Epithelial-mesenchymal transition (EMT) is a reversible cellular process in which epithelial cells lose polarity, tight junctions, and epithelial morphology. During this process, cells acquire mesenchymal traits that promote motility and invasion [1]. The reverse process, mesenchymal-epithelial transition (MET), reinstates epithelial characteristics such as polarity and cell-cell adhesion. Together, EMT and MET are governed by interconnected signaling pathways and gene regulatory networks. These networks orchestrate the suppression of epithelial markers like E-cadherin and upregulation of mesenchymal markers such as vimentin and fibronectin. EMT is classically categorized in three major biological contexts: development (Type 1), tissue repair and fibrosis (Type 2), and cancer progression (Type 3) [1, 2].

In cancer, EMT promotes tumor cell invasion, dissemination, and therapeutic resistance. This occurs through the activation of canonical signaling pathways, including TGF-$\beta$, Wnt/$\beta$-catenin, Notch, ERK/MAPK, as well as hypoxia-related factors, pro-inflammatory cytokines, and growth factors. These signals collectively converge on core EMT-associated transcription factors such as Snail, Slug, Twist, Fra1, and ZEB1/2 [3]. Notably, multiple, often partial or hybrid epithelial-mesenchymal states may coexist within tumors, highlighting the dynamic and reversible nature of EMT rather than a strict binary transition [4, 5]. On the other hand, MET contributes to the colonization of distant sites by restoring epithelial traits that support proliferation, tissue integration, and outgrowth at distant sites [6, 7, 8].

Despite its importance, the regulatory mechanisms of EMT remain highly complex and not yet fully delineated. Several key inquiries remain unresolved, including how the tumor microenvironment, extracellular vesicles, and epigenetic modifications influence EMT. Another unresolved question concerns how baseline epithelial and mesenchymal gene expression programs differ between normal and malignant tissues. Large gaps remain in defining the molecular networks that regulate EMT in cancer [6, 9].

Recent studies point to significant similarities between EMT programs and neural developmental processes such as neural crest migration and neurogenesis [2, 10]. Key EMT-associated transcription factors, including Snail and Slug, can activate neural-like gene expression and stemness pathways, especially in gliomas. High-grade gliomas often express both developmental and pluripotency markers, suggesting that cancer cells may activate neural stemness programs that mimic or overlap with EMT [2]. However, it remains unclear whether tumor cells transition fully into mesenchymal states or instead adopt intermediate hybrid phenotypes influenced by neural developmental signals.

Machine learning (ML) approaches have become essential for biomarker discovery in EMT research due to their ability to analyze complex, high-dimensional datasets [11]. Widely used models include Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Networks (ANNs) [12]. RF is particularly useful for both classification and feature selection, providing feature-importance scores that help identify candidate biomarkers from transcriptomic data [13]. SVM performs well in classifying categorical data [14], while ANNs can recognize complex patterns within large datasets. Collectively, these ML approaches offer strong potential for uncovering key regulators of EMT and improving diagnostic and therapeutic strategies.

The objective of this study was to systematically characterize transcriptional programs associated with EMT and MET across diverse mouse cell line and tissue models. To this end, we integrated large-scale transcriptome data with differential gene expression and functional enrichment analyses to identify genes and biological pathways linked to EMT-related transcriptional dynamics. Additionally, we utilized machine learning models to prioritize candidate genes in both non-malignant and cancerous contexts. By stratifying the datasets into cancer and normal groups, we also sought to explore potential molecular relationships between EMT, cancer progression, and neural-related pathways. Altogether, this integrative approach provides a comprehensive framework for discovering candidate genes and regulatory networks for further investigation in tumor biology and neuronal development.

Functional enrichment analysis of the DEGs across all datasets revealed that normal and cancer-associated EMT share core biological processes, yet exhibit distinct, context-specific regulatory features (Fig. 3 and Supplementary Table 2). In cancer tissues, EMT was predominantly related to the developmental and neural-associated pathways, including neurogenesis and regulation of cell adhesion. This pattern suggests a reactivation of embryonic programs that may facilitate tissue remodeling, invasion, and metastatic potential [23, 24]. Additional enriched terms, such as epithelial tube morphogenesis, regulation of apoptotic signaling, muscle tissue development, and response to peptide hormones. Together, these findings further highlight the convergence of developmental plasticity and tumor progression [25, 26].

