Machine Learning‐Optimized Single‐Atom Catalysts Enable Microenvironment‐Adaptive Chemodynamic‐Bioorthogonal Cancer Therapy

ABSTRACT

Chemodynamic therapy (CDT), which harnesses endogenous chemical energy within the tumor microenvironment (TME), has shown high potential for precise cancer treatment. However, its efficacy is often limited by the mildly acidic and reductive nature of the TME that compromises catalyst stability and activity. Developing catalysts capable of maintaining robust performance under such physiological constraints remains a key challenge. Herein, we report a programmable dual‐catalytic platform that integrates machine learning‐guided design with atomic‐level precision. Through predictive modeling, we establish quantitative structure‐performance relationships that guided the rational synthesis of iron single‐atoms (Fe‐N₅ SAs). The Fe‐N₅ SAs demonstrate exceptional chemodynamic reactivity and environmental stability within the complex TME, efficiently converting endogenous hydrogen peroxide into hydroxyl radicals for precise tumor ablation. Moreover, Fe‐N₅ SAs exhibit potent bioorthogonal catalytic activity, enabling in situ prodrug activation and localized synthesis of doxorubicin under physiological conditions. This synergistic CDT‐bioorthogonal dual‐catalytic mechanism achieves tumor‐selective, stimulus‐free, and combinatorial therapy, markedly enhancing overall antitumor efficacy. This study establishes a machine learning‐guided framework for single‐atom catalyst design and expands the frontiers of metal catalysis in biomedical applications.

A machine learning (ML) framework constructed from literature‐derived catalytic data uncovers quantitative structure‐activity relationships and identifies Fe single‐atoms as optimal catalysts with excellent chemodynamic therapy (CDT) activity and bioorthogonal catalysis. These catalysts generate reactive oxygen species by CDT and activate doxorubicin (DOX) in situ by bioorthogonal catalysis, achieving synergistically enhanced therapeutic efficacy.

1. Introduction

Chemodynamic therapy (CDT) has emerged as a promising modality for precise tumor treatment since its introduction in 2016 [1, 2, 3, 4, 5]. Distinguished by its ability to harness the intrinsic chemical energy of the tumor microenvironment (TME), CDT induces tumor cell death without the need for external stimuli [6, 7]. Unlike radiotherapy and phototherapy, which rely on exogenous energy sources, or chemotherapy, which employs nonspecific cytotoxic agents, CDT exploits the pathological characteristics of the TME (weak acidity and elevated endogenous hydrogen peroxide (H₂O₂)) and the presence of transition‐metals (Fe, Cu, Co, Mn, V, Ti, and Ag) ions to catalyze Fenton or Fenton‐like reactions [8, 9, 10, 11, 12, 13, 14]. These reactions generate reactive oxygen species (ROS) that damage nucleic acids and proteins, leading to selective tumor cell apoptosis while minimizing harm to normal tissues [15, 16, 17]. Recent advances have further demonstrated that integrating CDT with photothermal therapy, immunotherapy, or chemotherapy can enhance therapeutic efficacy and overcome the intrinsic limitations of single‐modality CDT, such as hypoxia‐induced resistance and incomplete tumor ablation [1, 18, 19]. However, clinical translation of CDT remains challenging due to the lack of systematic approaches for discovering and optimizing high‐performance catalysts. The Fenton and Fenton‐like chemistries are highly dependent on microenvironmental parameters [20, 21, 22, 23]. For instance, Fe²⁺ favors strongly acidic conditions, whereas Cu⁺ and Mn²⁺ catalyze hydroxyl radical generation more efficiently under mildly acidic conditions [24, 25, 26]. Developing microenvironment‐adaptable catalysts requires robust, data‐driven screening and predictive modeling.

