Tumors may harbor pre-existing drug-resistant genotypes or epigenetic phenotypes that gain a competitive advantage during treatment, ultimately leading to the manifestation of drug resistance. For example, a study found that before gefitinib treatment in epidermal growth factor receptor (EGFR)-mutated lung cancer, a cell subpopulation with amplification of the oncogene mesenchymal-epithelial transition factor (MET) was already present in the tumor. This subpopulation developed resistance to gefitinib when activated by hepatocyte growth factor (HGF) through the phosphatidylinositol 3-kinase (PI3K)/protein kinase B (AKT)/GRB2-associated binding protein 1 (GAB1) signaling pathway after drug-related selection^[33]. In an evolutionary model, events such as MET amplification can originate from a rare pre-existing subclone generated by neutral drift or early branched evolution. Upon EGFR-tyrosine kinase inhibitor (TKI) treatment, this specific subclone undergoes a pronounced clonal sweep driven by strict Darwinian selection, rapidly expanding to dominate the bulk of the tumor and manifesting as clinical resistance. Studies on leukemia have also indicated that pre-existing resistant subpopulations exist before treatment^[25]. Prior to therapy, certain leukemia subclones exhibit chromatin accessibility–mediated epigenetic upregulation of T-cell lymphoma invasion and metastasis-inducing protein 1 (TIAM1) and zinc finger protein 257 (ZNF257) promoters, conferring enhanced resistance^[25]. Similarly, in chronic lymphocytic leukemia (CLL), a latent ibrutinib-resistant subclone was identified before treatment. This subclone displayed MYC target gene upregulation, marked C-X-C motif chemokine receptor (CXCR)4 downregulation, and breakpoint cluster region (BCR) pathway activation, resembling lymph-node-like activated cells^[34]. During therapy, this pre-existing subclone underwent selective expansion, ultimately leading to ibrutinib resistance at relapse^[34].

In triple-negative breast cancer (TNBC), reprogramming of histone H3K27me3 sites and DNA methylation–mediated chromatin alterations may drive resistance to 5-fluorouracil (5-FU) and capecitabine^[27]. Throughout this process, intratumoral diversity progressively decreases, indicating pressure-driven selection among the initial subclonal populations^[27]. Resistant subclones subsequently emerge as dominant components of the drug-resistant tumor^[27]. Likewise, in muscle-invasive bladder cancer (MIBC), two distinct resistant subclones,S1 and S4,were already present before neoadjuvant immune checkpoint inhibitor (CPI) therapy^[35]. S1 was enriched for fibroblast growth factor receptor 3 (FGFR3)/Kirsten rat sarcoma viral oncogene homolog (KRAS) mutations and peroxisome proliferator-activated receptor gamma (PPARG) amplification, whereas S4 showed upregulation of MYC, cell cycle, and DNA damage response (DDR)-related pathways^[35]. These subclones constituted the initial resistance population and gained a competitive advantage during treatment, representing a major cause of pembrolizumab resistance^[35].

A study on sonic hedgehog (SHH)-type medulloblastoma revealed significant intratumoral spatial heterogeneity in pediatric medulloblastoma with chromothripsis^[36]. These tumors displayed enhanced proliferative and stem-like features with reduced immune infiltration, and contained two major drug-resistant clones: Clone A, an “ancestral” near-diploid clone, and Clone B, which underwent additional chromothripsis, exhibiting polyploidy and complex chromosomal rearrangements^[36]. These two clones coexisted across tumor regions both before and during treatment, thereby driving disease recurrence^[36].

Tumors may undergo genetic or epigenetic reprogramming during treatment and gradually acquire drug resistance. Before the acquisition of permanent genetic mutations, cancer cells often exploit non-genetic plasticity to enter a reversible, drug-tolerant persister (DTP) state^[28]. This early epigenetic reprogramming serves as a vital survival reservoir, allowing persister cells to withstand initial therapeutic pressure until more permanent resistance mechanisms evolve^[37].

In the study on MIBC discussed earlier, in addition to the selection of pre-existing subclones (S1 and S4) the same cohort also exhibited adaptive evolution during pembrolizumab treatment. Specifically, therapeutic pressure induced the emergence of two novel subclones, S6 and S7, which were entirely absent prior to therapy^[35]. S6 subclone activated the FGFR–mitogen-activated protein kinase (MAPK) pathway and upregulated transforming growth factor (TGF)-β1, SMAD2/3, zinc finger E-box binding homeobox 1 (ZEB1), Snail family transcriptional repressor 2 (SNAI2), and other transcription factors through histone deacetylation and methylation reprogramming, exhibiting epithelial–mesenchymal transition (EMT)-like characteristics. The S7 subclone, on the other hand, activated inflammatory signals such as interferon-gamma (IFN-γ), tumor necrosis factor-alpha (TNF-α), and interleukin (IL)-6, upregulated the MAPK/MAPK-ERK kinase (MEK)/protein kinase C (PKC) pathway, and demonstrated immune-tolerant properties^[35].

Measurable residual disease (MRD) after chemotherapy is a common cause of resistance and relapse in pediatric B-cell acute lymphoblastic leukemia (B-ALL)^[23]. Zhang et al. employed paired single-cell B-cell receptor sequencing (scBCR-seq) and single-cell RNA sequencing (scRNA-seq) in pediatric B-ALL patients treated with first-line conventional chemotherapeutic agents (cytarabine, daunorubicin, and vincristine) and used these paired single-cell approaches to track temporal tumor heterogeneity^[23]. The authors integrated data from three phases of disease progression in the same patient - diagnosis (Dx), day 19 of induction chemotherapy (D19; MRD), and relapse (Rel) - and identified distinct phenotypic differences across these phases. D19 samples exhibited widespread G0/G1 arrest, with differential enrichment analyses revealing significantly upregulated hypoxia-related pathways. In contrast, Rel samples showed a shift toward less differentiated or more primitive leukemic states, with increased proportions of earlier-stage components - such as pro-B or hematopoietic stem cell (HSC)/lympho-myeloid primed progenitor (LMPP)-like cells - and resolution of the G0/G1 arrest observed at D19^[23]. These findings indicate that, under chemotherapeutic pressure, tumor cells collectively transitioned from their original Dx state to a stress-response and survival state at D19, and subsequently to a low-differentiation, re-proliferative state in Rel^[23]. This temporal reprogramming ultimately led to chemotherapy failure.

