The BRCA1 is traditionally characterized as a nuclear tumor suppressor involved in the repair of DNA double strand breaks via homologous recombination. However, contemporary research has identified a significant fraction of phosphorylated BRCA1 protein localized within the mitochondria, specifically associated with mitochondrial nucleoids [20]. This mitochondrial localization is essential for the maintenance of mitochondrial DNA (mtDNA) stability and the regulation of mitophagy, which is the selective autophagic clearance of damaged mitochondria [21]. When BRCA1 function is lost, the mitochondria experience a profound metabolic trap involving the NAD⁺/NADH redox balance. Loss of BRCA1 induces chronic genomic instability, leading to the hyperactivation of Poly(ADP-ribose) polymerase 1 (PARP1) as it attempts to repair accumulated nuclear DNA damage [22]. Because PARP1 utilizes nicotinamide adenine dinucleotide (NAD⁺) as its primary substrate to synthesize poly(ADP ribose) chains, its overactivation causes a systemic depletion of the cellular NAD⁺ pool [22]. This depletion directly impacts the mitochondrial electron transport chain (ETC), specifically Complex I, which requires NAD⁺ as an electron acceptor to convert NADH back into NAD⁺. Consequently, the NAD⁺/NADH ratio drops significantly, leading to impaired OXPHOS and an increased reliance on glycolysis [23].

Furthermore, BRCA1 deficient cells exhibit elevated levels of mitochondrial reactive oxygen species (ROS), particularly hydrogen peroxide (H₂O₂), which further stabilizes Hypoxia Inducible Factor 1α (HIF1α) and drives a Reverse Warburg Effect [23]. In this scenario, the oxidative stress within BRCA1 mutated epithelial cells is transferred to the neighboring tumor stroma, inducing aerobic glycolysis in associated fibroblasts. These fibroblasts then feed the cancer cells with high energy metabolites such as lactate and pyruvate, which can be utilized by the cancer cells to fuel the TCA cycle despite their impaired intrinsic OXPHOS capacity [24]. This metabolic imprinting is a hallmark of the TNBC (basal-like) subtype, where BRCA1 mutations drive a systemic NAD⁺ depletion trap, distinguishing its metabolic requirements from those of receptor-positive subtypes.

The MYC oncogene acts as a master transcriptional regulator that synchronizes the nuclear and mitochondrial genomes to support rapid biomass accumulation. MYC directly promotes the expression of mitochondrial transcription factor A (TFAM) and other NEMGs involved in the assembly of the respiratory chain [12,14]. Central to this process is the PGC1α, which functions as the primary rheostat for mitochondrial biogenesis. In metastatic BC cells, the MYC/PGC1α axis is often upregulated to meet the high bioenergetic demands of migration and colonization [14,25]. Recent research highlights that PGC1α enhances OXPHOS and mitochondrial biogenesis, which is essential for the survival of circulating tumor cells (CTCs) and the formation of distant metastases [26]. While the Warburg Effect suggests that cancer cells move away from mitochondrial respiration, the MYC/PGC1α axis demonstrates that aggressive BC cells actually re-adopt OXPHOS to achieve metabolic robustness under microenvironmental stress [25,26]. This metabolic flexibility allows cells to survive in nutrient deprived environments by maximizing ATP yield per molecule of glucose.

The tumor suppressor p53 (TP53) plays a dual role in regulating mitochondrial health. Wild type p53 supports mitochondrial efficiency by transcriptionally activating synthesis of cytochrome c oxidase 2 (SCO2), which is crucial for the assembly of cytochrome c oxidase or complex IV [27,28]. It also upregulates TIGAR (TP53 induced glycolysis and apoptosis regulator) to inhibit glycolysis and divert glucose toward the pentose phosphate pathway for NADPH production, thereby maintaining redox homeostasis and limiting ROS damage [28]. In contrast, TP53 mutations, which are found in over 80% of TNBCs, drastically remodel the mitochondrial landscape. Mutant p53 loses its ability to promote SCO2 mediated OXPHOS, leading to a shift toward glycolysis [27,29]. Paradoxically, in basal like BCs, mutant p53 has been shown to upregulate the expression of carnitine palmitoyltransferase 1C (CPT1C), a key regulator of fatty acid oxidation [30]. This enables mutant p53 cells to utilize lipids as an alternative fuel source, providing a survival advantage during glucose deprivation and contributing to chemoresistance [27]. Furthermore, mutant p53 interacts directly with the mitochondrial inner membrane to alter the permeability transition pore, making cells more resistant to apoptosis induced by mitochondrial stress.

