2.1 A systematic review to identify relevant mechanisms and non-animal methods for inclusion into a breast cancer AOP

In an attempt to distill the extraordinary diversity of cancer into a common set of underlying core cellular parameters, the ‘Hallmarks of Cancer‘ concept was introduced [16,17,18]. It currently includes 14 underlying principles of cancer development that provide an actual description of the common characteristics of cancer and functional capabilities that are crucial for the ability of normal cells to form malignant tumors. However, the underlying causally-related molecular and cellular mechanisms are still actively researched. Here the AOP concept can be particularly valuable. An AOP is a hierarchical representation of a defined sequence of causally-related molecular and cellular events, whose disruption at different levels of biological organization can eventually lead to a defined adverse (pathological) outcome (AO) (Fig. 1B). For a specific adverse effect, the corresponding AOPs collate the existing information including research articles, clinical reports and public databases. To date, more than 400 AOPs (with varying levels of completeness) have already been proposed for various endpoints that are actively reviewed in a community-based approach and published online in a publicly available AOP-Wiki database (https://aopwiki.org/).

2.2 The breast cancer AOP 200—current status

The ‘Estrogen receptor activation leading to breast cancer’ AOP (AOP 200, https://aopwiki.org/aops/200) was originally published by Morgan et al. [19] and recently expanded by Del'haye et al. [20]. It starts with perturbation of ER activity as specific molecular initiating event (MIE), followed by multiple interconnected key events (KE) at the molecular, cellular, tissue and organ level, which may eventually lead to ER + breast cancer and cancer-related death (AO) (Fig. 1C; Table 1; Data Matrix, AOP overview tab) (https://aopwiki.org/aops/200#Events). At the molecular level, KEs include causally-linked changes in gene expression and protein production of breast cells, leading on a cellular level to the escape from cell cycle regulation as well as changes in apoptosis and motility. At the tissue level, these cellular changes translate to local hyperplasia, disruption of tissue architecture, invasion, and eventually metastasis to distant organs. In addition to effects in cancer cells of the primary tumor (Fig. 1C, solid lines), the AOP further covers changes in the local tumor microenvironment (TME) (Fig. 1C, dashed lines), including endothelial proliferation, angiogenesis, and local responses from tumor-associated macrophages and fibroblasts. Importantly, these KEs align with many of the proposed ‘Hallmarks of Cancer‘ categories (Table 1; Data Matrix, AOP overview tab) and are interconnected through weighted KE relationships, which are determined based on scientific evidence (from weak to high; https://aopwiki.org/aops/200#KE_relationships).

Table 1 Overview of Cancer Hallmarks, the breast cancer AOP 200, and proposed breast cancer key events

As the number of AOPs addressing cancer-related and other disease-relevant mechanisms is constantly increasing, so is the overlap of shared MIEs, KEs or AOs among them. This will ultimately lead to the combination and interconnection of AOPs into AOP networks [21, 22], which has recently been demonstrated for endocrine-mediated perturbations [23], thyroid hormone disruption [24], and carcinogenicity [25]. The breast cancer AOP 200 likewise interconnects with three other AOPs that address different modes-of-action leading to breast cancer, i.e., aryl hydrocarbon receptor activation (AOP 439) [26], increased DNA damage (AOP 293), and increased reactive oxygen and nitrogen species (AOP 294) (Fig. 1D; Data Matrix, AOP overview tab). This way, the increasing scientific knowledge on breast cancer mechanisms could be effectively integrated in a growing catalogue of defined MIEs, KEs, and AOs linked by specified key event relationships.

2.3 Identification of additional KEs for an updated breast cancer AOP

In order to investigate to what extent the existing MIE, KEs, and AO of the breast cancer AOP 200 are represented by published biomedical and toxicological research articles, we systematically reviewed and categorized 299 relevant publications from the PubMed database. With a particular focus on recent original studies (published within the last five years, no reviews) that used non-animal methods to study mechanisms and (environmental) stressors of breast cancer, we collected the retrieved articles and their categorizations in a Data Matrix (Data Matrix, Literature classification tab). The search strategy was based on MeSH term combinations that describe the study focus and methodologies used (Data Matrix, All queries tab), followed by a manual inclusion of high priority publications based on title and abstract review.

We next determined the main study focus and mapped these publications to individual categories representing key breast cancer mechanisms, which included the established MIE, KEs, and AO of the breast cancer AOP 200 (Data Matrix, Literature classification tab). These 35 categories consisted of (environmental) stressors, molecular and cellular responses in both primary tumor and stromal cells, responses at the tissue level, and the AO, i.e., breast cancer. We further grouped these studies into primary research fields, i.e. basic biomedical research (mechanisms of breast cancer; 160 studies, 50%) translational biomedical research (diagnostics, (pre-)clinical testing, drugs, therapy; 103 studies, 32%) and toxicology (environmental effects; 54 studies, 17%) (Data Matrix, Analysis tab). The analyzed studies from basic and translational biomedical research covered cellular (tumor and stroma) and tissue level responses to a similar extent (Fig. 2A; Data Matrix, Analysis tab). The AO, i.e. breast cancer (covering various tumor (sub-)types), was most strongly addressed by translational biomedical studies. Notably, the analyzed toxicological studies almost exclusively investigated effects on the tumor cell level. The main KEs that were investigated in the toxicological studies included the activation of the ER, effects on gene expression and proliferation. In contrast, biomedical studies covered a much broader spectrum of more complex KEs, including apoptosis, modulation of the TME, tumor growth and invasion (Data Matrix, Analysis tab).