Cancer cell models displayed a broadly similar enrichment profile. Over-represented pathways were related to calcium ion transport, cytoskeletal remodeling, and transmembrane receptor signaling. These pathways are consistent with enhanced motility, intracellular communication, and metabolic adaptation during EMT [27, 28, 29]. In contrast, normal epithelial tissues undergoing EMT were enriched for pathways related to physiological regulation and differentiation. These included axonogenesis, immune response-regulating signaling, lipid transport, autophagy, and myeloid cell differentiation, reflecting tightly regulated developmental and homeostatic programs [30, 31] (Fig. 3). Enrichment of epithelial tube morphogenesis, chemical synaptic transmission, cell-substrate adhesion, muscle system processes, and regulation of catalytic activity in normal epithelial cells. These patterns indicate a coordinated remodeling of cellular architecture and communication during physiological EMT [24] (Fig. 3).

Overall, these findings indicate that while EMT in both normal and cancer contexts engages conserved biological programs related to adhesion, morphogenesis, and cytoskeletal dynamics. However, cancer-associated EMT shows a relative enrichment of neural and developmental signaling pathways, suggesting a potential functional link between neurogenesis, cellular plasticity, and malignant transformation.

Supervised machine learning models were employed to identify potential biomarker genes across transcriptomic datasets representing four distinct groups. Three supervised learning algorithms - random forest [13], SVM [14], and ANN [32] - were employed, and each group was analyzed independently. To derive an informative cancer-associated EMT gene signature, models were trained on each group’s datasets. The top 50 predictive genes from each model were compared to assess concordance among algorithms. In general, the ML models could accurately prioritize the most significant genes, as evidenced by the performance metrics shown in Tables 1,2,3,4.

Of note, the SVM model exhibits comparatively modest performance across the evaluation metrics relative to the other ML approaches. This is likely attributed to the fact that SVM is primarily a supervised classification algorithm optimized for sample-level discrimination rather than a statistical framework for gene-level differential expression testing. Therefore, SVM is not ideally suited as a standalone method for DEG prioritization [33, 34].

In addition, the comparatively lower performance observed in Table 4 relative to that obtained for normal cells/tissues and cancer cell lines (Tables 1,2,3) likely reflects the inherent complexity of bulk tumor transcriptomes rather than a limitation of the algorithm itself. Unlike cell lines or relatively homogeneous normal tissues, bulk cancer tissues consist of a heterogeneous mixture of malignant cells, stromal fibroblasts, endothelial cells, and diverse immune cell populations [35]. Signals arising from stromal contamination and immune infiltration can dominate transcriptional profiles. These signals may obscure EMT-associated expression patterns intrinsic to cancer cells [36, 37], challenging in this way sample-level classification. Under these conditions, SVM decision boundaries may be driven primarily by variation in cellular composition rather than by biologically meaningful EMT-related differences, ultimately resulting in reduced discriminatory power.

Several strategies could help mitigate these limitations. Tumor purity adjustment or cell-type deconvolution prior to model training could reduce conflicting signals from non-malignant cells [38]. Incorporating feature selection or pathway-level aggregation [39, 40], instead of individual gene expression values, may further improve model robustness by reducing dimensionality and attenuating noise and biological heterogeneity. Finally, integrating single-cell and/or spatial transcriptomic data [41, 42] could yield more accurate representations of tumor-intrinsic EMT programs. Such integration may also support the development of improved tissue-level classifiers.

