Construction of a Glutathione‐Gated Intelligent Nanoclassifier for Spatioselective Visualization of Dual CircRNAs and Synergistic Photodynamic Therapy

ABSTRACT

Circular RNAs (circRNAs) are key endogenous regulators of tumorigenesis and progression, and their high interspecies diversity and strong sequence homology pose a great challenge for simultaneous, live‐cell differentiation of multiple circRNA variants. Herein, we construct an intelligent nanoclassifier (DB‐CHA@PAN) by assembling disulfide bond (DB)‐modified catalytic hairpin assembly (CHA) probes onto programmable cruciform framework‐based nanoparticles (PAN) for spatioselective visualization of dual circRNAs and synergistic photodynamic therapy. Upon entering cancer cells, GSH cleaves disulfide bonds to initiate multiple rounds of circRNAs‐fueled cyclic CHA cascades, inducing spatial separation of photosensitizer (PS)/BHQ3 and Cy3/BHQ2 pairs and consequently recovery of PS photodynamic activity and Cy3 fluorescence signal. Released PS can generate abundant singlet oxygen (¹O₂) upon light irradiation, inducing oxidative damage and apoptosis. This intelligent nanoclassifier enables attomolar‐level detection of circCDYL and circHIPK3 in vitro and simultaneous imaging of circCDYL and circHIPK3 in living cells. It can quantify circRNA levels at single‐cell sensitivity, discriminate circRNAs from mismatched variants with single‐base resolution, and even diagnose breast/lung cancer across entire clinical spectrum with 100% accuracy. Moreover, it can real‐time track circRNAs dynamics in living cells and MCF‐7 tumor‐bearing nude mice, and significantly enhance therapeutic efficacy via synergistic photodynamic activation, with promising applications in clinical diagnostics and therapeutics.

We construct an intelligent nanoclassifier by assembling disulfide bond‐modified catalytic hairpin assembly probes onto programmable cruciform framework‐based nanoparticles for spatioselective visualization of dual circRNAs and synergistic photodynamic therapy. This intelligent nanoclassifier not only demonstrates a programmable and cell‐selective approach to enhance PDT efficacy but also opens a new avenue for the application of nucleic acid‐based molecular machinery in precision cancer theranostics.

1. Introduction

CircRNAs are a novel class of covalently closed non‐coding RNA molecules characterized by the absence of free 5'‐cap and 3'‐tail [1, 2, 3]. They are derived from precursor mRNAs (pre‐mRNAs) through “back‐splicing,” a non‐canonical splicing mechanism that ligates a downstream 5' splice donor to an upstream 3' splice acceptor, thereby forming a circular structure [4, 5, 6]. CircRNAs exert diverse biological functions [7, 8], such as interacting with proteins [9], participating in protein translation [10], acting as miRNA sponges to adsorb microRNAs and regulate gene expression [11, 12], and inhibiting host gene expression by competing with canonical pre‐mRNA splicing [13]. Furthermore, circRNAs exhibit distinct expression patterns in tissues, cell types, and developmental stages and significant diversity across species [14, 15]. Aberrant circRNA expression is closely linked to the occurrence and progression of various diseases (e.g., cardiovascular, neurological, and oncological diseases) [16, 17] and cancers (e.g., gastric, lung, breast, colorectal, bladder, esophageal, and osteosarcoma cancers) [18]. Notably, the covalently closed structure of circRNAs confers strong resistance to ribonuclease (RNase) degradation, rendering them highly stable even in complex biological environments [19]. Consequently, circRNAs have emerged as promising cancer biomarkers, and their effective monitoring holds substantial importance for both disease diagnosis and therapy development [20].

Conventional approaches for circRNA detection primarily include Northern blotting [21, 22], circRNA sequencing (Circ‐seq) [23, 24], and quantitative reverse transcription‐polymerase chain reaction (qRT‐PCR) [25, 26]. Northern blotting distinguishes individual circRNAs based on electrophoretic mobility, but it suffers from radioactive labeling hazard, high sample consumption, and inherently non‐quantitative results [22, 27]. Circ‐seq facilitates high‐throughput screening of circRNAs, but it is hampered by high cost, specialized instrumentation, multi‐step workflow, and computationally intensive data analysis [28]. QRT‐PCR offers superior sensitivity, but its accuracy for circRNAs is compromised by the inherently circular conformation, which frequently introduces reverse transcription artifacts and consequent quantification biases [26]. To address these limitations, fluorescent methods based on isothermal nucleic acid amplification have been developed, such as reverse transcription‐rolling circle amplification (RT‐RCA) [29], loop‐mediated isothermal amplification (LAMP) [30], T7 nuclease‐assisted cycling signal amplification [31], and CRISPR–Cas12a signal amplification [32]. Despite improved accuracy, these methods depend on intricate RNA extraction, stringent temperature control and multiple exogenous components (e.g., primers, templates and enzymes), preventing their application for tracking dynamic changes of circRNAs within living cells. Due to the interspecies diversity, high sequence homology, and inherent expression heterogeneity [33], most of existing methods are limited to detecting a single type of circRNA, and thus fail to provide reliable and comprehensive diagnostic information in clinical applications. Owing to the conformationally restricted ribose moiety, fluorescent locked nucleic acids (LNA) probes have been applied for spatiotemporal imaging of endogenous microRNAs in living cells [34], but their efficacy relies on the transfection reagents with potential cytotoxicity, variable cell‐type delivery, and poor in vivo/clinical suitability. Organic, inorganic and carbon‐based nanomaterials have been explored as the versatile reaction probes/drug carriers with improved cellular delivery efficiency [35, 36, 37], but they are hindered by dose‐dependent cytotoxicity (e.g., accumulation, oxidation, and inflammation) and imprecise surface functionalization (e.g., batch inconsistency, altered biodistribution, and unpredictable bio‐interactions) [38]. Therefore, the development of well‐tolerated and robust methods capable of multiple circRNAs detection and imaging‐guided therapy remains a formidable challenge.

Characterized by nanoscale dimension, structural stability, excellent biocompatibility, and unparalleled programmability [39, 40], DNA nanostructures exhibit precise regulation of the spacing, valency, and spatial distributions of nucleic acids, peptides, and proteins at the molecular level [41, 42], with promising applications in the fields of molecular imaging [43], drug delivery [44], and tumor therapy [45]. Among them, cruciform DNA nanostructures have attracted increasing attentions in RNA detection [46, 47], cruciform nanostructure platform possesses several key characteristics: i) the cruciform structure exhibits restricted mobility within the confined cellular environment, making it highly suitable for in‐situ circRNA imaging [48]; ii) the modular “arms” of the cruciform structure can be programmably engineered with diverse functional modules, suitable for multi‐analyte detection [49]; iii) the stable and flexible architecture of cruciform DNA facilitates signal amplification through diverse mechanisms, significantly improving the detection sensitivity [46]. Because most biomarkers are present in both normal and tumor cells, cruciform structure‐based probes are susceptible to nonspecific activation, which severely compromises the detection specificity and accuracy. Photodynamic therapy (PDT) has the advantages of minimal invasiveness, rapid therapeutic efficacy, and drug resistance evasion, and it can induce oxidative damage and apoptosis in cancer cells by producing cytotoxic reactive oxygen species (ROS) and consuming oxygen (O₂) under laser irradiation [50, 51, 52, 53], but the widespread application of PDT faces the challenges of inherent hydrophobicity, poor targetability, and nonspecific activation of photosensitizers [54]. The tumor microenvironment (TME), driven by the fast proliferation and metabolism of tumors, exhibits weak acidity, reactive oxygen species, high‐level GSH, and overexpressed proteins (e.g., nucleolin) [55], and it has been widely exploited as an endogenous stimulus to control the targeting of tumors and the release of theranostic agents. Herein, we construct an endogenous GSH‐gated intelligent nanoclassifier (DB‐CHA@PAN) by assembling DB/PS‐embedded CHA probes onto cruciform DNA framework‐based programmable atom‐like nanoparticles (PAN) for spatioselective visualization of dual circRNAs and synergistic photodynamic therapy. Programmable atom‐like nanoparticles (PAN) are programmable cross‐shaped framework nanoparticles used to load CHA probes. Integrating target recognition, stimulus‐responsive activation, signal amplification, and photodynamic therapy, this all‐in‐one nanoplatform enables simultaneous circRNAs detection and tumor treatment, serving as an “intelligent nanoclassifier” for precise cancer cell identification and accurate staging classification of breast and lung cancers. This nanoclassifier can detect attomolar circCDYL and circHIPK3 simultaneously, distinguish circRNAs from mismatched variants with single‐base resolution, quantify circRNAs at single‐cell sensitivity, and even diagnose breast/lung cancer from stage I to IV with high accuracy. Moreover, it can real‐time track circRNAs dynamics in living cells and MCF‐7 tumor‐bearing nude mice, visually discriminate normal from malignant cells, and enhance tumor specificity and therapeutic efficacy via synergistic photodynamic activation.

