many of these processes can be captured through the use of ‐omics based screening methods such as transcriptomics.

The task of evaluating chemicals for potential impact on breast cancer risk is daunting, particularly when considering the large landscape of chemicals to which humans are exposed, as well as the diverse possible mechanisms underlying carcinogenesis. Recent toxicological efforts have been organized to more rapidly evaluate substances for potential toxicity and resulting disease hazard, while informing underlying biological mechanisms. For example, in vitro HTS assays, including some using MCF7 breast cancer cells, and computational models have been leveraged to investigate chemical bioactivity across many endpoints, including ER activity and genotoxicity (Harrill et al. 2021; Kavlock et al. 2012; Mansouri et al. 2016; Thomas et al. 2019). Furthermore, predictive modeling that incorporates machine learning algorithms has previously been used to predict carcinogenicity of chemicals based on structure, breast cancer status in human cohorts, and biomarkers of various diseases using transcriptomic data (Ahmed 2020; Fjodorova et al. 2010; Khatri and Bhasin 2020). While these studies reflect valuable contributions to the field, there remains a large data gap in chemicals screened for breast cancer hazard, and there is a further need to elucidate mechanisms by which certain chemicals may promote the initiation or progression of breast cancer.

This study is designed to leverage a recently published high‐throughput transcriptomic (HTTr) screen of 1751 chemicals in luminal A breast cancer‐derived MCF7 cells (Harrill et al. 2024b) with the novel lens of identifying chemicals in the environment that have the potential to impact molecular targets that may alter breast cancer progression and/or initiation using machine learning. Specifically, we aimed to integrate these NAMs‐based tools to identify understudied chemicals for further screening based upon in vitro HTS‐derived transcriptomic signatures similar to signatures predictive of known breast cancer associated chemicals. We hypothesized that genes involved in breast cancer etiology, including those regulating cell cycle, cell death, DNA integrity, and endocrine signaling, would serve as informative in vitro‐based predictors used in the classification of breast cancer relevant agents, while informing mechanisms through which known carcinogens and lesser studied compounds may influence breast cancer hazard in humans. Findings were further anchored to disease etiology using human tissue transcriptomic profiles collected from breast cancer patients and made available through The Cancer Genome Atlas (TCGA) (Koboldt et al. 2012). Implications of this research are also broader in scope, where this approach can be pivoted to address different environmental health questions based on similar in vitro transcriptomic signature data and disease association classifications to improve our general understanding of environmental influences on human health outcomes. Overall, this study set out to leverage recent transcriptomic screening data of hundreds of chemicals in MCF7 cells in a machine learning framework with the goal of identifying potential chemicals in the environment with the ability to interfere with molecular targets that may alter breast cancer progression and/or initiation.

2. Materials and Methods

2.1. Study Overview

In this study, we aimed to integrate novel in vitro high‐throughput transcriptomics (HTTr) screening data from MCF7 cells exposed to hundreds of chemicals, physicochemical property predictions, and chemical associations with breast cancer from various databases to develop machine learning algorithms capable of differentiating chemicals with and without a putative association to breast cancer. The best machine learning model was applied to understudied chemicals (UCs), that is, those with insufficient or conflicting data to assess breast cancer hazard and was also used to inform novel biomarkers of disease. A high‐level flow chart of these analyses and approaches is shown in Figure 1. Further details surrounding each step are provided below. All cell culture experimental steps, cell viability measures, and targeted RNA sequencing methods have been published previously (Harrill et al. 2024a, 2024b; Harrill et al. 2021) and are summarized below. New data analysis methods are described in detail afterwards to address this specific study's goals.

FIGURE 1.

A schematic diagram overviewing the methods of this analysis. They included (1) dataset generation and acquisition, (2) random forest modeling to identify understudied chemicals with properties more similar to chemicals with a putative association with breast cancer, and (3) pathway analysis run on informative genes from the random forest modeling and assessment of expression patterns of these genes in a human populations. The green check indicates chemicals that have been tested but do not show elevated risk for breast cancer risk, the pink tumor indicates chemicals that do have a known association with increased breast cancer risk, and the yellow question mark indicates chemicals that have not been thoroughly screened for breast cancer related outcomes.

2.2. Chemical Exposures and High‐Throughput Transcriptomics Screening

For the HTTr screening study, hormone receptor positive luminal A HTB‐22 MCF7 human breast adenocarcinoma cells were cultured and exposed to hundreds of individual chemicals, including those of environmental relevance as previously described (Harrill et al. 2021, 2024a).

In summary, 1751 unique chemicals were selected from the ToxCast chemicals collection for HTTr screening. Unique chemicals were identified using Distributed Structure‐Searchable Toxicity Database Substance Identifiers (DTXSIDs) (Grulke et al. 2019; Richard et al. 2016). A total of eight concentrations were assessed using 1/2 log₁₀ spacing, and assays were incubated for 6 h prior to the generation of cell lysates used for TempO‐Seq analysis. Upon cell viability assessment, a portion of substances was omitted from the analysis of the TempO‐Seq data, and assay wells with corresponding concentrations were excluded from the analysis of HTTr data, as previously described. Additionally, a random subset of 34 chemicals were selected as screening duplicates, and for these substances, samples labeled with EPAPLT were selected for the machine learning analysis as described below. This strategy was employed to only retain a single sample for the chemicals screened in duplicate. Notably, transcriptomic response signatures have been previously compared across screening duplicates, and responses to chemical exposures were identified as largely consistent (Harrill et al. 2024b). Full methods for chemical exposure, sample processing, and transcriptomic data generation were previously described (Harrill et al. 2021).

MCF7 cell lysates were analyzed using a custom‐attenuated version of the TempO‐Seq human whole transcriptome version 1 assay by BioSpyder (Yeakley et al. 2017) using previously described methods (Harrill et al. 2021, 2024a, 2024b). This assay covers 19,287 genes with 21,111 probes. Analysis was conducted with a target sequencing depth of three million mapped reads on average per sample.

2.3. RNA‐Sequencing Data Processing and Differential Expression Analysis

FASTQ files for each well containing raw TempO‐Seq reads were processed using a previously established bioinformatics pipeline (Harrill et al. 2021, 2024a). Moderated log₂ fold changes (L2FCs) were estimated for each probe and each concentration of each test chemical. As the TempO‐Seq whole transcriptome assay contains some probes that target the same gene, L2FC results were aggregated to the gene level by selecting the average L2FC of probes across all test conditions.