The 50 highest-ranked predictive genes derived from RF, SVM and ANN models across cancer cell datasets are shown in Fig. 4. Intersection analysis across the three supervised learning models revealed a core set of 15 genes consistently identified by all methods. These included CTSL, S100A9, BC1, HBA-A1, HBA-A2, KRT8, KRT18, LGALS1, and several mitochondrial genes, i.e., MT-CO3, MT-CYTB, MT-ND3, and MT-RNR1. These mitochondrial genes encode components of the mitochondrial respiratory chain (MT-CO3, MT-CYTB, MT-ND3) or mitochondrial rRNA (MT-RNR1) and are central to oxidative phosphorylation and mitochondrial translation. Altered expression or mutation of mitochondrial genes has been associated with metabolic reprogramming, ROS production, and hypoxia-related signaling, processes that may influence EMT-associated transcriptional programs. Pairwise overlap analysis (Fig. 4) showed substantial agreement between models. Eleven genes shared between RF and SVM, fourteen between RF and ANN, and seventeen between SVM and ANN, indicating complementary predictive performance. Within these intersections, several key EMT- and metastasis-related genes - such as S100A8, KRT5, KRT14, CTSK, LGALS3, and EEF1A1- were recurrently identified. In contrast, S100A4 and MT-ND4 were uniquely detected by the RF model, suggesting potential model-specific sensitivity in capturing cytoskeletal and mitochondrial features associated with EMT progression.

To identify biomarkers associated with cancer-related EMT in vivo, the top 50 predictive genes from the three supervised ML models were compared (Fig. 5). This analysis revealed a core set of 15 genes detected by all three models, representing candidates for EMT-related biomarkers. Additionally, each algorithm identified model-specific genes: RF contributed 15 unique genes, SVM 5, and ANN 5. Pairwise overlaps were also observed, highlighting genes shared between two models but not the third, reflecting both shared and distinct predictive features captured by each approach (Fig. 5).

Furthermore, the potential functional synergy among the individual signature genes within a protein-protein interaction network was explored (Fig. 6). The protein products of these genes form a functionally interconnected network, either through direct interactions or indirectly via five putative connector proteins, namely VIM (Vimentin), CD44, GYPC (Glycophorin C), RRAGC (Ras Related GTP Binding C), and LGALS9C (Galectin 9C).

Gene sets from RF, SVM and ANN models were compared to explore candidate biomarkers of normal cellular states (Fig. 7). Ten genes were shared across all models, indicating a subset of consistently selected features. Pairwise overlaps were substantial (RF-SVM: 10 genes; RF-ANN: 11 genes; SVM-ANN: 24 genes), including members of the CRY gene family and MT-ND genes. In addition, model-specific signatures were detected (RF: 19; SVM: 6; ANN: 5 genes). These findings highlight both core biomarkers and complementary model-specific candidates.

Comparison of gene sets from all three models (Fig. 8) identified nineteen core biomarkers of normal tissue. Pairwise overlaps revealed additional shared genes (RF-SVM: 4, RF-ANN: 5, and SVM-ANN: 25), while each model also yielded unique genes (RF: 22, SVM: 2, and ANN: 1). These findings highlight both shared and algorithm-specific biomarkers, reflecting the complementary strengths of different machine learning approaches.

In this study, we implemented an integrative framework to systematically compare EMT-associated transcriptional programs across normal and cancer datasets. By combining differential gene expression and functional enrichment analyses with machine learning-based classification, this approach enabled the identification of candidate EMT biomarkers and the delineation of regulatory networks and molecular patterns. Notably, our integrated analyses revealed recurrent enrichment of gene sets associated with neural-related processes in EMT, particularly neurogenesis, gliogenesis, and axonogenesis. Neurogenesis and gliogenesis proceed sequentially from common neural progenitors to ensure proper lineage specification and cell fate determination [43, 44], whereas axonogenesis facilitates the functional integration of newly generated neurons into pre-existing circuits [45, 46, 47].

The consistent enrichment of neural-related pathways observed across cancer datasets suggests that EMT-associated transcriptional reprogramming may extend beyond canonical epithelial and mesenchymal states. Rather, EMT appears to encompass neural-like characteristics that facilitate tumor-nerve interactions. Increasing evidence indicates that cancer cells undergoing EMT can engage molecular programs resembling those of neural progenitors or differentiated neural cells. Through this process, tumor cells acquire the capacity to sense, respond to, and reshape the neural microenvironment. This phenomenon is often referred to as “neuronal mimicry” [48]. Such neural-like plasticity may provide selective advantages during tumor progression by increasing cellular motility, enabling directed invasion along nerve fibers, and enhancing survival within neural-rich niches [49, 50].