2. Results and Discussion

2.1. Construction and Mechanism of Endogenous GSH‐Gated Intelligent Nanoclassifier

We construct an endogenous GSH‐gated intelligent nanoclassifier (DB‐CHA@PAN) for spatioselective visualization of dual circRNAs and synergistic photodynamic therapy based on assembling DB/PS‐embedded CHA probes onto cruciform DNA framework‐based PAN (Figure 1A). We selected circCDYL and circHIPK3 as the model targets because they are highly expressed in various tumor tissues and have been implicated in tumor progression and invasion [38, 56, 57]. In this nanoclassifier, PAN structure is self‐assembled from four single‐stranded DNAs (T₁, T₂, T₃, and T₄) via Watson–Crick base pairing (Figure S1). Serving as a delivery vehicle for the CHA probes, PAN not only markedly enhances the stability of imaging probes in physiological environments but also promotes efficient intracellular delivery of DNA probes [38, 46]. Furthermore, four CHA probes (H₁, H₂, H₃, and H₄) are ingeniously designed. H₁ and H₃ are GSH‐responsive molecular probes, with each comprising four functional regions (I–IV). Functional region I is a GSH‐responsive domain for recognizing disulfide bonds (DB); functional region II is a recognition domain for identifying circRNAs; functional region III is a structural scaffold for DB‐CHA@PAN assembly; and functional region IV is a binding domain for reporter molecular probes. H₂ and H₄ are reporter molecular probes, with each containing three functional regions (I–III). Functional region I is a structural scaffold for DB‐CHA@PAN assembly; functional region II is a binding domain for GSH‐responsive molecular probes (H₁ and H₃); and functional region III is a reporting domain for signal output (methylene blue, a widely used photosensitizer (PS) with photodynamic activity quenched by BHQ3 [58]; Cy3, a common fluorescent dye with fluorescent signal quenched by BHQ2). Notably, the anti‐nucleolin AS1411 aptamer is incorporated at the 5′ end of T₁ in PAN, facilitating the efficient intracellular delivery of DB‐CHA@PAN (Figure S1) [47].

FIGURE 1.

Workflow diagram of GSH‐gated intelligent nanoclassifier for spatioselective visualization of dual circRNAs and synergistic photodynamic therapy. An endogenous GSH‐gated intelligent nanoclassifier (DB‐CHA@PAN) is formed through the self‐assembly of disulfide bond (DB)‐modified catalytic hairpin assembly (CHA) probes onto cruciform DNA framework‐based programmable atom‐like nanoparticles (PAN).

As depicted in Figure 1B,C, upon encountering target cancer cells, the anti‐nucleolin AS1411 aptamer (orange color, Figure S1) in PAN interacts with nucleolin, a major receptor highly overexpressed in cancer cells [59]. Following this initial binding, the DB‐CHA@PAN complex is internalized into the cells via nucleolin receptor‐mediated endocytosis. In normal physiological conditions, the toehold regions in H₁ and H₃ are locked in the stem structures to prevent nonspecific triggering reactions. While in tumor microenvironments, the disulfide bonds on H₁ and H₃ can be cleaved by GSH (high‐level in cancer microenvironment [55]) (Figure S2) [60, 61] to dissociate short single‐stranded fragments (a*/c*), forming the H_1a/H_3a structures and leading to the exposure of the toehold regions (a/c). The target circCDYL binds to the toehold region (a) of the metastable probe H_1a, and concurrently, the target circHIPK3 binds to the toehold region (c) of the metastable probe H_3a, both triggering branch migration reactions to form the circCDYL/H_1a duplex with an overhang (b) and the circHIPK3/H_3a duplex with an overhang (d), respectively. The exposed b/d overhang then acts as a toehold to initiate strand migration of H₂/H₄, leading to the formation of the H_1a–H₂ complex/H_3a–H₄ complex and the activation of the CHA modules for circCDYL (CHA‐C) (Figure S3) and circHIPK3 (CHA‐H) (Figure S4). The CHA reactions spatially separate the PS/BHQ3 and Cy3/BHQ2 pairs, consequently restoring the photodynamic activity of PS and the fluorescence signal of Cy3. Meanwhile, the hybridized circCDYL/circHIPK3 is stripped off during the CHA‐C/CHA‐H reactions, which in turn triggers the next round of CHA, continuously driving the dissociation of the PS/BHQ3 and Cy3/BHQ2 pairs. Through multiple rounds of the “assembly‐release” cycling, large amounts of PS and Cy3 are ultimately accumulated, achieving simultaneous and highly efficient imaging of circCDYL and circHIPK3 in living cells (Figures S3 and S4). Importantly, the cyclic release of the PS from H₁ enables the abundant generation of cytotoxic singlet oxygen (¹O₂) upon light irradiation (e.g., 660 nm), achieving amplified PDT effects that induce oxidative damage and apoptosis in cancer cells [58]. Engineered with computational AND‐gate logic, this intelligent nanoclassifier requires sequential input of specific intracellular “keys” (AS1411, GSH, circCDYL, and circHIPK3), functioning as a molecular circuit with inherent error‐checking. This built‐in, multi‐stage unlocking mechanism ensures activation exclusively within target cells, drastically enhancing therapeutic precision and minimizing collateral damage.

2.2. In Vitro Feasibility of the GSH‐Gated DB‐CHA@PAN for Dual CircRNAs Detection

We first substituted the photosensitizer (PS) in H₁ with Cy5, as Cy5 is more suitable for detecting low‐abundance biomolecules and bioimaging due to its high sensitivity, low background, and photostability [58], whereas PS is primarily utilized for photodynamic therapy. To simulate endogenous GSH‐gated intelligent nanoclassifier responsive to circRNAs, we synthesized covalently closed circCDYL and circHIPK3 using 5′‐phosphorylated linear RNA as the original material with the assistance of T4 RNA ligase 2 and guide RNAs (Figure S5). To explore the stepwise self‐assembly of DB‐CHA@PAN, 12% non‐denaturing polyacrylamide gel electrophoresis (PAGE) was used to analyze reaction products (Figures 2A and S6). Upon sequential addition of T₁, T₂, T₃, and T₄, the electrophoretic mobility of the bands gradually decreases (Figure 2A, lanes 1–4), and a clear band with the slowest migration rate is observed in lane 4, indicating the successful assembly of PAN. After introducing the catalytic hairpin assembly probes (H₁, H₂, H₃, and H₄), a bright band with a slower migration rate appears in lane 5 compared to that in lane 4 (Figure 2A), confirming the successful assembly of DB‐CHA@PAN. When circCDYL + circHIPK3 + GSH + DB‐CHA@PAN are present, an ultra‐bright merged yellow band (Figure 2A, lane 6; Figure S6D, lane 6) from the co‐localization of red (Cy5, Figure S6B) signals and green (Cy3, Figure S6C) signals is observed, indicating that both circCDYL‐mediated CHA‐C and circHIPK3‐mediated CHA‐H modules are successfully activated, leading to the restoration of the Cy5/Cy3 fluorescence signals (Figure 1). We further verified the specific initiation of the CHA cycling reaction by target circRNA using 20% non‐denaturing PAGE. When GSH + circCDYL + H₁ + H₂ are present, the bands of H₁ (Figure 2B, lane 1) and H₂ (Figure 2B, lane 2) disappear, while a new red band corresponding to the H_1a‐H₂ complex emerges (Figure 2B, lane 3), confirming that circCDYL can hybridize with H_1a (generated by GSH cleavage of H₁) to initiate the strand migration of H₂, successfully activating the CHA‐C cycling reaction. Similarly, when GSH + circHIPK3 + H₃ + H₄ are present, the bands of H₃ (Figure 2C, lane 1) and H₄ (Figure 2C, lane 2) disappear, with the concomitant appearance of a new green band corresponding to the H_3a‐H₄ complex (Figure 2C, lane 3), demonstrating that circHIPK3 can also effectively activate the CHA‐H cycling reaction. Atomic force microscopy and dynamic light scattering measurements further confirm the successful construction of DB‐CHA@PAN (Figure S7). We employed single‐molecule fluorescence imaging to characterize circRNAs with single‐molecule precision. In the absence of circCDYL and circHIPK3, neither Cy5 spots (Figure 2D, I) nor Cy3 spots (Figures 2D, II) is observed, indicating that DB‐CHA@PAN cannot be activated without its specific targets. Conversely, when circCDYL and circHIPK3 are present, distinct Cy5 spots (Figure 2D, III) and Cy3 spots (Figure 2D, IV) are simultaneously detected, conveying that DB‐CHA@PAN can be successfully and specifically activated for dual‐target imaging at the single‐molecule level. Notably, the Cy5 (Figures 2D, III) and Cy3 (Figure 2D, IV) spots exhibit single‐step photobleaching (Figure S8), validating that each spot represents an individual Cy5/Cy3 molecule.