The 2024 Nobel Prizes in Physics and Chemistry recognized the foundational development in artificial neural networks and the groundbreaking application of machine learning (ML) in protein structure prediction, respectively. Harnessing neural network design, physical principle integration, and complex system prediction, ML redefined paradigms in fundamental algorithms and experimental discovery, and it also provides a powerful strategy to screen suitable catalysts [27, 28, 29]. Leveraging capabilities in big data analysis, pattern recognition, and predictive modeling, ML has transformed materials discovery and optimization across disciplines [30, 31, 32]. In biomedical material research, ML enables the establishment of quantitative structure‐activity relationship models that correlate physicochemical features with catalytic or biological performance, thus accelerating material screening and optimization [33]. For example, ML has been employed to identify therapeutic materials with optimal antioxidant capacity and to predict catalytic activity by linking adsorption energy to reaction kinetics [34, 35]. Moreover, integrating historical synthesis datasets allows ML to predict morphology, activity, and yield prior to experimental validation [36, 37, 38]. Collectively, these advances provide a rational and high‐throughput pathway for identifying high‐performance catalytic materials and elucidating the key determinants of their efficiency.

Bioorthogonal chemistry, the core achievement of the 2022 Nobel Prize in Chemistry, is characterized by its ability to proceed in living systems without perturbing native biochemical processes [39, 40]. Under the catalysis of transition metals, it enables the specific conversion of non‐toxic prodrugs to toxic therapeutic drugs at tumor sites, which represents a complementary approach to CDT for localized and selective tumor treatment [41, 42, 43]. Both strategies rely critically on the catalytic efficiency, stability, and environmental adaptability of the employed catalysts. In this study, we systematically mined and analyzed literature data on catalytic materials and constructed an ML‐based predictive framework to identify catalysts exhibiting high efficiency and robustness under complex TME conditions (Figure 1). A comprehensive dataset encompassing structural features, reaction conditions, and catalytic performance was curated to train the model, and SHapley Additive exPlanations (SHAP) analysis was used to interpret feature importance and elucidate key performance determinants. The model identified iron single‐atoms (Fe‐SAs) as optimal candidates, displaying outstanding chemodynamic activity and robust bioorthogonal catalytic capacity. This double catalytic mechanism enables in situ activation of doxorubicin (DOX) through bioorthogonal reactions while simultaneously modulating the TME via CDT to enhance local cytotoxicity. The synergistic integration of these two catalytic processes significantly amplifies therapeutic efficacy. This work not only validates the superior catalytic efficiency of single‐atom nanomaterials in catalytic therapy but also highlights the transformative potential of ML in guiding the rational design and discovery of next‐generation nanocatalysts for precision cancer therapy.

FIGURE 1.

Schematic of ML‐assisted discovery pipeline that nominates Fe‐N₅ SAs combining chemodynamic and bioorthogonal catalytic activity to achieve synergistic antitumor effects.

2. Results and Discussion

2.1. ML‐Driven Rational Design of Bifunctional Catalysts

We prioritized controllable in situ catalysis within the TME as the principal screening criterion for bifunctional metal catalysts. Because chemodynamic activity is strongly conditioned by lesion‐relevant TME indexes (notably pH and H₂O₂), the catalyst candidates featuring high peroxidase (POD)‐like activity under such conditions were preferable. To enable systematic data‐driven selection, we conducted a targeted literature survey to assemble a standardized database of POD‐mimetic materials, capturing both intrinsic descriptors and catalytic performance metrics. For structure‐property‐activity modeling, we selected descriptors that are consistently reported across studies, including material class, metal and non‐metal composition, particle size, micromorphology, surface functionalization, and support type. To enrich descriptor dimensionality, we derived secondary features (such as the number of distinct metal species and cumulative valence electron count) and included experimental covariates (such as solvent types and pH) that modulate reaction kinetics. We used the Michaelis‐Menten constant (K_m) as the primary activity metric, since the catalytic constant (K_cat) is ill‐defined for heterogeneous/ill‐characterized active sites, and maximum velocity (V_max) varies with catalyst loading and dispersion. By contrast, K_m more uniformly reflects apparent substrate affinity and permits robust cross‐study comparison.