2.4. Gene‐Level Concentration Responses and Area Under the Curve Calculations

Gene‐level concentration response curves were fit using the tcplfit2 package with the L2FCs as inputs (Sheffield et al. 2022). A constant model, a Hill model, 2 polynomial models, a power model, and 4 exponential models were all evaluated. The model with the lowest Akaike Information Criterion (AIC) was selected as the best fit model (Bozdogan 1987; Sheffield et al. 2022). This curve fitting pipeline returns four important parameters including (1) a continuous hitcall (a value between zero and one) which aims to evaluate the confidence that the observed concentration‐response effect is not due to assay noise, (2) a benchmark dose value (BMD), which is the concentration value where the response exceeds an estimate of the background noise (the cutoff), (3) a “top” which reflects the directional value of the largest modeled response, and (4) the top/cutoff (TC) value, which is a signal to noise ratio calculated by dividing the top value by the cutoff (Sheffield et al. 2022). Cutoffs were calculated as the 95th percentile of L2FC values for the lowest two concentrations across all chemicals for the corresponding gene (Harrill et al. 2021). For each chemical, genes with a continuous hitcall > = 0.9, a TC ≥ 1, and a BMD less than the highest tested concentration were deemed active. Next, area under the curve (AUC) values were calculated for all chemical/gene pairs as a way to summarize both potency and effect size into a single metric. Multiple AUC values were calculated as described in Harrill et al. (2024a) and (2024b), and this analysis specifically used the sign scaled AUC (ssAUC) as calculated by Equation (1). The ssAUC values provide information about the direction of the curve fit and thus were preferred to the non‐sign scaled AUC, as the ssAUC may be helpful in differentiating chemical categorizations, as described below.

$$\text{ssAUC}=3-{\log }_{10}{\left(\mathrm{BMD}\right)}^{*}{\left(\mathrm{TC}\right)}^{*}\left(\mathrm{sign}\left(\mathrm{top}\right)\right)$$ (1)

Specifically, in Equation (1), the direction, or sign, of the top (i.e., largest modeled response) is accounted for, along with the BMD and TC (i.e., signal to noise ratio) to calculate an AUC value that considers this directional element. From this equation, inactive genes received an ssAUC value of zero. Consequently, ssAUC values were used as surrogates for gene expression, indicating bioactivity, in the modeling efforts in this work and are often referred to as “gene expression” features.

2.5. Categorizing Evaluated Chemicals Based Upon Associations to Breast Cancer

With the end goal of developing a model that could differentiate substances with and without an association with molecular events leading to tumor progression relevant to breast cancer, we leveraged a previously published categorization schema which binned ~6800 chemicals into three categories reflective of their putative association with breast cancer (Koval et al. 2022). These categories include chemicals with known associations with breast cancer (and thus labeled BCs); chemicals with a demonstrated lack of relationship to breast cancer (non‐breast cancer chemicals [NBCs]); and chemicals that remain understudied for this risk (understudied chemicals [UCs]). The classifications were assigned using the US EPA's Toxicity Reference Database (ToxRefDB) and the Silent Spring Institute's Mammary Carcinogens Review Database (Koval et al. 2022; Rudel et al. 2007; Watford et al. 2019). ToxRefDB is an opensource database that houses in vivo study results, largely of data from animal studies with a focus on pharmaceutical, agrochemical, and industrial chemicals (Watford et al. 2019). The Mammary Carcinogens Review Database aggregates findings from the International Agency for Research on Cancer Monographs, Carcinogenic Potency Database, U.S. National Toxicology Program (NTP), NTP 11th Report on Carcinogens, and the Chemical Carcinogenesis Research Information System Database and contains 216 chemicals of concern (Rudel et al. 2007). Additionally, a 2024 publication by the Silent Spring Institute provides an updated list of potential mammary carcinogens based on genotoxicity and endocrine activity in vitro, and these substances were also considered in this analysis (Kay et al. 2024). Specifically, chemical DTXSIDs identified as “MCs,” or as having met “criteria for evidence of inducing mammary tumors” within Kay et al. (2024) were added to the list of chemicals considered BCs in the current study. Chemicals originally identified as NBCs or UCs from Koval et al. (2022) that appeared as breast cancer relevant according to Kay et al. (2024) were omitted from this analysis due to conflicting categorizations.

2.6. Physicochemical Property Data Organization

Physicochemical property data have previously demonstrated utility toward improving predictive modeling of chemical‐induced toxicity (Klaren et al. 2019; Mansouri et al. 2018; Ring et al. 2021). As such, we incorporated physicochemical property data for chemicals included in the machine learning analysis in the form of OPEn structure–activity Relationship App (OPERA) predictions. OPERA predictions were specifically acquired through a batch search of DTXSIDs through the US EPA's CompTox Chemicals Dashboard (EPA 2024). Predicted properties included atmospheric hydroxylation rate, bioconcentration factor, biodegradability half‐life, boiling point, fish biotransformation half‐life, Henry's Law constant, melting point, octanol/air partition coefficient, octanol/water partition coefficient, soil adsorption coefficient, vapor pressure, and water solubility, which we hypothesize may affect pharmacokinetic properties and thus may be useful as predictors in this analysis (Mansouri et al. 2018).

2.7. Data Integration and Filtering

After establishing the gene expression dataset, the chemical categorization dataset, and the physicochemical property dataset, the data were filtered and pre‐processed. As described below, filters were applied to the datasets both before and after merging. Data organization and cleaning steps were performed in R version 4.3.2 using the dplyr package version 1.1.4. First, the ssAUC data and the categorization data were merged and filtered to only retain chemicals that had an ssAUC > 0 for at least 200 genes. This threshold was selected to retain a sufficiently large set of chemicals for analysis while omitting substances that did not induce significant biological activity in the MCF7 cells. Next, a filter was applied to remove genes that did not demonstrate activity after exposure in at least 10% of the remaining chemicals. This filter was applied to focus our predictive model on genes that show consistent responses to at least a moderate subset of chemicals. Finally, the physicochemical property predictions were merged into the dataset. Using the base R scale function, both physicochemical properties and ssAUCs were scaled for all chemicals (BCs, NBCs, and UCs) such that each feature had a mean of 0 and standard deviation of 1. The scaled feature data for the BCs and NBCs were carried forward for predictive model development. As one of the goals of this analysis was to predict whether certain UCs exhibited bioactivity and/or physicochemical properties more similar to BCs or NBCs, the scaled features for the UCs were saved for later evaluation.

2.8. Building Machine Learning Models to Predict BCs vs. NBCs Classifications

Random forest (RF) classifiers were trained to discriminate between BCs and NBCs using the transcriptomic and physicochemical property features as predictor variables and breast cancer classification as the outcome variable. A RF algorithm specifically develops an ensemble of individual decision trees using random sets of features from the predictor space; then final classification for an observation is determined by majority vote of the individual trees (Parmar et al. 2019; Wang et al. 2020). RF was selected for its demonstrated utility in evaluating HTS data as well as the available variable importance calculations, which were used to inform biological significance (Ring et al. 2021; Riniker et al. 2014). To further detail, three different RF modeling scenarios were trained using (1) only the gene expression ssAUC data, (2) only the physicochemical property data, and (3) both the gene expression ssAUC data and physicochemical property data. All models were trained and tested using the R caret package version 6.0‐94, which leverages the randomForest package (Kuhn 2008; Liaw and Wiener 2002).

A 70‐30 train‐test split of the chemicals was generated using the createDataPartition function in the caret package. This function splits the observations in a dataframe into groups that should be used for training and testing. Once identified, the same split was used in each of the modeling scenarios described above. For all models, majority class downsampling was used to account for the observed class imbalance in the training data, with roughly two times more NBCs than BCs. Downsampling occurred during the training by setting the sampling hyperparameter in the trainControl function to “down.” Additionally, leave‐one‐out cross‐validation was employed during the training process. This approach allows each observation to serve as the test set during model training and validation, thus reducing bias by letting all data points contribute to model development (Yates et al. 2023). Finally, a grid search of p, p/2, and sqrt(p) features were tested, where p reflects the total number of features considered, yielding three potential models per modeling scenario. Seeds for random number generation were continually set throughout this process to ensure reproducibility.