FIGURE 2.

(A) 12% non‐denaturing PAGE analysis of the stepwise self‐assembly and activation of DB‐CHA@PAN. (B) Verification of circCDYL‐induced CHA‐C cycling reaction by 20% non‐denaturing PAGE. (C) Verification of circHIPK3‐induced CHA‐H cycling reaction by 20% non‐denaturing PAGE. (D) Single‐molecule imaging without circCDYL (I) and circHIPK3 (II), and with circCDYL (III) and circHIPK3 (IV). Scale bar: 5.5 µm. (E) Fluorescence spectral analysis of DB‐CHA@PAN and TDB‐CHA@PAN in response to circRNAs without GSH. (F) Fluorescence counts induced by DB‐CHA@PAN in response to GSH, circCDYL, or circHIPK3. (G) Fluorescence counts generated by nDB‐CHA@PAN, DB‐mCHA₁@PAN, DB‐mCHA₂@PAN, and DB‐CHA@PAN in response to GSH or circCDYL. (H) Fluorescence counts generated by nDB‐CHA@PAN, DB‐mCHA₁@PAN, DB‐mCHA₂@PAN, and DB‐CHA@PAN in response to GSH or circHIPK3. (I) Fluorescence spectra in response to circRNAs (100 nM) with various concentrations of GSH. (J) Cy5/Cy3 counts induced by the control, UDG, RNase H, APE1, and GSH. (K) Linear relationship between counts and the logarithm of circRNA concentration. (L) Radar diagrams of counts generated by the control, piRNA‐36026, MALAT1, HOTAIR, circMTO1, circCDYL, circHIPK3, and circCDYL + circHIPK3. The concentration of each RNA is 100 nM. (M) Degradation rate of DPBF after incubation with DB‐CHA@PAN in the presence of GSH, circCDYL, and circHIPK3 under 660‐nm light irradiation. ***P< 0.001. Error bars denote mean ± SD (n = 3).

We further conducted fluorescence spectroscopy measurements to evaluate the feasibility of this intelligent nanoclassifier (DB‐CHA@PAN) for simultaneous profiling of dual circRNAs. For comparison, a control nanoclassifier (TDB‐CHA@PAN) was designed with a traditional continuously active CHA probe (H₁ without the blocking strand a* and disulfide bond; H₃ without the blocking strand c* and disulfide bond). Upon the addition of circCDYL and circHIPK3, TDB‐CHA@PAN exhibits significant Cy5 and Cy3 fluorescence enhancement (Figure 2E). In contrast, DB‐CHA@PAN displays negligible Cy5 and Cy3 fluorescence signals in the presence of both targets, indicating that the designed endogenous GSH‐gated module effectively suppresses the non‐specific CHA cycling reaction. Furthermore, distinct Cy5 and Cy3 fluorescence signals are exclusively observed when all three inputs (GSH, circCDYL, and circHIPK3) are present concurrently (Figure 2F), suggesting the stringent AND‐gate logic of the GSH‐gated CHA system. This AND‐gate dependency yields a dramatic signal enhancement, with Cy5 and Cy3 counts being 17.77‐fold and 16.51‐fold higher, respectively, than those in control groups without target circRNAs (Figure 2F). These results reveal the successful application of this nanoclassifier for the simultaneous detection of dual circRNAs. Moreover, another three negative control nanoclassifiers were designed (Figure S9), one nanoclassifier lacking the disulfide bonds in H₁ and H₃ (termed nDB‐CHA@PAN), one nanoclassifier with mutated bases in the circCDYL‐recognizing sequence of H₁ (termed DB‐mCHA₁@PAN), and one nanoclassifier with mutated bases in the circHIPK3‐recognizing sequence of H₃ (termed DB‐mCHA₂@PAN). nDB‐CHA@PAN displays no significant Cy5 and Cy3 signals even with GSH and targets present (Figure 2G–H), indicating that the disulfide bond is the core switch for the activation of the endogenous GSH‐gated nanoclassifier. DB‐mCHA₁@PAN and DB‐mCHA₂@PAN show only minimal Cy5 and Cy3 signals in response to circCDYL (Figure 2G) and circHIPK3 (Figure 2H), respectively, implying that the respective fluorescence signals originate specifically from their corresponding circRNA‐induced CHA reactions.

Guided by the pivotal role of GSH as the molecular initiator, we identified 8 mM as the optimal concentration based on the signal plateau of both Cy5 and Cy3 (Figure 2I). We then evaluated the specificity of DB‐CHA@PAN for GSH. As shown in Figure 2J, only GSH induces a significant increase in both Cy5 and Cy3 counts, which is clearly distinguishable from those induced by interfering enzymes (i.e., APE1, RNase H, and UDG) and the control, suggesting the high specificity towards GSH. Furthermore, under the optimized experimental parameters (Figure S10), the Cy5 and Cy3 counts exhibit a linear correlation with the logarithmic concentrations of circCDYL and circHIPK3, respectively, across a broad range from 10⁻¹⁶ M to 10⁻⁷ M (Figure 2K). The corresponding regression equations are N = 571.53 + 32.95 log₁₀ C (R ² = 0.9998) for circCDYL and N = 471.78 + 27.20 log₁₀ C (R ² = 0.9999) for circHIPK3. The calculated limits of detection (LOD) are 18.6 aM for circCDYL and 24.5 aM for circHIPK3 based on the mean signal of the control plus 3‐fold the standard deviation, which are superior to those of fluorescence spectroscopy (Figure S11, LOD: 110 aM for circCDYL, 126 aM for circHIPK3). To assess the specificity of the DB‐CHA@PAN for circCDYL and circHIPK3, we employed circMTO1, piRNA‐36026, MALAT1, and HOTAIR as interfering RNAs (Figure 2L). The circCDYL and circHIPK3 induce robust Cy5 and Cy3 signals, whereas the interfering RNAs yield only background‐level signals (Figure 2L), underscoring the exceptional selectivity of the nanoclassifier toward circCDYL and circHIPK3. We next evaluated the photosensitizing capability of DB‐CHA@PAN by monitoring singlet oxygen (¹O₂) generation under 660‐nm light irradiation, using commercially available 1,3‐diphenylisobenzofuran (DPBF) for photooxidation analysis. In the presence of GSH, circCDYL, and circHIPK3, DB‐CHA@PAN exhibits the time‐dependent decrease in DPBF absorption (Figure 2M), signifying DPBF decomposition by singlet oxygen ¹O₂ photogenerated from DB‐CHA@PAN.

2.3. Thermodynamics and Kinetics Analysis of DB‐CHA@PAN

We evaluated the standard Gibbs free energy change (ΔG) and the theoretical conversion efficiency for the DB‐CHA activation process. The photosensitizer in H₁ was first replaced with Cy5. CHA is initiated via a toehold‐driven strand displacement process. Theoretically, H₁ and H₃ should have dual characteristics of stability for resisting nonspecific reaction and flexibility for priming subsequent amplification cycles. The Gibbs free energy change (ΔG) of this activation process is calculated through the two steps: (1) Hybridization of circCDYL with either H₁ or H_1a and circHIPK3 with either H₃ or H_3a (Figure 3Aa–d); (2) hybridization of T₁‐H_1a with H₂ as well as T_2–H_3a with H₄ (Figure 3Ae,f). The ΔG of each reaction is calculated using NUPACK (http://www.nupack.org/) and is listed in Table 1. The ΔG of the whole activation reaction is calculated based on Hess's law.

$$\mathrm{\Delta }{\mathrm{G}}_1\left(\mathrm{circCDYL}\right)=\mathrm{\Delta }{\mathrm{G}}_{1-2}+\mathrm{\Delta }{\mathrm{G}}_{1-3}$$ (1)

$$\mathrm{\Delta }{\mathrm{G}}_2\left(\mathrm{circHIPK}3\right)=\mathrm{\Delta }{\mathrm{G}}_{2-2}+\mathrm{\Delta }{\mathrm{G}}_{2-3}$$ (2)

FIGURE 3.