In this context, the enrichment of pathways linked to neurogenesis and gliogenesis may reflect the reactivation of evolutionarily conserved developmental signaling cascades, including Notch, Wnt/$\beta$-catenin, and Hedgehog pathways. These pathways are known to promote cellular adaptability, lineage plasticity, and stemness - key attributes of EMT [51, 52].

A clinically significant consequence of tumor-nerve interaction is perineural invasion (PNI), in which EMT programs facilitate tumor cell infiltration and spread along neural structures [53, 54]. EMT has been closely associated with PNI in multiple cancer types, including pancreatic, prostate, colorectal, and head and neck cancers. In these contexts, neurotrophic signaling axes such as NGF-Trk, BDNF-TrkB, and GDNF-RET contribute to directional tumor cell migration and invasive behavior [53, 55, 56].

More broadly, these observations are consistent with the rapidly emerging field of cancer neuroscience, which focuses on the bidirectional interactions between the nervous system and tumor biology [57]. Within this framework, neural activity plays an active role in tumor progression. Neurotransmitters (e.g., acetylcholine, norepinephrine, and glutamate), as well as neurotrophic factors are gaining recognition as key components of the tumor microenvironment that influence tumor growth, immune modulation, angiogenesis, and metastatic dissemination [58, 59, 60].

Collectively, the consistent enrichment of neural-related pathways across cancer EMT datasets supports a framework in which EMT-associated plasticity intersects with neural developmental and signaling programs. This convergence promotes tumor-nerve crosstalk and contributes to malignant progression. Future studies integrating single-cell and spatial transcriptomics with functional assays will be essential to elucidate the causal roles of these pathways and to explore their therapeutic potential in EMT-driven malignancies.

Focusing on pathological EMT, machine learning analyses identified a set of genes associated with neurodevelopmental processes that appeared recurrently across EMT-related cancer datasets [61]. Among these, BCYRN1 (the human ortholog of mouse Bc1), a neuronal long non-coding RNA involved in translational regulation and synaptic plasticity, was consistently detected across all cancer cell and tissue EMT datasets but was absent from normal EMT profiles [62]. This selective enrichment suggests potential reactivation of neurodevelopment-associated programs during cancer-associated EMT. Such reactivation may be linked to increased cellular plasticity and stem-like characteristics observed during cancer progression [2]. Clinically, elevated BCYRN1 expression correlates with both overall and disease-free survival [63].

In parallel, the identification of immune- and microenvironment-associated factors, including B2M, points to an immunomodulatory dimension of EMT in cancer [64]. B2M is frequently upregulated across multiple malignancies and has been reported to promote cancer cell survival, invasion, and metastasis through PI3K/AKT, MAPK, and PKA/CREB signaling; these findings suggest a context-dependent role in modulating EMT-related states and tumor-microenvironment interactions [65]. Consistently, elevated B2M expression has been associated with poor prognosis, including reduced progression-free survival [66, 67].

Additional mediators, S100A8 and S100A9 showed context-dependent patterns, with S100A8 detected only in cancer-associated EMT datasets and S100A9 only in normal EMT profiles; this divergence suggests differential involvement of inflammatory signaling in malignant versus physiological EMT [68, 69]. Notably, high expression of S100A8/A9 is generally associated with poor prognosis, metastasis, and advanced disease stage across several tumor types, including colorectal, breast, and gastric cancers [70].

SRGN (serglycin) was consistently detected in cancer-associated EMT datasets but not in normal EMT. Prior studies have linked SRGN to EMT-like transcriptional states, invasiveness, and microenvironmental responsiveness; these findings indicate that cancer-associated EMT may preferentially engage extracellular matrix- and developmental-related regulatory programs [71, 72]. Elevated SRGN expression is also associated with adverse clinical outcomes across multiple cancers [71, 73, 74]. In breast cancer, SRGN contributes to chemoresistance by sustaining stemness through crosstalk with YAP-dependent transcriptional programs [75].

