Fe metal–organic framework–derived heterojunction for metabolic diagnosis of thymic epithelial tumor

Significance

Thymic epithelial tumors (TETs) are difficult to diagnose due to complex tissue features, and current metabolic tools have poor sensitivity and require complex sample preparation. We developed a Fe-MOF-derived heterojunction using smart pyrolysis, combining Fe₃ O₄ nanoparticles and an S-scheme structure. This material boosts metabolic signals by 1,000-fold in seconds for LDI MS without preprocessing. Its special design prevents electron–hole recombination, improving LDI efficiency by controlling charge separation. It accurately identifies TETs from benign controls (AUC = 0.960) and classifies risk levels (AUC = 0.856), outperforming the current imaging methods (AUC = 0.627). This tool enables fast and reliable metabolic diagnosis to improve TET classification and support personalized treatment.

Abstract

Thymic epithelial tumors (TETs), rare yet clinically significant malignancies, face diagnostic challenges due to their occult presentation and lack of noninvasive risk-stratification tools, leading to systemic overtreatment and poor prognoses for high-risk subtypes. To address this unmet need, we developed a Fe₃O₄@Fe metal–organic framework heterojunction-enhanced laser desorption ionization mass spectrometry (LDI MS) platform for the efficient analysis of serum metabolic fingerprints (SMFs). Engineered through gradient pyrolysis, this nanomaterial synergizes ultraviolet absorption and photothermal conversion from its two constituent components with enhanced charge separation, achieving 1,000-fold improvement in sensitivity and thus enabling direct SMF acquisition from 1 μL of serum. Coupled with machine learning, the platform demonstrates robust diagnostic performance, yielding area under the curve (AUC) of 0.960 for distinguishing TETs from benign control and AUC of 0.856 for hierarchical risk stratification, outperforming clinical workflows. Beyond advancing material design for LDI MS, this work establishes a clinically translatable framework for rapid, large-scale screening, addressing critical gaps in TET management through metabolic-driven stratification.

Thymic epithelial tumors (TETs), encompassing thymomas and thymic carcinomas (1), represent the predominant neoplasms of the anterior mediastinum, characterized by a complex pathological phenotype (2). While low-risk (LR) TETs generally exhibits favorable prognosis, with a 5-y survival rate reaching 78%, high-risk (HR) TETs demonstrates worse prognosis due to their propensity for lymph node metastasis and distant dissemination (5-y survival rates of only approximately 40%) (3). Current management relies on imaging modalities like computed tomography (CT) for initial screening; however, their limited specificity necessitates labor-intensive multidimensional evaluations to differentiate thymic masses (4). In addition, histopathological biopsies, though serving as gold standard, encounter practical limitations including patient resistance to invasive sampling and histomorphological ambiguities between TETs and non-neoplastic lymphocytes (5). The time-consuming detection methods and low specificity often lead to the overtreatment of low-risk TETs. Despite being confined locally and potentially requiring only adjuvant therapy, these patients are sometimes subjected to unnecessary surgical interventions, which imposes substantial healthcare burdens (6).

Metabolites, representing terminal small-molecule products of biological pathways, provide real-time insights into dynamic physiological states (7). In TETs, aberrantly proliferating thymocytes exhibit tumor intrinsic endocrine activity, disrupting systemic metabolic homeostasis, and manifesting in paraneoplastic syndromes like myasthenia gravis. This tumor-driven metabolic reprogramming positions metabolic profiling as a promising tool for both accurate diagnosis and mechanistic exploration (8). Matrix-assisted laser desorption/ionization mass spectrometry (MALDI MS) has emerged as a promising omics platform, enabling rapid, high-throughput analysis with minimal sample preparation. Yet, its application has been limited to proteomics due to challenges in small-molecule detection (9). Conventional organic matrices, such as α-cyano-4-hydroxycinnamic acid (CHCA) and 2,5-dihydroxybenzoic acid (DHB), suffer from high background interference in the low molecular weight range (100 to 1,000 Da), which are critical limitations for metabolite analysis (10).