Exploratory statistics revealed uneven feature distributions, in which most amples were monometallic, and Pt was the most frequently reported metal (Figure S1). Valence electron counts clustered predominantly in the 7–9 range (Figure 2a). Material class showed no significant univariate dependence on ln(1 + K_m) (Figure 2b), and pairwise correlations among descriptors and between descriptors and K_m were weak (Figure 2c), indicating latent, nonlinear relationships that require multivariate modeling. After encoding and preprocessing, the skewed and discrete K_m distribution was log‐transformed and binned to form classification labels, which stabilize small‐sample learning and enhance interpretability. Four classifiers including Random Forest, LightGBM, XGBoost, and Logistic Regression, were benchmarked using accuracy and F1‐score, and XGBoost achieved the best overall performance (Figure 2d,e; Figure S2), demonstrating that the assembled descriptors contain predictive signal for K_m and that gradient‐boosted trees effectively capture complex dependencies.

FIGURE 2.

ML‐driven screening of biocatalysts. (a) Circular bar chart showing representative feature distributions in the curated POD‐mimic dataset. (b) Raincloud plot of material classes versus ln‐transformed K_m. (c) Feature correlation heatmap (Pearson) that visualizes pairwise relationships among engineered descriptors. (d) Confusion matrix of classification outcomes (test set) for the evaluated algorithms. (e) Comparison bar plot of accuracy and F1‐score for the four benchmarked classifiers. (f) SHAP summary plot for the XGBoost model. (g) Average positive SHAP values for each descriptor, used to mitigate frequency bias and rank net positive contributions. (h) Predicted score comparison across material classes. (i) Predicted score comparison for selected metal elements.

For model interpretation, we employed SHAP to quantify each feature's marginal contribution. SHAP analysis identified SAs as the most influential material class, and carbon‐based architectures also scored highly, reflecting their prevalence in SAs constructs (Figure 2f). To correct for frequency bias, we computed average positive SHAP values per feature, which ranked SAs and Fe as the top contributors (Figure 2g). Partial dependence analysis corroborated these findings. SAs attained the highest predicted performance among material classes, and among metals, Fe > Ni > Pt (Figure 2h,i). These results nominate Fe‐based SAs as a potential platform for chemodynamic catalysis. Fe‐SAs combine maximal atomic utilization, atomically uniform and well‐defined active sites, and a finely tunable electronic environment, attributes that favor high intrinsic activity and selectivity. Moreover, iron‐centered catalysts have documented compatibility with bioorthogonal reduction chemistries, suggesting a strategic advantage for integrating chemodynamic and bioorthogonal functions.

2.2. Synthesis and Characterization of Fe‐N₅ SAs

Iron single‐atoms (Fe‐N₅ SAs) were synthesized through a two‐step pyrolysis strategy (Figure 3a). Transmission electron microscopy (TEM) revealed well‐defined rhombic dodecahedral particles with a diameter of ∼50 nm featuring subtle surface concavities (Figure 3b; Figure S3). High‐angle annular dark field‐scanning electron microscopy (HAADF‐STEM) further confirmed this morphology (Figure 3c). Energy‐dispersive X‐ray spectroscopy (EDS) elemental mapping demonstrated homogeneous distributions of C, N, O, and Fe elements throughout the structure (Figure 3d). Aberration‐corrected HAADF‐STEM image exhibited isolated bright spots corresponding to atomically dispersed Fe centers (Figure 3e). X‐ray diffraction (XRD) patterns of Fe‐N₅ SAs displayed two broad peaks at 25° and 43°, characteristic of graphitized carbon, and showed no detectable Fe‐related reflections, strongly supporting the absence of metallic Fe clusters (Figure S4) [44].

FIGURE 3.