(A) (a and b) Procedural diagram of the hybridization of circCDYL (T₁) with H₁ or H_1a. (c and d) Procedural diagram of the hybridization of circHIPK3 (T₂) with H₃ or H_3a. (e) Procedural diagram of the hybridization of T₁‐H_1a with H₂. (f) Procedural diagram of the hybridization of T₂‐H_3a with H₄. (B) Workflow diagram of nT‐DB‐CHA@PAN for detecting dual circRNAs. (C) Real‐time monitoring of the Cy5 fluorescence signal produced by nT‐DB‐CHA@PAN + circCDYL (red curve) and nT‐DB‐CHA@PAN (blue curve), respectively. (D) Real‐time tracking of the Cy3 fluorescence signal produced by nT‐DB‐CHA@PAN + circHIPK3 (green curve) and nT‐DB‐CHA@PAN (orange curve), respectively. (E) Single‐molecule imaging of nT‐DB‐CHA@PAN in response to GSH, circCDYL, and circHIPK3. (F) Cy5/Cy3 counts generated by nT‐DB‐CHA@PAN in response to GSH, circCDYL, and circHIPK3. (G) Workflow diagram of DB‐CHA@PAN for detecting dual circRNAs. (H) Real‐time monitoring of Cy5 fluorescence signal produced by DB‐CHA@PAN + circCDYL (red curve) and DB‐CHA@PAN (blue curve), respectively. (I) Real‐time monitoring of Cy3 fluorescence signal produced by DB‐CHA@PAN + circHIPK3 (green curve) and DB‐CHA@PAN (orange curve), respectively. (J) Single‐molecule imaging of DB‐CHA@PAN in response to GSH, circCDYL, and circHIPK3. (K) Cy5/Cy3 counts generated by DB‐CHA@PAN in response to GSH, circCDYL, and circHIPK3. Scale bar is 5.5 µm. Error bars denote mean ± SD (n = 3).

TABLE 1.

ΔG of each decomposition reaction.

	circCDYL		circHIPK3
ΔG1‐1 /kcal mol−1	> 0	ΔG2‐1 /kcal mol−1	> 0
ΔG1‐2 /kcal mol−1	−13.72	ΔG2‐2 /kcal mol−1	−12.97
ΔG1‐3 /kcal mol−1	−5.1	ΔG2‐3 /kcal mol−1	−2.62
ΔG1 (circCDYL) /kcal mol−1	−18.82	ΔG2 (circHIPK3) /kcal mol−1	−15.59

Notably, the ΔG_1‐1 of the reaction process between circCDYL and H₁ as well as the ΔG_2‐1 of the reaction process between circHIPK3 and H₃ are both positive, indicating that both H₁ and H₃ maintain their initial self‐complementary state. However, the ΔG₁ (circCDYL) (–18.82 kcal mol⁻¹) and the ΔG₂ (circHIPK3) (–15.59 kcal mol⁻¹) are both negative, indicating that circCDYL and circHIPK3 initiate the CHA‐C and CHA‐H cycling reactions in the H_1a and H_3a systems, respectively. The theoretical conversion efficiency (E) of CHA is estimated by using the aforementioned thermodynamic parameters to Equations ((3), (4)).

$$\begin{array}{cc}& \mathrm{\Delta }{\mathrm{G}}_1\left(\mathrm{circCDYL}\right)=\mathrm{\Delta }\mathrm{G}\left({\mathrm{T}}_1\right)+\mathrm{\Delta }\mathrm{G}\left({\mathrm{H}}_{1\mathrm{a}}-{\mathrm{H}}_2\right)-\mathrm{\Delta }\mathrm{G}\left({\mathrm{H}}_{1\mathrm{a}}\right)\hfill \\ & -\mathrm{\Delta }\mathrm{G}\left({\mathrm{H}}_2\right)-\mathrm{\Delta }\mathrm{G}\left({\mathrm{T}}_1\right)=\mathrm{\Delta }\mathrm{G}\left({\mathrm{H}}_{1\mathrm{a}}-{\mathrm{H}}_2\right)-\mathrm{\Delta }\mathrm{G}\left({\mathrm{H}}_{1\mathrm{a}}\right)-\mathrm{\Delta }\mathrm{G}\left({\mathrm{H}}_2\right)\hfill \end{array}$$ (3)

$$\begin{array}{cc}& \mathrm{\Delta }{\mathrm{G}}_2\left(\mathrm{circHIPK}3\right)=\mathrm{\Delta }\mathrm{G}\left({\mathrm{T}}_2\right)+\mathrm{\Delta }\mathrm{G}\left({\mathrm{H}}_{3\mathrm{a}}-{\mathrm{H}}_4\right)-\mathrm{\Delta }\mathrm{G}\left({\mathrm{H}}_{3\mathrm{a}}\right)\hfill \\ & -\mathrm{\Delta }\mathrm{G}\left({\mathrm{H}}_4\right)-\mathrm{\Delta }\mathrm{G}\left({\mathrm{T}}_2\right)=\mathrm{\Delta }\mathrm{G}\left({\mathrm{H}}_{3\mathrm{a}}-{\mathrm{H}}_4\right)-\mathrm{\Delta }\mathrm{G}\left({\mathrm{H}}_{3\mathrm{a}}\right)-\mathrm{\Delta }\mathrm{G}\left({\mathrm{H}}_4\right)\hfill \end{array}$$ (4)

where T₁, H_1a, H₂, H_1a‐H₂, T₂, H_3a, H₄, and H_3a–H₄ represent target circCDYL, hairpin probe 1a, hairpin probe 2, the hybrid of hairpin probe 1a and hairpin probe 2, target circHIPK3, hairpin probe 3a, hairpin probe 4, the hybrid of hairpin probe 3a and hairpin probe 4, respectively. Moreover, we introduced four tunable variables (c₁, c₂, c_3, and c₄) and two variables (x ₁, x ₂) into the equations. The equilibrium concentrations of all nucleotides can be obtained by solving equations ((5), (6)).

$${\mathrm{K}}_1={\mathrm{e}}^{-\mathrm{\Delta }{\mathrm{G}}_1/\mathrm{RT}}=\frac{\left[{\mathrm{H}}_{1\mathrm{a}}-{\mathrm{H}}_2\right]\left[{\mathrm{T}}_1\right]}{[{\mathrm{T}}_1][{\mathrm{H}}_{1\mathrm{a}}][{\mathrm{H}}_2]}=\frac{\left[{\mathrm{H}}_{1\mathrm{a}}-{\mathrm{H}}_2\right]}{[{\mathrm{H}}_{1\mathrm{a}}][{\mathrm{H}}_2]}=\frac{{\mathrm{x}}_1}{({\mathrm{c}}_2-{\mathrm{x}}_1)({\mathrm{c}}_2-{\mathrm{x}}_1)}$$ (5)

$${\mathrm{K}}_2={\mathrm{e}}^{-\mathrm{\Delta }{\mathrm{G}}_2/\mathrm{RT}}=\frac{\left[{\mathrm{H}}_{3\mathrm{a}}-{\mathrm{H}}_4\right]\left[{\mathrm{T}}_2\right]}{[{\mathrm{T}}_2][{\mathrm{H}}_{3\mathrm{a}}][{\mathrm{H}}_4]}=\frac{\left[{\mathrm{H}}_{3\mathrm{a}}-{\mathrm{H}}_4\right]}{[{\mathrm{H}}_{3\mathrm{a}}][{\mathrm{H}}_4]}=\frac{{\mathrm{x}}_2}{({\mathrm{c}}_3-{\mathrm{x}}_2)({\mathrm{c}}_4-{\mathrm{x}}_2)}$$ (6)

where c ₁ = [H_1a]₀, c ₂ = [H_2a]₀,c ₃ = [H_3a]₀, c ₄ = [H_4a]₀, x ₁ = [H_1a − H₂], and x ₂ = [H_3a − H₄]. K ₁ is the equilibrium constant for circCDYL‐activated CHA‐C reaction, and K ₂ is the equilibrium constant for circHIPK3‐activated CHA‐H reaction. Hess's law is used to calculate the K ₁/K ₂ value for the entire activation process. When T = 310.15 K and R = 1.987 × 10⁻³ kcal·K⁻¹·mol⁻¹, we calculated E ₁/E ₂ for the initiation of CHA‐C/CHA‐H, with E ₁/E ₂ being defined as the ratio of the balanced hybrid concentration (x ₁ = [H_1a − H₂]/x ₂ = [H_3a − H₄]) to the initial H_1a/H_3aconcentration (c ₁ = [H_1a]₀/c ₃ = [H_3a]₀).

$${\mathrm{E}}_1=\frac{{\mathrm{x}}_1}{{\mathrm{c}}_1}\times 100\%$$ (7)

$${\mathrm{E}}_2=\frac{{\mathrm{x}}_2}{{\mathrm{c}}_3}\times 100\%$$ (8)

As presented in Table 2, the theoretical conversion efficiency is largely insensitive to circCDYL/circHIPK3 concentration across a wide range (1 fM to 100 nM). These results confirm that the high theoretical conversion efficiency facilitates the toehold‐mediated strand displacement and even a few copies of target can efficiently initiate the CHA circuits. Notably, at a target concentration of 100 nM, the theoretical E ₁/E ₂ ratio is 1.01, whereas the corresponding experimental ratio derived from single‑molecule detection (Cy5/Cy3 counts) is 1.20 (Figure 2K). This slight discrepancy arises from multiple objective factors involved in practical reactions and imaging processes: (1) Different fluorophores possess distinct photophysical properties, resulting to fluorophore‐dependent detection efficiency [62]; (2) Intrinsic kinetic differences exist in nucleic acid hybridization and cascade amplification within the actual experimental system, thereby affecting the final signal output ratio [63]; (3) Imaging biases caused by background noise, photobleaching, and instrumental detection sensitivity can also induce minor fluctuations in single‐molecule counting results [64].