Using scRNA-seq, Miao et al. reported that in squamous cell carcinoma (SCC) a population of TGF-β-responsive α6hi/cluster of differentiation (CD) 34⁺/CD44⁺ tumor-initiating stem cells (tSCs) selectively acquired high CD80 expression to preferentially survive adoptive cytotoxic T-cell transfer (ACT)-based immunotherapy and drive recurrence^[38]. During this process, CD80 expressed by tumor cells bound directly to cytotoxic T-lymphocyte-associated protein 4 (CTLA4) on cytotoxic T lymphocytes (CTLs), and inhibited CD8⁺ T-cell proliferation and effector molecule production,such as granzyme B (GZMB), IFN-γ, and TNF-α, thereby promoting T-cell exhaustion and establishing adaptive immune tolerance^[38]. Correspondingly, knocking out CD80 or blocking CTLA4 restored CTL activity, increased tSC apoptosis, and reduced recurrence^[38]. Similarly, in multiple myeloma, resistance to bispecific antibodies targeting G-protein-coupled receptor family C group 5 member D (GPRC5D) could be gained through gene inactivation or through long-range epigenetic silencing of its promoter and enhancer regions^[39]. Whole-genome duplication (WGD) also plays a significant role in the development of acquired drug resistance in tumors. In the TRAcking Cancer Evolution through therapy (Rx) (TRACERx) project, researchers found that in many cases of lung SCC and adenocarcinoma, cancer cells first undergo extensive loss of heterozygosity (LOH), followed by WGD^[40]. This allows them to evade the control of tumor suppressor genes and also the lethal consequences of defects in survival-related genes, demonstrating the sophisticated strategies employed by tumor cells to gain drug resistance.

Specific tumor subclones, stromal cells, and other cellular components can shape distinct TMEs, thereby generating local conditions that ultimately contribute to drug resistance. In the interaction between the TME and drug therapy, the TME frequently alters the entire tumor tissue’s response to drug treatment through mechanisms such as rebuilding the extracellular matrix (ECM) physical barrier, reshaping the immunosuppressive microenvironment, and modulating paracrine signaling.

In urothelial carcinoma treated with gemcitabine plus cisplatin, Kikuchi et al. observed that drug-stimulated tumor cells upregulated IL-8 secretion, which in turn acted on tumor endothelial cells to activate nuclear factor kappa-light-chain-enhancer of activated B cells (NF-κB)^[41]. This activation led to increased ATP-binding cassette sub-family B member 1 (ABCB1) expression in tumor endothelial cells, conferring drug-efflux capacity that not only enabled tumor tissues to resist first-line chemotherapeutic agents such as gemcitabime and cisplatin, but also reduced the efficacy of second-line drugs such as paclitaxel^[41].

Under the selective pressure of immunotherapy, the tumor immune microenvironment undergoes extensive immunoediting. Tumor sites with low intrinsic resistance and poor immunogenicity are preserved due to their strong survival capacity, allowing them to establish a “cold” TME that ensures their long-term persistence. In metastatic melanoma, immune cells and cancer-associated fibroblasts (CAFs) jointly shaped a microenvironment resistant to anti-programmed cell death protein 1 (PD-1) and anti-CTLA-4 immunotherapy. In this microenvironment, CD8⁺ T cells decreased in number and accumulated at the tumor periphery, whereas B cells and plasma cells became enriched within the tumor and formed lymphoid follicle-like structures^[42]. At the same time, CAFs upregulated chemokines and genes associated with ECM remodeling, thereby establishing both physical and immune barriers^[42]. Similarly, intraductal papillary mucinous neoplasms (IPMN), the precursor lesions of pancreatic ductal adenocarcinoma (PDAC), also exhibited an immunosuppressive TME^[43]. High-grade IPMN was found to secrete large amounts of abnormally O-glycosylated mucins that formed physical and immunosuppressive barriers. These impeded natural killer (NK)/T-cell recognition and infiltration and promoted immune escape, ultimately leading to resistance to the combination of gemcitabine and nab-paclitaxel^[43]. Liu et al. also identified specific tumor immune barrier (TIB) structures at the tumor periphery during immune checkpoint inhibitor (ICI) treatment of hepatocellular carcinoma (HCC)^[44]. These structures manifested as secreted phosphoprotein 1-positive (SPP1⁺) macrophage-CAF colocalization, in which SPP1⁺ macrophages highly expressed ligands such as TGFB1, SPP1, and interleukin-1beta (IL1B), while CAFs highly expressed their corresponding receptors. Subsequent activation of downstream collagen genes (COL1A1/2/3A1/4A1/5A1), matrix metalloproteinases (MMP) and their inhibitors, and chemokines [C-C motif chemokine ligand (CCL) 3/4/5, CXCR4] led to regulation of ECM organization, cell adhesion, fibrotic responses, and chemotactic signaling, and ultimately to resistance to CD8⁺ T-cell infiltration^[44].