Mitochondria are active signaling organelles that communicate their functional status to the nucleus via retrograde signaling. In BC, mitochondrial dysfunction, which is characterized by mtDNA depletion or reduced membrane potential, triggers the release of mitochondrial calcium into the cytoplasm [31]. This influx activates the Calcineurin NFAT signaling pathway, which induces the expression of genes associated with the epithelial mesenchymal transition (EMT), such as Snail, Slug, and Vimentin [31,32,33]. This retrograde loop effectively converts metabolic stress into a pro invasive genetic program, generating BC stem cells that are inherently resistant to traditional therapies [31]. Additionally, the accumulation of oncometabolites like succinate and fumarate due to TCA cycle dysregulation acts as a critical signal [34,35]. These metabolites can exit the mitochondria and competitively inhibit alpha ketoglutarate dependent dioxygenases in the nucleus, including the TET family of DNA demethylases and the JmjC domain containing histone demethylases [34,36]. This leads to widespread epigenetic deregulation, specifically the hypermethylation of tumor suppressor promoters, and the stabilization of HIF1α under normoxic conditions [34,37]. Such metabolic epigenetic crosstalk reinforces the aggressive phenotype and ensures that the cancer cell remains in a perpetually glycolytic and invasive state, even in the presence of adequate oxygen. The physical proximity of mitochondria to the nucleus in aggressive BC cells further facilitates this rapid exchange of signaling metabolites and ions. The specific mitochondrial phenotypes and associated therapeutic vulnerabilities for each breast cancer subtype are summarized in Table 1.

The mitochondrial genome or mtDNA is highly susceptible to damage due to its proximity to the site of reactive oxygen species production and its lack of protective histones and robust repair mechanisms. In BC, mtDNA mutations often exhibit heteroplasmy, which refers to the co-existence of mutant and wild type mitochondrial genomes within a single cell. The dynamic shift in heteroplasmy levels is recognized as a significant driver of tumor evolution and metabolic adaptation. Research has established that the D loop region, a non-coding displacement loop, is a mutation hotspot in BC mtDNA [40]. This region is critical because it contains essential regulatory elements for the replication and transcription of the entire mitochondrial genome. Specific mutations in the D loop, such as the T16189C and A10398G variants, have been strongly correlated with increased BC susceptibility and malignant progression [41]. Large scale analyses of clinical samples demonstrate that the burden of mtDNA mutations increases in direct correlation with the clinical stage of the disease. Specifically, patients in stage III and stage IV exhibit significantly higher levels of mitogenomic diversity compared to those in early stages, suggesting that these variations provide a selective advantage to tumor cells operating under severe microenvironmental stress [17].

In addition to point mutations, mitochondrial DNA copy number or mtDNAcn serves as a quantitative indicator of mitochondrial mass and bioenergetic capacity. In TNBC, it has been observed that tumor tissues frequently exhibit lower mtDNAcn relative to adjacent normal tissues. However, in cases where residual tumor cells survive neoadjuvant chemotherapy, there is often a compensatory resurgence in mtDNAcn [42]. This ‘compensatory resurgence’ is quantitatively defined as a statistically significant increase in mtDNAcn within chemoresistant clones, typically represented by a ≥1.5-fold increase relative to pre-treatment biopsies or the restoration of mtDNA levels to match adjacent normal tissues [42]. The scientific rationale for this threshold lies in the metabolic plasticity of surviving cells, which must upregulate mitochondrial mass to compensate for drug-induced oxidative stress and fulfill the high energetic demands of therapeutic escape. To ensure the rigor of these measurements, clinical studies utilize normalized dPCR or qPCR assays, where mtDNAcn is calculated relative to nuclear reference genes to account for variations in cell number and DNA quality. This recovery reflects the metabolic plasticity of the cells as they adapt to therapeutic pressure, positioning mtDNAcn as a valuable biomarker for monitoring treatment response and potential recurrence.

