Assessment of chemical risk factor for breast cancer based on gene expression profiles

Abstract

Breast cancer is one of the leading types of cancer diagnosis worldwide. Although the relationship between hormone exposure and breast cancer has been established, the role of environmental chemicals that induce alterations in mammary gland development and function remains understudied. This highlights the need for additional research into chemicals that may be involved in breast cancer progression. In this study, we analyzed transcriptomics data on MCF7 cell lines exposed to a set of 161 suspected mammalian carcinogens chemicals. From these data, we identified genes and pathways that impact the regulation of cellular proliferation and death processes, as well as cellular repair systems (regulation of the apoptotic process and cellular response to DNA damage stimuli). We also observed significant up-regulation of signaling pathways crucial to cancer development, such as the negative regulation of the MAPK cascade and the regulation of signal transduction by p53 class mediator. Furthermore, through an in vitro/ex vivo comparison, using RNA sequencing from 1,092 breast cancer tissues obtained from the TCGA-BRCA dataset, we identified a set of chemicals that share similar gene deregulation patterns with at least one patient. These compounds should be investigated further as they could be involved in breast cancer progression.

Introduction

Breast cancer is one of the leading type of cancer diagnosis worldwide with more than 2.3 millions people and the second leading cause of cancer death in women after lung cancer^1,2. It is expected that over 3 million new cases of breast cancer and 1 million deaths will occur each year worldwide by the year 2040³.

The initiation and progression of breast cancer are known to be influenced by a wide range of risk factors, including obesity, genetic factors, family history and endocrine factors^4,5. It is an heterogeneous and evolving disease, marked by various somatic mutations and changes in gene and protein expression⁶. Based on the expression modulation of the progesterone receptor (PR), estrogen receptor (ER), and human epidermal growth factor receptor 2 (HER2), breast cancer can be classified in 5 subtypes: luminal A, luminal B, HER2-enriched, normal-like and basal, with the majority of the latter class being triple-negative breast cancer (TNBCs)^7,8. The relationship between hormone exposure and breast cancer has been established in animal models and it can be assumed that chemicals that induce mammary tumors, of which breast cancer is the malignant form, or activate estrogen or progesterone signaling, are likely to increase breast cancer (BC) risk⁷.

However, chemicals that induced alterations of the mammary gland (MG) development and function remain understudied highlighting the need for additional research into chemicals that can increase breast cancer risk. Identifying such potential carcinogenic or mammary tumor-inducing chemicals could help in the prevention and the protection in human health.

Several lists of suspected chemicals involved in breast cancer progression have been reported including a list of 268 chemicals with key characteristics of stimulating progesterone or estrogen signaling^9,10. Many of them have been identified as endocrine disrupting chemicals (EDCs), affecting androgen, estrogen, thyroid and steroidogenic (EATS) mechanism of action, and associated to alterations in breast development, lactation, or cancer in both human cohorts and experimental studies in rodents¹¹.

However, multiple additional pathways influence MG development and function are not necessarily captured by standard endocrine-sensitive endpoints. For example, multiple stages of MG development are influenced by the aryl hydrocarbon receptor, prolactin, insulin-like growth factors, transforming growth factor-, and placental lactogens^12–14.

Previous RNA sequencing analysis of breast cancer individuals have revealed many others molecular processes that could potentially impact breast cancer, showing the heterogeneity of mechanisms associated with breast cancer^15–25.

In this context, the objective of our study was to characterize through a computational approach, a global gene expression profile for chemicals inducing mammary tumors using a large in vitro toxicogenomics dataset, LINCS²⁶. Considering breast cancer tissue from patients, available in The Cancer Genome Atlas (TCGA), and the Genotype-Tissue Expression (GTEx)²⁷, we collected sets of genes significantly differentiated between breast cancer and normal tissues. We, then, developed a mathematical score matching the gene signature for any chemicals tested with the 1,092 gene’s signature for breast cancer patients. Interestingly, we observed for some chemicals a similar gene signature suggesting the potential carcinogenicity or mammary tumors-inducing of these chemicals. Genes and pathways involved in breast cancer in our analysis are further discussed below.

Overall, our analysis could be of interest as a new alternative approach (NAM) in the identification of mammary carcinogens in a large scale of chemicals with RNA sequencing.

Results

L1000 analyses

Known and suspected carcinogens dysregulate cancer-relevant pathways in MCF7 cells

We conducted Over-Representation Analysis (ORA) using the L1000 dataset, focusing on two subsets of compounds: mammalian carcinogens (MC) and suspected mammalian carcinogens (SMC). Analyses were performed with landmark genes alone (978 genes) and with the combined set of landmark and best inferred genes (10,174 genes).

Mammalian Carcinogen (MC) From the 268 reported mammalian carcinogens in the literature, gene expression for MCF7 treated with 34 chemicals have been collected in the L1000 dataset. Doing an over representation analysis (ORA), our analysis revealed significant pathway and process deregulations linked to cancer.

Up-regulated Pathways As expected we observed that using the landmark genes set MC compounds associated signatures showed through the ORA that mostly up-regulated pathways were linked to a panel of cancer associated pathways, including “Breast Cancer”. This aligned with the fact that the majority of the compounds tested were considered as mammalian carcinogens. We observed that the pathways were not specific to one type of cancer. This was evident within explicitly cancer-related pathways such as “Pathways in Cancer” and “Breast Cancer” (through genes: CDKN1A; GADD45A; E2F2; BRCA1; DDB2) but also in less straightforward terms concerning mechanisms implicated in the onset or progression of cancers. For example, our analysis characterized the “ErbB signaling pathway” through genes: STAT5B; CDKN1A; PAK1, and the “MAPK signaling pathway”, through genes: HSPA8; PAK1; GADD45A; NFATC3; PDGFA; MAP3K4; HSPA1A. These pathways have already been reported in the literature to play a role in the onset of cancer and the tumorigenesis^28,29 (Fig. 1a; Supplementary Table S1 online).

Fig. 1.