The types and frequency of cell-cell interactions between stromal cells and tumor cells are also dynamically reshaped under therapeutic pressure, resulting in connections that are more conducive to tumor survival. Zhou et al. reported that in colorectal cancer samples from patients receiving oxaliplatin chemotherapy, thrombospondin-2 (THBS2) positive CAFs localized close to oxaliplatin-resistant malignant cells and engaged in collagen-mediated interactions^[45]. THBS2⁺ CAFs specifically secreted COL8A1, which directly bound to integrin subunit beta 1 (ITGB1) on the surface of resistant malignant cells, activated the PI3K-AKT pathway, induced EMT, and ultimately led to drug resistance^[45]. Additionally, by integrating The Cancer Genome Atlas Program (TCGA) bulk RNA-seq data, pan-cancer protein datasets, pan-cancer cell line expression data, pan-cancer scRNA-seq data, and public spatial transcriptomics datasets from ovarian, breast, pancreatic, and head and neck squamous cell carcinomas (HNSCC), THBS was found to be enriched in CAFs across multiple cancer types, suggesting that such tumor-stroma interactions might be associated with poor chemotherapy response across a broad spectrum of cancers^[45]. Similarly, carboplatin-resistant ovarian clear cell carcinoma (OCCC) subclusters frequently colocalized with CAFs, forming a unique resistance niche in which tumor cell subclusters activated CAFs via platelet-derived growth factor (PDGF). These activated CAFs, in turn, induced tumor cells to upregulate hypoxia inducible factor (HIF)-1α/HIF-2α, thus promoting chemotherapy resistance^[46]. In drug-resistant breast cancer gene (BRCA)1/2-mutated breast cancer treated with selective poly (ADP-ribose) polymerase inhibitors (PARPi) like olaparib and AZD5305, a specific tumor-associated macrophage (TAM) subset, TAM_C3, was enriched within the TME. These macrophages exhibited enhanced 40S ribosomal protein S19 (Rps19)–complement component 5a 1eceptor 1 (C5aR1) signaling with tumor cells, thereby suppressing effector molecules such as GZMB and perforin 1 (PRF1) in CD8⁺ T cells, inducing immune suppression, and ultimately promoting resistance to PARPi therapy^[47]. Similarly, during osimertinib treatment for non-small cell lung cancer (NSCLC), M2-TAM exosomes carrying lncRNA MSTRG.292666.16 were upregulated and delivered to tumor cells, thus activating the MSTRG.292666.16–miR-6836-5p–C-jun-amino-terminal kinase-interacting protein 3 (MAPK8IP3) axis and consequently activating MAPK signaling. The resulting enhanced tumor cell survival and growth in the presence of osimertinib led to the development of acquired resistance^[48].

The TME functions as a dynamic system that co-evolves with tumor clones via bidirectional signaling and metabolic coupling, rather than acting as a static barrier. In study using osimertinib to treat NSCLC, Zhang et al. identified a collagen triple helix repeat containing 1 (CTHRC1) positive CAF subpopulation that promoted TGF-β1 binding to TGF beta receptor 2 (TGFBR2) by secreting CTHRC1, thereby enhancing cancer cell glycolysis through the TGF-β/SMAD3/hexokinase 2 (HK2) axis and leading to EGFR-TKI resistance^[49]. Conversely, tumor cells increased lactate concentrations in the TME through glycolysis, which in turn upregulated CTHRC1 expression in CAFs via p300-mediated H3K18la, thus completing a positive feedback loop^[49].

Treatment-induced permanent cell cycle arrest in cancer cells is known as therapy-induced senescence (TIS) and is often regarded as a highly successful outcome of cancer therapy because it irreversibly halts the proliferation of malignant cells^[50]. However, these senescent cells remain metabolically hyperactive. They secrete a complex proteome comprising pro-inflammatory cytokines (such as IL-6, IL-1α/β), chemokines [such as C-X-C motif chemokine ligand (CXCL) 8 and 11], MMPs, and small extracellular vesicles (sEVs), collectively known as the senescence-associated secretory phenotype (SASP)^[50,51]. The SASP shows dynamic changes over the course of cancer treatment. Under prolonged therapeutic pressure, the chronic SASP signals shift from an initial tumor-suppressive role to a highly pro-inflammatory and matrix-degrading effect, becoming one of the primary contributors to drug resistance and tumor recurrence^[51]. For example, Raynard et al. found in breast and prostate cancer models that NF-κB-dependent SASP factors secreted by senescent cells can directly induce neuroendocrine transdifferentiation in adjacent epithelial cancer cells, endowing them with strong invasiveness and intrinsic resistance to treatment^[52]. These findings illustrate the dynamic changes in TME that occur during therapy, gradually leading to the development of a treatment-tolerant phenotype.

Genetic mutations or epigenetic modifications occurring at the target site itself may alter the target’s expression or conformation, thereby preventing the drug from effectively binding to its intended target.A classic example of this type of targeted therapy resistance mechanism is the EGFR-TKI-based targeted therapy for NSCLC. A large number of off-target mutations have been identified in EGFR to date, with considerable interpatient variability, including C797S^[56-60], G796S/R/D^[61-63], L792 and L718 mutations^[61,64], G724S^[65-67], and EGFR amplification^[68,69]. Heterogeneity adds an extra layer of complexity: different subclones within the same patient may harbor diverse resistance mutations simultaneously. For example, during crizotinib therapy, patients exhibited different resistance mechanisms, and even a single patient could harbor multiple resistance-associated mutations^[70].

BRAF inhibitors such as vemurafenib or dabrafenib are commonly used to treat melanoma harboring BRAF V600E/K mutations. However, BRAF amplification and abnormal alternative splicing can confer resistance to the cytotoxic effects of BRAF inhibitors by increasing BRAF kinase activity levels or generating truncated BRAF variants that are more prone to dimerization, with the types of resistance varying among different patients^[71-73]. In targeted therapies for chronic myeloid leukemia involving BCR–Abelson murine leukemia viral oncogene homolog (ABL) inhibition, different BCR–ABL mutant clones, such as Y253H, E255K/V, F317L, and T315I, may coexist, or even form compound mutations, and be selected under drug pressure, thereby leading to resistance^[74,75]. The resistance patterns also vary with specific mutations. For example, Y253H and E255K/V were resistant to imatinib but remained sensitive to dasatinib; F317L was resistant to both imatinib and dasatinib but was sensitive to nilotinib; and T315I conferred resistance to all three inhibitors^[74,75].