While mtDNA mutations are primary indicators, the metabolic phenotype of a tumor is largely governed by nuclear encoded mitochondrial genes or NEMGs. Integrating large scale genomic datasets from sources such as The Cancer Genome Atlas or TCGA and the Gene Expression Omnibus or GEO allows for the identification of metabolic gene signatures with high prognostic value. The development of these signatures typically employs advanced bioinformatics pipelines and machine learning algorithms, including Least Absolute Shrinkage and Selection Operator or LASSO regression and random forest models. To ensure model stability and reproducibility, these frameworks utilize rigorous hyperparameter optimization; for instance, LASSO regression models frequently employ 10-fold cross-validation to identify the optimal penalty parameter that minimizes partial likelihood deviance. Similarly, random forest models used in these signatures typically specify a standardized number of trees (ntree = 500) and a fixed number of variables randomly sampled at each split (mtry) to prevent overfitting and ensure consistent feature importance rankings across diverse datasets. For instance, recent studies have validated a 14 gene mitochondrial related prognostic signature that can accurately predict overall survival at one, three, and five year intervals with an area under the curve or AUC exceeding 0.90 in external validation cohorts [19,43]. Another 18-gene diagnostic model developed in 2025 through the evaluation of over 113 machine learning combinations demonstrated a 0.947 accuracy rate in differentiating between malignant and healthy breast tissues [18]. These metabolic panels provide more than just survival predictions. They also offer insights into the tumor microenvironment. Higher expression levels of certain NEMGs are associated with reduced infiltration of natural killer cells and T cells, implying that a more oxidative metabolic state may facilitate immune evasion by the tumor [44,45]. By utilizing these metabolic fingerprints, clinicians can classify patients into specific metabolic subtypes that traditional genomic subtyping might overlook, enabling the delivery of more targeted interventions.

Liquid biopsy technology offers a non-invasive method to monitor the metabolic evolution of BC in real time. Circulating tumor mitochondrial DNA or ct-mtDNA has emerged as a promising marker for early diagnosis and the detection of minimal residual disease or MRD. Ct-mtDNA consists of fragmented mitochondrial DNA released into the systemic circulation by necrotic or apoptotic tumor cells. Clinical evidence confirms that the concentration of ct-mtDNA in the plasma of BC patients is significantly higher than in healthy controls [46]. Beyond simple quantification, ct-mtDNA allows for the real time assessment of the tumor metabolic state without the need for repetitive tissue biopsies. One of the most significant applications of ct-mtDNA is in monitoring MRD following surgical intervention. Ultrasensitive sequencing techniques can detect MRD signals in the blood years before a clinical recurrence is visible through standard imaging. In one cohort study, ct-mtDNA detection anticipated late relapse in early-stage BC by an average of 5.7 years [47]. This comprehensive approach provides a dynamic molecular portrait of the disease, allowing for early intervention with metabolic therapies before the transition to macro metastatic recurrence occurs. However, the clinical translation of these metabolic indicators requires the rigorous validation of assays across diverse patient populations and the establishment of clear regulatory guidelines [4]. Current technical challenges involve the need for standardized detection protocols for ct-mtDNA to ensure consistency in quantification and sensitivity across different laboratory settings [46]. The transition to metabolic markers in the clinical setting requires the standardization of detection protocols and the validation of these markers in diverse patient populations. However, the current trajectory of research indicates that mitochondrial and metabolic signatures may serve as valuable adjuncts to the next generation of BC diagnostics, pending further clinical validation.