Synthesis and characterization of Fe‐N₅ SAs. (a) Schematic of the synthesis process of Fe‐N₅ SAs. (b) TEM image of Fe‐N₅ SAs. (c) HAADF‐STEM image of Fe‐N₅ SAs. (d) Elemental mapping of Fe‐N₅ SAs. (e) Aberration‐corrected HAADF‐STEM image of Fe‐N₅ SAs. (f) Fe K‐edge XANES spectrum of Fe‐N₅ SAs. (g) EXAFS Fourier‐transformed spectrum of Fe‐N₅ SAs. (h, i) EXAFS curve‐fitting results in k‐space (h) and r‐space (i). (j) WT analysis of the Fe K‐edge.

To elucidate the electronic configuration and coordination environment of the Fe sites, Fe K‐edge X‐ray absorption spectroscopy (XAS) was performed. The X‐ray absorption near‐edge structure (XANES) spectra positioned the absorption edge of Fe‐N₅ SAs between those of FeO and Fe₂O₃, indicating an average Fe valence state between +2 and +3 (Figure 3f). This result is consistent with X‐ray photoelectron spectroscopy observations, where Fe 2p3/2 peaks were located at 710.3 and 713.2 eV, corresponding to Fe²⁺ and Fe³⁺ species, respectively. (Figure S5). Extended X‐ray absorption fine structure (EXAFS) analysis revealed a dominant peak at 1.5 Å, assignable to Fe‐N coordination, with no Fe–Fe scattering peak near 2.2 Å (Figure 3g). Wavelet transform (WT) analysis corroborated these findings, showing a single intensity maximum near 4.5 Å similar to iron phthalocyanine (Fe‐Pc), and no signal corresponding to Fe–Fe coordination (Figure 3j), further confirming atomic dispersion of Fe. Quantitative EXAFS curve fitting in both k‐ and r‐space yielded a coordination number of approximately 4.8 for the Fe centers, indicating a predominant Fe‐N₅ configuration (Figure 3h,i; Table S1). Based on these results, the atomically dispersed species were unambiguously identified as Fe‐N₅ SAs sites. Fe‐N₅ SAs exhibited remarkable physiological stability in H₂O, phosphate‐buffered saline (PBS), and cell culture medium. No visible aggregation or precipitation was observed after 3 days, and dynamic light scattering analysis showed a stable hydrodynamic diameter of 170–185 nm (Figures S6 and S7). TEM images further confirmed the absence of structural degradation (Figure S8). These results demonstrate the robust stability of Fe‐N₅ SAs, providing a solid basis for subsequent catalytic and biological evaluations.

2.3. Chemodynamic and Bioorthogonal Dual Catalysis of Fe‐N₅ SAs

The chemodynamic activity of Fe‐N₅ SAs was investigated using 3,3′,5,5′‐tetramethylbenzidine (TMB) as a chromogenic probe. In the presence of H₂O₂, Fe‐N₅ SAs rapidly oxidized TMB to blue‐colored oxidized TMB (ox‐TMB), producing a characteristic absorbance at 652 nm (Figure 4a,b). TMB alone or TMB + H₂O₂ showed negligible absorption at 652 nm, confirming that oxidation is catalyzed by Fe‐N₅ SAs. Addition of the ·OH scavenger tert‐butanol (TBA) markedly suppressed the 652 nm signal, implicating ·OH as the active oxidant. Electron paramagnetic resonance (EPR) spectroscopy further detected the 1:2:2:1 signature consistent with ·OH formation, corroborating H₂O₂ decomposition to ·OH by Fe‐N₅ SAs (Figure 4c). Kinetic analysis was performed by measuring time‐dependent absorbance changes at 652 nm with varying H₂O₂ concentrations to construct Michaelis‐Menten and Lineweaver‐Burk plots (Figure 4d,e; Figure S9). The apparent K_m and V_max were determined to be 40.45 mm and 0.25 µm s⁻¹, respectively. The pH‐dependent assays revealed the enhanced chemodynamic catalytic activity under weakly acidic conditions (Figure 4f), consistent with the acidic TME and implying preferential activation in lesions while limiting activity in normal tissues.

