From Molecular Escape to Clinical Relapse: Convergent Adaptive Programs and the Case for Predictive Oncology in Breast Cancer

1. Introduction

Breast cancer remains the most commonly diagnosed cancer among women worldwide, and although major advances in early detection and systemic treatment have significantly improved patient outcomes, disease recurrence and metastatic progression continue to be the leading causes of breast cancer–related mortality [1,2]. Importantly, this ongoing burden persists despite the growing range of therapeutic options, including endocrine therapies, CDK4/6 inhibitors, HER2-targeted antibodies and antibody–drug conjugates, PARP inhibitors, and immune checkpoint inhibitors [3,4]. This highlights an important challenge in breast cancer: treatment failure is not simply a consequence of ineffective drugs, but also reflects the remarkable ability of tumour cells to adapt, evolve, and survive therapeutic pressure [5,6].

Clinically and molecularly, breast cancer is broadly classified into three major subtypes: hormone receptor–positive (HR+), HER2-positive, and Triple-negative breast cancer (TNBC) [1,3]. These subtypes differ in their receptor expression, biological behaviour, disease progression, and response to treatment, and therefore require distinct therapeutic strategies [3,4]. This classification has been transformative for treatment selection, but it also frames resistance as a subtype-specific problem: ESR1 mutation and PI3K/AKT/mTOR activation in HR+ disease [7,8], HER2 loss and receptor-trafficking alteration in HER2-positive disease [9,10], and genomic instability with early clonal selection in TNBC [11,12]. What has received comparatively less attention is the extent to which these distinct escape routes converge on a shared set of adaptive programs such as epithelial-mesenchymal plasticity, cancer stem cell enrichment, metabolic rewiring, and microenvironment-mediated immune evasion—that recur across all three subtypes regardless of the genetic lesion that initiates resistance [13,14,15,16,17,18,19].

This review takes a closer look at this shared landscape of resistance. We begin by outlining the clinical course of breast cancer across different stages and identifying where treatment most commonly fails within each subtype [3,4]. We then explore four recurring adaptive programs that appear to drive resistance across tumours, regardless of receptor status. Finally, we examine how large-scale genomic studies, single-cell transcriptomics, circulating tumour DNA (ctDNA) profiling, and computational modelling have improved our understanding of when resistance emerges and whether it can be predicted, while also considering the practical challenges that continue to limit their use in routine clinical care [6,20,21].

Finally, and most importantly we ask why this mechanistic knowledge has translated so incompletely into practice. Current treatment algorithms remain anchored to a narrow biomarker set ER, PR, HER2, Ki67, and a handful of actionable mutations that captures only a fraction of a tumour’s biological state, and resistance is typically identified only after it manifests as radiological or symptomatic progression, by which point resistant populations are already well established. We argue that closing this gap requires reframing resistance biomarkers around the convergent adaptive programs described here rather than subtype-exclusive genetic catalogues, and in Section 6 we present exploratory survival-stratification data suggesting that different programs dominate the resistant phenotype in different subtypes—a pattern that, if validated, argues for subtype-anchored rather than universal resistance panels [6,20,21].

2. Clinical Progression and Point of Failure by Subtype

Treatment for all three subtypes is stage-dependent, escalating from local control in in-situ disease to biomarker-guided systemic therapy in metastatic disease (Figure 1) [3,20]. Despite very different drug classes, the point at which each subtype most commonly fails therapy follows a recognisable pattern: early, receptor-driven adaptation in HR+ disease [7,8]; compensatory signalling rewiring after initially deep responses in HER2-positive disease; and rapid clonal selection under chemotherapy or immunotherapy pressure in TNBC, where no receptor-directed option exists [11,12]. Figure 1 summarises standard of care sequencing and the resistance mechanism most strongly implicated at each stage. Full staging detail follows current NCCN guidance [2].

The fact that these different treatment trajectories ultimately converge at relapse provides an important starting point for this review. Although receptor status guides the choice of initial therapy, it does not, on its own, determine how tumour cells adapt to treatment or what enables them to survive and eventually develop resistance [5].