TABLE 2.

Theoretical conversion efficiency (E) of H_1a/H_3a activated by circCDYL/circHIPK3 at different concentrations.

Concentration of circCDYL/circHIPK3	Conversion efficiency E1 (%) for CHA‐C	Conversion efficiency E2 (%) for CHA‐H
100.0 nM	99.95	99.29
1.0 nM	99.95	99.29
10.0 pM	99.95	99.29
100.0 fM	99.95	99.29
1.0 fM	99.95	99.29

To further investigate the amplification kinetics of GSH‐gated DB‐CHA@PAN, we constructed a control nanoclassifier without target cycling function (termed nT‐DB‐CHA@PAN) (Figure 3B). The kinetics of the CHA reactions were monitored using time‐dependent fluorescence measurements. In the absence of circCDYL/circHIPK3, both nT‐DB‐CHA@PAN and DB‐CHA@PAN show negligible Cy5 (Figure 3C,H, blue curves) and Cy3 fluorescence signals (Figure 3D,I, orange curves). In contrast, upon addition of circCDYL/circHIPK3, DB‐CHA@PAN exhibits a rapid and pronounced increase in Cy5/Cy3 signals (Figure 3H,I), which are substantially higher than the weak responses from nT‐DB‐CHA@PAN (Figure 3C,D). The initial reaction rates (V ₀) for nT‐DB‐CHA@PAN are 67.67 min⁻¹ for circCDYL (Figure 3C) and 59.60 min⁻¹ for circHIPK3 (Figure 3D). Comparatively, the initial reaction rates (V ₀) for DB‐CHA@PAN are 136.96 min⁻¹ for circCDYL (Figure 3H) and 105.80 min⁻¹ for circHIPK3 (Figure 3I). These findings are directly attributed to the ability of DB‐CHA@PAN to promote the cyclic targets release, thereby accelerating the cascade CHA amplification reaction. We also verified the CHA efficiency of DB‐CHA@PAN using single‐molecule detection. In the absence of GSH, no Cy5/Cy3 spots are observed in response to either nT‐DB‐CHA@PAN (Figure 3E) or DB‐CHA@PAN (Figure 3J), regardless of circCDYL/circHIPK3 presence. Upon addition of GSH, distinct Cy5 spots are detected with circCDYL alone (Figure 3E,J; red color), obvious Cy3 spots are observed with circHIPK3 alone (Figure 3E,J; green color), and both Cy5 and Cy3 spots are detected simultaneously with the coexistence of circCDYL and circHIPK3 (Figure 3E,J). The Cy5 and Cy3 counts induced by DB‐CHA@PAN (Figure 3K) are significantly higher than those obtained by nT‐DB‐CHA@PAN (Figure 3F), respectively. These results demonstrate that the target recycling function is the primary driver of the CHA cycling in DB‐CHA@PAN.

2.4. Spatioselective Imaging of Dual CircRNAs in MCF‐7 Cells Using the GSH‐Gated DB‐CHA@PAN

We first replaced the photosensitizer in H₁ with Cy5. To evaluate the capability of DB‐CHA@PAN for selective and efficient monitoring of intracellular circRNAs, we employed three nanoclassifiers (i.e., nDB‐CHA@PAN (lacking disulfide bonds in H₁ and H₃), DB‐mCHA₁@PAN (with mutated bases in circCDYL‐recognizing sequence of H₁), and DB‐mCHA₂@PAN (with mutated bases in circHIPK3‐recognizing sequence of H₃)) as the controls (Figures 4A and S9). The confocal laser scanning microscopy (CLSM) imaging was performed to verify Cy5 and Cy3 fluorescence signals in live cells under various conditions (Figures 4B‐D). Significant Cy5 and Cy3 fluorescence signals are visualized in MCF‐7 cells treated with DB‐CHA@PAN (Figure 4Ba,Ca,Da). In contrast, extremely low fluorescence signals are observed in MCF‐7 cells treated with nDB‐CHA@PAN (Figure 4Bb,Cb,Db), implying that only the disulfide bonds must first be cleaved by intracellular GSH to activate CHA‐C/CHA‐H reactions for subsequent restoration of Cy5 and Cy3 fluorescence signals, confirming the critical role of disulfide bonds in specific activation of DB‐CHA@PAN. Furthermore, a markedly reduced Cy5 fluorescence signal is detected in MCF‐7 cells treated with DB‐mCHA₁@PAN (Figure 4Bc,Cc,Dc), indicating that the introduced base mutations in H₁ prevent circCDYL from recognizing the toehold region (a), blocking the activation of the CHA‐C cycling reaction, underscoring the essential role of target circCDYL in specific CHA‐C activation. Similarly, very weak Cy3 signals are observed in MCF‐7 cells treated with DB‐mCHA₂@PAN (Figure 4Bd,Cd,Dd), suggesting that the base mutations in H₃ disrupt the recognition of the toehold region (c) of H_3a by circHIPK3, consequently failing to activate the CHA‐H cycling reaction, confirming the indispensable role of target circHIPK3 in specific CHA‐H activation. Next, we quantitatively assessed the Cy5 and Cy3 fluorescence signals using flow cytometry under the same set of conditions (Figures 4E,F and S12), which strongly correlates with the results obtained from corresponding CLSM imaging (Figure 4B–D). These results verify that DB‐CHA@PAN is controlled by the sequential unlocking of intracellular “keys” (GSH, circCDYL, and circHIPK3), thereby underlining its high precision and specificity. Moreover, to assess the contribution of target cycling, we employed one nanoclassifier without target cycling functionality (i.e., nT‐DB‐CHA@PAN) as the control (Figure 4G). As shown in Figure 4H–J, significantly higher Cy5 and Cy3 fluorescence signals are detected in MCF‐7 cells treated with DB‐CHA@PAN by CLSM imaging (Figure 4Ha,Ia,Ja), whereas minimal fluorescence signals are measured in cells treated with nT‐DB‐CHA@PAN (Figure 4Hb,Ib,Jb). This pronounced difference is further quantified and validated by flow cytometry (Figure 4K,L), revealing the pivotal role of the target cycling in the signal amplification mechanism of DB‐CHA@PAN.

FIGURE 4.

(A) Procedural diagram of the designs of DB‐CHA@PAN, nDB‐CHA@PAN, DB‐mCHA₁@PAN, and DB‐mCHA₂@PAN. (B) CLSM imaging of MCF‐7 cells treated with DB‐CHA@PAN, nDB‐CHA@PAN, DB‐mCHA₁@PAN, and DB‐mCHA₂@PAN, respectively. The corresponding diagrams show the fluorescence intensity distribution on the arrows drawn in the images. (C and D) Mean fluorescence intensity (MFI) of MCF‐7 cells corresponding to different treatment groups shown in (B). (E and F) Flow cytometric analysis of Cy5 (E)/Cy3 (F) fluorescence intensity in MCF‐7 cells corresponding to different treatment groups shown in (B). (G) Procedural diagram of dual circRNAs imaging in MCF‐7 cells treated with DB‐CHA@PAN and nT‐DB‐CHA@PAN. (H) CLSM imaging of MCF‐7 cells treated with DB‐CHA@PAN and nT‐DB‐CHA@PAN. The corresponding diagrams show the fluorescence intensity distribution on the arrows drawn in the images. (I‐J) MFI of MCF‐7 cells treated with DB‐CHA@PAN and nT‐DB‐CHA@PAN. (K and L) Flow cytometric analysis of Cy5 (K)/Cy3 (L) fluorescence intensity in MCF‐7 cells treated with DB‐CHA@PAN and nT‐DB‐CHA@PAN. Scale bar is 10 µm. Error bars denote mean ± SD (n = 3).