By inducing genetic or epigenetic alterations in bypass pathways or downstream targets, bypass activation or downstream reactivation can be established in tumor subclones, leading to targeted drug failure. Using targeted therapy for NSCLC as an example, bypass activation, including abnormal amplification of MET and the human EGFR (HER2), has long been recognized as one of the mechanisms underlying resistance to EGFR-TKIs^[76,77]. When EGFR was inhibited, its downstream signaling pathways, particularly the PI3K/AKT and MAPK pathways, could still be activated via MET or HER2, thereby bypassing EGFR dependency and leading to treatment failure with EGFR-TKIs^[78]. Similarly, when EGFR-mutated NSCLC was treated with the third-generation EGFR-TKI osimertinib, some cells did not undergo cell death but instead entered a DTP state^[79]. These cells exhibited unique characteristics, including slowed proliferation, reduced dependence on EGFR signaling, and enhanced antiapoptotic capacity^[79]. In DTP cells Izumi et al. discovered that B-cell lymphoma 2-like protein 1 (BCL2L1) was significantly upregulated, whereas pro-apoptotic genes such as induced myeloid leukemia cell differentiation protein (MCL1), BCL2-associated X protein (BAX), and BCL2 homologous antagonist/killer 1 (BAK1) were downregulated. This enabled resistance to EGFR-TKI–induced apoptosis and prolonged survival, which in turn provided a time window for the accumulation of new mutations in drug-resistant clones^[79]. Conversely, knockout or pharmacological inhibition of BCL2L1 restored NSCLC sensitivity to osimertinib, suggesting its potential as a therapeutic target^[79].

During tamoxifen treatment, breast cancer acquired a novel epigenetic landscape characterized by loss of H3K27me3 at polycomb-group (PcG) protein-targeted genes and genes associated with basal-like mammary epithelial signatures driven by the selective evolution of pre-existing drug-resistant subclones. These alterations led to the upregulation of genes such as EGFR, insulin-like growth factor binding protein 3 (IGFBP3), and activated leukocyte cell adhesion molecule (ALCAM), established novel survival and proliferation pathways, and ultimately drove the emergence of tamoxifen resistance^[80].

Bevacizumab is an anti-vascular endothelial growth factor (VEGF) pathway antibody. During bevacizumab therapy for glioblastoma (GBM), reactive activation of MET-related signaling pathways occurred, thereby leading to targeted treatment resistance^[81]. Using paired tumor specimens from GBM patients before and after bevacizumab treatment, together with a mouse intracranial GBM model, Lu et al. demonstrated that prior to treatment, VEGF recruited the protein tyrosine phosphatase PTP1B by binding to the MET/VEGF receptor 2 (VEGFR2) heterodimer complex, thereby inhibiting HGF-induced tyrosine phosphorylation of MET and cell migration^[81]. When VEGF was blocked, this inhibitory mechanism was relieved, thereby activating MET-related pathways and triggering EMT, which resulted in increased invasiveness, distant metastasis, and targeted therapy resistance^[81]. Similarly, when gefitinib and erlotinib were used as EGFR-TKIs to treat GBM, targeted resistance accompanied by MET activation also occurred^[82]. Jun et al. found that activated MET sustained downstream survival signaling, such as AKT, thereby shifting the response to EGFR-TKIs from cytotoxicity to cell quiescence (G1 arrest) and conferring tolerance^[82]. In contrast, combined inhibition of EGFR and MET restored apoptosis and reduced survival, overcoming this tolerant phenotype^[82].

Tumor tissues may undergo various phenotypic alterations, such as EMT or metabolic reprogramming, during subclone variation and selection, rendering targeted drugs ineffective. As observed in the aforementioned study on MIBC, drug resistance can manifest through phenotypic changes. Specifically, the emergent S6 subclone displayed EMT features in tumor tissues and showed associated drug resistance^[35]. Smoothened inhibitors (SMOi) target the Hedgehog pathway and are commonly used as first-line treatments for basal cell carcinoma (BCC) associated with Gorlin syndrome (nevus BCC syndrome)^[83]. Sporadic BCC exhibits a high rate of resistance, whereas tumors arising in Gorlin syndrome patients with germline patched-1 (PTCH1) mutations show uniform suppression with inhibitor therapy^[83]. Jussila et al. combined 10× Genomics scRNA-seq with 10× Genomics Visium spatial transcriptomics to investigate why a small subset of Gorlin syndrome patients treated with long-term inhibitors developed drug-resistant tumor clones and progressed rapidly^[83]. They found that tumors in resistant Gorlin syndrome patients underwent basal-to-squamous transdifferentiation (BST), building SCC-like regions that exhibited a Hedgehog-independent phenotype^[83]. During this process, GLI family zinc finger 1 (GLI1) signaling was significantly downregulated, whereas the phosphatidylethanolamine synthesis pathway was activated. Mutations in the tumor regions undergoing BST were found enriched in phosphate cytidylyltransferase 2 (PCYT2) and ethanolamine kinase 1 (ETNK1), ultimately leading to increased expression of squamous differentiation marker genes, such as cytokeratin 5/6 (KRT5/6), tumor protein p63 (TP63), and lymphocyte antigen 6 family member D (LY6D), in tumor cells. This resulted in a SCC phenotype, loss of dependence on SMO signaling, and subsequent targeted therapy resistance^[83]. Similarly, in BCC treated with SMOi, Li et al. used scRNA-seq, single-cell assay for transposase-accessible chromatin sequencing (scATAC-seq), and cleavage under targets and release using nuclease (CUT&RUN) to identify and define an NF-κB-driven tumor epithelial state known as the basal-to-inflammatory transition (BIT)^[84]. During this process, triggering receptor expressed on myeloid cells 1 (TREM1) positive myeloid cells within a specialized intratumoral niche secreted IL-1 and oncostatin M (OSM), activating the NF-κB/signal transducer and activator of transcription 3 (STAT3) pathway in the tumor epithelium, directly upregulating BIT markers such as chitinase-3-like protein 1 (CHI3L1) and vascular cell adhesion molecule-1 (VCAM1) while downregulating Hedgehog/GLI1 signaling, thereby reducing sensitivity to SMOi^[84]. Notably, Li et al. detected both BIT and BST sites coexisting in a non-overlapping manner: BIT resided in the inflammation-enriched, myeloid-dominated superficial-epithelial niche, whereas BST was more tumor-core-oriented and matrix-isolated^[84]. This further highlights the complexity of tumor-intrinsic heterogeneity and demonstrated how local heterogeneity within the tumor–immune environment shapes drug responses.