FIGURE 4.

Characterization of the catalytic performance of Fe‐N₅ SAs. (a) Schematic of ·OH detection using the chromogenic probe TMB. (b) UV–vis absorption spectra of TMB under different reaction conditions. (c) EPR spectrum of ·OH trapped by DMPO. (d) Michaelis‐Menten curve of steady‐state kinetic analysis with H₂O₂ as substrate (n = 3). (e) Lineweaver‐Burk plot of steady‐state kinetic analysis with H₂O₂ as substrate (n = 3). (f) pH dependence of POD‐like activity, monitored as absorbance at 652 nm (n = 3). (g) Schematic of bioorthogonal uncaging of Pro‐RH110 by Fe‐N₅ SAs. (h, i) Fluorescence spectra of Pro‐RH110 reaction solutions (h) without and (i) with Fe‐N₅ SAs (Ex = 488 nm). (j) CLSM images showing intracellular uncaging of Pro‐RH110 in 4T1 cells after different treatments (n = 3). (k) Quantification of intracellular fluorescence from (j) (n = 3). (l) Flow cytometry analysis of 4T1 cells after diverse treatments. Data are presented as mean values ± SD. Statistical analysis in (k) was performed using a two‐tailed Student's t‐test; significance is indicated as ^*** p < 0.001.

We next assessed bioorthogonal catalytic ability using an azide‐caged rhodamine 110, Pro‐RH110. Catalytic uncaging of Pro‐RH110 by Fe‐N₅ SAs produced rhodamine 110 (RH110, λ_em = 525 nm) (Figure 4g). In solution, the fluorescence increased continuously upon co‐incubation with Fe‐N₅ SAs, the reaction rate and final fluorescence scaled with Fe‐N₅ SAs concentration (Figure 4h,i; Figure S10). The fluorescence signal increased over time and gradually reached a plateau, exhibiting typical catalytic kinetic behavior (Figure S12). Compared with other iron‐based control materials, Fe‐N₅ SAs showed significantly enhanced catalytic activation efficiency toward Pro‐RH110, demonstrating their advantage in bioorthogonal in situ drug synthesis (Figure S11). To probe activity in a complex physiological milieu, Fe‐N₅ SAs and Pro‐RH110 was conducted on 4T1 murine breast cancer cells. Confocal imaging revealed strong intracellular fluorescence in the Fe‐N₅ + Pro‐RH110 group, whereas control groups showed minimal fluorescence signal (Figure 4j). Quantification indicated 4.5‐fold higher mean intracellular fluorescence versus Pro‐RH110 alone (Figure 4k), and flow cytometry confirmed the pronounced catalytic uncaging in treated cells (Figure 4l).

2.4. In Vitro Anti‐Tumor Efficacy of Fe‐N₅ SAs

Motivated by the potent chemodynamic and bioorthogonal dual catalysis activities of Fe‐N₅ SAs, we evaluated their antitumor effects in vitro. Cell counting kit 8 (CCK‐8) assays on human umbilical vein endothelial cells (HUVECs) and 4T1 cells revealed a marked selectivity, in which Fe‐N₅ SAs showed negligible toxicity toward HUVECs (more than 80% viability at 50 µg·mL^{− 1}), whereas 4T1 cells' viability decreased in a dose‐dependent manner to 42.6% at the same concentration (Figure 5a,b). This differential toxicity was consistent with elevated H₂O₂ levels and mild acidosis in tumor cells, which promote Fe‐N₅‐mediated H₂O₂ decomposition to cytotoxic ·OH radicals.

FIGURE 5.