The translation of mitochondrial metabolic indicators into standardized clinical practice requires the rigorous validation of assays across diverse patient populations and the establishment of clear regulatory guidelines [4]. Current technical challenges involve the need for standardized detection protocols for ct-mtDNA to ensure consistency in quantification and sensitivity across different laboratory settings [46]. Furthermore, the reliable detection of minimal residual disease (MRD) through liquid biopsy remains dependent on addressing critical pre-analytical variables, including specimen type, tumor fraction, and heteroplasmy thresholds, to ensure that low-frequency mitochondrial signals are distinguishable from biological background noise [46,47]. In the context of nuclear-encoded mitochondrial gene (NEMG) signatures, the application of machine learning necessitates strategies to account for inter-platform variability and batch effects that can adversely impact the reproducibility of prognostic models [17,18]. Developing and implementing robust normalization strategies is therefore a prerequisite for the clinical deployment of these metabolic panels [18]. Ultimately, the transition to metabolic precision medicine is contingent upon large-scale prospective validation and the creation of regulatory frameworks that address the significant costs and accessibility of multi-omics integration [4].

The ETC is the primary site of ATP production via OXPHOS. Inhibiting specific complexes within the ETC represents a strategy to starve cancer cells of energy and biosynthetic precursors.

Mitochondria play a central role in the intrinsic apoptotic pathway, which is regulated by the B cell lymphoma 2 (Bcl-2) family of proteins. Many BCs, particularly estrogen receptor (ER) positive subtypes, overexpress anti apoptotic proteins like Bcl-2 to evade cell death induced by oncogenic stress or chemotherapy.

Synthetic lethality occurs when the simultaneous disruption of two pathways leads to cell death, while the disruption of either alone is compatible with survival. In the context of mitochondrial metabolism, this involves targeting compensatory pathways or specific genetic dependencies.

Endocrine therapy resistance is a major obstacle in the management of ER positive BC. While agents like tamoxifen and aromatase inhibitors successfully disrupt estrogen signaling, a subset of cancer cells survives by undergoing a profound metabolic shift. These resistant cells frequently exhibit a transition from a glycolytic phenotype to one characterized by high mitochondrial activity and increased OXPHOS capacity [15]. Recent studies published in 2025 have confirmed that this metabolic conversion is driven by the upregulation of mitochondrial biogenesis markers such as the translocase of outer mitochondrial membrane 20 or TOMM20 and the SIRT3 protein [7]. High mitochondrial mass in these cells provides the necessary energy to bypass the growth inhibitory effects of estrogen deprivation, effectively facilitating a state of metabolic persistence.

In the context of chemotherapy, metabolic plasticity enables the survival of drug tolerant persister cells. These cells often exit the cell cycle and adopt a dormant state that is less sensitive to traditional cytotoxic agents like paclitaxel or doxorubicin. Research indicates that the survival of these persisters is critically dependent on mitochondrial respiration and fatty acid oxidation [58,62]. In these aggressive subtypes, mutant TP53 has been shown to upregulate the expression of CPT1C, enabling cancer cells to utilize lipids as an alternative fuel source during glucose deprivation [30]. In TNBC, chemoresistant clones have been shown to overexpress PGC1α, which not only enhances OXPHOS but also increases the production of antioxidants like glutathione [63]. This dual function allows the cells to maintain energy production while neutralizing the reactive oxygen species generated by chemotherapeutic stress, thereby preventing apoptosis.

Furthermore, mitochondrial dynamics, specifically the balance between mitochondrial fusion and fission, play a key role in resistance. Resistant BC cells often exhibit an elongated mitochondrial morphology promoted by fusion proteins like MFN2. This elongated network is more efficient at producing ATP and sequestering pro apoptotic factors, thereby raising the threshold for mitochondrial outer membrane permeabilization and cell death [64]. Targeting these structural adaptations represents a potential strategy to re sensitize resistant tumors to conventional therapies. These structural adaptations, particularly MFN2-mediated fusion, often exhibit subtype-specific characteristics [64]. While MFN2 facilitates the formation of elongated networks for energy production and resistance, this dynamic has been identified as a critical therapeutic target specifically in TNBC [65]. Such findings underscore the distinct mitochondrial bioenergetic profiles observed between Luminal and Basal subtypes, which dictate their unique metabolic escape routes during therapy.