3. Convergent Adaptive Programs Across Subtypes

Although HR+, HER2-positive, and TNBC differ in receptor status, molecular drivers, and therapeutic strategy, resistance across these subtypes increasingly appears to arise from a common, limited set of adaptive programs rather than isolated subtype-specific escape mechanisms: epithelial-mesenchymal plasticity, cancer stem cell (CSC) enrichment, metabolic rewiring, and microenvironment-mediated immune evasion (Figure 2) [22].

This four-program scope is intentionally limited and is not a complete list. The broader cancer-resistance literature also implicates drug-efflux transporters, enhanced DNA-damage repair, apoptosis evasion, epigenetic reprogramming, and therapy-induced senescence as resistance mechanisms [23]. We focus on EMT, CSC enrichment, metabolic rewiring, and immune evasion for two reasons. First, these four are the ones most consistently reported, independently, across all three breast cancer molecular subtypes, specifically the pattern this review is built around. Several of the alternatives above are documented mainly within a single subtype or model system and have not been shown to converge across HR-positive, HER2-positive, and TNBC disease in the way Section 2 and Section 3 describe [24]. Second, several of the excluded mechanisms function mechanistically as downstream effectors of the four selected here rather than as independent axes alongside them: drug efflux and apoptosis evasion are frequently outputs of stemness and metabolic reprogramming rather than separable causal programs, and therapy-induced senescence is closely intertwined with EMT rather than clearly distinct from it [25].

3.1. Epithelial-Mesenchymal Plasticity

Epithelial-mesenchymal transition (EMT), along with the ability of tumour cells to occupy intermediate or hybrid epithelial-mesenchymal states, has emerged as an important mechanism of therapy resistance across all three breast cancer subtypes [13]. Rather than representing a simple, reversible change in cell phenotype, EMT reflects a broader increase in cellular plasticity that enables tumour cells to adapt to therapeutic stress [26]. Cells that acquire mesenchymal-like or hybrid characteristics often become more tolerant to treatment, invasive, and better equipped to survive and disseminate. In hormone receptor-positive (HR+) breast cancer, endocrine-resistant tumours commonly show increased activity of EMT-associated pathways. Similarly, HER2-positive and triple-negative breast cancers (TNBC) can develop EMT-like features following anti-HER2 treatment or chemotherapy, which are associated with reduced treatment sensitivity and increased metastatic potential.

3.2. Cancer Stem Cell Enrichment

Another feature shared across breast cancer subtypes is the enrichment of cancer stem-like cells (CSCs) or tumour-initiating populations following treatment. These cells are characterised by their ability to self-renew, withstand cellular stress, and remain in a relatively dormant state, allowing them to persist despite therapy and contribute to both early and late disease recurrence [27]. Therapeutic pressure may preferentially eliminate the more treatment-sensitive bulk tumour cells while allowing stem-like subpopulations to survive. As a result, residual disease following endocrine therapy, anti-HER2 treatment, or chemotherapy can become enriched in cells expressing stemness-associated markers such as ALDH1 and CD44 [28]. These surviving populations can subsequently expand and regenerate the tumour, contributing to the development of increasingly therapy-refractory disease.

3.3. Metabolic Rewiring

Metabolic plasticity provides another important mechanism through which breast cancer cells adapt to therapeutic stress. Tumour cells can remodel their use of glycolysis, oxidative phosphorylation, fatty-acid metabolism, and redox pathways to maintain energy production and survival under conditions of hypoxia, nutrient limitation, and drug exposure. In HR+ breast cancer, resistance to endocrine therapy and PI3K/mTOR-targeted treatments is often accompanied by increased mitochondrial activity and changes in lipid metabolism [29]. HER2-positive and TNBC tumours similarly undergo metabolic adaptations that help them maintain survival during targeted therapy, chemotherapy, or immune-mediated stress. Thus, metabolic rewiring does not simply support tumour growth; it provides resistant cells with the flexibility needed to survive changing environmental and therapeutic conditions.

3.4. Microenvironment-Mediated Protection and Immune Evasion

The tumour microenvironment facilitates treatment escape across all subtypes. Stromal cells, cancer-associated fibroblasts, tumour-associated macrophages, extracellular matrix remodelling, and hypoxic niches create protective ecosystems that shield residual tumour cells from both therapy and immune surveillance [30]. In HR+ tumours, resistant clones often emerge within stromal, hypoxic, immune-suppressed microenvironments; in HER2-positive and TNBC disease, chronic inflammatory signalling and stromal interaction contribute to T-cell exhaustion and regulatory T-cell expansion, limiting the effectiveness of targeted therapy and immunotherapy.