Recent advances in nanoscale matrices address these limitations through engineered material interfaces. Two-dimensional metal–organic framework nanosheets (2D MOF NSs), for instance, exhibit unfavored self-decomposition under laser ablation, thereby reducing the generation of background interference peaks (11). In addition, carboxyl groups in aromatic acid ligands within MOFs further enhance cation adduction efficiency through proton transfer mediation (12). Nevertheless, pristine MOFs still face practical constraints from rapid electron–hole recombination and suboptimal energy dissipation during the LDI process. Current strategies primarily aim to construct noble metal/metal oxide heterostructures through extrinsic metallophilic functional group modifications or conventional thermal agitation methods to establish synergistic host–guest interfaces. However, these approaches are constrained by either complex multistep procedures or weak interfacial coupling, hindering the efficient design of robust heterojunction architectures with precise interfacial control (13–16). A particularly promising approach involves the controlled partial pyrolysis of MOFs to engineer interfacial heterojunctions and defects, a method that enhances photogenerated charge carrier separation and may significantly improve LDI performance for precision medical applications (16, 17).

In this work, we developed a gradient pyrolysis protocol (350 to 500 °C) to construct Fe MOF NSs (denoted as Fe-MOF) derived heterojunction. These heterostructures incorporate magnetite nanoparticles (Fe₃O₄ NPs) with high photothermal conversion efficiency, enhancing interfacial ionization performance (Fig. 1). The optimized Fe₃O₄@Fe-MOF heterojunction demonstrated a maximum 1,000-fold signal amplification in metabolite detection, enabling direct acquisition of serum metabolic fingerprints (SMFs) with ultralow sample consumption (1 μL). Through machine learning of SMF patterns, we constructed a robust diagnostic platform for distinguishing TETs from benign control (BC) with an area under the curve (AUC) of 0.960. Furthermore, a panel comprising 12 metabolites was identified for hierarchical risk stratification with AUC of 0.856. This work not only advances the rational design of advanced materials for LDI MS but also establishes a transformative platform for clinical metabolomics toward precision oncology.

Fig. 1.

Schematic workflow for metabolic diagnosis of thymic epithelial tumors (TETs) patients. (A) Thermally engineered Fe₃O₄@Fe metal–organic frameworks (MOF, FM-450). (B) The diagnostic workflow involved the extraction of serum metabolic fingerprints (SMFs) from TET cohort using FM-450 as matrix in laser desorption/ionization mass spectrometry (LDI MS), and machine learning of SMFs for TET diagnosis.

Results

Design and Characterization of Materials.

As a precursor material, Fe-MOF was first synthesized via a modified sonication-assisted water bath method (Fig. 2A, see detailed in Materials and Methods). We then characterized the crystalline structure of the prepared Fe-MOF. The powder X-ray diffraction (PXRD) pattern of Fe-MOF showed well-defined diffraction peaks centered at 9°, which aligned with the simulated structure (SI Appendix, Fig. S1A). Subsequently, transmission electron microscopy (TEM) and atomic force microscopy images indicated that the as-prepared Fe-MOF exhibited an irregular sheet shape with a thickness of ca. 1.2 nm (Fig. 2 B and C). These observations are consistent with prior literature and confirm the successful synthesis of the precursor material (11). Additionally, X-ray photoelectron spectroscopy (XPS) experiment was carried out to study the elemental composition and surface chemical state of Fe-MOF. The survey spectra (SI Appendix, Fig. S1B) verified the existence of Fe, C, and O elements. In particular, the Fe 2p spectra (SI Appendix, Fig. S1C) revealed that Fe-MOF exhibited a mixed valence state of Fe (II) and Fe (III), which is crucial for the growth of Fe₃O₄ NPs during the subsequent pyrolysis. It was worth noting that the presence of Fe (III) may arise from the partial fracture of Fe-O on the surface of the Fe-MOF structure during ultrasound and centrifugation processing, resulting in the coordination of unsaturated sites (18).

Fig. 2.

Preparation and characterization of Fe-MOF and FM-X (X = 350, 400, 450, 500). (A) Schematic illustration of the preparation of Fe-MOF and its partial pyrolysis to form Fe₃O₄@Fe-MOF. (B) TEM, (C) atomic force microscopy images, and (D) thermogravimetric analysis curve of Fe-MOF. (E) TEM and (F) high-resolution TEM (HRTEM) of FM-450. (G) Powder X-ray diffraction (PXRD) patterns, (H) Fourier transform infrared spectra, (I) Fe 2p XPS spectra, and (J) O 1 s XPS spectra of Fe-MOF and FM-X (X = 350, 400, 450, 500).