Over representation analysis conducted on deregulated genes from MC subset of L1000 gene expression profiles. (a) ORA representing up-regulated KEGG pathways, based on landmark genes only. (b) ORA representing up-regulated KEGG pathways, based on landmark and best inferred set of genes. (c) ORA representing down-regulated KEGG pathways, based on landmark genes only. (d) ORA representing up-regulated KEGG pathways, based on landmark and best inferred set of genes.

Looking at biological processes deregulated using the Gene Ontology database as a reference, we observed that most of them were related to cell proliferation, cell death, cell organization and cell migration (Supplementary Table S2 online). These phenomena are known to be involved in the appearance or development of cancer cells³⁰.

Adding best inferred genes to the landmarks genes for the ORA of MC compounds yielded similar results. Many up-regulated genes were involved in cancer-associated pathways. However, we observed that the significance of the deregulated pathways identified increased with much lower adjusted p-values. Also, we noticed that two pathways associated with atherosclerosis were up-regulated with very low p-values and could play a role into the apparition or the development of breast cancer. Explicit cancer-related pathways such as “Pathways in Cancer” and “Breast Cancer” were deregulated, but also in a less straightforward mechanisms implicated in the onset or progression of cancers such as “TNF signaling pathway”³¹ (Fig. 1b; Supplementary Table S3 online).

Similarly, for biological processes, genes up-regulated by MCs were found to impact regulation of cellular proliferation and death processes, as well as cellular repair systems. including processes like “regulation of apoptotic process” and “cellular response to DNA damage stimulus.” We also observed significant up-regulation in signaling pathways crucial to cancer development, such as the “negative regulation of MAPK cascade” and “regulation of signal transduction by p53 class mediator.” The deregulation across these pathways indicates potential cellular stress and disruptions in key mechanisms associated with tumorigenesis.³² (Supplementary Table S4 online).

Down-regulated Pathways

Similarly, landmark genes down-regulated by MC compounds impacted biological pathways and processes involved in cancer. In particular, the cell cycle and certain DNA repair systems were significantly impacted as well as signaling pathways known to be linked to cancer development, such as the “p53 signaling pathway”. We also observed that certain pathways associated with the female reproductive system such as “Oocyte meiosis”, and even with endocrine systems such as “Progesterone-mediated oocyte maturation” are impacted by these compounds (Fig. 1c). Lists of genes deregulated and found involved in the different pathways mentioned can be accessed in Supplementary Table S5 online.

Applying ORA including the best inferred genes with landmark genes, we observed that almost all down-regulated pathways and biological processes were identical to those when only considering landmark genes (Fig. 1d; Supplementary Table S6 online).

By comparing results obtained using landmark genes alone versus the combined set with best inferred genes, we identified key similarities and additional specificities that enhance our understanding of the underlying biological mechanisms. Indeed, the inclusion of best inferred genes enhanced the analysis by lowering adjusted p-values and increasing the significance of these deregulations.

Suspected Mammalian Carcinogens (SMC)

From the 1626 compounds defined as suspected mammalian carcinogens, we removed the 75 compounds included in the set of MC compounds. This step was done in order to not have redundancy in the further analyses and get results specific to the group of SMC compounds.

Thus, from the 1551 remaining compounds, 127 were found in our L1000 dataset. Considering these 127 SMC compounds we applied the same ORA analyses on the corresponding gene expression data.

Up-regulated Pathways

Using landmarks genes set for ORA, we observed, the impact through the up-regulation of cancer related pathways such as “Pathways in cancer”, “Cell cycle”, “MicroRNAs in cancer”, but also through signaling pathways that could play a role in apparition or progression of cancer such as “HIF-1 signaling pathway”³³ , “NF-kappa B signaling pathway”³⁴, or “IL-17 signaling pathway”³⁵ (Fig. 2a; Supplementary Table S7 online). ORA on SMCs compounds also highlighted up-regulated signaling pathways of various hormones such as “Relaxin signaling pathway”, “Estrogen signaling pathway”, “Progesterone-mediated oocyte maturation” or “Oocyte meiosis”. This is probably due to the nature of the compounds that have been identified as potential endocrine disruptors, and highlights a potential link between disruption of hormone secretion and the female reproductive system. Interestingly Hepatitis B was also on the top list. It is in agreement with some studies that report that chemotherapy-induced immunosuppression drugs can lead to HBV reactivation³⁶.

Fig. 2.

Over representation analysis conducted on deregulated genes from SMC subset of L1000 gene expression profiles. (a) ORA representing up-regulated KEGG pathways, based on landmark genes only. (b) ORA representing up-regulated KEGG pathways, based on landmark and best inferred set of genes. (c) ORA representing down-regulated KEGG pathways, based on landmark genes only. (d) ORA representing up-regulated KEGG pathways, based on landmark and best inferred set of genes.

The addition of best inferred genes to the genes set used for the ORA revealed similar up-regulated pathways but with lower p-values. A larger number of pathways were significantly up-regulated, and we noticed that in addition to cancer related pathways and endocrine system related pathways a group of pathways involved in metabolism was up-regulated (Fig. 2b; Supplementary Table S8 online).

Down-regulated Pathways

In the same way ORA showed that pathways down-regulated were involved in cancer onset or progression. In particular, the cell cycle and certain DNA repair systems were significantly impacted as well as signaling pathways known to be linked to cancer development, such as the “p53 signaling pathway”³⁷. We also observed that certain pathways associated with the female reproductive system such as “Oocyte meiosis”, and even with endocrine systems such as “Progesterone-mediated oocyte maturation” are impacted by these compounds as we observed for the MC compounds (Fig. 2c, d; Supplementary Tables S9, S10 online). There was almost no difference in the results between the two types of gene sets from which ORA was performed, except for the p-values, which were smaller when the best inferred genes were included.

Interestingly, these compounds impacted the mechanisms associated with different types of cancer, as well as on different disrupted mechanisms involved in cancer development. These included deregulation of the cell cycle, breast cancer, and biological processes linked to apoptosis and cell proliferation. In addition to cancer-related deregulations, we observed that signaling pathways associated with the endocrine and reproductive systems were also deregulated by these compounds.