Evaluation of combined therapeutic efficacy in vitro. (a) Cell viability of HUVECs treated with different concentrations of Fe‐N₅ SAs (n = 4). (b) Cell viability of 4T1 cells treated with various concentrations of Fe‐N₅ SAs (n = 4). (c, d) Cell viability of 4T1 cells treated with (c) free DOX and (d) Pro‐DOX at various concentrations (n = 4). (e) Cell viability of 4T1 cells after different treatments (n = 4). (f) Calcein‐AM/PI staining fluorescence images of 4T1 cells after various treatments. (g) CLSM images of intracellular ROS in 4T1 cells following various treatments. (h) CLSM images of mitochondrial membrane potential assessed by JC‐1 staining after different treatments (n = 3). (i) Quantification of JC‐1 green/red fluorescence ratios from h) (n = 3). (j) CLSM images of LPO using BODIPY 581/591 C11 in treated 4T1 cells. (k) Immunofluorescence images of γ‐H2AX showing DNA damage across treatment groups. (l) Schematic diagram of combined treatment with Fe‐N₅ SAs and Pro‐DOX. Statistical analysis in (e, i) was performed using a two‐tailed Student's t‐test; significance is indicated as ^** p < 0.01 and ^*** p < 0.001.

To combine CDT with on‐demand chemotherapy, we prepared an azide‐caged doxorubicin prodrug (Pro‐DOX) (Figure S13). The cytotoxicities of Pro‐DOX and free DOX were evaluated in both 4T1 cells and HUVECs. Pro‐DOX remained inactive in vitro toward 4T1 tumor cells, where 20 µm Pro‐DOX produced minimal cytotoxicity compared with free DOX, confirming effective masking of DOX activity (Figure 5c,d). Similarly, Pro‐DOX exhibited negligible toxicity toward HUVECs, while free DOX caused obvious cytotoxicity (Figure S14). These results indicate that Pro‐DOX remains largely inert prior to activation, whereas free DOX shows nonselective toxicity. Compared with using Fe‐N₅ SAs alone, co‐treatment of 4T1 cells with Fe‐N₅ SAs and Pro‐DOX enhances the killing effect, reducing cell viability to 32.6%. (Figure 5e). Live/dead calcein‐acetoxymethyl ester (AM)/propidium iodide (PI) staining corroborated the enhanced cell death in the combined treatment group (Figure 5f), supporting a model in which Fe‐N₅ SAs both generate ROS and catalytically unmask Pro‐DOX to potentiate cytotoxicity.

The mechanistic assays link cell death to oxidative stress and mitochondrial dysfunction. 2′,7′‐Dichlorodihydrofluorescein diacetate (DCFH‐DA) staining revealed robust intracellular ROS in Fe‐N₅ and Fe‐N₅ + Pro‐DOX groups (Figure 5g), indicating that Pro‐DOX activation didn't impair chemodynamic ROS generation. 5,5′,6,6′‐Tetrachloro‐1,1′,3,3′‐tetraethyl‐imidacarbocyanine (JC‐1) imaging showed loss of mitochondrial membrane potential following Fe‐N₅ SAs treatment, with more pronounced disruption upon combined therapy (Figure 5h,i), plausibly reflecting additive mitochondrial damage from ROS and generated DOX [45].

Because excessive ROS can deplete intracellular glutathione and provoke lipid peroxides (LPO), we then probed the ferroptosis markers. BODIPY 581/591 C11 staining displayed a substantial shift from red to green in Fe‐N₅ and Fe‐N₅ + Pro‐DOX groups, indicating elevated LPO accumulation consistent with ferroptotic processes (Figure 5j). Concurrently, DNA damage assessed by γ‐H2AX immunofluorescence was markedly increased in the Fe‐N₅ + Pro‐DOX group, reflecting combined oxidative strand breaks and DOX‐mediated topoisomerase II inhibition (Figure 5k). These data indicate that Fe‐N₅ SAs trigger tumor cell death via two synergistic catalytic pathways, in which chemodynamic generation of ROS drives mitochondrial dysfunction and LPO/ferroptosis, and local bioorthogonal uncaging of Pro‐DOX produces DNA damage and augments cytotoxicity (Figure 5l).