LGALS1 was detected in cancer-associated EMT datasets and has been linked to tumor progression, angiogenesis, immune modulation, and therapy resistance. While its role in EMT may be indirect, elevated LGALS1 expression consistently correlates with increased recurrence risk and poorer survival in multiple cancers, including colorectal cancer, often through EMT-linked signaling pathways and immune regulatory mechanisms [76, 77, 78].

Finally, LARS2, primarily studied in neuronal contexts, has been associated with mitochondrial dysfunction and neurodegenerative disease; however, its relevance to cancer or EMT remains unclear [79, 80, 81]. CST3 has been associated with tumor invasion and poor prognosis in several cancers and may be influenced by hormonal regulation, although its specific role in EMT requires further investigation. Although no direct association has yet been reported between MIR6236 and EMT or TME regulation, limited evidence suggests a potential tumor-suppressive role in endometrial cancer.

Overall, these findings suggest that cancer cells may preferentially engage neural developmental programs and context-specific gene regulators to promote EMT and malignant phenotypes. In contrast, normal EMT processes appear more reversible and tightly regulated. The ML-identified genes in cancer tissues - particularly B2M, CST3, LARS2, SRGN, S100A8/A9, LGALS1, BCYRN1, and MIR6235 - were prioritized as candidate markers. Genes involved in functionally related disease processes tend to be interconnected within biological networks and are frequently co-regulated. Hence, it is plausible that these EMT-related genes participate in shared co-expression networks and are governed by common epigenetic regulatory programs [82, 83, 84].

Furthermore, the protein products of the identified signature genes form an interconnected interaction network, linked either directly or through intermediate putative nodes. This suggests coordinated physical and/or functional associations that collectively modulate EMT and TME dynamics. One of the key connector nodes is Vimentin, a canonical mesenchymal marker and structural effector of EMT. Another prominent connector node, CD44, functions as a central regulator of EMT and cancer stemness by suppressing epithelial markers (such as E-cadherin) and inducing mesenchymal markers (e.g., N-cadherin and vimentin), thereby enhancing invasion; CD44 silencing prevents or reverses EMT, supporting its causal role in EMT regulation [85, 86]. LGALS9C, another connector node within the network, belongs to the Galectin-9 family of $\beta$-galactoside-binding lectins. Galectin-9 family members modulate EMT-relevant processes by regulating cell-cell and cell-matrix adhesion, immune-tumor crosstalk, and migratory signaling pathways [87].

The coordinated activity of these genes/proteins suggests convergence of interconnected programs governing EMT dynamics and tumor microenvironment remodeling. In particular, inflammatory mediators (S100A8/S100A9, SRGN, LGALS1, LGALS9C) [88, 89] and immune interface components (B2M, CD44) [90, 91] can establish cytokine- and chemokine- dependent signaling. This signaling promotes EMT-associated transcriptional plasticity and may sensitize tumor cells to neural-derived signals. At the same time, factors involved in cell-extracellular matrix interactions and cytoskeletal organization (including CD44, VIM, and CST3) [92, 93, 94] are likely to promote directed migration and invasion along nerve-associated structures. These factors may also facilitate tumor cell engagement with the surrounding stroma. In parallel, metabolic and stress-adaptation pathways (RRAGC, LARS2) [95, 96] may support the energetic demands of EMT, including tumor cell survival within nerve-rich microenvironments. Together, these coordinated programs provide a plausible mechanistic link between EMT, extracellular matrix remodeling, and reorganization of the neural niche within the tumor microenvironment. These interactions may contribute to the stabilization of EMT states and the advancement of tumor progression.

To further delineate the differences between physiological and cancer-associated EMT, we found that SPP1 is the only gene shared between both contexts. In contrast, several other recurrently detected genes (BC1, HBA-A2, HBA-A1, KRT18, GM26035, GM28437, MT-CO3, MT-CYTB, MT-ND3, KRT8, and MT-RNR1) currently lack clear evidence linking them to cancer, EMT, or neurogenesis. Nevertheless, their consistent detection across the analyzed datasets suggests potential biological relevance and highlights the need for targeted in vitro and in vivo experimental studies to clarify their roles.