Transcriptomic similarity reveals potential risk profiles among suspected carcinogens

To assess the risk factor of breast cancer progression associated with new molecules, we first compared in vitro transcriptomic profiles perturbed by compounds defined as presenting risk in the context of breast cancer (MC subset) with those of molecules suspected to be at risk (SMC subset). This comparison was performed by calculating the cosine similarity between each profile from the two subsets. This approach provided a general mapping of the similarity between profiles within both subsets. Through this analysis, we identified a subset of profiles exhibiting high similarity, while the majority showed no particular resemblance (Fig. 3; Supplementary Table S11 online).

Fig. 3.

Cosine similarity between L1000 MC and L1000 SMC gene expression profiles. Each L1000 profile has been obtained considering a time exposure of 24h. In x-axis, we have 74 SMC L1000 profiles corresponding to 50 distincts compounds. In y-axis we have 24 MC L1000 profiles for a number of 14 distincts compounds. NDGA is short for ‘Nordihydroguaiaretic acid’.

To focus on these highly similar profiles, we filtered the analysis data by retaining only those with a cosine similarity greater than 0.4. This refinement facilitated the visualization of strongly similar profiles (Fig. 3), as well as the experimental conditions for which the compounds have been tested.

We observed that a set of MC compounds (including among other doxorubicin, metronidazole, and dacarbazine) affects the gene expression of cells in the same way as a subset of SMC compounds (including among other guanethidine, tert-butylhydroquinone, naringenin and pyrazinamide). Indeed, we measured high cosine similarity (up to 0.843) for ornidazole 0.08 M and indomethacin 0.03 M).

Some of these SMC compounds appear to have, at several concentrations, strongly similar effects to the set of MC compounds mentioned just above. For example we can mention ornidazole (0.01M, 0.08M), equol (0.04 M, 10 M) or tert-butylhydroquinone (0.66 M, 4M, 12M). Others have specific concentrations that are more likely to deregulate genes expression than others in the same way that the set of MC compounds mentioned above. For example, luteolin-treated cells showed a much higher similarity of gene expression profiles, with the set of MC compounds mentioned earlier, when treated at 0.08 M than at 0.74 M and 3.33 M. Similarly, fluvoxamine showed a favorable concentration at 0.125 M as opposed to 0.04 M and 1.11 M, where a lower cosine similarity is observed.

TCGA-GTEx: tissue level analysis

Validation of a distinct pathological signature in tumor tissues for comparative analysis

With the transcriptomics analysis performed on the MCF7 cell line, we were able to observe that some chemicals were involved in breast cancer. In a second step, we were interested to detect if the gene’s deregulation of a breast cancer cell line (in vitro) by chemicals was similar to the gene’s deregulation in breast cancer tissues from patients. If this is the case, we could make the hypothesis that the chemicals might have an impact on breast cancer progression.

First, we verified that the RNA sequencing of patients collected in TCGA were different from a set of ‘healthy’ breast tissue obtained in GTEx. With an UMAP visualization of RNA Seq data from both databases, we could confirm that profiles from healthy tissue are similar to each other and different from profiles from pathological tissue. Only 2 breast cancer tissues and 1 healthy breast tissue overlapped but didn’t change the global difference between healthy and tumoral tissues. These two distinct clusters has allowed to assert that, following DEG analysis, the vector of z-score obtained would be specific to pathological gene expression.

Furthermore, within the profiles from diseased tissue, the distribution of profiles projected onto 2D UMAP space did not allow them to be grouped by pathological stage (Supplementary Fig. S1 online). This allows to hypothesize that there will be no obvious specific characterization of the different pathological states, and that, we will instead focus on the whole range of diseased tissue profiles without distinction.

Pathway dysregulation in patient tissues mirrors in vitro chemical perturbations

We obtained in a first step a global breast cancer tissue gene expression profile by conducting a DEG analysis using TCGA and GTEx data. Then we performed ORA based on deregulated genes without considering the pathological stage. Deregulated genes were selected when was greater than 2 as for previous ORAs on L1000 data.

By reproducing the same analysis as for the L1000 data, but this time using transcriptomic profiles from breast cancer tissues, we observed that genes over-expressed were significantly involved in cellular mechanisms strongly associated with cancer. The terms “Cell cycle”, “Chemical Carcinogenesis” or “IL-17 signaling pathway” were found similarly to the analysis on L1000 data from carcinogenic compounds in breast cancer (Supplementary Table S3 online). Moreover, we also found pathways impacted by SMC compounds on MCF7 cells from L1000 as previously observed. Indeed, the terms “Oocyte meiosis” and “Progesterone-mediated oocyte maturation” stood out when enriched from over-expressed genes in diseased tissues (Fig. 4a; Supplementary Table S12 online).

Fig. 4.

Over representation analysis conducted on deregulated genes from TCGA-GTEx DEG analysis gene expression profile. (a) ORA representing up-regulated KEGG pathways. (b) ORA representing down-regulated KEGG pathways.

Under-expressed genes were found to be involved in a wider range of biological mechanisms and processes. These include a range of biological processes involved in cardio/muscular contraction, such as “Adrenergic signaling in cardiomyocytes”, or “Cardiac muscle contraction”, but also mechanisms involved in metabolism, such as “Regulation of lipolysis in adipocytes”, “Pyruvate metabolism” or “Glycolysis/Gluconeogenesis” confirming previous reported results^38–40 (Fig. 4b; Supplementary Table S13 online).

A consistent similarity pattern with subtle inter-individual variations connects tumor and chemical profiles

Given the diversity of forms of breast cancer, characterized in particular by differential expression of hormone receptors (ER, PR) and the HER2 receptor, it is plausible that exposure to chemical compounds induces variable biological effects in different individuals, resulting in specific transcriptomic signatures. Disease progression and response to chemical exposures may indeed vary according to genetic and epigenetic factors, and the characteristics of the tumor microenvironment.

To explore this variability, we performed DEG analysis on each tumor tissue sample, keeping the healthy tissue sample as control. The transcriptomic profiles thus obtained were then compared with in vitro profiles from the LINCS database, enabling to assess the similarities between individual transcriptomic responses and those induced by chemical compounds on cell models.

Heatmap analysis revealed a strong similarity among individual gene expression profiles from tumor tissues. When compared to profiles from the LINCS database, most individuals exhibited highly consistent similarity scores, as indicated by vertical bands of similar values.