Some EGFR-mutated lung adenocarcinomas (LUAD) undergo transformation into small cell lung cancers (SCLC) after EGFR-TKI treatment, resulting in drug resistance^[85]. Using 10× Visium-FFPE, Li et al. revealed that neurodifferentiation-related pathways were upregulated during this process, whereas NSCLC-specific, apoptosis, cell-adhesion, and immune-related pathways were downregulated^[85]. This was accompanied by reduced histone deacetylase 1 (HDAC10) expression and sustained activation of fibroblast growth factor (FGF) signaling pathways^[85]. Consequently, the epidermal growth factor (EGF)–EGFR axis and its downstream signaling were significantly attenuated following transformation, whereas the FGF–FGFR axis was markedly upregulated and maintained sustained activity consistent with pathway dependency^[85]. This allowed tumor cells to escape EGFR dependency and resulted in resistance to EGFR-TKIs.

Single-cell transcriptomics and genomics enable high-throughput mapping of gene expression and genetic variations, allowing precise identification of resistant subclones and their evolutionary pathways under therapeutic pressure. One of the most widely adopted single-cell transcriptomic approaches is 10× Genomics scRNA-seq^[15,131]. For example, Jerby-Arnon et al. applied this technology, combined with in situ protein imaging, to melanoma treated with ICI. The authors identified an immune-resistant tumor-cell state characterized by high expression of cell cycle and transcription factor genes [cyclin-dependent kinase (CDK) 4/6–E2F–MYC axis], low expression of IFN-γ-response and major histocompatibility complex class I (MHC-I) pathways, a blunted response to IFN-γ stimulation, impaired antigen presentation, and absent chemotactic signaling, resulting in an immune-rejecting phenotype that ultimately led to ICI treatment failure^[132].

In docetaxel-resistant prostate cancer (CRPC), Schnepp et al. analyzed the parental and docetaxel-resistant DU145 and PC3 cell lines using scRNA-seq^[133]. They identified coexisting sensitive and resistant regulatory states within the same cell line, accompanied by large-scale rearrangements of transcription factor-target gene connectivity in resistant cells, which were characterized by an increased G1-phase fraction and a decreased S/G2M-phase fraction^[133]. Gene Ontology clusters related to cytoskeleton/mitosis, signaling, and metabolism were driven by distinct transcription factor combinations within the resistance networks^[133]. Based on these findings, the researchers successfully overcame single-agent resistance by combining trichostatin A with docetaxel^[133]. Similarly, Taavitsainen et al. combined scRNA-seq and scATAC-seq in enzalutamide (ENZ)-resistant CRPC and revealed extensive chromatin remodeling at open chromatin sites in post-treatment tumors, which initiated distinct transcriptional programs^[134]. Among these cells, an ENZ-induced cluster closely associated with resistance displayed activation of the mechanistic target of rapamycin complex 1 (mTORC1) pathway, enhanced expression of MYC target genes, and re-expression of cell cycle drivers, thus leading to a therapy escape state^[134]. As ENZ exposure extended to weeks or months, some clusters in the single-cell transcriptome began expressing neuroendocrine-like genes such as enhancer of zeste homolog 2 (EZH2), aurora kinase A (AURKA), paternally expressed 10 (PEG10), and SRY-box transcription factor 2 (SOX2), transitioning from androgen-dependent glandular-like cells to neuroendocrine-like lineages^[134]. Under ENZ stimulation, CRPC progressively evolved from an initial short-term, induced-resistance phenotype to a long-term, consolidated-resistance phenotype, significantly compromising therapeutic efficacy^[134].

Complementary to genetic profiling, single-cell epigenomics enables analysis of regulatory mechanisms associated with phenotypic plasticity and adaptive drug tolerance, beyond stable genetic alterations. Using scATAC-seq-based single-cell resolution determinations of chromatin accessibility, Satpathy et al. identified potential epigenomic drivers of T-cell exhaustion, a major cause of anti-programmed death-ligand 1 (PD-L1) immunotherapy failure^[135]. Similarly, Derrien et al. performed single nucleus ATAC-seq (snATAC-seq) and RNA sequencing (RNA-seq) and found that multiple myeloma could gain resistance to bispecific antibodies targeting GPRC5D through genetic or epigenetic silencing of the corresponding gene^[39]. Moreover, snATAC-seq also revealed that activation of AP-1 binding site–related motifs and enhancers, together with an increased proportion of mesenchymal cells, were primary drivers of drug resistance in recurrent GBM^[136].

The formation of 5-methylcytosine and its derivatives, hydroxymethylcytosine and formylcytosine, constitutes a major source of cancer epigenetic heterogeneity. Accumulation of 5-methylcytosine in CpG islands - promoter regions with high CpG content - inhibits gene expression. Similarly, accumulation within gene bodies or enhancers modulates cis-regulatory activity, leading to aberrant gene expression in cancer, with suppression of tumor suppressors, and activation of oncogenes^[137]. Several techniques are already available that use single-cell reduced-representation bisulfite sequencing (scRRBS-seq) to detect the methylation status of multiple CpG islands at single-cell level^[93-96]. Additionally, single cell CpG island sequencing (scCGI-seq), a methylation-detection method that utilizes restriction enzymes and is bisulfite-independent, has been developed and implemented^[97].