2.5. Real‐Time Tracking of CircRNAs Dynamics in Living Cells Using the GSH‐Gated DB‐CHA@PAN

To enable real‐time tracking of circRNAs dynamics in living cells, we first replaced the photosensitizer in H1 with Cy5 (Figure 5A). The cellular uptake ability of DB‐CHA@PAN as delivery carriers was evaluated using one control nanoclassifier without the AS1411 aptamer. As shown in Figure 5B, pronounced Cy5 and Cy3 fluorescence signals are visualized in MCF‐7 cells treated with DB‐CHA@PAN (Figure 5Bb,C,D), whereas minimal fluorescence signals are observed in MCF‐7 cells treated with DB‐CHA@PAN without AS1411 aptamer (Figure 5Ba,C,D), indicating that the efficient transport of the DB‐CHA@PAN into living cells is achieved through nucleolin/AS1411 aptamer‐mediated micropinocytosis. Additionally, we pretreated MCF‐7 cells with excess free AS1411 aptamer to block nucleolin on the cell membrane surface. Figure S13 shows that both Cy5 and Cy3 fluorescence signals are significantly reduced after blocking nucleolin, confirming that DB‐CHA@PAN enters tumor cells primarily via the AS1411‑nucleolin‑mediated endocytosis. We further employed CLSM to track the internalization process of DB‐CHA@PAN (Figure S14). The Pearson's colocalization coefficients between DB‐CHA@PAN and lysosomes increase markedly from 0.07/0.05 at 0.5 h to 0.80/0.78 at 1 h, then decline to 0.24/0.21 at 2 h, indicating the reduced lysosomal colocalization at later time points, consistent with possible cytosolic redistribution [65, 66]. CLSM imaging reveals that Cy5/Cy3 fluorescence intensity in MCF‐7 cells increases with prolonged incubation, and reaches a plateau at 2 h (Figure S15), suggesting that DB‐CHA@PAN displays efficient internalization kinetics within a 2 h incubation period. Furthermore, significantly stronger fluorescence signals are observed in MCF‐7 cells treated with DB‐CHA@PAN compared to cells treated with free DB‐CHA (Figure S16). The performance of PAN (Figure 1) was also compared with the commercial Lipofectamine 3000 in delivering CHA probes. DB‐CHA@PAN achieves 2.07‐fold higher Cy5 and 2.26‐fold higher Cy3 fluorescence intensities in MCF‐7 cells than the Lipofectamine 3000 (Figure S17). These results demonstrate that DB‐CHA@PAN achieves enhanced cellular delivery efficiency, consistent with the superior membrane permeability of structurally rigid DNA nanostructures.

FIGURE 5.

(A) Procedural diagram of intracellular circCDYL and circHIPK3 detection in diverse human cell lines treated with DB‐CHA@PAN. (B) CLSM imaging of MCF‐7 cells treated with DB‐CHA@PAN with and without the AS1411 aptamer. (C and D) MFI of MCF‐7 cells treated with DB‐CHA@PAN with and without the AS1411 aptamer shown in (B). (E) CLSM imaging of circCDYL and circHIPK3 in MCF‐7, HepG2, HeLa, A549, and MCF‐10A cells. (F and G) MFI of Cy5 and Cy3 in MCF‐7, HepG2, HeLa, A549, and MCF‐10A cells shown in (E). (H and I) Flow cytometric analysis of Cy5 (H)/Cy3 (I) fluorescence intensity in MCF‐7 cells in various cells treated with DB‐CHA@PAN. (J) CLSM imaging of intracellular circCDYL and circHIPK3 in MCF‐7 cells treated with PBS, anti‐circCDYL, anti‐circHIPK3, and anti‐circCDYL + anti‐circHIPK3, respectively. (K and L) MFI of MCF‐7 cells treated with PBS, anti‐circCDYL, anti‐circHIPK3, and anti‐circCDYL + anti‐circHIPK3. (M and N) Flow cytometric analysis of Cy5 (M)/Cy3 (N) fluorescence intensity in MCF‐7 cells treated with PBS, anti‐circCDYL, anti‐circHIPK3, and anti‐circCDYL + anti‐circHIPK3, respectively. Scale bar is 10 µm. Error bars denote mean ± SD (n = 3).

We next investigated whether DB‐CHA@PAN could discriminate circCDYL/circHIPK3 expression levels among diverse human cell lines. CLSM imaging shows variable fluorescence intensities corresponding to circCDYL (Cy5) and circHIPK3 (Cy3) in different cells (Figure 5E). The Cy5 fluorescence intensity decreases in the order of MCF‐7> HepG2> HeLa > A549> MCF‐10A cells (Figure 5F). Similarly, the Cy3 fluorescence intensity follows the order of HeLa > HepG2> A549> MCF‐7> MCF‐10A cells (Figure 5G), aligned with the in vitro measurements and qRT‐PCR assay (Figures 6B and S21). Flow cytometry independently confirms the above observations, showing strong Cy5/Cy3 fluorescence signals in the cancer cell lines (i.e., MCF‐7, HepG2, HeLa, and A549 cells) but only minimal background signals in the non‐cancerous MCF‐10A cells (Figures 5H,I and S18). These results demonstrate that DB‐CHA@PAN enables accurate profiling of circCDYL and circHIPK3 expressions in different types of living cells. We further performed real‐time tracking of circRNAs expression fluctuations in living cells. Two different antisense oligonucleotides (ASOs) (i.e., anti‐circCDYL and anti‐circHIPK3, Table S1) were synthesized to specifically knock down their respective circRNAs. As shown in Figure 5J, robust Cy5 and Cy3 fluorescence signals are observed in PBS‐treated MCF‐7 cells (Figure 5Ja,Ka,La), corresponding to basal levels of circCDYL and circHIPK3 (Figure 5Bb,C,D). In contrast, anti‐circCDYL treatment selectively diminishes Cy5 fluorescence signal while leaving Cy3 fluorescence signal unchanged (Figure 5Jb,Kb,Lb), and anti‐circHIPK3 treatment specifically reduces the Cy3 fluorescence signal without affecting Cy5 fluorescence signal (Figure 5Jc,Kc,Lc). Combined treatment with both ASOs leads to a marked reduction in both Cy5 and Cy3 fluorescence signals (Figure 5Jd,Kd,Ld). Flow cytometry analysis also validates these results (Figure 5M,N), demonstrating that the nanoclassifier can specifically and effectively monitor circCDYL and circHIPK3 dynamics in living cells.

FIGURE 6.

(A) Procedural diagram of the GSH‐gated intelligent nanoclassifier for profiling circCDYL and circHIPK3 in clinical samples. (B) Heat map analysis of circCDYL (I)/circHIPK3 (II) expression across different cell types by this nanoclassifier. (C and D) Correlation analysis between the nanoclassifier and qRT‐PCR for detecting circCDYL (C)/circHIPK3 (D) in the above various human cells. (E and F) Linear correlation between fluorescence counts and the logarithm of MCF‐7 (E)/HeLa (F) cell numbers. RNA extracted from an equivalent of 10⁵ cells was used in all experiments. (G and H) Correlation analysis between the nanoclassifier and qRT‐PCR for detecting circCDYL (G) and circHIPK3 (H) in breast tissues. (I and J) Correlation analysis between the nanoclassifier and qRT‐PCR for detecting circCDYL (I) and circHIPK3 (J) in lung tissues. (K–M) ROC curves for differentiating breast cancer (BC) patients at different stages from healthy individuals (HD) by the nanoclassifier. (N–P) ROC curves for differentiating lung cancer (LC) patients at different stages from healthy individuals (HD) by the nanoclassifier. (Q and R) PCA‐based pattern recognition for distinguishing breast (Q)/lung (R) cancer at different stages. (S and T) Identification and classification of unidentified breast (S)/lung (T) cancer at different stages. Student's t‐test, ∗∗∗p < 0.001. Error bars denote mean ± SD (n = 3).