2.5. RNA‐seq Reveals Pathways Underlying Fe‐N₅ SAs‐Mediated Dual Catalysis

To elucidate the molecular basis of tumor suppression by Fe‐N₅ SAs combined with Pro‐DOX, we performed RNA‐seq on 4T1 cells treated with PBS or Fe‐N₅ + Pro‐DOX. Transcriptome profiling revealed that 2,534 differentially expressed genes (DEGs) were detected, of which 1,832 were upregulated and 702 downregulated in the treated group (Figure 6a,b). Kyoto encyclopedia of genes and genomes (KEGG) pathway enrichment analysis highlighted perturbations in glutathione metabolism, chemical carcinogenesis, DNA adducts, TNF signaling, and notably ferroptosis, which showed the strongest enrichment (Figure 6c). Construction of a protein–protein interaction (PPI) network and centrality analysis identified hub genes tightly connected to ferroptotic regulation, including Hmox1, Sqstm1, Nqo1, Plin2, Slc7a11 and Gclc (Figure 6d). Guided by these results, we interrogated expression changes of canonical ferroptosis genes such as Hmox1, Slc7a11, Gclc, Ftl1, Slc3a2, Fth1 and Gclm were significantly upregulated, whereas Cp and Acsl6 were downregulated (Figure 6e).

FIGURE 6.

Mechanistic dissection of cell death induced by combined catalytic therapy. (a) Heatmap showing global transcriptional remodeling after combined treatment. (b) Volcano plot of DEGs between the treatment and control groups. (c) KEGG pathway enrichment of DEGs between the treatment and control groups. (d) PPI network of DEGs. (e) Correlation network map of canonical ferroptosis genes. (f) Circular representation of GO enrichment showing major biological processes perturbed by treatment. (g) Western blot analysis of GPx4 protein levels in 4T1 cells after various treatments.

Given the pivotal role of GPx4 in ferroptosis suppression, we evaluated its protein level by Western blot. GPx4 was markedly decreased after Fe‐N₅ and Fe‐N₅ + Pro‐DOX treatment, consistent with transcriptomic indications of impaired lipid peroxide detoxification (Figure 6g). Complementary gene ontology (GO) analysis revealed enrichment of terms related to DNA replication and transcriptional regulation involving sequence‐specific DNA binding by RNA polymerase II and DNA‐binding transcription factor activity, implicating concomitant genotoxic stress (Figure 6f, Table S2). The multi‐omic data support a dual‐mechanism model in which Fe‐N₅ SAs‐mediated H₂O₂ conversion amplifies intracellular ROS and lipid peroxidation, driving ferroptosis via GPx4 suppression and dysregulated glutathione metabolism, while in situ uncaging of DOX provokes DNA replication/transcriptional perturbation and genotoxic damage.

2.6. In Vivo Antitumor Efficacy of Fe‐N₅ SAs

Guided by the potent in vitro performance, we evaluated the therapeutic efficacy of Fe‐N₅ SAs combined with Pro‐DOX in 4T1 tumor‐bearing mice. Prior to therapy, comprehensive biosafety screening indicated excellent biocompatibility. The hemolysis rate of Fe‐N₅ SAs remained below 5% at 50 µg·mL⁻¹ (Figure S15). In addition, complete blood counts and serum biochemistry analyses after treatment with Fe‐N₅ SAs, Pro‐DOX, and their combination showed no treatment‐related abnormalities (Figure S16). Furthermore, hematoxylin and eosin (H&E) stained sections of major organs revealed no evident histopathological changes (Figure S17). Meanwhile, the in vivo distribution and metabolic behavior of Fe‐N₅ SAs were further investigated. Fe‐N₅ SAs efficiently accumulated in tumor region, reaching a peak level of 0.039 µg/mg at 12 h post‐administration, followed by gradual metabolism and clearance (Figure S18). These results support the suitability of Fe‐N₅ SAs and Pro‐DOX for in vivo administration. Furthermore, given that sodium ascorbate (NaAsc) was used as the reducing agent for the bioorthogonal reaction in vivo, a control experiment including PBS and NaAsc groups was conducted. The results showed that NaAsc alone exhibited no tumor inhibitory effect (Figure S19).