The tumor microenvironment or TME is a site of intense competition for limiting nutrients, where the metabolic needs of the cancer cell directly conflict with those of infiltrating immune cells. Aggressive BC cells utilize aerobic glycolysis at a high rate, a process that depletes the local glucose supply and generates a large volume of lactate [26,66]. This glucose deprivation is particularly detrimental to CD8⁺ T cells and natural killer cells, which require rapid glycolytic flux to fuel their effector functions, such as the secretion of perforin and granzyme B. Beyond glucose, the competition for amino acids like glutamine and fatty acids is a critical factor in immune evasion [67]. Glutamine is essential for both cancer cell proliferation and T cell activation. In the TME of TNBC, cancer cells often overexpress glutamine transporters like ASCT2, effectively starving the T cells and inducing a state of metabolic exhaustion [18]. Additionally, the accumulation of lactate leads to the acidification of the TME, with pH levels often dropping below 6.5. This acidic environment inhibits the mammalian target of rapamycin or mTOR signaling in T cells while promoting the recruitment and activation of myeloid derived suppressor cells or MDSCs and regulatory T cells or Tregs [68,69,70]. The spatial distribution of nutrients further complicates this competition. In poorly vascularized regions of the tumor, hypoxia induces the expression of HIF1α in both cancer and immune cells. However, cancer cells adapt more effectively by switching to alternative metabolic pathways, such as acetate or ketone body utilization, which immune cells are less equipped to exploit [65]. Metabolic escape routes differ significantly across subtypes. While ER⁺ BCs often shift toward an OXPHOS-high phenotype via TOMM20 upregulation to evade endocrine therapy [15], TNBC cells frequently exhibit a more glycolytic and glutaminolytic dominance to thrive in the nutrient-deprived tumor microenvironment [18]. This metabolic plasticity allows TNBC to effectively outcompete infiltrating immune cells for essential amino acids like glutamine. This metabolic dominance allows the tumor to create a sanctuary where it can proliferate shielded from the host immune system.

When primary metabolic pathways are pharmacologically targeted, BC cells demonstrate a remarkable ability to activate compensatory loops. This phenomenon, known as metabolic switching, is a fundamental cause of the failure of metabolic monotherapies. For instance, the inhibition of glycolysis through lactate dehydrogenase A or LDHA targeting often leads to a rapid increase in mitochondrial glutaminolysis and fatty acid oxidation to maintain the TCA cycle flux [71,72]. A particularly sophisticated form of this plasticity is metabolic symbiosis between different cell populations within the tumor. In many BCs, glycolytic cells located in hypoxic regions export lactate via the monocarboxylate transporter 4 or MCT4. This lactate is then taken up by oxidative cancer cells in well oxygenated regions through MCT1, where it is converted back to pyruvate to fuel the TCA cycle [73,74]. To accurately distinguish these metabolic subpopulations within the complex tumor architecture, researchers utilize spatial transcriptomics and single-cell RNA sequencing, which enable the visualization of metabolic niches based on the expression of markers such as LDHA and MCT4 for glycolytic cells versus TOMM20 and MCT1 for oxidative cells. These high-resolution approaches effectively exclude stromal cell contamination (such as cancer-associated fibroblasts) by employing cell type specific markers or bioinformatic deconvolution to isolate the malignant epithelial transcriptome for independent metabolic flux analysis. This adaptive rewiring allows tumor cells to bypass the glycolytic blockade by utilizing alternative carbon sources to maintain energy production and biosynthetic requirements. This symbiotic relationship allows the tumor as a whole to maximize nutrient utility and thrive under heterogeneous conditions.