However, some individuals displayed slight variations in their similarity to LINCS profiles, suggesting a degree of inter-individual variability in transcriptomic deregulation ex vivo. This variability might reflect differences in tumor biology, genetic background, or microenvironmental influences. Indeed, the impact of the environment is significant; exposure to a wide spectrum of molecules, even non-carcinogenic ones, and physical factors can alter the cellular state by inducing epigenetic changes or modulating local inflammatory processes. (Fig. 5; Supplementary Table S14 online).

Fig. 5.

Cosine similarity between L1000 mammalian compound and tumoral tissues gene expression profiles. In x-axis, the 220 L1000 MC gene expression profiles corresponding to 34 different compounds. In y-axis, the 1092 ex vivo individual gene expression profiles obtained from DEGs analysis.

Identification of chemicals that induce a breast cancer-like transcriptomic signature

We compared signatures from the LINCS L1000 dataset with the transcriptomic signature obtained from the DEG analysis performed using the TCGA and GTEx databases.

Thirteen Known Carcinogens are Found to Induce a Breast Cancer-Like Signature

We first made these comparisons on a L1000 subset from MC compounds (220 signatures), using landmark genes only (S1) and landmark in addition to best inferred genes (S2). As expected, we observed that the cosine similarity values were impacted by the applied z-score threshold, particularly with an increase in positive and negative values when a high threshold is applied. This was due to the number of genes being compared, which decreases significantly when high z-score thresholds were used (Fig. 6).

Fig. 6.

Gene expression cosine similarity between L1000 MC compounds and TCGA-GTEx tissue reference. In x-axis, we have 220 L1000 MC profiles corresponding to 34 distinct compounds. Dash lines represent for each z-score cutoff the cosine similarity threshold determined empirically. CP is short for “Cyclophosphamide”; MPA is short for “Medroxyprogesterone acetate”.

We also noticed that the number of values exceeding the similarity thresholds empirically determined using our method described above was limited when the analysis was performed considering the LINCS landmark genes only. The number of genes being too small to specifically characterize the transcriptomic signature, we have considered the best inferred genes in addition to the landmark ones to select similar signatures.

When the best inferred genes were added to characterize the transcriptomic profiles, comparison of the LINCS signatures with the reference signature derived from DEG analysis using the TCGA and GTEx databases showed that our method identifies a greater number of similar signatures. In fact, 52 signatures passed the empirically determined cosine similarity threshold of 0.0845 in the case of a z-score cutoff of 1. In the case of a z-score of 2, 20 signatures passed the threshold of 0.178. 17 L1000 transcriptomics signatures were identified as similar as TCGA-GTEx transcriptomic signature for the two z-score cutoffs, corresponding to 13 different MC compounds. The most found compound is Amsacrine appearing for three different concentration (10 M, 2.22 M, 3.33 M), followed by Chlorambucil (0.01 M, 0.04 M) and Diethylstilbestrol (0.01 M, 0.74 M). The highest cosine similarity value, equal to 0.369, was found when comparing the transcriptomic signature of 10 M griseofulvin-treated cells with the reference signature using a z-score cutoff of 2 (Fig. 6; Supplementary Table S15 online.

A Large Fraction of High-Risk Suspected Carcinogens Induces a Breast Cancer-Like Profile

Similarly to the previous analysis with MC chemicals, with the SMC chemicals, there was a big difference in the number of signatures identified as similar between those based solely on landmark genes and those including best inferred in addition to landmark genes. With the same logic we included the best inferred genes to the landmark ones for the analysis. We identified a much pertinent set of similar transcriptomic signatures. Indeed, there were 70 signatures (corresponding to 48 unique compounds) that exceed the cosine similarity threshold determined empirically for each of the two z-score cutoffs. These two thresholds were: 0.0846 for cutoff and 0.1772 for cutoff.

We also observed that the highest number of occurrences for a single compound is 3, and this applies to three distinct compounds: Naringenin (2.22 M, 3.33 M, 10 M), Escitalopram (0.03 M, 0.125 M, 3.33 M), and Cytarabine (0.01 M, 0.04 M, 0.74 M). The highest cosine similarity value, equal to 0.367, is associated with Pyrazinamide tested on MCF7 cell line, treated at a concentration of 0.125 M with a z-score cutoff of 2 (Supplementary Table S16 online).

A large proportion of molecules that exceed the thresholds and thus alter gene expression similarly to ex vivo perturbations were also identified in the previous analysis comparing the profiles of cells treated with SMC and MC compounds. Indeed, 32 of the 50 molecules identified in the previous analysis (Fig. 3) were found here among the 48 compounds identified for inducing similar gene expression deregulation to the reference ex vivo signature.

A network analysis to identify hub genes and key modules associated to breast cancer

In the aim to identify hub genes and key modules associated with breast cancer, a WGCNA analysis was performed on the 1092 breast cancer tissues. As a result, co-expressed genes were grouped into 13 modules which in combination with a topological overlap measure (TOM) allowed the visualization of significant genes co-deregulated by MC and SMC chemicals (highlighted in yellow in Supplementary Fig. S3 online and presented in in Supplementary Table S18 online.. We can observe that the two larger clusters (green-grey and purple-red) contain a large proportion of co-expressed genes deregulated by MC and SMC chemicals. While a mixture of up and down regulated genes are captured in the purple-red cluster, it is almost exclusively upregulated genes that are impacted by MC and SMC in green-grey cluster. Interestingly, many of them (such as MELK, E2F2, AURKA, TPX2, KIF4A, FOXM1, CENPA…) have been reported to be associated with breast cancer. Regarding the chemical compounds, our analysis focused on the gene deregulation observed at a standardized concentration of 10 M. This concentration was selected due to its prevalence across the dataset, ensuring a more robust and less biased comparison between different chemicals. Within this context, we notice that some SMCs such as citalopram, fluoxetine, or crizotinib impact some of these genes. We also observe that these chemicals are among those that share the most deregulated genes with the two clusters. This suggests that these chemicals might be a risk factor of breast cancer progression. (Supplementary Tables S19–S26 online).