Beside DNA methylation and chromatin-accessibility regulation, histone modifications represent another major pathway of cellular epigenetic regulation^[27]. A broad diversity of histone modifications exists in tumor cells, making single-cell approaches uniquely advantageous^[138]. Single-cell chromatin immunoprecipitation sequencing (scCHIP-seq) provides valuable insights into single-cell level heterogeneity of histone modifications in tumor cells^[98]. For example, Grosselin et al. utilized scSHIP-seq to reveal that breast cancer acquired resistance to capecitabine and tamoxifen during chemotherapy through selective evolution of pre-existing drug-resistant subclones and modifications at H3K27me3 sites^[80].

Additionally, at the single-cell level, in situ cleavage (CUT&RUN) or tagging (cleavage under targets and tagmentation, CUT&Tag) techniques enable single-cell histone modification analysis with improved signal-to-noise ratios and higher throughput^[99,100,139]. Using CUT&Tag technology, specific cell populations can be detected based on chromatin states defined by histone modifications on adjacent nucleosomes^[138]. Using this technology in gliomas, Wu et al. identified a Verhaak_GBM_proneural gene set that was specifically suppressed by PcG-mediated H3K27me3 in drug-resistant tumor subclusters. These subclusters were selectively enriched during treatment, driving tumor evolution toward a more drug-resistant mesenchymal phenotype^[139].

New technologies such as cytometry by time-of-flight (CyTOF), ultrahigh-throughput combined tagmenting enrichment for multiple epigenetic proteins (uCoTarget), and multiplexing antibodies by barcode identification (MAbID) enable the simultaneous detection of multiple histone modification patterns within cells with higher efficiency^[101-103]. These approaches offer effective solutions for refined measurement and classification of tumor heterogeneity, as well as for investigating mechanisms underlying drug resistance.

Linking transcriptomic profiles to functional phenotypes requires single-cell proteomics, which enables detection of surface markers, signaling proteins, and post-transcriptional modifications involved in resistance mechanisms. In GBM, Pombo Antunes et al. combined cellular indexing of transcriptomes and epitopes by sequencing (CITE-seq) with scRNA-seq to systematically map spatiotemporal heterogeneity. They found that, following standard treatment with surgery plus radiotherapy and temozolomide (TMZ), the predominant population of microglia-derived TAMs (Mg-TAMs) was replaced by monocyte-derived TAMs (Mo-TAMs). This shift not only conferred TMZ resistance, but also established an innate tolerance foundation for resistance to PD-1/PD-L1 ICI^[140]. Similarly, the combined use of CITE-seq and scRNA-seq in metastatic melanoma revealed that spatiotemporal heterogeneity of tumor and stromal cells jointly promoted resistance during combined anti-PD-1 and anti-CTLA-4 immunotherapy. More specifically, microphthalmia-associated transcription factor (MITF) positive, secreted protein acidic and rich in cysteine-like 1 (SPARCL1) positive and centromere protein F (CENPF) positive tumor subclones emerged and their proportion increased, reducing antigen-processing and antigen-presentation pathways while exhibiting EMT features^[42]. Tian et al. utilized CITE-seq and revealed that pediatric GBM underwent selective amplification and immune escape under therapeutic pressure due to heterogeneous expression of tumor-associated antigens during chimeric antigen receptor T-cell (CAR-T) single-targeted therapy^[141]. This led to insufficient CAR-T persistence and increased exhaustion, ultimately resulting in treatment failure^[141], Cadot et al. applied CITE-seq to CLL treated with the Bruton’s tyrosine kinase (BTK) inhibitor ibrutinib to identify a pre-treatment-existing, progression-enriched subclone characterized by BCR activation, MYC target gene upregulation, and CXCR4 downregulation^[34]. The selective expansion of this subclone, coupled with BTK mutation accumulation and signaling pathway reprogramming, ultimately led to ibrutinib resistance^[34].

Single-cell metabolomic approaches rely primarily on mass spectrometry (MS) and facilitate characterization of tumor cell biochemical phenotypes, supporting the investigation of drug resistance mechanisms associated with metabolic reprogramming. Irinotecan (IRI) is widely used in chemotherapy regimens for metastatic colorectal cancer (mCRC), but its efficacy is limited by acquired resistance^[106]. Chen et al. analyzed the single-cell metabolomics of IRI-resistant HCT-116 cells using single-probe single-cell MS (SCMS)^[106]. They found that the IRI-resistant subpopulation displayed cancer stem cell-like characteristics and significant metabolic reprogramming, including markedly upregulated sphingomyelins, downregulated tricarboxylic acid-cycle intermediates, and substantially elevated lactic acid^[106,142]. Notably, lipid composition showed a marked increase in de novo fatty acid synthesis, a process directly correlated with fatty acid synthase (FASN) activity^[106]. Similarly, combined metformin and IRI treatment significantly reduced FASN activity, leading to substantial downregulation of lipids such as phosphatidylcholine (PC), phosphatidylserine (PS), phosphatidylinositol (PI) and triglycerides (TG), and multiple fatty acids (e.g., palmitic, oleic and stearic acids), thereby restoring anti-CRC activity^[106]. Sun et al. also applied SCMS to IRI-resistant CRC samples and identified lipid-metabolism reprogramming^[143]. In IRI-resistant cells, multiple unsaturated phosphatidylcholines [PC(33:4), PC(34:4), PC(38:5), PC(34:3), PC(36:3), PC(34:2), PC(34:1), PC(36:6), etc.] were significantly elevated^[143]. The total amounts of the major fatty acids C16:0, C16:1, C18:0, and C18:1 increased, and the ratios of monounsaturated to saturated fatty acids, including C16:1/C16:0 and C18:1/C18:0, were markedly elevated^[143]. Further studies revealed that these changes were associated with significant upregulation of stearoyl-CoA desaturase-1 (SCD1) mRNA and protein in drug-resistant cells, resulting in elevated synthesis of unsaturated fatty acids, altered membrane fluidity and topology, reduced drug-lipid interactions, and impaired drug entry or membrane-associated activity^[143].