Together, these four programs indicate that therapy resistance in breast cancer is driven not only by subtype-specific genetic alterations but by broader adaptive states that promote plasticity, survival, and immune escape a shared evolutionary toolkit that different subtypes draw on to different degrees (Figure 2).

4. Computational and Data-Driven Approaches

Resistance to systemic therapy is rarely a single abrupt event; rather it reflects a continuous evolutionary process shaped by tumour heterogeneity, therapeutic pressure, and tumour-microenvironment interaction [31,32]. Over the past decade, large-scale bulk genomics, single-cell transcriptomics, liquid biopsy, and artificial intelligence (AI)-driven modelling have substantially advanced mechanistic understanding of resistance and begun to translate into clinical tools [33,34].

4.1. Bulk Genomics and Multi-Omic Integration

Large cohort studies, including integrative multi-omic clustering into molecular subgroups, have mapped recurrent resistance-associated alterations, most notably ESR1 mutation and PI3K pathway reactivation in endocrine-resistant HR+ disease [35]. Gene-expression signatures increasingly stratify recurrence risk within receptor-defined subtypes rather than merely confirming receptor status, and integrative multi-omic analyses are beginning to show that the same convergent adaptive programs described in Section 3 recur across independent cohorts and platforms [36].

4.2. Single-Cell Transcriptomics and Liquid Biopsy

Single-cell RNA sequencing has revealed the extent of intra-tumour heterogeneity that is obscured when gene expression is measured across the tumour as a whole, and pseudotime/trajectory analyses have begun to reconstruct the order in which resistant subclones emerge under treatment. Circulating tumour DNA (ctDNA) provides a complementary, longitudinal readout: quantitative ctDNA dynamics track treatment response earlier than imaging in several settings, and ctDNA-based minimal residual disease (MRD) monitoring is increasingly used to flag molecular relapse before it becomes radiologically apparent [37].

4.3. AI/ML and Network Modelling

Machine learning models trained on genomic, transcriptomic, and imaging data are being developed to predict resistance onset, and radiomics-genomics integration aims to extract resistance-relevant signal from routine imaging [38,39]. In parallel, network and pathway modelling including in silico perturbation, synthetic-lethality screening, and dynamic mathematical models of tumour evolution is being used to anticipate which combination strategies might pre-empt, rather than merely respond to, the convergent adaptive programs described above.

4.4. Limitations and Data Requirements

These approaches remain limited by cross-sectional, single-modality datasets that cannot fully capture a dynamic, multi-dimensional process [38]. Longitudinal sampling (serial biopsy, ctDNA, and clinical data at baseline, on-treatment, and at progression) is needed to distinguish causal drivers from passenger alterations, and multi-modal integration frameworks combining genomics, imaging, and clinical data are needed to achieve clinically actionable accuracy. Open, well-annotated, FAIR (Findable, Accessible, Interoperable, Reusable) datasets, supported by initiatives such as the Global Alliance for Genomics and Health, the Human Tumour Atlas Network, and the Cancer Research Data Commons, remain a prerequisite for reproducible, generalisable resistance models at the scale this problem requires [40].

5. Why Mechanistic Insight has Not Yet Changed Practice

The translational picture differs sharply by subtype. HER2-positive disease is oncology's clearest precision-medicine success: sequential HER2-directed agents, from dual blockade through antibody drug conjugates, have converted what was once the poorest-prognosis subtype into one of the more favourable ones, though adaptive signalling rewiring still erodes benefit over time. HR-positive disease, by contrast, remains managed largely reactively: ESR1 and PI3K-pathway status are tested mainly after progression rather than used to anticipate it, despite growing evidence that these alterations are detectable in ctDNA months before radiological relapse. TNBC shows the widest gap between biological knowledge and clinical translation, thoughimmunotherapy and PARP inhibition have improved outcomes in biomarker-selected patients, but no routine assay yet captures the EMT/stemness axis that Section 3 identifies as this subtype's dominant resistance program.