Discussion

In this study, we explored the gene expression deregulation induced by chemicals characterized as mammalian carcinogens (MC) or suspected mammalian carcinogens (SMC), integrating both in vitro and ex vivo transcriptomics data. The association of large-scale chemical perturbation datasets with gene expression profiles from breast cancer cell lines and, healthy and tumor breast tissues allowed to bridge the gap between controlled experimental conditions and real biological systems. By comparing chemical-induced gene expression signatures with transcriptomic profiles from breast cancer tissues, we were able to assess the extent to which these chemicals trigger molecular responses similar to those observed in tumors.

Our analyses confirmed that MC and SMC compounds disrupt several key pathways involved in cancer development, particularly cell cycle regulation, p53 signaling, and MAPK pathways. These findings are consistent with previous research showing that many carcinogens interfere with processes that control cell growth and survival. Interestingly, we also observed that SMC compounds were strongly associated with endocrine-related pathways, supporting the idea that endocrine disruption plays a role in carcinogenesis. This is particularly relevant given that many endocrine disruptors are not officially classified as carcinogens but may still contribute to cancer risk through indirect mechanisms.

One interesting observation from our analysis is that gene expression changes appear to be more continuous than discrete. By comparing in vitro gene expression data (L1000) with ex vivo gene expression data (TCGA-GTEx), we were able to pinpoint that the deregulation of genes induced by chemicals in vitro mirrored that observed in clinical samples, confirming the potential role of chemicals in cancer progression. UMAP visualization confirmed distinct clustering of healthy and pathological tissues, ensuring the reliability of our differential expression analysis. However, diseased tissues did not cluster by pathological stage, suggesting a continuous rather than stage-specific molecular shift. We also noticed that genes down-regulated in breast cancer tissues were often linked to metabolic mechanisms and could promote breast cancer progression through metabolic reprogramming, tumor microenvironment modification and cancer cells adaptation^41,42.

By incorporating cosine similarity analysis, we ranked chemicals based on their resemblance to breast cancer gene expression profiles. As expected, several well-known carcinogens, such as Amsacrine, Chlorambucil, and Diethylstilbestrol, showed strong transcriptomic similarity to breast cancer tissues. About the SMC compounds, some of them showed high similarities with the transcriptome of some patients. Among them, we can notice some antidepressants such as escitalopram and fluvoxamine. Interestingly, a study on a large breast cancer cohort reported that selective serotonin reuptake inhibitors (SSRIs) were associated with a 27% increase in breast cancer⁴³. More surprisingly, some chemicals used in breast cancer protection such as raloxifene⁴⁴ or letrozole⁴⁵ were also ranked in the top hits. It could be explained by the fact that the compounds have been tested on a breast cancer cell line (MCF7). Although MCF-7 possess estrogen, progesterone, glucocorticoid and Human Epidermal Growth Factor Receptor-2 (HER2) receptors, which are features commonly seen in luminal subtype breast cancers⁴⁶, the main regulatory pathways are already highly deregulated. So, chemicals that modulate gene’s deregulation in MCF7 could be assumed to be as much protective as aggressive by our approach. Since we translate the gene’s deregulation, it is possible that our method detects as many chemo-preventive agents as risk factor chemicals observed in vitro to ex vivo, considering breast cancer tissues from patients. Having transcriptomics data on a “healthy” cell line like, for example, hTERT-HME1, exposed to the same set of MC and SMC chemicals would be interesting to characterize the chemo-prevention or the risk factor potential of these agents.

Other pitfalls have to be considered in this analysis. The LINCS datasets are based on 978 directly measured landmark genes while the 9196 best inferred genes are gene expression defined from a robust computed approach. Therefore, the molecular profile for MCF7 exposed to chemicals considered in this study might be a little different from the reality. However, the over representation analysis based solely on landmarks or including the inferred genes detected, in general, the same set of deregulated pathways. Also, it is known that KEGG can detect very broad pathways (like cell cycle) and so some of the pathway’s results have to be considered cautiously. Finally, our study does not consider directly the gene’s susceptibility to mutation that could predispose an individual to develop breast cancer. It is known that mutations in genes like BRCAC1 and BRCA2, among others, increase the risk of breast cancer development⁴⁷. Having access to the RNA sequencing data for 1092 individuals from the Breast Invasive Carcinoma project, it could be possible to assess if the genes deregulated by chemicals in vitro are characterized as susceptible genes with critical genetic variations involved in the progression of breast cancer.

Conclusion

Overall, our study demonstrates that transcriptomic profiling, combined with breast cancer tissue expression data, offers a powerful strategy for identifying and prioritizing potentially hazardous chemicals as a risk factor in breast cancer progression. By leveraging cosine similarity analysis and multi-source gene expression comparisons, we introduce a novel way to assess the carcinogenic potential of chemicals based on their molecular impact rather than their historical classification. This approach enhances existing risk assessment strategies and highlights the need for continued toxicological monitoring of chemicals with unrecognized cancer-related effects. The conclusions derived from this study are only suggestive and further molecular level validation is required in animal models or tumor models to validate them.

Methods

Chemical data

We have gathered different sets of compounds categorized as known and suspected to be mammalian carcinogens. First, a set of 268 of potential Mammalian Carcinogens (MC) chemicals, known to increase the synthesis of estradiol or progesterone, was established based on the studies by Borrel and Rudel¹⁰. Secondly, a set of 1626 endocrine disruptors chemicals (EDCs) and defined as suspected to be mammalian carcinogens (SMC) was compiled. For this dataset, we considered chemicals defined as endocrine disruptors in one of the three databases (DEDUCT, TEDX, EU) as many of them are also present in the mammalian carcinogens dataset (75 chemicals in common with the MC dataset)^48–51.

In total we assembled 1819 chemicals (Supplementary Table S17 online), all of them encoded with a unique identifier.

Transcriptomics data

The differential gene expression of cells exposed to our set of chemicals were collected, when available. We used the Library of Integrated Network-Based Cellular Signatures (LINCS) Program which includes the differential gene expression of cells exposed to more than 30000 various perturbing agents²⁶. Subsequently, we used data from The Genotype-Tissue Expression (GTEx) Portal and The Cancer Genome Atlas (TCGA) to explore and observe breast tissue gene’s expression specificity (cancer/non cancer tissue).