2.6. Clinical Profiling of Dual CircRNAs in Human Cells and Tissues Using the GSH‐gated DB‐CHA@PAN

We first replaced the photosensitizer in H₁ with Cy5. To validate the clinical applicability of the intelligent nanoclassifier (Figure 6A), we spiked varying concentrations of circCDYL and circHIPK3 into phosphate‐buffered saline (PBS) and human serum (1% and 5%), respectively. The recovery ratio of circRNAs in PBS samples ranges from 99.11% to 102.78%, with a relative standard deviation (RSD) of less than 4.30% (Figure S19A,B; Tables S2,S5). For serum samples, the recovery ratio is 99.06%−102.80%, with an RSD of less than 4.99% (Figure S19A,B; Tables S3, S4, S6, and S7). These results demonstrate that the intelligent nanoclassifier is capable of accurately quantifying circRNAs in complex biological matrices. Importantly, DB‐CHA@PAN also exhibits excellent stability even in FBS or serum (Figure S20). Furthermore, we applied this nanoclassifier to profile the expression levels of circCDYL and circHIPK3 across a panel of cell lines, including MCF‐7, HepG2, HeLa, A549, and MCF‐10A cells. Heat map was employed to visualize the expression profiles of circCDYL and circHIPK3. As shown in Figure 6B (I), the Cy5 counts generated by MCF‐10A cells are significantly lower than those generated by MCF‐7, HepG2, HeLa, and A549 cells, revealing the up‐regulation of circCDYL expression level in breast, liver, cervical, and lung cancer cells [56, 67]. The circCDYL expression level follows an order: MCF‐7> HepG2> HeLa > A549> MCF‐10A cells, which is 4.91‐, 4.03‐, 2.96‐, and 1.98‐fold higher than that obtained by MCF‐10A cells, respectively, consistent with previous studies [67]. We further quantified the relative expression of circCDYL in above cell lines using standard qRT‐PCR (Figure S21(I)). The obtained results (Figure S22A, blue rectangles) are highly concordant with those obtained by this nanoclassifier (Figure S22A, red rectangles). Moreover, a high‐degree linear correlation between this nanoclassifier and the qRT‐PCR method is achieved, with a Pearson's coefficient of 0.9848 (Figure 6C). Similarly, as shown in Figure 6B (II), high Cy3 counts are generated by HeLa, HepG‐2, A549, and MCF‐7, indicating the up‐regulation of circHIPK3 in cervical, liver, lung, and breast cancers [32]. The circHIPK3 expression order is determined as: HeLa > HepG2> A549> MCF‐7> MCF‐10A cells, which is 5.49‐, 4.32‐, 3.19‐, and 2.10‐fold higher than that induced by MCF‐10A cells, respectively, in line with the earlier observations [32]. The relative expression of circHIPK3 was further quantified in the above cell lines by qRT‐PCR (Figure S21 (II)). The results (Figure S22B, purple rectangles) agree well with those measured by this nanoclassifier (Figure S22B, green rectangles). Moreover, a strong linear correlation is confirmed between the two methods, with a Pearson's coefficient of 0.9971 (Figure 6D), further validating the reliability of this nanoclassifier. We also investigated the dependence of fluorescence counts on the cell numbers. For circCDYL, Cy5 counts (N) increase proportionally with the MCF‐7 cell numbers (X) ranging from 1 to 10⁵ cells, with an equation of N = 45.62 + 45.07 log₁₀ X (R ² = 0.9999) (Figure 6E). The LOD is 1 cell. For circHIPK3, Cy3 counts (N) increase proportionally with HeLa cell numbers (X) ranging from 1 to 10⁵ cells, with an equation of N = 41.12 + 39.68 log₁₀ X (R ² = 0.9998) (Figure 6F). The LOD is 1 cell. These findings demonstrate that this nanoclassifier can distinguish between circCDYL and circHIPK3 across diverse cell types, and enables precise quantification of circRNAs at the single‐cell level.

To further estimate the potential of this nanoclassifier for clinical applications, we quantitatively profiled the expression levels of circRNAs in tumor tissues from 24 breast/lung cancer patients (12 cases at stages I‐II and 12 cases at stages III‐IV) and 10 adjacent normal counterparts (healthy controls) (Table S8). The Cy5/Cy3 fluorescence signals are significantly enhanced in patient cohorts compared with healthy counterparts and reach the highest signals in stages III‐IV of breast and lung cancers (Figure S23A,C), indicating the up‐regulation of circCDYL/circHIPK3 in both breast and lung cancers [32, 56, 68]. Furthermore, qRT‐PCR is used to measure the relative expressions of circCDYL and circHIPK3 in the above breast and lung tissue samples (Figure S23B,D). The average expression levels of circCDYL and circHIPK3 in the breast/lung patient cohort are statistically higher than those in healthy counterparts, consistent with those of this nanoclassifier (Figure S23A,C). Importantly, a high‐degree linear correlation is observed between the nanoclassifier and the qRT‐PCR method across both tissue types. Specifically, in breast tissues, the Pearson's coefficients are 0.9947 for circCDYL (Figure 6G) and 0.9961 for circHIPK3 (Figure 6H). Similarly, in lung tissues, the Pearson's coefficients achieve 0.9995 for circCDYL (Figure 6I) and 0.9604 for circHIPK3 (Figure 6J).

Moreover, we examined the diagnostic performance of the dual‐biomarker combination for breast cancer (BC)/lung cancer (LC) by analyzing the receiver operating characteristic (ROC) curves [69]. As shown in Figure 6K–M, the combined detection of circCDYL and circHIPK3 exhibits markedly superior performance compared with single‐biomarker detection across all breast cancer diagnostic scenarios. Specifically, for distinguishing early‐stage (Stage I/II) from late‐stage (Stage III/IV) breast cancer patients, the area under the curve (AUC) of the dual‐biomarker combination detection reaches 1.00, substantially higher than that of single‐biomarker detection (0.85 for circCDYL and 0.88 for circHIPK3) (Figure 6K); for differentiating Stage I/II breast cancer patients from healthy individuals (HD), the AUC of the combination detection is 1.00, superior to single‐biomarker detection (0.84 for circCDYL and 0.85 for circHIPK3) (Figure 6L); for distinguishing Stage III/IV breast cancer patients from healthy individuals, the combination detection yields an AUC of 1.00, significantly exceeding that of single‐biomarker detection (0.89 for circCDYL and 0.90 for circHIPK3) (Figure 6M). Similarly, the dual‐biomarker detection achieves outstanding diagnostic capacity for both early‐ and late‐stage lung cancer patients, with performance considerably better than single‐biomarker detection (Figure 6N–P). Based on the differential expression levels of circCDYL and circHIPK3 across various stages, principal component analysis (PCA) can effectively distinguish healthy individuals from breast/lung cancer patients at different stages (Figure 6Q,R). Additionally, we performed a blinded analysis of 34 unidentified tissue samples. The classification results further verify that the dual‐biomarker combination detection achieves superior sensitivity, good specificity and high accuracy (each reaching 100%) for the differential diagnosis of breast/lung cancer compared with single‐biomarker detection strategy (Figures 6S,T and S24), confirming the exceptional clinical diagnostic reliability of this nanoclassifier.

2.7. Validation of the PDT Efficacy for Cancer Treatment

To investigate the PDT efficacy of DB‐CHA@PAN in cancer therapy, we employed the oxidant‐sensing probe 2',7'‐dichlorodihydrofluorescein diacetate (DCFH‐DA) to validate the ROS generation. Following cellular uptake, DCFH‐DA is hydrolyzed by esterases to DCFH, and then oxidized by ROS to yield the 2′,7′‐dichlorofluorescein (DCF) with strong green fluorescence [58]. As shown in Figure 7A, under 660‐nm light irradiation, a strong DCF fluorescence signal is observed in MCF‐7 cells treated with DB‐CHA@PAN, markedly exceeding that in nT‐DB‐CHA@PAN‐treated cells. Comparatively, only a minimal DCF fluorescence signal is detected in MCF‐10A cells treated with DB‐CHA@PAN. Without light irradiation, negligible green fluorescence is detected in both MCF‐7 and MCF‐10A cells treated with either DB‐CHA@PAN or nT‐DB‐CHA@PAN. These findings indicate that DB‐CHA@PAN effectively induces intracellular ROS generation in cancer cells. Next, we systematically evaluated the PDT‐induced cytotoxicity of DB‐CHA@PAN using the Cell Counting Kit‐8 (CCK‐8) assay. As displayed in Figure 7B, without light irradiation, no obvious cytotoxicity is measured in MCF‐7 cells treated with increasing DB‐CHA@PAN concentration, whereas a pronounced, concentration‐dependent phototoxic effect is observed upon 660‐nm light irradiation. Notably, under identical light dosage, the cytotoxicity induced by DB‐CHA@PAN‐treated MCF‐7 cells is 1.55‐fold higher than that obtained by cells exposed to nT‐DB‐CHA@PAN (Figure 7C), conveying that the target recycling function enhances PDT efficacy to produce a stronger cell‐killing effect. In contrast, only a slight viability decrease is observed in normal MCF‐10A cells treated with either DB‐CHA@PAN or nT‐DB‐CHA@PAN under 660‐nm light irradiation (Figure 7D). This can be attributed to the low basal expressions of GSH and circRNA, as well as minor non‐enzymatic dissociation between the photosensitizer and quencher in MCF‐10A cells [58]. Importantly, these results demonstrate that the toxicity of this system toward normal cells is far weaker than its significant killing effect on tumor cells, fully proving that DB‐CHA@PAN‐based PDT possesses excellent cancer cell specificity. We further conducted apoptosis assessment by flow cytometry using Annexin V‐FITC/PI staining. As indicated in Figure 7E, without light irradiation, no significant growth inhibition is shown in DB‐CHA@PAN/nT‐DB‐CHA@PAN‐treated MCF‐7 cells. While under 660‐nm light irradiation, a significantly higher apoptosis/necrosis rate (59.4%) is detected in MCF‐7 cells treated with DB‐CHA@PAN compared to that (33.8%) treated with nT‐DB‐CHA@PAN, aligning with the cytotoxicity trend observed in the CCK‐8 assay (Figure 7C,D). Meanwhile, we performed the live/dead staining to assess cell viability, with viable and non‐viable cells stained by calcein AM (green) and PI (red) fluorescence, respectively. As indicated in Figure 7F, strong green fluorescence signals are observed in MCF‐10A cells treated with DB‐CHA@PAN both with and without light irradiation, evidencing the excellent biocompatibility and low toxicity of DB‐CHA@PAN toward normal cells. In contrast, a low red fluorescence signal is shown in MCF‐7 cells treated with nT‐DB‐CHA@PAN, reflecting inefficient cell killing due to the lack of target recycling. The strongest red fluorescence signal and the largest area of dead cells are visualized in MCF‐7 cells treated with DB‐CHA@PAN, indicating that its target recycling function drives highly efficient PDT and potent tumor cell killing. These results demonstrate that this nanoclassifier can induce tumor‐selective apoptosis, offering promising applications in activatable synergistic cancer therapy.