Two major structural limitations contribute to this gap. First, current treatment decisions are still largely guided by a relatively narrow set of biomarkers, including ER, PR, HER2, Ki67, and a limited number of actionable mutations. Although these markers are clinically valuable, they capture only a small part of the dynamic biological state of a tumour. Second, resistance is generally recognised only after it becomes clinically evident through radiological or symptomatic progression. By this stage, resistant subclonal populations may already be well established, making them considerably more difficult to target or reverse. These challenges are particularly pronounced in resource-limited settings, where comprehensive next-generation sequencing and repeated circulating tumour DNA (ctDNA) monitoring may not be readily accessible. In such settings, relatively inexpensive targeted panels could offer a practical alternative, provided that the key biological processes driving resistance can be represented by a sufficiently small and informative set of genes.

Several emerging approaches may help shift resistance detection towards an earlier stage. Low-cost targeted ctDNA panels, single-timepoint transcriptomic signatures, and AI-assisted imaging are increasingly providing feasible ways to capture treatment-induced changes before overt clinical progression. Importantly, the EMT and immune-exhaustion mechanisms discussed in Section 3 could potentially be translated into such biomarker strategies. Achieving this, however, requires a shift from the current reactive approach where resistance biomarkers are assessed primarily after treatment failure to a more predictive framework. In this model, the activity of convergent adaptive programmes would be monitored alongside established clinical biomarkers, allowing emerging resistance to be identified before it develops into clinically apparent relapse.

6. Discussion

When the stage-wise clinical trajectories described in Section 2 are considered alongside the mechanistic and computational evidence reviewed in Section 3, Section 4 and Section 5, a consistent pattern emerges. Breast cancer does not appear to develop treatment resistance through a single dominant mechanism. Instead, resistance repeatedly converges on a relatively small set of adaptive programmes, including EMT, cancer stem cell enrichment, metabolic rewiring, and immune evasion, across HR-positive, HER2-positive, and TNBC subtypes. What varies between subtypes is less the presence of these programmes than the extent to which each contributes to the resistant phenotype and the stage of disease at which it becomes clinically important.

This perspective shifts the central question from identifying subtype-specific resistance genes to understanding which adaptive programme is most active within an individual tumour. The more clinically relevant challenge may therefore be to determine whether EMT, stemness, metabolic adaptation, or immune evasion is driving resistance in a particular patient and, importantly, whether that biological state can be detected early enough to guide treatment before overt progression occurs.

6.1. What Subtype-Wise Survival Stratification Adds to the Mechanistic Narrative

Methods: gene-program stratification and survival analysis

To extend this argument beyond a purely mechanistic synthesis, we examined overall survival within each breast cancer subtype according to the gene-expression programme most strongly enriched in the tumour. These patterns were compared with a reference framework based solely on the standard clinical biomarker panel (Figure 3). Although the findings should be interpreted cautiously given the statistical limitations of the analysis, they revealed a notable subtype-specific pattern. EMT appeared to have the strongest association with survival in TNBC, immune evasion in HR-positive disease, and metabolic rewiring in HER2-positive disease.

Importantly, this pattern provides more than a simple survival association. It suggests that the four convergent adaptive programmes discussed in Section 3 may not represent interchangeable descriptions of the same underlying biology. Instead, they may function as distinct routes through which tumours adapt to therapeutic pressure, with their relative contribution varying across breast cancer subtypes. In this context, subtype-specific dominance of an adaptive programme could help explain why similar resistance mechanisms ultimately produce different clinical trajectories across breast cancer.

Patients were drawn from TCGA-BRCA (PanCancer Atlas study, cBioPortal study ID: brca_tcga_pan_can_atlas_2018) and assigned to HR-positive (n = 310), HER2-positive (n = 33), and triple-negative (n = 86) subtypes using immunohistochemical receptor status (ER, PR, HER2) as annotated in the clinical metadata, corroborated by PAM50 intrinsic subtype calls provided in the same dataset. For each of the four programs defined in Section 3 (EMT, CSC, metabolic rewiring, immune evasion), tumours were scored for pathway enrichment using the mean z-score of RNA-seq expression (RSEM, log2-transformed, batch-normalized) across the subtype-specific gene panel for TNBC, EMT (ZEB1, SNAI1, SNAI2, TWIST1, VIM, CDH1, FN1, CDH2), CSC (ALDH1A1, CD44, SOX2, PROM1, NANOG, POU5F1, KLF4), immune evasion (CD274, IDO1, LAG3, B2M, HLA-A/B/C, CTLA4, TIGIT), metabolic rewiring (SLC2A1, LDHA, GLS, IDH1, MYC, HIF1A); for HR+/HER2, EMT (ZEB1, SNAI2, VIM, CDH1), CSC (ALDH1A1, CD44, SOX2), immune evasion (CD274, IDO1, TIGIT), metabolic rewiring (SLC2A1, LDHA, MYC); for HER2+, EMT (ZEB1, TWIST1, VIM, CDH1, CDH2), CSC (CD44, SOX2, PROM1), immune evasion (CD274, LAG3, B2M), f’DDfGene panels were curated from the subtype-specific mechanistic literature cited in Section 4 and Section 6 of this review.