This study has several limitations that should be acknowledged: (i) dataset heterogeneity, (ii) the analysis was restricted to canonical epithelial/mesenchymal states, with intermediate or hybrid E/M phenotypes excluded due to the intrinsic limitations of bulk transcriptomic data; this represents a major limitation, given the established biological and clinical relevance of hybrid E/M states in cancer, and underscores the need for future single-cell-based studies to more comprehensively resolve EMT heterogeneity, (iii) potential tissue contamination, (iv) the sample sizes of the normal (n = 17) and cancer (n = 12) tissue groups are relatively small, which may limit statistical power and generalizability; therefore, validation in larger, independent cohorts will be necessary to confirm and strengthen the robustness and accuracy of the findings derived from these groups, (v) the absence of experimental validation.

Nevertheless, despite these limitations, the findings presented herein may serve as a foundation for the rational design of future experimental and translational studies. The cancer tissue-associated signature genes identified in the present study could be incorporated into clinical settings to improve diagnostic strategies. These genes may complement and refine currently established EMT-related biomarkers, especially in tumors exhibiting pronounced mesenchymal features. Beyond their diagnostic utility, these genes could represent promising anti-cancer therapeutic targets. This potential arises either from their direct involvement in EMT-relevant signaling pathways or from their function as non-coding epigenetic regulators that modulate EMT-associated protein-coding genes within complex regulatory networks. This dual contribution highlights the multifaceted nature of EMT regulation during cancer progression. Notably, targeting specific components of this signature, such as SRGN, has been reported to sensitize tumor cells to chemotherapeutic agents, suggesting that EMT-linked molecular vulnerabilities may be therapeutically exploitable. Furthermore, the expression profiles of BCYRN1, B2M, S100A8/A9, SRGN, and LGALS1 are associated with poor prognostic outcomes in EMT-high tumors, underscoring their potential value as prognostic biomarkers and potential predictors of therapeutic response.

Herein, an integrative computational strategy was applied to explore EMT-associated transcriptional programs across normal and cancer-related mouse RNA-seq datasets. While EMT processes - such as cell adhesion, cytoskeletal remodeling, and tissue morphogenesis - were shared across contexts, cancer-associated EMT showed additional enrichment of developmental and neural-related pathways. This pattern suggests that malignant cells may rely on a broader range of plasticity-associated programs compared to normal EMT. The application of complementary machine learning models enabled the prioritization of candidate genes associated with EMT across heterogeneous datasets, revealing both shared and context-specific features. Collectively, the results of this study provide a comprehensive overview for understanding the transcriptional differences between physiological and cancer-associated EMT and provide the foundation for future targeted experimental studies.

ANN, artificial neural network; EMT, epithelial-mesenchymal transition; MET, mesenchymal-epithelial transition; SVM, support vector machine.

Publicly available data were utilized in this study. The relevant database sources and the corresponding links are provided below: PubMed (NCBI): https://pubmed.ncbi.nlm.nih.gov/; NCBI Gene Expression Omnibus (GEO): https://www.ncbi.nlm.nih.gov/gds/; STRING Protein-Protein Interaction Database: https://string-db.org/. All necessary data have been provided in the Supplementary Material.

Conceptualization, ŞB, HA, AGG and AP; methodology, ŞB, HA, AGG and AP; software ŞB and AP; validation, ŞB, HA and AP; formal analysis, ŞB, AGG and AP; data curation, ŞB; writing—original draft preparation, ŞB, HA, AGG and AP; writing—review and editing, ŞB, HA, AGG and AP; supervision, AP; project administration, AP. All authors contributed to editorial changes in the manuscript. All authors read and approved the final manuscript. All authors have participated sufficiently in the work and agreed to be accountable for all aspects of the work.

Not applicable.

We would like to express our sincere gratitude to the peer reviewers for their constructive comments and valuable suggestions, which helped improve the quality of this work.

This research received no external funding.

Given his role as the Editorial Board member, Alexandros G. Georgakilas had no involvement in the peer-review of this article and has no access to information regarding its peer review. Full responsibility for the editorial process for this article was delegated to Nguyen Quoc Khanh Le and Yudong Cai.

Supplementary material associated with this article can be found, in the online version, at https://doi.org/10.31083/FBL48085.