LINCS: cell level

We loaded the LINCS dataset (L1000) and used the z-scores computed for each sample. Basically, in L1000, from the raw data acquired through flow cytometry using the Luminex platform, differential gene expression signatures between a chemical versus a vehicle (usually DMSO) are computed for a set of 978 landmark genes and transformed into a z-score. The higher is the z-score, the more is the deregulation of the gene. A positive z-score corresponds to an up-regulation instead of a negative z-score that means a down regulation of the gene. Landmark genes represent 82% of information in the full transcriptome. In addition to the landmark genes, a similar study including 9,196 genes (defined as best inferred) approximated through an inferred model, as described by Subramanian et al., was performed. It allows to assess if landmark genes contains enough information to characterize breast cancer risk factors²⁶.

L1000 contains 1,201,944 signatures, for which each signature corresponds to the differential expression of genes for a specific compound, cell line, concentration and exposure time. Depending on the cell lines, there are replicates i.e measurements with all these parameters have been performed multiple times.

From these data, we filtered those relevant to our study. First, we removed replicates by retaining only the signature with the highest transcriptional activity score (TAS) which is an aggregate measure of signature strength and replicate correlation²⁶. A high TAS represents a chemical that lead to high gene’s deregulation. We, then, selected signatures corresponding to the MCF7 cell line (breast cancer) with an exposure time of 24 hours only. This resulted in a dataset of 47,654 signatures for 12,441 compounds. We identified 52 different concentrations, although not all concentrations were tested for each compound. Each compound is associated with a range of different concentrations, including a set of common concentrations. Finally, by filtering the data using our list of compounds (MC and SMC), we retained a dataset of 1060 gene expression profiles for 161 chemical compounds.

TCGA-GTEx: tissue level

To investigate the differences in gene expression between ‘healthy’ and breast cancer tissue, we collected publicly available data on gene expression in non-diseased human tissues and tissues affected by breast cancer. We obtained (reads count) data from two publicly accessible databases. The first, the GTEx (Genotype-Tissue Expression) database, includes gene expression profiles from nearly 1,000 non-diseased individuals across 53 different tissue types, including specific data for breast tissue. Given the different formats of gene expression data available in GTEx, we chose to use gene counts directly, allowing for a straightforward comparison with data from The Cancer Genome Atlas (TCGA).

The second database, TCGA, offers comprehensive genomic, epigenomic, transcriptomic, and proteomic data, covering more than 20,000 primary cancers categorized into various projects or studies. Our analysis focused on the transcriptomic data from the Breast Invasive Carcinoma project (ID: TCGA-BRCA), which is adapted for our study. A differential gene expression analysis was performed between the transcriptomics data from TCGA and GTEx in order to characterize genes with expression levels found to be significantly different between the two conditions.

In TCGA, pathological tissues are categorized into four distinct groups: Stage I, Stage II, Stage III, and Stage IV. These groups are an indicator of cancer progression representing how large the tumor is and if it has spread as defined by the National Cancer Institute⁵². They are created using metadata to consolidate associated subgroups; for example, Stage IA and Stage IB are combined under Stage I. This categorization enables to perform analyses taking into account the stage of the disease. As a result, we had 5 distinct datasets for TCGA data, data grouped regardless of disease state, and the 4 sub-data sets corresponding to each disease state.

Uniform manifold approximation and projection (UMAP)

In order to visualize the data and represent them in an easily interpretable space, we performed a Uniform Manifold Approximation and Projection (UMAP) on the RNAseq data from GTEx and TCGA. This method is a nonlinear dimension reduction technique that aims to capture both the global and local structure of the data. We used the python package umap-learn in version 0.5.4 to apply this method and python packages matplotlib version 3.7.2 and seaborn version 0.12.2 to visualize the projection in 2 dimensions⁵³.

Over representation analysis (ORA)

From the filtered data, we created subsets corresponding to different categories of compounds (MC, SMC). To do this, we first selected deregulated genes with a z-score > 2 or < -2. Given the global scope of our approach, we aimed to create conditions that allowed for the detection of subtle variations sensitive to the specific nature of chemical compounds disrupting the cells. Too high or too low concentrations may not accurately reflect the disturbances naturally induced by these compounds. Therefore, in the next step, we selected a range of concentrations to minimize noise in subsequent analyses. So, we decided to retain only the concentrations between 0.1 and 1.3 M. This concentration range was designed to maximize the number of available L1000 signatures that aligned with the conditions relevant to our study’s chosen approach. As a result, we obtained two subsets of L1000 data, each comprising signatures with similar experimental conditions corresponding to the two types of compounds described previously. Each of these subsets have then be used to perform an Over-Representation Analysis (ORA).

To facilitate the characterization of pathways modulated by a set of chemicals, we developed a method able to identify set of genes highly and frequently deregulated by each subset. This scoring method consists to evaluate the overall tendency of each gene through an ensemble of gene expression profiles. This score is described in equation (1).

1

Equation (1) : Formula of the Gene specific Score () representing the overall deregulation tendency according to a z-score cutoff. s: Number of signatures exceeding the z-score cutoff for this gene. ns: Number of signatures not exceeding the z-score cutoff for this gene. : Average of values exceeding the z-score cutoff for this gene. : Average of values not exceeding the z-score cutoff for this gene.

Given that each subset contained a different number of signatures, it was important to establish a cutoff that would allow the selection of genes to be evaluated in each case, taking into account the composition of each subset. Accordingly, a cutoff score was established, as described in equation (2). This score cutoff expression represented the case where, for a given gene, there were as many signatures that pass the predefined cutoff as there were signatures where the cutoff was not exceeded. And in order to have an expression that is applicable to all genes in a global approach, we weighted this case by the average value of all genes exceeding the predefined z-score cutoff and by the average value of all genes not exceeding it.

2

Equation (2) : Formula of the cutoff. This equation represents the value of score from which we consider a specific gene as deregulated. : Average of values exceeding the z-score cutoff across all genes. : Average of values not exceeding the z-score cutoff across all genes.