When tumors reached around 100 mm³, the tumor‐bearing mice were randomized to four groups: I) PBS, II) Pro‐DOX, III) Fe‐N₅, and IV) Fe‐N₅ + Pro‐DOX (Figure 7a). Compared with PBS, Fe‐N₅ slowed the tumor progression, whereas the combined Fe‐N₅ + Pro‐DOX regimen produced the most pronounced tumor growth suppression (Figure 7b,c), with tumor inhibition rates showing 41.3% for Fe‐N5, 70.1% for the combined regimen, and no significant inhibition in the Pro‐DOX group (Figure S20). Ex vivo tumor and tumor weights also corroborated this trend (Figure 7d,e). Histopathology and molecular staining confirmed enhanced antitumor activity after combination treatment. Tumors from the Fe‐N₅ + Pro‐DOX group displayed extensive necrosis and apoptosis, markedly reduced proliferative index (Ki‐67), and substantially decreased GPx4 expression, which is indicative of ferroptotic engagement (Figure 7f). Survival analysis demonstrated that combined catalytic therapy significantly prolonged median survival relative to controls (Figure 7g). Importantly, no significant body‐weight loss occurred in any group (Figure 7h), and organ histology remained unremarkable (Figure S21), indicating favorable systemic tolerability.

FIGURE 7.

In vivo evaluation of combined catalytic therapy. (a) Therapeutic schedule for 4T1 tumor‐bearing mice. (b) Individual tumor growth trajectories for mice in each treatment group. (c) Grouped tumor‐volume curves (n = 5). (d) Representative photographs of excised tumors collected at study end after various treatments. (e) Tumor weights of mice in different groups at the end of treatment (n = 5). (f) H&E, Ki‐67 immunohistochemical, TUNEL immunofluorescence, and GPx4 immunohistochemical staining of tumor sections after different treatments. (g) Body weight changes during treatment. (h) Survival curves of mice in various treatment groups (n = 5). Data are presented as mean values ± SD. Statistical analysis in (c) was performed using two‐way analysis of variance (ANOVA), and that in (e) was performed using two‐tailed Student's t‐test; significance is indicated as ^** p < 0.01.

3. Conclusion

This work integrates ML with experimental data mining to uncover structure‐activity relationships of transition metals for antitumor applications. A high‐precision predictive model, interpreted via SHAP analysis, identified key features governing catalytic performance, highlighting Fe‐SAs as the optimal candidates. The experimental validation of Fe‐N₅ SAs demonstrated dual catalytic functionality in the tumor microenvironment, including bioorthogonal uncaging of DOX for in situ tumor‐targeted chemotherapeutic activation and efficient H₂O₂ decomposition to generate ·OH for oxidative stress‐mediated tumor cell damage. These synergistic processes enhance both the precision and efficacy of antitumor therapy. The atomically defined active sites of Fe‐N₅ SAs maintain high catalytic efficiency while maximizing metal utilization and minimizing nanomaterial‐associated toxicity. This study showcases the power of ML‐assisted catalyst design to accelerate discovery, reduce development costs, and establish single‐atom catalysts as promising platforms for next‐generation, high‐efficiency, low‐toxicity antitumor therapies.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting File: adma73689‐sup‐0001‐SuppMat.docx.

Data Availability Statement

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