This switching is coordinated by global metabolic sensors such as the AMP activated protein kinase or AMPK and the mTOR complex. Under conditions of energy stress, AMPK is activated to inhibit biosynthetic processes and stimulate catabolic pathways like mitophagy and fatty acid oxidation [65,75]. Conversely, if the cell is supplied with abundant nutrients, mTOR stimulates the production of proteins and lipids required for cell division. The crosstalk between these sensors ensures that the cancer cell can pivot its metabolic strategy almost instantaneously upon sensing a change in nutrient availability or therapeutic pressure. Furthermore, recent evidence suggests that BC cells can acquire mitochondria from neighboring stromal cells through tunneling nanotubes or extracellular vesicles. This horizontal mitochondrial transfer allows cancer cells with damaged or inhibited mitochondria to restore their OXPHOS capacity and survive metabolic crises [76,77]. Such mechanisms of extreme plasticity underscore the need for multi target metabolic inhibitors that can block both the primary pathways and the potential escape routes used by the tumor.

While recent AI models trained on patient-derived organoid data have reported predictive accuracies exceeding 92%, it is essential to evaluate these findings within the context of model interpretability and clinical calibration [18,36]. To ensure the reliability of these predictions and mitigate the risk of model overfitting, these frameworks typically utilize a 10-fold cross-validation strategy, often repeated across 100 iterations to stabilize performance metrics and provide a robust estimate of model generalizability. In specific instances involving smaller or highly specialized patient cohorts, Leave-One-Out Cross-Validation (LOOCV) is employed to maximize the utility of the available training data. The high performance of mitochondrial gene signatures often relies on retrospective datasets where cohort sizes may be limited, potentially increasing the risk of overfitting and feature leakage if rigorous cross-validation and external validation are not employed [19,40]. Furthermore, the transition of these models into clinically realistic workflows requires addressing the “black box” nature of deep learning through interpretability frameworks to ensure clinicians can understand the biological rationale behind AI-generated insights [17,18]. Current studies frequently lack prospective testing in real-world settings, which remains a critical barrier to establishing the true clinical utility and reliability of metabolic digital twins [4].

A significant limitation of current clinical trials in BC is the reliance on protein expression or genetic mutations as the sole inclusion criteria. The failure of broad spectrum mitochondrial inhibitors in early trials highlights the need for a more nuanced approach to patient selection. Future clinical trials should incorporate metabolic phenotyping as a central component of their design [4]. The novelty of integrating AI-driven metabolic modeling lies in its ability to refine clinical trial selection by shifting inclusion criteria from static genetic markers to dynamic metabolic phenotypes. By utilizing the metabolic gene signatures and liquid biopsy markers discussed in previous sections, researchers can stratify patients into metabolic subtypes such as OXPHOS high or glycolytic high. This stratification ensures that metabolic therapies are tested in populations that are biologically predisposed to respond. For instance, a trial targeting Mitochondrial Complex I would specifically enroll patients whose tumors show high dependence on oxidative phosphorylation and low metabolic flexibility [14]. This paradigm shift in trial design is hypothesized to potentially improve the success rate of metabolic agents in the clinical pipeline by refining patient selection.

Genome scale metabolic models or GEMs provide a comprehensive computational framework for analyzing the entire biochemical network of a cancer cell. GEMs like Recon3D and the more recent Recon4D integrate thousands of metabolic reactions and gene associations into a single mathematical matrix [78]. In the context of BC, GEMs are used to perform in silico gene essentiality screens to identify novel therapeutic targets. By simulating the metabolic requirements for biomass production in specific BC subtypes, GEMs can pinpoint enzymes that are essential for tumor survival but redundant in healthy tissues [79]. This approach has already led to the identification of several non-obvious targets in glutamine and lipid metabolism that are currently entering preclinical validation. As GEMs become more personalized through the incorporation of patient specific data, they will serve as indispensable tools for the development of the next generation of metabolic precision therapies.