Within each subtype, patients were assigned to the program with the highest relative enrichment z-score ("dominant program"), and overall survival was compared across program groups, benchmarked against a comparator group defined by the standard clinical panel (ER/PR/HER2/Ki67), using Kaplan-Meier estimation with log-rank testing for univariate comparison, followed by multivariable Cox proportional hazards regression adjusting for AJCC stage and PAM50 subtype, implemented in R (v4.3) using the survival and survminer R package used to make survival-analysis. Follow-up time was measured from date of initial diagnosis to death from any cause or last known follow-up, with patients alive at last contact right-censored at that date. Given the multiple pairwise comparisons performed across program groups and subtypes, Benjamini-Hochberg false discovery rate (FDR) correction was applied to log-rank p-values; uncorrected nominal p-values are also reported for transparency, consistent with the exploratory framing already noted in Section 6.5.

6.2. HER2-Positive Disease as a Case Study: Metabolic Rewiring as a Candidate Maintenance Target

The HER2-positive stratum illustrates why this distinction has practical consequences. Standard anti-HER2 therapy dual blockade with chemotherapy in the first line, or antibody drug conjugates thereafter eliminates the ERBB2-dependent bulk of the tumour with remarkable efficacy, which is precisely why this subtype is often held up as a precision-oncology success. Yet ERBB2 amplification constitutively activates PI3K-AKT-mTOR signalling, which upregulates GLUT1 (SLC2A1)-mediated glucose uptake and glycolytic flux; MYC, frequently co-amplified on the same 17q amplicon, reinforces this glycolytic and lipogenic phenotype independently of continued HER2 signalling; and HIF1A stabilisation sustains a pseudo-hypoxic, therapy-tolerant state even in normoxic tumour regions. This provides a biologically plausible route by which a metabolically rewired subpopulation could survive adequate HER2 blockade, persist as radiologically invisible residual disease, and seed the early relapses that Figure 3 suggests are concentrated in exactly this subgroup.

If confirmed in larger cohorts, this would argue for testing metabolic-program enrichment, to be assessed at diagnosis or on-treatment as a stratification biomarker for maintenance therapy layered onto the anti-HER2 backbone, rather than reserved for use after relapse. Candidate agents span a spectrum of clinical maturity: metformin, an AMPK activator with indirect suppressive effects on mTOR signalling and glycolytic flux, is already in clinical use and has been evaluated in the adjuvant setting, making it the most immediately testable option for a biomarker-selected maintenance trial; FASN inhibitors target the lipogenic arm of the same program and have early-phase data in other FASN-high tumours; GLUT1/SLC2A1 inhibitors and HIF1A-pathway-directed strategies remain earlier in development but map directly onto the effectors enriched in this subgroup. We propose this as a testable model rather than a therapeutic recommendation: metabolic-program enrichment at diagnosis may identify a HER2-positive subset at elevated relapse risk despite adequate ERBB2 suppression, defining a population for a biomarker-stratified maintenance trial analogous in design logic to how immune-checkpoint enrichment already guides maintenance immunotherapy in other cancers.