Once the genes have been selected as up-regulated or down-regulated, an Over-Representation Analysis (ORA) was performed using two reference databases: KEGG Human 2021^54–56 and Gene Ontology Biological Process 2021^57,58. This allowed us to identify the biological processes and cellular mechanisms in which the highly deregulated genes, caused by the different families of compounds, were involved. For each compound subsets (MC, SMC) we performed ORA taking into account first landmarks genes and then best inferred genes in addition to landmarks genes in a second part. At the end, deregulated pathways were ranked according to their adjusted p-value.

The package gseapy⁵⁹ version 0.9.5 in python was used to perform ORA through Enrichr web-server and to generate the different enrichment figures⁶⁰.

Differential expression gene (DEG) analysis

As for GTEx and TCGA, we had access to the reads count of gene expression (gene counts), we performed a Differential Expression Gene (DEG) analysis, to identify the changes in gene expression associated with breast cancer in humans. We adapted and conducted a DEG for breast cancer using the previously mentioned GTEx data (healthy tissues) and TCGA data (cancerous tissues), following and adapting the protocol proposed by Huey-Miin Chen et al.⁶¹. We downloaded all the files required to ensure the proper execution of the protocol. Both transcriptomics data from GTEx and TCGA were gathered using R version 4.4.1 with the UCSCXenaTools package version 1.4.8⁶². We thus had gene counts for 178 healthy breast tissue samples that served as references and 1,092 breast cancer tissue samples for comparison.

TCGA and GTEx gene counts data downloaded from toilHub through UCSCXenaTools were previously log2(x+1) transformed. Therefore, it was necessary to back-transform our gene expression data before performing the DEG, which was carried out using the edgeR package, version 4.2.2. We ensured that genes had enough large counts to be kept for further steps of the analysis. We checked that all samples were assumed to have a similar range and distribution of logCPM values. Then we normalized the data to scale the raw library sizes and performed gene-wise quasi-likelihood (QL) F-tests using Generalized Linear Models (GLMs). Once the QL dispersions were estimated for all genes, we performed the actual test to obtain the differential expression of the genes. Then, we kept only the DE genes that were above a 0.05 false discovery rate (FDR) in order to retain the most biologically relevant genes. Finally, to align with the L1000 data for further analysis, we calculated z-scores from the log fold changes (logFC) and p-values.

From the differentially expressed genes (DEGs), we conducted an Over Representation Analysis (ORA) using the same approach as applied to the L1000 data but this time not as a holistic approach with the custom Gene score developed. This time we only applied the cutoff to select the deregulated genes in order to conduct the ORA.

Cell-tissue cosine similarity

Having collected and generated similar data revealing the impact on genes at the cellular level through L1000 data and at the tissue level through GTEx and TCGA data, it is insightful to compare the transcriptomic profiles. To this end, we estimated the similarity between these two types of transcriptomic profiles by measuring cosine similarity.

3

Equation(3): : Formula of Cosine Similarity. A and B are two n-dimensional vectors of attributes. and are their i-th components respectively.

Cosine similarity evaluates how closely two-dimensional vectors align by measuring the cosine of the angle between them. Unlike Euclidean distance, which quantifies the difference in vector positions, cosine similarity focuses on the direction of the vectors⁶³. Using cosine similarity, two transcriptomic profiles with similar expression patterns will be considered similar, even if their overall expression levels differ. Thus, we identified this method as effective for our case and conducted a comparative analysis.

Following the Differential Expression Gene (DEG) Analysis conducted on the TCGA-GTEx data, we obtained differential expression data for 14,701 genes (z-scores) with a . On the other hand, we had gene expression data from MCF7 cell lines exposed to compounds identified as mammalian carcinogens (MC). For this analysis, we considered only the profiles for a 24-hour chemical exposure, resulting in a dataset of 220 distinct signatures, corresponding to 34 compounds for an ensemble of 17 doses. Additionally, we categorized the genes into two distinct sets: Set 1 (S1), which includes 978 landmark genes, and Set 2 (S2), comprising both landmark genes and best inferred genes for a total of 10,174 genes. When not applying any z-score cutoff (meaning when considering genes with ), we identified in S1, 888 genes that overlap with the tissue data, while in S2, we found 8,833 overlapping genes. For both scenarios, we evaluated the cosine similarity for all corresponding common genes across two cutoffs: , . This analysis allowed us to assess how similar or different each L1000 signature was compared to the tissue transcriptomic profile obtained from the DEG.

Next, we reproduced the same analysis, but considering SMC compounds, in order to assess the potential risk associated with exposure to these compounds, in the context of the onset or development of breast cancer. Filtering the L1000 data using the same filters except for the chemical compounds, leads to a set of 840 different signatures. In the same way, we carried out the analysis considering both sets of genes (landmark and landmark in addition to best inferred). Figure 7 provides a workflow illustration that highlights the various steps and analyses conducted throughout this study.

Fig. 7.

Overview representation of the study workflow. Transcriptomic data from healthy and tumor breast tissue samples were collected from the GTEx and TCGA databases, respectively. A differential gene expression analysis was performed to establish a representative transcriptomic profile of breast cancer tissue. Based on this reference profile, an Over-Representation Analysis (ORA) was conducted. In parallel, transcriptomic data from L1000 cell culture experiments were filtered and grouped into different datasets based on chemical compounds and genes defining their transcriptomic profiles. ORA was also performed on these subsets. Finally, transcriptomic profiles from the different datasets were evaluated using a cosine similarity-based approach and an empirical methodology to assess the breast cancer risk associated with specific compounds.

Empirical threshold determination for similarity estimation

Once two transcriptomic profiles composed of a set of z-scores have been compared, it is interesting to estimate the cosine similarity value at which they can be considered similar. This will make it possible, for example, to assess whether a compound represents a risk of breast cancer by estimating whether it deregulates a cell in a similar way to what can be found in breast cancer conditions ex vivo. In addition to the difficulty of estimating which value is a good threshold for assessing the similarity of two transcriptomic profiles, it is important to consider the number of genes being compared, as the number of dimensions has a significant impact on the cosine similarity values measured. Thus, we developed a method that enable to define an empirical and adaptive threshold for the number of genes being compared, allowing to assess the similarity of two transcriptomic profiles as a function of the number of common genes they share.