FIGURE 7.

(A) Intracellular ROS fluorescence imaging after treatment with DB‐CHA@PAN and nT‐DB‐CHA@PAN under light irradiation or light‐free conditions. Scale bar is 40 µm. (B) Cell viability of MCF‐7 cells treated with varying‐concentration DB‐CHA@PAN with and without light exposure. (C and D) Cell viability of MCF‐7 (C) and MCF‐10A (D) cells treated with DB‐CHA@PAN or nT‐DB‐CHA@PAN. **p < 0.01, and n.s. (not significant). (E) Assessment of MCF‐7 cells apoptosis triggered by DB‐CHA@PAN or nT‐DB‐CHA@PAN through flow cytometry with Annexin V/PI staining. (F) CLSM imaging of MCF‐7 cells dual‐stained with Calcein‐AM and PI following DB‐CHA@PAN or nT‐DB‐CHA@PAN treatments. Scale bar is 100 µm. Error bars denote mean ± SD (n = 3).

2.8. GSH‐Activated DB‐CHA@PAN for in Vivo circRNA Imaging and Antitumor Therapy

We further investigated the in vivo imaging capability and therapeutic efficacy of DB‐CHA@PAN in mouse models (Figure 8A). Nude mice bearing MCF‐7 xenograft tumors are randomly assigned to receive intravenous injections of PBS, nT‐DB‐CHA@PAN, or DB‐CHA@PAN, and in vivo fluorescence imaging is performed at predetermined time points (Figures 8B, C). As displayed in Figure 8B,C, the Cy5/Cy3 fluorescence signals in tumor regions are markedly enhanced in the DB‐CHA@PAN group over time, whereas only a slight fluorescence elevation is observed in the nT‐DB‐CHA@PAN group. Quantitative fluorescence analysis reveals that the intratumoral Cy5 and Cy3 fluorescence intensities of the DB‐CHA@PAN group are 2.33‐ and 2.76‐fold higher than those of the nT‐DB‐CHA@PAN group at 3 h post‐injection (Figure 8D,E), confirming that DB‐CHA@PAN achieves in vivo cyclic signal amplification. Subsequently, the in vivo PDT effect of DB‐CHA@PAN is systematically evaluated using MCF‐7 tumor‐bearing nude mice. When the tumor volume reaches approximately 100 mm³, the mice are divided into six groups and intravenously injected with PBS, nT‐DB‐CHA@PAN or DB‐CHA@PAN, with or without 660‐nm laser irradiation on the tumor regions. The entire theranostic experiment lasted 21 days, with mice receiving injections every 3 days. During this period, changes in mouse body weight and tumor volume were recorded regularly to assess the antitumor effects. As shown in Figure 8F, body weights in all groups remain stable without significant differences, indicating negligible systemic toxicity of the two nanoclassifiers. In the absence of light irradiation, the tumor growth trends of the nT‐DB‐CHA@PAN and DB‐CHA@PAN groups are consistent with those of the PBS group, further demonstrating that both nanoclassifiers possess low dark toxicity (Figure 8G,H). Notably, under the same dosage and irradiation conditions, DB‐CHA@PAN exerts much stronger tumor growth inhibition than nT‐DB‐CHA@PAN, which is attributed to the fact that the target recycling function of DB‐CHA@PAN can efficiently amplify PDT efficacy. After treatment, all mice were euthanized, and H&E staining and TdT‐mediated dUTP nick‐end labeling (TUNEL) assays were carried out to evaluate the therapeutic efficacy of DB‐CHA@PAN. Histological results confirm that the tumor tissue damage in the DB‐CHA@PAN + light group is more severe than that in the nT‐DB‐CHA@PAN + light group, further validating the superior PDT performance of DB‐CHA@PAN (Figure 8I). Additionally, routine blood biochemical analysis (Figure S25) and H&E staining of major organs (Figure S26) reveal no obvious tissue damage or necrosis, manifesting the favorable biocompatibility of DB‐CHA@PAN. Collectively, DB‐CHA@PAN possesses potent in vivo antitumor efficacy and satisfactory biosafety, holding great promise for imaging‐guided PDT applications.

FIGURE 8.

(A) Procedural diagram of the treatment protocol for nude mice bearing MCF‐7 xenograft tumors. (B and C) Fluorescence imaging of MCF‐7 tumor‐bearing mice treated with PBS, nT‐DB‐CHA@PAN, or DB‐CHA@PAN (n = 3). (D and E) Quantitative analysis of Cy5 (D)/Cy3 (E) fluorescence intensities in tumor sites corresponding to (B and C). (F) Relative body weight of mice in various treatment groups. (G) Average tumor volume of mice in various treatment groups. (H) Optical photo of tumor excised from mice in various treatment groups. (I) H&E (top) and TUNEL (bottom) staining images of tumor tissues in various treatment groups. Scale bar is 50 µm. Error bars denote mean ± SD (n = 3).

3. Conclusion

In this research, we construct an endogenous GSH‐gated intelligent nanoclassifier (DB‐CHA@PAN) by assembling disulfide bond (DB)‐modified CHA probes onto cruciform DNA framework‐based programmable atom‐like nanoparticles (PAN) for spatioselective visualization of dual circRNAs and synergistic photodynamic precision therapy. Taking advantage of specific recognition, enhanced biostability, high imaging contrast, efficient internalization, rapid kinetics, and potent photodynamic activity, this intelligent nanoclassifier can achieve attomolar sensitivity with an LOD of 18.6 aM for circCDYL and 24.5 aM for circHIPK3 as well as a 9‐order dynamic range in vitro. Moreover, it can effectively discriminate circRNAs from mismatched variants with single‐base resolution, accurately quantify circRNA levels at single‐cell sensitivity, and even precisely diagnose breast/lung cancer across the entire clinical spectrum (from stage I to IV) with 100% accuracy. Furthermore, it facilitates real‐time tracking of circRNAs dynamics in living cells and MCF‐7 tumor‐bearing nude mice, visually distinguishes malignant cells from normal cells, and markedly enhances tumor specificity and therapeutic efficacy via synergistic photodynamic activation. Compared with the reported circRNA methods (Table S9), this intelligent nanoclassifier exhibits distinctive advantages: (1) the anti‐nucleolin aptamer AS1411 is integrated into cruciform framework to enhance specific target recognition, facilitating efficient internalization of DB‐CHA@PAN via nucleolin receptor‐mediated endocytosis; (2) the disulfide bond act as a intracellular GSH‐responsive switch, which prevents “off‐tumor” signal leakage that is inherent to “always‐on” activation designs; (3) the target‐recycling‐based CHA circuit facilitates cascaded signal amplification with attomolar sensitivity and efficient photodynamic therapy; (4) the cruciform nanostructure precisely integrates multiple functional modules, enabling in situ simultaneous tracking of trace‐level circRNAs; (5) this intelligent nanoclassifier exploits the tumor microenvironment (e.g., nucleolin, elevated GSH, and target circRNAs) as the energy boosters and leverages the spatial confinement of DNA framework to locally concentrate reactants, which greatly accelerates overall reaction kinetics without any external stimulus; (6) this intelligent nanoclassifier integrates with computational AND‐gate logic to guarantee the therapeutic precision and minimize the off‐target effects; (7) the combination of circCDYL and circHIPK3 can serve as a robust diagnostic biomarker, enabling accurate diagnosis and staging of breast/lung cancer with an AUC of 1; (8) this intelligent nanoclassifier can be adapted to detect a wide range of RNA targets (e.g., mRNA, miRNAs, piRNAs, and lncRNAs) by simply changing the recognition sequences, holding significant promise for clinical diagnostics and precision medicine.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Detailed experimental procedures and Supporting Figures. Supporting File 1: anie73571‐sup‐0001‐SuppMat.pdf.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.