For each modeling scenario, a final model was selected from the respective grid search of p, p/2, and sqrt(p) features by identifying the model with the highest balanced accuracy. Balanced accuracy was selected as the main performance metric due to its utility in assessing datasets with a class imbalance, as was observed in our dataset (Jiao and Du 2016). To identify which of the three modeling scenarios produced the best performing model overall, the three final models were compared to one another with balanced accuracy serving as the primary metric of comparison, though additional consideration was given to the sensitivity and specificity (Jiao and Du 2016). A comparison of the results of the final model from each scenario was visualized using the R package ggplot2 version 3.5.0. The best overall model was then applied to the UCs with the goal of predicting additional chemicals that alter the same biological mechanisms as the BCs.

2.9. Evaluation of Genes With Highest Importance for Model Predictions

To investigate biological mechanisms and identify potential biomarkers of disease, we selected the genes that were most informative in differentiating the BCs and NBCs for further investigation. During the training paradigm, the mean decrease in accuracy for each feature was calculated as an indicator of variable importance and visualized for the best overall model. Since not all features are included in all trees during training, the mean decrease in accuracy informs which features were most important for successfully differentiating the outcome classes across the ensemble of trees (Archer and Kimes 2008). The variable importance data was filtered to retain only genes that had a mean decrease in accuracy greater than zero, resulting in a list of top‐ranking predictor genes.

Next, using the original ssAUC values prior to scaling, the mean ssAUC was computed for the top‐ranking predictor genes for the BCs and the NBCs. The difference between the mean for the BCs and NBCs was calculated. This delta informed whether expression was generally upregulated or downregulated in the BCs compared to the NBCs and was considered during pathway enrichment analysis. Specifically, the top‐ranking genes and corresponding deltas were analyzed for relationships to canonical pathways and upstream regulators according to the QIAGEN Ingenuity Pathway Analysis Knowledgebase (Kramer et al. 2014). A modified Fischer's Exact test was considered for statistical enrichment with multiple testing correction through the Benjamini‐Hochberg (BH) procedure (Benjamini and Hochberg 1995). Pathways and upstream regulators were deemed statistically enriched at a BH adjusted p value < 0.05. For upstream regulators specifically, substances identified as “chemical drug” were omitted to allow for a more focused screen of endogenous regulators. These methods parallel those implemented in our recent studies (Carberry et al. 2022; Ring et al. 2021).

2.10. Sensitivity Analyses to Validate Model Performance

To validate the findings from our best overall model, two sensitivity analyses were run. First, to ensure that the best overall model did not achieve the reported performance metrics by chance, 100 models were trained using the training set from the best performing model, but with the outcome variables randomly shuffled. This approach effectively serves as a control as it provides an estimate of the performance that can be achieved by chance on the given data. Notably worse performances from the shuffled outcome models increase confidence in the results of our best performing model (Bender et al. 2022; Bundy et al. 2022). Additionally, to verify measures of feature importance which informed genes for which we ran pathway‐level analyses, three new features were generated by randomly sampling with replacement data points from the full distribution of scaled ssAUCs and physicochemical property predictions from all chemicals in the training set. These “dummy” features were integrated with the original features from the best overall model and a new model was trained. The resulting feature importance based on mean decrease in accuracy was compared to the best overall model, ensuring the dummy features were not the most informative. This approach was used to increase our confidence that the top‐ranking features that could most effectively differentiate between BCs and NBCs in the best overall model were indeed informative and not just appearing by chance. The caret package was used for model training, and the base R sample function was used to generate the random noise variables. Results were visualized using ggplot2.

2.11. Relationships to Disease Outcomes

To evaluate findings in the context of disease outcomes, expression patterns of the most informative genes from the MCF7 cell analysis were examined in The Cancer Genome Atlas (TCGA) breast cancer data. Using the TCGAbiolinks package v2.30.0, RNA sequencing (RNA‐seq) data from the Breast Invasive Carcinoma (TCGA‐BRCA) project was acquired along with corresponding sample demographic and molecular subtype data (Colaprico et al. 2016; Koboldt et al. 2012). The acquired TCGA RNA‐seq data were processed through the Spliced Transcripts Alignment to a Reference (STAR) workflow (Dobin et al. 2013) and filtered to only include the normalized gene expression data for the genes of interest. These values were pseudo‐log₂ transformed (log₂(x + 1)), then the dataset was split by molecular subtype (luminal A, luminal B, Her2, and basal).

For each gene, an analysis of variance (ANOVA) was performed evaluating gene expression versus subtype to see if the genes identified as important in the analysis of the luminal A, MCF7 cells were differentially expressed among the subtypes in a human cohort. This analysis was conducted using the aov function in the stats package with the formula expression ~ subtype. Results were adjusted for multiple testing using the BH procedure (Benjamini and Hochberg 1995). Next, a principal components analysis (PCA) was run using the pseudo‐log₂ transformed data for the statistically significant genes. The prcomp function in the stats package was used, and results were visualized using ggplot2 and the fviz_pca_var function in the factoextra package v1.0.7. Additionally, a secondary ANOVA was run using the formula expression ~ subtype + race to rule out race as a confounder. This required an additional filter of the input data. Specifically, races/ethnicities with fewer than 5 individuals represented were dropped, as well as individuals with unknown/non‐reported race/ethnicities. This filter was applied to remove groups for which robust statistical analysis could not be adequately performed. A PCA was then conducted on this subset of the data for the corresponding significant genes using the same methods as above.

3. Results

3.1. Integration of Transcriptomic and Chemical Categorization Data

A total of 1751 chemicals were previously screened in human MCF7 breast adenocarcinoma cells at eight different concentrations for 6 h, and gene expression changes were captured in a HTTr assay covering ~20,000 genes. A total of 9120 unique genes were sufficiently expressed in this cell type at levels that allowed for meaningful analysis (non‐zero ssAUC for at least 10% of chemicals used in model training). Cell viability and quality control results are detailed in full in Harrill et al. (2024a) and (2024b). In brief, 29 of the 1751 chemicals resulted in decreased cell viability, and 255 samples out of 45,266 were removed from the HTTr assay due to failing QC criteria. Concentration‐response modeling was performed for each chemical‐gene pair, and a best fit model was identified. An ssAUC value was then calculated. The minimum and maximum ssAUCs across all chemical‐gene pairs were −63.94 and 81.26, respectively, with a mean of −0.0389. The vast majority (~96%) of ssAUCs were zero, indicating no concentration‐dependent change in expression for the chemical‐gene pair. Still, nearly 60,000 chemical‐gene pairs demonstrated a non‐zero ssAUC.

A total of 6793 chemicals were classified according to their putative association to breast cancer in Koval et al. (2022). Of these chemicals 78 were categorized as BCs, or having an association with increased breast cancer risk, 409 were categorized as NBCs, or having been studied, but lacking an association with increased breast cancer risk, and the remaining 6306 were categorized as UCs, or lacking data or being understudied with respect to breast cancer risk. A total of 279 chemicals were identified as having met “criteria for evidence of inducing mammary tumors” from Kay et al. (2024). When combined and filtered, a total of 281 chemicals were classified as BCs, 262 chemicals were classified as NBCs, and the remaining 631 were classified as UCs.