In order to determine the terms of our equation, we simulated a set of pairs (100,000 pairs) of random transcriptome profiles comprising a predefined number of genes. From these simulations we calculated the cosine similarity for each pair in order to estimate the cosine similarity distribution expected for non-similar profiles. This enabled to estimate the mean and standard deviation for different transcriptomic profile lengths. Supplementary Fig. S2 accessible online, shows that the threshold value following the proposed equation (3) is indeed impacted by the number of genes in the transcriptomic profiles. In order to estimate the constant “k”, we need to determine a threshold from experimental transcriptomic profiles, corresponding to a predefined number of genes.

4

Equation (4) : Formula defining Threshold Cosine Similarity. N is the number of dimensions of the two compared arrays. Here, N corresponds to the number of genes. represents the mean of cosine similarity distribution for random arrays of N genes. k is a constant determined empirically and independent from the number of genes. represents the standard deviation of cosine similarity distribution for random arrays of N genes.

We, therefore, studied the LINCS transcriptomic profiles in order to identify sets of similar profiles. To do this, we generated cosine similarity heatmaps between transcriptomic profiles of the same compound, at the same exposure time but different concentrations. Indeed, we assumed that if a compound had a specific effect on gene expression in a cell, its effects would be measurable for similar conditions (exposure time and concentration). To minimize noise and retain specificity when calculating cosine similarity, we set all gene z-score values between -2 and 2 to 0.

In this way, we were able to identify three compounds (Hexachlorophene, Ellagic Acid, Orlistat) for which a pattern revealing higher cosine similarities was observed for similar conditions (Fig. 8). Indeed, we could observe for Hexachlorophene that at the range of concentrations comprising 0.12 M, 0.125 M, and 0.25 M, we identified closer cosine similarity values. We assumed that at this set of concentrations, the effect induced by exposure to hexachlorophene for 24 hours was similar. Thus, the cosine similarity values found will be representative of what can be expected between two similar transcriptome profiles for a predefined set of genes. In the same way, similar conditions were observed for Ellagic Acid (0.08 M, 0.125 M, 0.25 M) and for Orlistat (0.03 M, 0.04 M, 0.08 M, 0.125 M).

Fig. 8.

Heatmaps of cosine similarity scores between L1000 gene expression profiles for different compounds. L1000 transcriptomic profiles were grouped by compound and kept for only 24 h exposure time. Comparison of these profiles using cosine similarity highlights similar cellular response to chemical perturbation. In red circles. (a) Cosine similarity values for Hexachlorophene. (b) Cosine similarity values for Ellagic Acid. (c) Cosine similarity values for Orlistat.

After collecting the cosine similarity values of those profiles identified as similar, we got the minimum cosine similarity value calculated for each pair of similar profiles, for each of the 3 cases. By averaging these three minimum values, we then obtained the threshold value determined empirically for a number of genes “N” equal to 978: . From this threshold we could estimate the constant “k” where

Finally, in order to determine the and values for a gene set equal to 978 with greater precision, we simulated the cosine similarities of 100, 000 pairs of random transcriptomic profiles. The mean of s and s were retained as the values to be used in the equation. This gave a value for k equal to 3.595

In this way, we could reuse the values of , and k when comparing them with profiles derived from LINCS on the landmark genes set. We could also recalculate and values for different transcriptomic profile sizes.

Network analysis based on WGCNA

Weighted Gene-expression Correlation Network Analysis (WGCNA) is a method originally designed to identify groups of highly connected genes. The principle is based on correlation networks: a network is constructed based on gene expression profiles across samples, representing the interconnectivity between genes. The network is coded by an adjacency matrix, which can be linked to a correlation matrix between the genes. Clustering of this matrix allows for the identification of clusters of highly correlated genes, also known as modules. In order to use this method, we prepared tumor samples RNA-seq data from TCGA-BRCA dataset. To prepare the data for co-expression network analysis, the raw expected counts required normalization and transformation. The integer-rounded count matrix was normalized to account for differences in library size among samples. Subsequently, a Variance Stabilizing Transformation (VST) was applied. This critical step transforms the data to a log-like scale, where the variance becomes approximately independent of the mean expression level. We prepared the data using the pydeseq2 package (version 0.5.2) delivered by Muzellec B et al.⁶⁴, and implemented the code in python (version 3.15.5). This ensures that the downstream correlation calculations, which form the basis of the network analysis, are not biased by the high variance typically associated with highly expressed genes. The resulting VST-transformed matrix, being appropriately normalized and stabilized, was used as the input for WGCNA. Then we used the pyWGCNA package (version 2.2.1) delivered by Rezaie N et al.⁶⁵ and implemented in python (version 3.13.5) to create modules of genes associated with breast cancer using the 1092 breast cancer tissues. With WGCNA, 13 modules were identified. Modules are, by default, named by color. Then, to generate a network of genes highly coderegulated, a topological overlap measure (TOM) approach was developed. The TOM assess reliably how interconnected two genes are within a network. Instead of only looking at the direct connection between them, the TOM also incorporates the degree to which they share connections with the same set of neighboring genes, making it a robust indicator of network proximity. In our study, a TOM of 0,13 was considered to avoid too many interconnected genes and to limit a high number of small hubs of 2 or 3 genes. Finally, genes deregulated by MC and SMC chemicals that overlap with genes presented in the network are depicted in yellow. The protocol is highlighted in Supplementary Fig. S4 accessible online.

Author contributions

R.A performed to the data collection, data analysis, developed the scoring functions, prepared figures and tables and wrote the original manuscript. O.T initiated the study and wrote the manuscript. All authors reviewed the manuscript.

Data availability

The datasets used and/or analysed during the current study are available within the article and its supplementary materials. Transcriptomics data from LINCS L1000 have been deposited in the Gene Expression Omnibus (GEO; https://www.ncbi.nlm.nih.gov/gds/advanced) repository under accession number GSE92742. RNA-Seq data for breast healthy tissues from GTEx database and RNA-Seq data for breast cancer tissues from TCGA database have been deposited in The database of Genotypes and Phenotypes (dbGaP; https://www.ncbi.nlm.nih.gov/gap/advanced_search/) under accession numbers phs000424 and phs000178 respectively.

Declarations

Competing interests

The authors declare no competing interests.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-27732-0.