6.3. Why a Single Biomarker Panel Will Not Suffice

The divergence in dominant program across subtypes has a direct implication for how resistance biomarkers should be built. A panel optimised to detect EMT activation, however well validated in TNBC, would have limited value in a HER2-positive tumour whose principal vulnerability is metabolic, just as an immune-exhaustion signature calibrated to HR-positive disease may not capture the biology driving relapse in a metabolically rewired HER2-positive tumour. This argues against a universal resistance signature and toward subtype-anchored, program-specific panels deployed alongside, rather than in place of, the standard clinical panel. Practically, this need not require full genomic infrastructure: a modest, targeted expression panel built around the effector genes identified for each program (Figure 2) could in principle be implemented as a low-cost, PCR-based assay, which matters directly for the resource-limited settings noted in Section 5, where comprehensive NGS and ctDNA monitoring remain largely inaccessible.

6.4. The Need for Multiple Approaches

Combining multi-omics data with multiple experimental model systems gives a much richer, more validated picture than relying on any single layer or single model. Each experimental model captures a different part of the biology. On the measurement side, genomics, transcriptomics, and proteomics answer different questions rather than the same question at different resolutions. In preclinical research, no single model can fully capture tumour biology while also allowing high-throughput experimentation. The balance between biological fidelity and experimental throughput therefore becomes particularly important when studying the four convergent resistance programs. Two-dimensional cell lines are cheap, scalable, and genetically tractable, but grown without stroma, immune cells, or a three-dimensional architecture, they are poorly suited to studying EMT plasticity, immune evasion, or microenvironment-mediated protection. Patient-derived xenografts (PDX) preserve genomic, histological, and drug-response fidelity to the original tumour more faithfully than cell lines, and remain a standard for testing whether a resistance mechanism holds in vivo, but they are slow and expensive to generate at scale, and because they are grown in immunodeficient mice, they cannot be used to study immune-evasion biology at all. Patient-derived organoids sit between these two extremes: they preserve three-dimensional architecture and patient-specific drug sensitivity with useful fidelity to clinical outcome, and can be generated and screened far faster than PDX, but standard culture conditions still lack a native stromal and immune compartment unless deliberately co-cultured. Ex vivo tumour-slice or patient-derived cell cultures briefly preserve the most complete native microenvironment of any model, at the cost of viability windows too short for most longitudinal resistance studies.

The practical implication is that no single platform whether an omic layer or an experimental model can validate a convergent-program hypothesis on its own; each captures one program well and the others poorly.

6.5. Limitations

Several limitations temper these conclusions. The gene-program subgroups particularly within HER2-positive disease (n = 33) and the individual program strata within TNBC and HR-positive disease are small, and the corresponding survival estimates carry wide, overlapping confidence intervals; only the TNBC stratification reached conventional significance, and even there, multiple comparisons were performed without independent replication. The analysis is correlative and retrospective, drawing on transcriptional enrichment at a single timepoint rather than longitudinal profiling, and does not adjust for stage, treatment received, or line of therapy, any of which could confound the association between program enrichment and survival. The proposed metabolic-maintenance strategy in HER2-positive disease is accordingly a mechanistic hypothesis grounded in a numerically consistent but statistically non-significant signal, not a validated clinical recommendation, and would require confirmation in larger, prospectively collected HER2-positive cohorts ideally with serial biopsy or ctDNA-based tracking of metabolic-program activity before any maintenance-therapy strategy could reasonably be tested in a trial setting.

6.6. Concluding Perspective

Why does therapy fail in breast cancer? The evidence discussed throughout this review, together with the survival patterns shown in Figure 3, suggests on the one hand, treatment resistance across HR-positive, HER2-positive, and TNBC disease appears to converge on a relatively small and recognisable set of adaptive programmes. EMT, stem-cell enrichment, metabolic rewiring, and immune exhaustion are therefore not simply features of resistant tumours; they represent potentially actionable biological states that could, in principle, be therapeutically targeted.

On the other hand, most of these adaptive states remain largely invisible to routine clinical decision-making until they have already translated into radiological or symptomatic progression. The real challenge, then, isn't just discovering new resistance mechanisms it is finding ways to measure them clinically before resistance actually takes hold. By treating EMT status, stem-cell burden, metabolic phenotype, and immune exhaustion as measurable, subtype-specific variables alongside the markers we already rely on (ER, PR, HER2, and Ki67), we could get a far more dynamic picture of how a tumor is behaving. This kind of shift would move breast cancer treatment away from a reactive approach, where we respond only after resistance has emerged, and toward a predictive one, where the goal is to catch and intervene against resistance before it ever leads to clinical progression.