The gene expression and chemical categorization information were merged together using DTXSID as the primary chemical identifier, resulting in a total of 914 DTXSIDs (91 BCs, 192 NBCs, 631 UCs) with data spanning both domains. After removing chemicals that exhibited a non‐zero ssAUC in < 200 genes, 218 DTXSIDs remained (29 BCs, 56 NBCs, 134 UCs). We also removed 3325 out of 9120 genes that did not exhibit activity after exposure to at least 10% of the remaining chemicals, yielding 5793 genes carried forward into the analysis. After merging in the physicochemical property predictions (n = 12), 21 DTXSIDs were removed for not having complete coverage, yielding a total of 200 DTXSIDs (23 BCs, 51 NBCs, 126 UCs) that were scaled and carried forward into the analysis. A full list of chemicals and classifications can be found in Table S1. ssAUC values for the 200 chemicals used in the machine learning analysis are shown in Table S2, and complete physicochemical property predictions can be found in Table S3.

3.2. Results of RF Models to Predict Chemical Categorization

Three RF modeling scenarios were trained to differentiate 23 BCs and 51 NBCs using (1) only the ssAUC data, (2) only the physicochemical property predictions, and (3) both the ssAUC data and physicochemical property predictions combined. For consistency, the same chemicals served in the training and testing sets for all modeling scenarios. Additionally, majority class downsampling was used during training to account for the fact that there were more than double the number of NBCs than BCs. Leave‐one‐out cross‐validation was also applied during model training to reduce bias, and a grid search was used to optimize the number of features in each model, with balanced accuracy serving as the primary metric to evaluate model performance.

For each modeling scenario, a final model was selected, and the sensitivity, specificity, and balanced accuracy were recorded. Results are visualized in Figure 2. The sensitivity, or percentage of BCs successfully classified, was 0.83 for ssAUC predictors only, 0.67 for physicochemical property predictors only, and 1.00 for ssAUC and physicochemical property predictors combined. The specificity, or ability to successfully classify NBCs, was 0.47 for ssAUC predictors only, 0.67 for physicochemical property predictors only, and 0.60 for both ssAUC and physicochemical property predictors. Finally, the balanced accuracy, which offers a more comprehensive look at a model's ability to accurately classify both BCs and NBCs, was 0.65 for ssAUC predictors only, 0.67 for physicochemical property predictors only, and 0.80 for both ssAUC and physicochemical property predictors combined. Since it resulted in the highest balanced accuracy, the model trained on both the ssAUC and physicochemical properties was deemed the best performing model that was carried forward to investigate biological significance and was applied to the UCs to predict whether they exhibit properties more like BCs or NBCs.

FIGURE 2.

Results of the final models from each of the three modeling scenarios. Modeling scenarios reflect a random forest model trained on (1) only ssAUC gene expression data (purple circle), (2) only physicochemical property data (yellow square), and (3) both ssAUC gene expression and physicochemical property data (green triangle) to predict whether a chemical was a BC or NBC. The specificity, or ability of the model to predict NBCs, sensitivity, or ability of the model to predict BCs, and balanced accuracy, or the average of sensitivity and specificity, are shown for each scenario.

The best performing model was applied to the previously mentioned 126 UCs to predict whether they exhibited ssAUC and physicochemical property patterns more similar to BCs or NBCs, and thus identify chemicals that may be of interest for additional screening related to breast cancer risk. In total, 97 of the 126 UCs were predicted to demonstrate attributes more analogous to BCs. These include the biocides benfuracarb and terbutryn, the dye C.I. Acid Yellow 23, the plasticizer bisphenol A (BPA), and the polycyclic aromatic hydrocarbon (PAH) benz[a]anthracene. The remaining 29 were predicted to be more comparable to NBCs and included the dyes Basic Blue 7, Allura Red C.I. 16035, Methylene Blue, and Gentian Violet as well as the plasticizer di‐n‐octyl phthalate. Full results are shown in Table 1.

TABLE 1.

Prediction results from the overall best model trained on ssAUC gene expression and physicochemical property data when applied to the UCs.