Fig. 2

Use of non-animal breast cancer methods in biomedical research and toxicology. A-D Visualization of the collected publications and their classifications according to the addressed KE levels (tumor cell, stromal cell, tissue, adverse outcome) or the experimental categories (in silico, in chemico, in vitro, in vivo), experimental types (e.g. cells, biopsies, AI), in vitro methods (e.g. cell culture, scaffolds, lab-on-the-chip), in vitro dimensions (2D, 3D), and high-throughput capacity used in different research fields (basic and translational biomedical research, toxicology) (see Data Matrix, Analysis tab for more details)

Based on the analyzed publications, we further identified additional key breast cancer mechanisms that were not yet established as KEs (Table 1; Data Matrix, AOP overview tab). These studies are mainly from the biomedical domain (Data Matrix, Literature classification tab) and represent key breast cancer mechanisms that should be considered for timely inclusion into AOPs. In particular, important breast cancer mechanisms that were studied in many of the analyzed publications but are not yet covered by the current breast cancer AOPs relate to the decrease of cell stiffness and cell adhesion when transforming from non-malignant to metastatic states. For example, it has been shown that estrogens determine the organization of the essential cell-cell adhesion molecule E-Cadherin at adherens junctions as well as stiffness and motility of breast cancer cells [27]. The higher mechanical elasticity and deformability of cancer cells links to a reorganization of the actomyosin cytoskeleton and strongly correlates with cell malignancy and metastatic potential [28]. In addition, metastatic sites differ between individual cancer types [29] with variations in cytoskeletal organization and stiffness of breast tumor subpopulations matching the biomechanical properties of the metastasized organs being a possible mechanical indicator explaining metastatic site preferences (organotropism) [30]. In addition to changes in biomechanical properties, invasive breast tumor cells do further show a reduced cell-cell and cell-matrix adhesion, which increases their ability to detach from the primary tumor, a process mimicking the developmental epithelial-to-mesenchymal transition (EMT) program [31], and to invade the surrounding stroma. Along this line, the Cadherin- and Integrin-family transmembrane adhesion proteins have been identified to play a major role in the transition to metastatic states in breast cancer [32, 33].

These findings emphasize the relevance of decreased cell stiffness and cell adhesion as disease mechanisms (KE) in metastatic breast cancer progression, and could be included into the existing breast cancer AOPs or used to establish new AOPs. These developments would establish connections between alterations of ER activity (MIE 1181) with metastatic breast cancer (AO 1982) through reduction of cell stiffness (new KE) by modulation of actomyosin contractility as well as reduction of cell-cell adhesion (new KE) by modulation of E-Cadherin localization. Additionally, they would consider induction of EMT (KE 1457), increased cancer cell motility (KE 1241), and increased invasion (KE 1196) (Fig. 1E) as part of the disease mechanism.

2.4 Systematic collection of available non-animal methods in a breast cancer AOP

In addition to organizing available knowledge on cancer mechanisms, AOPs could further serve as information sources for selection of suitable non-animal methods (or combinations thereof) that adequately recapitulate particular physiological conditions and disease states. However, with regard to breast cancer research, this potential of AOPs is rather unexplored. Therefore, we further categorized the collected publications according to experimental categories, experimental types, in vitro methods, in vitro dimensions, and high-throughput capacity (Fig. 2; Data Matrix, Literature classification tab).

The collected publications encompass models such as simple biochemical and cell-based models as well as more advanced co-culture models, organoids, microfluidic lab-on-the-chip, and computational models. The main KEs that were investigated using in vitro and ex vivo methods at the tumor and stromal cell level included alterations of gene expression, increased proliferation, decreased apoptosis, and activation of tumor-associated fibroblasts. The two newly proposed KEs, i.e., decreased cell stiffness and decreased cell adhesion, were addressed by a smaller number of methods (Data Matrix, Analysis tab). At the tissue level, methods focused on modulation of ECM composition, increased tumor growth, and invasion. In chemico and in silico methods were mainly used to investigate mechanisms leading to activation of the ER, e.g. by xenobiotic exposure, and alterations of gene expression at the tumor cell level. In addition to the collection of available methods for individual KEs, this analysis further highlights KEs for which limited or no methods have been extracted from the retrieved publications and, thus, could be prioritized when developing non-animal methods. These method gaps include increased second messenger production, alterations of the circadian clock, occurrence of cancer-related exosomes, increased proliferation and migration of endothelial cells, induction of tumor cell intra-/extravasation, and occurrence of circulating tumor cells (Data Matrix, Analysis tab).