Similarly, Zhu et al. employed surface-activated slow rate tandem charge single-cell MS (sSRTC-scM), a single-cell MS technique based on intact living-cell electrolaunching ionization MS (ILCEI-MS), to analyze the mechanisms underlying secondary resistance to the first-generation EGFR-TKI gefitinib in EGFR-mutant NSCLC^[107]. They found increased metabolic heterogeneity in PC9 cells during treatment, with gefitinib-resistant cells (PC9GR) showing lipid metabolism alterations centered on glycerophospholipid reprogramming, with significant increases in PC, PE, and sphingomyelin (SM)^[107]. These changes reflected alterations in membrane lipid composition, signaling lipids, and lipid synthase activity, which were closely correlated with the resistant phenotype and increased with the degree of resistance^[107].

Obtaining genotypes and gene expression patterns across distinct regions within an organ is crucial for identifying functional structures and their specialized roles within tumors. Unlike single-cell sequencing alone, integrating single-cell technologies with spatial omics enables more comprehensive analysis of cellular localization within tissues and the regulatory networks that shape cell–environment relationships^[91]. For these reasons, spatial transcriptomics was selected by Nature Methods as “Method of the Year 2020”^[172].

Kiviaho et al. used 10× Visium spatial transcriptomics data to successfully identify a subclone of club-like cells closely associated with resistance in CRPC^[173]. They found that this cell population highly expressed genes such as polymeric immunoglobulin receptor (PIGR), secretoglobin family 3A member 1 (SCGB3A1), mucin 1 (MUC1), lipocalin-2 (LCN2), CXCL1/2/8, and CCL20, significantly activating the IL6–Janus kinase (JAK)–STAT3, TNF–NF-κB, p53–SASP, the inflammatory response, and chemokine signaling pathways^[173]. As a result, these cells recruited and activated polymorphonuclear myeloid-derived suppressor cells, which expressed arginase-1 (ARG1), calprotectin (S100A8/A9), CXCR2, IL1B, IL10, TNF-alpha-induced protein 6 (TNFAIP6), and other molecules, creating an immunosuppressive microenvironment^[173]. Similarly, Mori et al. jointly employed 10× Visium spatial transcriptomics and snRNA-seq to identify a unique drug-resistant site in OCCC, which was formed by a group of resistance-associated tumor cells and CAFs^[46]. They further revealed a positive feedback signaling circuit within this site involving tumor cells and CAFs, which ultimately led to treatment failure with carboplatin^[46].

Romero et al. applied 10× Visium spatial transcriptomics combined with snRNA-seq to prostate cancer to study the lineage transdifferentiation from prostate adenocarcinoma (PRAD) to androgen receptor signaling inhibitor (ARSI)-resistant neuroendocrine prostate cancer (NEPC)^[174]. Their research revealed that the achaete-scute homolog 1 (ASCL1)-positive NEPC cell population originated from a neuroendocrine cell population within KRT8⁺ luminal progenitor cells, which were shown to be spatially dependent on the TME, thereby providing a potential spatial marker for therapeutic resistance^[174]. Rubinstein et al. utilized 10× Visium spatial transcriptomics to describe the spatiotemporal heterogeneity of tumor tissue across different stages of dabrafenib plus trametinib combination therapy for BRAF V600E-mutated melanoma^[175]. The study revealed remarkable spatiotemporal tumor heterogeneity during treatment. Spatially, a center-to-periphery pathway gradient was observed, with drug-resistant lineages located closer to the tumor margin^[175]. Temporally, programmed transcriptional alterations emerged early during treatment, including upregulation of aerobic oxidative phosphorylation and invasiveness, with concomitant cell-cycle suppression, and persisted stably in MRD, potentially contributing to chemotherapy resistance^[175].

Agostini et al. employed GeoMx digital spatial profiling (DSP) whole-transcriptome analysis to study IPMN, the precursor lesion of PDAC, and identified an expression gradient of the mucin O-glycosylation pathway from low-grade to high-grade IPMN^[43]. High-grade IPMN secreted increased O-glycosylated mucins with significantly upregulated glucosaminyl (N-acetyl) transferase 3 (GCNT3), thereby shaping a unique immunoevasive microenvironment that drove gemcitabine and nab-paclitaxel chemotherapy resistance^[43]. Shiau et al. analyzed the mechanisms of resistance to neoadjuvant chemotherapy [8-12 cycles of FOLFIRINOX followed by fractionated radiotherapy (30-50 Gy equivalent dose in 2 Gy fractions) concurrent with 5-FU or capecitabine] in PDAC samples using CosMx spatial molecular imager (SMI) single-cell spatial transcriptomics^[111]. Based on these data, the authors developed spatially constrained optimal transport interaction analysis (SCOTIA), an optimal transport model featuring a cost function that incorporates spatial distance and ligand-receptor gene expression, which revealed significant alterations in ligand-receptor interactions between CAFs and malignant cells under therapeutic stress^[111]. The analysis confirmed that pathways associated with chemokines, cytokines, matrix remodeling, and immune regulation were significantly enhanced^[111]. Notably, signaling within the IL-6 family, e.g., cardiotrophin-like cytokine factor 1 (CLCF1)–ciliary neurotrophic factor receptor (CNTFR) and leukemia inhibitory factor (LIF)–IL-6 cytokine family signal transducer (IL6ST), was enriched in the CAF-to-cancer-cell direction and was accompanied by JAK/STAT activation, ultimately conferring chemotherapy resistance to tumor cells^[111].

Spatial epigenomics is a major application of spatial omics technologies that enables analysis of the spatial organization of gene expression and associated epigenetic modifications across tissues at varying resolutions^[91]. As previously mentioned, detecting epigenetic modifications at single-cell level plays a crucial role in the identification of tumor resistance mechanisms and the development of therapeutic approaches. Spatially resolving epigenomics in situ within tissue sections not only provides high resolution but also yields spatial distribution information. Additionally, combining multiple tissue sections can fully reveal the spatiotemporal tumor heterogeneity caused by epigenetic modifications. A growing number of technologies have recently emerged to enable spatially resolved analysis of epigenetic modifications. Although these approaches have not yet been applied to the study of tumor drug resistance, exploring spatial epigenetic heterogeneity may provide valuable insights for future research in this area^[55].