DTXSID	Chemical name	Prediction
DTXSID0020365	Cyclosporin A	BC
DTXSID0020523	2,4‐Dinitrophenol	BC
DTXSID0022777	Celecoxib	BC
DTXSID0023296	Methylene blue	NBC
DTXSID0024052	Dimethipin	BC
DTXSID0024553	Azodicarbonamide	BC
DTXSID0025523	Maltol	BC
DTXSID0026913	Dodecane	BC
DTXSID0026915	1‐Undecanol	BC
DTXSID0034192	17beta‐Trenbolone	BC
DTXSID1021455	C.I. Acid Yellow 23	BC
DTXSID1021821	2‐Butanone oxime	BC
DTXSID1021952	Triphenyl phosphate	NBC
DTXSID1021956	Di‐n‐octyl phthalate	NBC
DTXSID1023996	Cyclohexylamine	BC
DTXSID1024049	Diflubenzuron	BC
DTXSID1024126	Heptachlor epoxide B	BC
DTXSID1024209	Naled	BC
DTXSID1024835	3‐Phenylprop‐2‐enal	BC
DTXSID1025805	Octhilinone	BC
DTXSID1026900	Dodecyltrimethylammonium chloride	BC
DTXSID1026908	2‐Hexyloxyethanol	BC
DTXSID1027263	Pentyl acetate	BC
DTXSID1033664	17‐Methyltestosterone	BC
DTXSID1034187	Tamoxifen	NBC
DTXSID1040320	Bifenox	NBC
DTXSID1040619	Bexarotene	NBC
DTXSID1040792	Pyridoxine hydrochloride	BC
DTXSID2020341	Codeine	BC
DTXSID2020505	Dimethylaminoethanol	BC
DTXSID2020636	Fluvastatin	BC
DTXSID2021947	1‐Tridecanol	BC
DTXSID2021993	Dibenzofuran	BC
DTXSID2023224	Loratadine	NBC
DTXSID2024955	Di‐tert‐butyl peroxide	NBC
DTXSID2025264	Cinoxate	BC
DTXSID2026602	2,4‐Di‐tert‐butylphenol	NBC
DTXSID2026781	4‐Methyl‐2‐pentanol	BC
DTXSID2029874	Fenofibrate	BC
DTXSID2034259	2‐Methyl‐3(2H)‐isothiazolone	BC
DTXSID2037502	Bitertanol	NBC
DTXSID2041622	Benfuracarb	BC
DTXSID2044559	Profluralin	BC
DTXSID3020207	Dinoseb	BC
DTXSID3020548	Dodine	BC
DTXSID3021277	2E,4E‐Hexadienoic acid	NBC
DTXSID3021774	Acenaphthene	BC
DTXSID3022536	4‐Cumylphenol	BC
DTXSID3024318	Terbutryn	BC
DTXSID3025178	1,3‐Diphenylguanidine	BC
DTXSID3026564	Triethylene glycol bis(2‐ethylhexanoate)	BC
DTXSID3026728	(2Z)‐3,7‐Dimethylocta‐2,6‐dien‐1‐ol	NBC
DTXSID3027320	Sodium 4‐nitrophenolate	BC
DTXSID3029154	Cyclopentanone	BC
DTXSID3034165	5,5‐Dimethyl‐3‐(alpha,alpha,alpha‐trifluoro‐4‐nitro‐m‐tolyl)hydantoin	BC
DTXSID3036525	Androsterone	BC
DTXSID3042390	Sodium (2‐pyridylthio)‐N‐oxide	NBC
DTXSID3042423	Sucrose octaacetate	BC
DTXSID3047477	dl‐Norgestrel	BC
DTXSID3052725	Furathiocarb	BC
DTXSID4020292	Chloroacetaldehyde	BC
DTXSID4020870	2,2′‐Methylenebis(4‐methyl‐6‐tert‐butylphenol)	NBC
DTXSID4023541	Pyridoxine	BC
DTXSID4024436	Allura Red C.I.16035	NBC
DTXSID4026214	Triclocarban	BC
DTXSID4035209	Oxypurinol	BC
DTXSID4040713	Octyl gallate	NBC
DTXSID4042121	Mepanipyrim	NBC
DTXSID5020366	Cyproterone acetate	BC
DTXSID5020653	Gentian Violet	NBC
DTXSID5020784	Lovastatin	BC
DTXSID5021126	Phenothiazine	BC
DTXSID5021388	2,4,5‐Trichlorophenoxyacetic acid	BC
DTXSID5023902	Benz(a)anthracene	BC
DTXSID5024134	delta‐Hexachlorocyclohexane	NBC
DTXSID5025364	Glyoxal	BC
DTXSID5026625	alpha‐Terpineol	BC
DTXSID5026918	1‐Dodecanol	BC
DTXSID5027061	2‐Naphthalenol	BC
DTXSID5032523	1,2‐Benzisothiazolin‐3‐one	BC
DTXSID5033836	4‐Nonylphenol	BC
DTXSID5038888	Basic Blue 7	NBC
DTXSID6020690	Hexachlorophene	NBC
DTXSID6020856	N‐Methyl‐2‐pyrrolidone	BC
DTXSID6022927	2′,3′‐Dideoxyinosine	BC
DTXSID6027921	Di(propylene glycol) dibenzoate	BC
DTXSID6032437	Naptalam	BC
DTXSID6037564	Mefenpyr‐diethyl	NBC
DTXSID6041688	Bupirimate	BC
DTXSID6041927	Ethirimol	BC
DTXSID7020182	Bisphenol A	BC
DTXSID7020970	Nitrofen	BC
DTXSID7021992	Sodium saccharin hydrate	BC
DTXSID7024112	Fomesafen	BC
DTXSID7024291	Quinalphos	BC
DTXSID7026156	Toluene 2,4‐diisocyanate	BC
DTXSID7029871	Clotrimazole	NBC
DTXSID7029879	Ketoconazole	BC
DTXSID7034250	Chlorfenvinphos	NBC
DTXSID7034674	Ipconazole	NBC
DTXSID7037185	Amiodarone hydrochloride	BC
DTXSID7040285	2‐Hydroxyacetophenone	BC
DTXSID7052529	2‐Methoxy‐4‐vinylphenol	BC
DTXSID8021519	2‐(2‐Butoxyethoxy)ethanol	NBC
DTXSID8021642	Octadecanoic acid	BC
DTXSID8022666	Betaine	BC
DTXSID8023216	3,5,3′‐Triiodothyronine	BC
DTXSID8024284	1‐Methoxy‐2‐propanol	BC
DTXSID8027034	2‐(Hydroxymethyl)‐2‐nitro‐1,3‐propanediol	BC
DTXSID8029868	Atorvastatin	BC
DTXSID8031077	Curcumin	BC
DTXSID8040270	Dithianon	BC
DTXSID8042478	Tolylfluanid	BC
DTXSID8044676	2‐(Dodecyloxy)ethanol	BC
DTXSID9020110	Astemizole	BC
DTXSID9020114	Auramine hydrochloride	NBC
DTXSID9020241	Capsaicin	BC
DTXSID9020582	Ethoxyquin	BC
DTXSID9020877	1‐Methylnaphthalene	BC
DTXSID9021348	Thiourea	BC
DTXSID9022312	4‐Octylphenol	BC
DTXSID9022360	4‐(1,1,3,3‐Tetramethylbutyl)phenol	NBC
DTXSID9022364	5alpha‐Dihydrotestosterone	BC
DTXSID9024063	2,6‐Dimethylphenol	BC
DTXSID9026689	1,1′,1″,1‴‐(Ethanediylnitrilo)tetrakis(2‐propanol)	BC
DTXSID9047542	Picoxystrobin	NBC

Note: Model predictions should be interpreted as chemicals exhibiting properties more similar to BCs or NBCs and not as definitive classifications or associations.

3.3. Evaluating Biological Significance of Top‐Ranking Genes

In order to better understand biological mechanisms that may be involved in breast cancer hazard and progression from chemical exposure, the variable mean decrease in accuracy for the best performing model was assessed. Specifically, the mean decrease in accuracy for each input feature served as an indicator of how well the model performed when including versus excluding that feature and was used to gauge which features were most informative for differentiating BCs and NBCs in the model. The top 15 features with the highest mean decrease in accuracy are shown in Figure 3, and full results can be found in Table S4. Specifically, the top three features were the genes Claspin (CLSPN), Runt‐related Transcription Factor 2 (RUNX2), and Ubinuclein 2 (UBN2). Other genes of interest include Cytochrome P450 Family 1 Subfamily A Member 1 (CYP1A1), Interleukin 20 (IL20), and Phospholipase D Family Member 6 (PLD6). In total, 172 of 5778 features demonstrated a mean decrease in accuracy > 0, indicating these features were useful in classifying BCs versus NBCs in the training dataset. Notably, only two of those features were physicochemical properties, which were the octanol/air partition coefficient and the soil adsorption coefficient.

FIGURE 3.

Variable importance for the 15 most influential genes from the best overall model according to the mean decrease in accuracy. The best overall model was trained on both the ssAUC gene expression data and the physicochemical properties. In total, 172 features, including 170 genes and 2 physicochemical properties, demonstrated a mean decrease in accuracy > 0 (Table S4). These features indicate an increased predictive capacity for differentiating BCs and NBCs.

To evaluate enriched canonical pathways and upstream regulators, the Ingenuity Knowledgebase was leveraged to investigate the 170 genes with a mean decrease in accuracy > 0. At a BH adjusted p value of 0.05, only 1 pathway was enriched in this gene set: ferroptosis signaling. However, 995 upstream regulators were identified as likely regulators of the observed differences in gene expression patterns predictive of BC status at a BH adjusted p value of 0.05. A total of 61 upstream regulators were predicted to be upregulated in BCs, 99 upstream regulators were predicted to be downregulated in BCs, and the remaining 835 upstream regulators were predicted to be generally dysregulated. Significantly upregulated upstream regulators in BCs included Glioma‐Associated Oncogene Family Zinc Finger 1 (GLI1), lactacystin, and Tumor Necrosis Factor (TNF). Significantly downregulated upstream regulators in BCs included Interleukin 6 (IL6), Nuclear Protein 1 (NUPR1), and PBX/Knotted 1 Homeobox 2 (PKNOX2). Other enriched upstream regulators of note are β‐estradiol and Estrogen Receptor 2 (ESR2). Full results for pathways and upstream regulators are located in Tables S5 and S6, respectively.