When comparing the experimental types performed in the three research fields, toxicology is dominated by the use of 2D cell culture methods using single cell lines (Fig. 2B; Data Matrix, Analysis tab). Single cell lines have also been used most frequently in basic and translational biomedical research, but were closely followed by co-culture or spheroid models. Although 2D cell culture remains a common in vitro model for studying tumor development and progression, the frequently observed altered growth characteristics or drug responses between 2D and 3D cell cultures have led to specific recommendations for more complex models in drug discovery [34]. Three-dimensional spheroid, organoid, scaffold-based, and microfluidic lab-on-the-chip systems currently represent the most advanced in vitro models recapitulating physiological and clinically-relevant breast cancer disease conditions to a large extent [35]. Based on our analysis, these complex non-animal methods are represented to a considerably higher extent in basic and translational biomedical research compared to toxicology (Fig. 2B; Data Matrix, Analysis tab). With regard to the investigated breast cancer KEs, the simpler 2D models were mainly used to study mechanisms at the tumor and stroma cell level, whereas 3D models were applied to study more complex responses at the tissue level and the AO (Fig. 2C; Data Matrix, Analysis tab).

Simpler 2D cell culture methods are often considered to have higher compatibility with robotic HTS applications, which can provide data for thousands of tests in little time. Interestingly, this notion does not necessarily apply to the breast cancer studies we analyzed, in which 2D- and more complex 3D-based methods were used equally in HTS-related projects (Fig. 2D; Data Matrix, Analysis tab). The majority of the analyzed HTS methods were used in toxicology and translational biomedical research (Data Matrix, Analysis tab). However, current advances in 3D culture methods promise an increased integration of 3D methods into HTS platforms in the biomedical domain [36].

Notably, translational biomedical research provided the highest proportion of published articles using ex vivo methods, particularly, cells isolated from primary material (biopsies). Other methods such as cultivation of organ-like structures derived from dissociated (organotypic) or intact, non-dissociated (explants) primary material have rather rarely been used (Fig. 2B; Data Matrix, Analysis tab). These ex vivo methods mainly focused on investigating different aspects of breast cancer at the AO level, in particular to study tumor invasion using patient-derived material in spheroid and organoid models (Data Matrix, Analysis tab).

With regard to the use of in silico methods to study breast cancer-related mechanisms, about every fourth article from the toxicological research field used computational tools, particularly structural analyses to investigate the MIE, i.e. ER activation (Fig. 2B; Data Matrix, Analysis tab). For example, the quantitative structure-activity relationship (QSAR) modeling method and molecular docking simulations are frequently used in toxicology to analyze receptor-ligand interactions, e.g., to predict ligand binding affinity of environmental stressors to ERα and, thus, to identify putative ER-mediated endocrine disrupting chemicals stimulating breast cancer [37]. The resulting data are the basis for so-called grouping and read across approaches, which are currently the most commonly applied non-animal methods for identification and characterization of chemical hazards without animal testing [38]. In the biomedical domain, structural analyses are successfully employed in the drug discovery process, e.g., to explore potential targetable sites on ERα and key structural traits to develop ERα inhibitors for breast cancer therapy [39, 40].

In recent years, artificial intelligence (AI) and computer vision have made considerable progress. AI involves machine learning (ML) and deep learning (DL) algorithms, which extract knowledge from sample data, known as training data, without being explicitly programmed. Among our dataset, AI-based models have exclusively been used in basic and translational biomedical research papers related to breast cancer (Fig. 2B; Data Matrix, Analysis tab). AI, ML, and DL become promising tools supporting or even exceeding the performance of human experts with regard to analysis [41], diagnosis [42, 43], and surgery [44] of breast cancer. For example, He et al. reported the development of a DL algorithm, which has been trained on spatial gene expression data and breast tumor morphologies in order to predicting local breast cancer biomarker expression levels directly from clinical histopathology images [41]. Another recent study has introduced the ML tool MFmap, which matches cell lines to tumor and cancer subtypes and thus can aid biomedical researchers with the selection of suitable methodologies to address their research question [45]. The rise of AI models in biomedicine has further triggered the establishment of the community-driven AIMe registry, which allows developers to easily register their AIs and helps researchers identify available AI systems suitable for their use cases [46]. In combination with advanced microscopic imaging, ML-based image analysis methodologies facilitate the automatic detection and quantification of cell morphologies. One such example is the ML-based analysis of the estrogen-dependent organization of E-Cadherin at cell-cell contacts in breast cancer cells [47], which correlates with changes in cell stiffness and cell motility of breast cancer cells [27] and which we propose as relevant KEs for an updated breast cancer AOP (see Fig. 1E). Along this line, cell morphologies are generally regarded as a holistic readout reflecting the biomechanical and physiological properties of single cells and, with regard to breast cancer, have been analyzed at high throughput in image-based phenotypic screening approaches using 2D and 3D methods in toxicology and biomedical research [48,49,50,51]. Moreover, cell morphologies can aid the differentiation of cancer from non-cancer cells and provide information on their tumorigenic and metastatic potentials [52, 53]. AI might also be important to facilitate the development of even more advanced in silico prediction tools but also of more comprehensive AOP Networks.