While the analysis of protein expression and spatial distribution using traditional staining and immunofluorescence has long been a central approach for characterizing tissue architecture, technical limitations such as spectral overlap have constrained high-throughput analyses of protein localization^[91]. Recently developed technologies are capable of simultaneously detecting multiple proteins, enabling spatial proteomics detection and analysis.

Kulasinghe et al. performed proteomic analysis of TNBC tissues treated with adjuvant chemotherapy comprising 5-FU, epirubicin, and cyclophosphamide using Nanostring GeoMx DSP^[176]. The authors identified numerous protein markers associated with tumor response to chemotherapy and their underlying mechanisms. In tumors, granzyme A (GZMA), stimulator of interferon response cGAMP interactor (STING), and fibronectin were positively correlated with chemotherapy response, whereas CD80 was negatively correlated. In the tumor stroma, estrogen receptor alpha (ER-α) expression was positively correlated with chemotherapy response, whereas 4-1BB and melanoma antigen recognized by T-cells-1 (MART1) were negatively correlated. Similarly, Li et al. utilized RNA-seq combined with fluorescence in situ hybridization (FISH)/immune histochemistry (IHC) and DSP in a Phase II neoadjuvant clinical trial for HER2-positive breast cancer to characterize treatment resistance arising from HER2 and Erb-B2 receptor tyrosine kinase 2 (ERBB2) expression heterogeneity within tumor tissues^[177]. Spatial-CITE-seq extends the CITE-seq technology to individual spatial spots within tissue sections. Although it has not yet been applied to drug resistance-related research, this technology holds considerable promise in such studies^[118,178].

Bouchard et al. analyzed the spatial heterogeneity in LUAD treated with erlotinib using spatial multiplexed immunofluorescence on the PhenoCycler. This study revealed that CAF-mediated spatial reorganization and tumor–stroma alignment could persist or regenerate under EGFR inhibition and were correlated with drug resistance^[179]. Jhaveri et al. analyzed the spatial proteomics of HNSCC tissues treated with pembrolizumab using co-detection by indexing (CODEX) with a 101-antibody panel^[180]. The study identified four metabolically distinct tumor regions and six spatially distinct neighborhoods enriched in different immune subpopulations. It was also found that M2 macrophages localized closer to tumor cells in glucose-6-phosphate dehydrogenase (G6PD)- and MMP9-overexpressing regions, where they exhibited immunosuppressive activity, providing new insights into TME features associated with response and sensitivity to ICI therapy^[180].

A study by Fan et al. employed laser capture microdissection (LCM)-MS-based spatial proteomics together with scRNA-seq to analyze cervical squamous cell carcinoma (CSCC) with poor response to ICI^[120]. They identified bidirectional interactions between epithelial-cytokeratin tumor state malignant epithelial cells and mannose-6-phosphate (MP6) positive and MMP11⁺ CAFs in key non-responsive regions, which formed an immune-rejecting microenvironment through FABP5-mediated TGF-β signaling^[120]. Spatially, MP6 co-localized with immune-depleted and CAF-enriched regions, while TGF-β activity showed peripheral ring-like enrichment along MP6 margins, thus defining a unique resistance niche^[120].

Spatial metabolomics supports the characterization of histological and functional properties of intact tissue sections and facilitates comprehensive analysis of the metabolic profiles of tumor cells and stromal regions^[181]. In HER2-positive advanced gastric cancer treated with trastuzumab, Wang et al. analyzed the spatial metabolome of pre-treatment biopsy samples from 42 patients using matrix-assisted laser desorption/ionization and Fourier transform ion cyclotron resonance imaging MS (MALDI-FT-ICR IMS) and identified extensive metabolic heterogeneity within tumor tissues^[182]. Quantification based on Simpson’s diversity index revealed a significant correlation between high metabolic heterogeneity and increased sensitivity to trastuzumab^[182]. Clustering based on metabolic status showed that multiple pathways, including nucleotide, carbohydrate, and amino acid metabolism, were collectively downregulated in the subclusters most strongly associated with drug resistance^[182]. Conversely, carbohydrate and amino acid metabolism were upregulated in the subclusters exhibiting the highest sensitivity to trastuzumab^[182]. In NSCLC, Shen et al. analyzed the metabolic status of platinum-resistant samples following neoadjuvant chemotherapy using MALDI-FT-ICR MS^[183]. They found that elevated levels of PC typically correlated with energy enrichment, enhanced proliferation, and poor prognosis in tumor regions, whereas in the stroma, glycerophospholipid metabolism exerted the strongest influence on prognosis and chemotherapy response^[183]. Although this study did not deeply explore the mechanisms linking altered metabolic states to drug resistance, the tumor and stroma classifiers developed based on metabolic signatures still presented high accuracy and practical value in patient risk prediction^[183].

Ji et al. utilized spatial metabolomics coupled with spatial transcriptomics to analyze key metabolic differences between treatment-sensitive and treatment-resistant samples from nasopharyngeal carcinoma (NPC) patients receiving cisplatin-based chemoradiotherapy and PD-1 immunotherapy^[184]. The authors observed elevated activity in branched-chain amino acid, glutamine, and fatty acid metabolism in treatment-sensitive samples, whereas the same metabolic pathways exhibited significantly reduced activity in treatment-resistant samples^[184]. Additionally, the authors identified spatial microdomains within the same NPC sample (tumor, fibroblast, immune, tumor-immune mixed, and normal epithelial zones) and significant variations in branched-chain amino acid-, glutamine-, and lipid-related metabolites across regions^[184]. Moreover, the spatial distribution patterns of metabolites differed between treatment-sensitive and treatment-resistant patients within these zones, suggesting widespread metabolic heterogeneity within tumors that directly correlated with treatment sensitivity^[184].