3.4. Sensitivity Analysis Results

Two sensitivity analyses were run to gauge model robustness and validate findings. The first sensitivity analysis included 100 RF models trained using the training set for the best performing model, but with the outcome variables randomly shuffled to provide an estimate of the performance that can be achieved by chance on the given data. These models yielded a median sensitivity of 0.5, a median specificity of 0.53, and a median balanced accuracy of 0.48. Notably, the highest achieved balanced accuracy of any of the 100 models was 0.76, compared to 0.8 of the best performing model. The results of this analysis are visualized in comparison to the best performing model in Figure 4A.

FIGURE 4.

Results of the two sensitivity analyses to gauge primary model robustness. (A) The sensitivities, specificities, and balanced accuracies from 100 iterations of the overall best performing model with the outcome variable randomly shuffled. This analysis was conducted to ensure best performing model results were not achieved by chance alone. Best performing model results are also shown (pink stars). (B) The top 15 most important variables according to mean decrease in accuracy for the analysis shown with the mean decrease in accuracy for the three random noise variables, indicated by the dashed vertical lines. These results validate the importance of the top‐ranking features in model performance.

The second sensitivity analysis was a RF model trained on the same training set as the best performing model but with three additional dummy features comprised of values randomly selected with replacement across the full distribution of the training set. The variable importance as determined by the mean decrease in accuracy was assessed to ensure that reported top‐ranking features from the best performing model were truly informative. Though not identical, the top 15 features from this sensitivity analysis strongly resemble those from the best performing model and are shown in Figure 4B. Additionally, the three dummy variables did not place highly in the variable importance ranking. Together, these results indicate that our best performing model trained on both the gene expression and physicochemical property data performs better than random chance, and the genes we identified as being particularly informative in differentiating BCs and NBCs are meaningful towards model performance.

3.5. Relationships to Human Cancer Disease Signatures

To relate in vitro findings to a human population and investigate whether specific chemicals are potentially more influential to the progression of breast cancer for certain molecular subtypes than others, the 170 top‐ranking genes contributing to BC versus NBC predictive model performance, as described in the above section Evaluating Biological Significance of Top‐Ranking Genes, were further assessed in TCGA breast cancer samples using a statistical model (Koboldt et al. 2012). A total of 1034 tumor samples with RNA‐seq data and of molecular subtypes of luminal A (n = 555), luminal B (n = 206), Her2‐enriched (n = 80), and basal‐like (n = 193) were identified in the TCGA‐BRCA project. The race/ethnicity breakdown of these samples is as follows: White (n = 719), Black or African American (n = 167), not reported/unknown (n = 87), Asian (n = 60), American Indian or Alaska Native (n = 1). Additionally, 156 of the 170 genes from the MCF7 analysis could be mapped to genes in the TCGA data. Normalized gene expression data processed through the STAR pipeline for the 156 genes were isolated for these 1034 samples and then pseudo‐log₂ transformed in preparation for downstream analyses.

For analysis, first, ANOVAs were run to determine which, if any, of the 156 genes were differentially expressed between subtypes in the TCGA dataset. This resulted in 149 genes that were statistically significant at a BH‐adjusted p value < 0.05 between cancer subtypes. Additionally, a PCA was performed on the 1034 samples and 149 significant genes to visualize the variability of expression across subtypes (Figure 5A). The first two principal components captured 34.2% of the variation in the dataset and Figure 5B shows which genes were the largest drivers of the variability. Next, the samples corresponding to a race/ethnicity of American Indian or Alaska Native were removed due to a low sample size. Samples with race/ethnicity listed as “not reported/unknown” were also removed to allow for a more clear interpretation of results. ANOVAs were run on the remaining 946 samples to assess differential expression across subtypes while controlling for race. This resulted in 148 statistically significant genes at a BH‐adjusted p value < 0.05. The PCA run on the 946 samples and 148 genes returned similar results to the analysis that did not consider race, indicating that race was not a confounder. These results are shown in Figure S1, and full results for both ANOVAs can be found in Table S7.

FIGURE 5.

PCA visualization of the TCGA breast cancer gene expression profiles for genes identified as top‐ranking in terms of predicting BCs versus NBCs using in vitro data and identified as differentially expressed across cancer patient molecular subtypes. (A) Dimensionally reduced expression profiles are shown via the first two PCs across all 149 genes that were both identified as top‐ranking in the predictive modeling analysis and differentially expressed across cancer subtypes. (B) The top 25 of 149 genes determined to drive the majority of variation across the first two PCs. Color reflects the strength of the contribution of the gene towards the variability of the given PCs with red reflecting stronger contributions. Full equations for the calculation of the contribution metric can be found in the factoextra documentation (Kassambara and Mundt 2017); (Factoextra 2017).

When comparing the results of the ANOVAs, all 148 of the genes that were significant in the race‐controlled analyses were also significant in the non‐race‐controlled analyses. The single gene unique to the non‐race‐controlled analysis was Splicing Factor 3a Subunit 1 (SF3A1), which has been previously linked to pancreatic cancer in a two‐stage case control study (Tian et al. 2015). Additionally, when examining the genes driving the variability observed across the subtypes in the TCGA data, ATP Binding Cassette Subfamily C Member 8 (ABCC8), Cytochrome P450 Family 4 Subfamily B Member 1 (CYP4B1), Interleukin 20 (IL20), and Neuropeptide Y Receptor Y1 (NPY1R) were among the strongest contributing genes. Upon further investigation, ABCC8 and CYP4B1 demonstrated a notably lower average pseudo‐log₂ expression across the basal subtype compared to the Her2‐enriched, luminal A, and luminal B subtypes in the TCGA dataset. IL20 and NPY1R demonstrated a higher average pseudo‐log₂ expression across the luminal subtypes in comparison to Her2‐enriched and basal subtypes. These differentially expressed genes may therefore serve as important linkages between in vitro‐derived cancer predictions and in vivo human observations, informing whether chemicals may alter BC related mechanisms across model systems.

4. Discussion

Overall, this work integrated HTTr data with physicochemical property information to build a machine learning algorithm capable of differentiating chemicals with and without a putative association with breast cancer. The model was then applied to chemicals understudied in regard to breast cancer risk to identify substances that warrant further investigation, and gene‐ and pathway‐level analyses were run to examine the biological alterations that may be relevant to chemically induced breast cancer progression. Four key take‐home points from this study, detailed further below, were the following: (1) high‐throughput transcriptomic screening was demonstrated to predict chemical carcinogenicity outcomes in vivo with a high balanced accuracy; (2) additional understudied chemicals were identified that merit further study regarding breast cancer risk; (3) potential biomarkers of chemicals that may increase breast cancer hazard were identified in the form of predictive genes that were involved in hormone signaling, inflammation, ferroptosis signaling, and cell proliferation processes (among others); and (4) highly predictive genes also demonstrated transcriptional dysregulation across breast cancer tissues from patient subgroups aggregated in TCGA. Collectively, findings exemplify further utilities of in vitro HTTr data to inform which chemicals may influence breast cancer hazard, and the mechanisms by which these chemicals may contribute to cancer progression and/or initiation.

In the past two decades, in vitro screening has contributed greatly to assessing human disease risk attributable to toxicants present in the environment. Examples include research aimed at identifying neurotoxicants with the potential to alter neurodegenerative disease risk such as Parkinson's and Alzheimer's (Schmidt et al. 2017; Tong et al. 2017) as well as certain cancer endpoints such as liver and lung cancer (Jia et al. 2019; Peyre et al. 2014). This work expands on these studies by using in vitro screening data to inform which chemicals may be affecting biological processes involved in breast cancer. Importantly, the model trained jointly on the gene expression and physicochemical properties yielded a higher balanced accuracy when compared to just gene expression or just physicochemical properties. This highlights the importance and utility of considering multiple data streams when predicting complex disease outcomes.

Specific chemicals, including those used as pesticides, dyes, and plasticizers, were predicted to exhibit gene expression and physicochemical properties more similar to BCs upon applying the best performing model to the UCs. These substances have been investigated to varying degrees and subsequently have variable information available to inform their potential toxicity and relation to cancer. For example, BPA, which was predicted to be a BC, has been studied with respect to breast cancer using various approaches and has, in total, yielded inconclusive results (Salamanca‐Fernández et al. 2021; Wang et al. 2017; Weber Lozada and Keri 2011). The chemical 2,4‐dinitrophenol has not yet been linked to breast cancer, to our knowledge, via in vivo or in vitro studies. However, a 2023 exposomics analysis evaluating environmental chemical exposure to pregnant women in a breast cancer cohort using banked serum samples found this substance was detected at higher levels in the women who developed breast cancer than those who remained breast cancer free (Go et al. 2023). Similarly, little is known surrounding benz[a]anthracene and breast cancer specifically. However, benz[a]anthracene is a PAH, and PAHs are a class of chemicals of concern for cancer risk, potentially functioning as endocrine disruptors and genotoxicants (Ewa and Danuta 2017; Zhang et al. 2016). It is worth mentioning, however, that many PAHs require bioactivation, often through Cytochrome P450 (CYP) enzymes, for carcinogenicity to be a concern (Baird et al. 2005; Shimada and Fujii‐Kuriyama 2004). While some level of CYP activity is present in MCF7 cells, it may be difficult to fully assess the breast cancer hazard of PAHs in this cell line (Sneha et al. 2021). Still, for other substances like the pesticide benfuracarb and dye C.I. Acid Yellow 23, there is limited research evaluating their potential impacts on breast cancer risk or underlying cancer etiologies. Finally, we have previously prioritized the pesticide, terbutryn, and the aromatic fragrance and flavoring agent, 2‐methoxy‐4‐vinylphenol, for further study based on an analysis integrating exposure potential, chemical structural attributes, and physicochemical properties (Koval et al. 2022). These chemicals therefore reflect candidates that should be first prioritized for further toxicology and epidemiology studies in the field of cancer research.

A set of informative genes was derived by assessing the relative importance of features in the best performing model based on the mean decrease in accuracy when each feature was excluded. These genes have roles in regulating hormone signaling, inflammation, and cell proliferation. Subsequently, pathway enrichment and upstream molecular regulator predictions were carried out to further understand the roles of these genes in regulating biological processes and ferroptosis signaling was found to be enriched. In terms of hormone signaling, enriched upstream regulators included β‐estradiol and ESR2 highlighting the importance of estrogen and ER signaling in hormone receptor positive tumor cells. ER signaling is known to influence cancer outcomes (Johnson et al. 2021; Tsang and Tse 2020). Enrichment of these upstream regulators was expected due to the estrogen sensitivity of MCF7 cells and thus supports the findings of this study. In terms of inflammation, IL6 and TNF are known pro‐inflammatory cytokines that were identified as enriched upstream regulators. IL6 not only helps regulate inflammation, but also plays an important role in tissue metabolism and immune response (Chen et al. 2022). The IL6 protein is highly expressed in tumors of many different cancers, including breast cancer, and is correlated with poor prognosis and metastasis (Chen et al. 2022; Salgado et al. 2003; Zhang and Adachi 1999). TNF regulates over 400 inflammatory genes (Hayden and Ghosh 2014; Liu et al. 2017) and can promote stemness in breast cancer, a key characteristic of cancer progression (Aponte and Caicedo 2017; Liu et al. 2020). Ferroptosis signaling was also enriched at the pathway level, a finding further supported through the prediction of NUPR1, a known regulator of ferroptosis (Liu et al. 2021), as a likely upstream regulator of cancer‐predictive gene expression signatures. Ferroptosis is a form of programmed cell death originally proposed in 2012. It is iron dependent and characterized by the prevalence of lipid reactive oxygen species (Dixon et al. 2012; Li et al. 2020). While this mechanism has the ability to suppress tumor growth, ferroptotic damage may also induce inflammation and immunosuppressive tendencies in the tumor microenvironment thus promoting tumor growth (Chen et al. 2021; Zhou et al. 2024). Additionally, elevated capacity for antioxidants commonly present in tumor cells can lead to mechanisms of ferroptosis evasion (Chen et al. 2021; Zhou et al. 2024). This complex phenomenon has been observed in many cancers, including breast cancer, but further research into ferroptosis is necessary to more fully understand this mechanism and its effects.

The top three genes with the highest importance in model performance, CLSPN, RUNX2, and UBN2, all demonstrate roles pertaining to cell proliferation. CLSPN is known to play a role in cell cycle signaling and DNA damage repair (Azenha et al. 2017). This gene has been further associated with roles in regulating tumor immune‐related microenvironments, tumor drug resistance, and patient prognosis for various cancers, including breast cancer (Azenha et al. 2017; Chen et al. 2023). Additionally, a recent review discussed RUNX2's associations with proliferation, stemness, metastasis, and chemoresistance in various cancers (Lin 2023). Interestingly, it has been demonstrated that silencing RUNX2 in both in vitro and in vivo models inhibited proliferation of triple negative breast cancer cells (Lin 2023; Si et al. 2022). Finally, UBN2 has published implications in cancer etiology, specifically colorectal cancer (Zhao et al. 2019). In a 2019 study of colorectal cancer, UBN2 was shown to act through the Ras/MAPK pathway, which regulates cell processes including cell growth and proliferation (Bahar et al. 2023; Zhao et al. 2019). In terms of breast cancer, UBN2 has been predicted to be targeted by multiple breast cancer‐associated microRNAs through molecular analysis of human tissues, highlighting the additional role of epigenetics in this disease outcome (Miglioli et al. 2022).

With the goal of anchoring in vitro‐based findings to human cancer patient data, the top‐ranking genes contributing to chemical breast cancer association prediction modeling were further evaluated in the primary tumor samples from TCGA. Of the 170 top‐ranking genes, 149 were identified as differentially expressed among the molecular subtypes luminal A, luminal B, Her2‐enriched, and basal‐like. As the top‐ranking genes were identified from toxicant induced gene expression changes based on toxicants known to increase breast cancer risk, continuing to investigate similar gene expression perturbations in molecular subtypes from human populations may be valuable in understanding which subtypes may be susceptible to toxicant‐induced alterations that can promote cancer progression, and the mechanisms by which these alterations may occur. Additionally, this analysis highlights the importance of considering molecular subtype when leveraging in vitro data to identify biomarkers of disease as the heterogeneity of breast cancer may not result in uniform expression patterns across subtypes. For example, ABCC8 was the greatest contributor to the differences observed among the subtypes in the TCGA data, and it has been previously demonstrated in a human cohort that this gene was expressed at higher levels in ER‐positive and PR‐positive patients and was associated with higher grade tumors (Hlaváč et al. 2013). Similarly, IL20, which encodes a cytokine involved in tissue remodeling and homeostasis, was an important driver of molecular subtype differences, and this gene is regulated by estradiol (Su et al. 2016).

While this study provides a novel examination of breast cancer association predictions using in vitro HTTr data across hundreds of chemicals, it is not without limitations. First, we recognize that the chemical categorization schema that we implemented to classify chemicals as BCs, NBCs, and UCs is imperfect and requires assumptions about breast carcinogenesis from rodent models in addition to the incorporated human studies‐based literature. Many chemical classifications lack consensus across institutions, and breast cancer etiology is highly complex; thus, it is difficult to identify chemicals that do (and do not) perturb any possible mechanism that affects disease initiation or progression. These categories will continue to be improved upon as further studies are conducted. We were also limited by the number of chemicals that had a chemical categorization, comparable in vitro gene expression data, and complete physicochemical property data/predictions. Ideally, we would have liked to identify chemicals that operate primarily through particular mechanisms (i.e., endocrine disruption, mutagenic/genotoxic, pro‐inflammatory, etc.) and train models specifically on these sets of chemicals, yet the current dataset did not contain enough chemicals (namely BCs) to meaningfully parse out into the sub‐categories. As a result, all BCs were grouped for model development regardless of mechanism of action. Future work should attempt to isolate chemicals acting by a particular mechanism to better elucidate biomarkers of disease and perturbed biological pathways relevant to breast cancer as datasets continue to expand. Finally, we note that MCF7 cells are of the luminal A molecular subtype and are commonly used in breast cancer research, but a number of other breast cell lines exist (both normal and cancerous) (Dai et al. 2017). The sensitivity of MCF7 cells to estrogen may limit the extension of findings to other breast cancer etiologies, particularly ER‐negative cancers. It is worth mentioning, however, that during this experiment's cell culture preparations, cells were not deprived of estrogens prior to chemical exposures through use of phenol‐red‐free media and charcoal‐stripped serum, and that chemical exposures were acute (6 h). While these experimental design considerations may influence expression profiles of estrogen‐responsive genes, a recent study from our group evaluated varying exposure durations and the use of estrogen stripping media prior to exposure and showed consistent results with exposure conditions used in the current study when comparing transcriptomic points of departure (Harrill et al. 2024a, 2024b). Furthermore, while we found a few studies that evaluate gene expression of non‐tumorigenic MCF‐10A and MCF‐12A breast epithelial cells after chemical exposure, most of these experiments only screened a limited number of substances (Mesnage et al. 2022; Zimta et al. 2024; van Duursen et al. 2003). Expanding the chemical domain evaluated in MCF‐10A, MCF‐12A, or other models for normal breast tissue will be valuable for comparing results between tumorigenic and non‐tumorigenic models. For this work, all results from the described in vitro experiments should be interpreted with the knowledge that these analyses are leveraging cancer cells, and thus cancer progression events may have been captured to a greater extent in comparison to initiation events, though there does exist some overlap in molecular targets (i.e., ESR1 and PPARγ) (Miyoshi et al. 2010; Augimeri et al. 2020). Future studies will incorporate transcriptional disruptions across different cell types and disease stages to continue expanding our understanding of chemical‐impacted disruptions in molecular homeostasis and their relations to human disease outcomes.

Overall, this study set out to leverage recent transcriptomic screening data of hundreds of chemicals in MCF7 cells in a machine learning framework with the goal of identifying potential chemicals in the environment with the ability to alter breast cancer disease risk. Results highlight the utility of integrating in vitro data, physicochemical properties, and machine learning for the purposes of assessing human disease outcomes. Additionally, 97 chemicals that have not been sufficiently studied regarding breast cancer were predicted to exhibit attributes more similar to chemicals with a putative association with breast cancer and therefore reflect candidate chemicals for further study for this disease. Finally, potential gene biomarkers of disease and resulting enriched pathways yield insight into biological mechanisms relevant to the impacts of chemical exposure on breast cancer risk and progression. Though this work contributes to the growing body of literature aimed at understanding environmentally driven causes of breast cancer, it is imperative that future studies in toxicology, exposure science, and epidemiology continue to identify substances of concern to better protect public health and reduce the burden of disease.

Author Contributions

L.E.K. conceptualization, data curation, formal analysis, investigation, methodology, software, validation, visualization, writing – original draft, review and editing. R.J. conceptualization, data curation, investigation, methodology, validation, writing – review and editing. J.E.R. conceptualization, data curation, formal analysis, funding acquisition, investigation, methodology, software, supervision, validation, visualization, writing – original draft, review and editing.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: PCA visualization of the TCGA breast cancer gene expression profiles for genes identified as top‐ranking in terms of predicting BCs versus NBCs using in vitro data and identified as differentially expressed across cancer patient molecular subtypes when controlling for race. Analysis was run on the same set of samples as those presented in Figure 5. (A) Dimensionally reduced expression profiles are shown via the first two PCs across all 148 genes identified as top‐ranking in the predictive modeling analysis and differentially expressed across cancer subtypes. (B) The top 25 of 148 genes determined to drive the majority of variation across the first two PCs. Importantly, this analysis aimed to assess race as a confounder.

Table S1: The chemicals considered in the predictive modeling analyses and their putative association with breast cancer. Chemicals were categorized as breast cancer associated (BC), non‐breast cancer associated (NBC), or unknown (UC).

Table S2: The ssAUC values for the 200 chemicals and the 5793 genes considered in the predictive modeling analyses.

Table S3: Physicochemical properties in the form of OPERA predictions for the chemicals considered in the predictive modeling analyses.

Table S4: Feature importance ranking based on mean decrease in accuracy for features in best overall model trained on ssAUC gene expression and physicochemical property data. A larger mean decrease in accuracy indicates a more informative feature for differentiating between BCs and NBCs.

Table S5: Full results of the pathway analysis run on the 170 genes identified as being the most informative in differentiating BCs and NBCs in the best overall model. Analysis was conducted using the Ingenuity Knowledgebase.

Table S6: Full list of upstream regulators identified for the 170 genes identified as being the most informative in differentiating BCs and NBCs in the best overall model. Analysis was conducted using the Ingenuity Knowledgebase.

Table S7: Results of the ANOVAs evaluating differences in gene expression across molecular subtypes in the TCGA data with and without controlling for race. Analyses focused on the top‐ranking genes contributing to model performance when differentiating chemicals with associations to breast cancer versus no known associations.

Koval, L. E. , Judson R., and Rager J. E.. 2025. “Identifying Gene Predictors of Chemicals Linked With Breast Cancer: A Machine Learning Analysis of MCF7 Cellular Transcriptomic Screening Data.” Environmental and Molecular Mutagenesis 66, no. 8: 425–441. 10.1002/em.70034.

Data Availability Statement

All data, analysis scripts, and results from this analysis are publicly available. Analysis scripts, datasets and results are organized on the Rager lab Github site (https://github.com/Ragerlab). FASTQ files and aligned counts have been deposited in the NCBI Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo/) with accession number GSE272548. Additionally human transcriptomic signatures were acquired through The Cancer Genome Atlas (https://www.cancer.gov/ccg/research/genome-sequencing/tcga).