The pyrolysis of Fe-MOF was carried out under incremental temperatures (350 to 500 °C) to obtain the final matrix materials. Thermogravimetric analysis of Fe-MOF (Fig. 2D) revealed that the MOF decomposed within 350 to 500 °C, prompting further investigation into the underlying chemical changes. The pyrolysis behavior was subsequently characterized by TEM and HRTEM. Below 350 °C, Fe-MOF retained its structural integrity under argon atmosphere (SI Appendix, Fig. S2). As the temperature increased from 350 to 400 °C, NPs of 5 to 10 nm formed on the MOF surface (Fe₃O₄@Fe-MOF, denoted as FM-350/400, SI Appendix, Fig. S3 A and B), creating unique host–guest heterointerfaces. These NPs exhibited lattice fringes with d-spacings of 0.253, 0.297, and 0.148 nm, corresponding to the (311), (220), and (440) planes of Fe₃O₄ (18, 19). Critically, Fe₃O₄ NPs reached to a maximum size of 20 nm at 450 °C while maintaining the heterojunction (denoted as FM-450, Fig. 2 E and F). However, above 500 °C, MOF decomposition occurred, leading to the aggregation of Fe₃O₄ NPs (denoted as FM-500, SI Appendix, Fig. S3C).

The material structure is closely associated with ionization efficiency, which determines the performance of the matrix. To investigate the influence of the pyrolysis temperature on the crystalline structure, the PXRD pattern of the Fe-MOF derivatives were measured (Fig. 2G). In FM-350 and FM-400, five new weak peaks appeared at 30°, 35°, 43°, 57°, and 62°, that can be well-indexed to inverse spinel structured Fe₃O₄ (JCPDS card No. 99-0073) (20). For FM-450, the five characteristic peaks were sharper and the first diffraction peak broadened, confirming the growth of Fe₃O₄ NPs and the partial decomposition of the MOF structure. In contrast, only Fe₃O₄ NPs were observed in the FM-500, with the MOF structure completely disappearing. These changes were further characterized by Fourier transform infrared (FTIR) spectra (Fig. 2H). The typical MOF peaks corresponding to O=C–O (1,374 cm⁻¹) and 1,4-substituted benzene core (747 cm⁻¹) which are initially observed in Fe-MOF and FM-X (X = 350, 400, 450), disappeared in FM-500. Only the Fe–O stretching vibration (550 cm⁻¹) remained, confirming complete MOF degradation and exclusive Fe₃O₄ formation.

Additionally, XPS experiments were conducted to investigate the chemical states and coordination environment pyrolyzed at different temperatures. In the Fe 2p spectra (Fig. 2I), peaks with binding energies of 709.6 and 722.7 eV were assigned to Fe²⁺, while 710.8 and 723.9 eV were attributed to Fe³⁺ (20, 21). Notably, below 450 °C, as the pyrolysis temperature increased, the binding energies of both Fe²⁺ and Fe³⁺ gradually shifted toward lower values, implying the increase of oxygen defects to balance the decrease of the Fe valence state in pristine MOF (22, 23). However, when the temperature reached 500 °C, the binding energies of Fe²⁺ and Fe³⁺ increased. We attributed this shift to the collapse of the MOF structure, which unexpectedly led to a reduction in oxygen defects. The O 1s XPS spectra of Fe-MOF (Fig. 2J) at 530.07, 531.20, 531.89, and 532.52 eV corresponded to Fe-O, oxygen vacancies (O_v), O=C–O, and O–H, respectively (24, 25). The reduced area of the characteristic peak of O=C–O indicated the partial destruction of the carboxyl group in MOF, while the O_v characteristic peak area initially increased and then decreased, consistent with previous results (SI Appendix, Table S1).

Collectively, the TEM, PXRD, FTIR, and XPS characterization elucidated the structural-chemical evolution of MOF derivatives during the pyrolysis process. These analyses demonstrated a coherent interplay between Fe-MOF degradation and Fe₃O₄ NPs crystallization. Notably, the synthetic strategy combining ultrasound-assisted exfoliation with mild hydrothermal conditions achieves precise control over NS architectures while significantly lowering energy consumption compared to conventional solvothermal methods. This approach not only avoids energy-intensive high-temperature processes but also enables the tailored formation of hierarchical nanostructures with well-defined interfaces and defect distributions. As a result, the presence of two-phase interfaces, multiple valence states, and oxygen defects could enhance ion/electron transport kinetics and reduce charge transfer resistance, thereby facilitating the subsequent LDI process (11, 26).

Performance of FM-450 in LDI MS Analysis.

To evaluate material performance and identify the optimal candidate, we tested Fe-MOF and FM-X (X = 350, 400, 450, 500) using five standard metabolites [alanine (Ala), lysine (Lys), arginine (Arg), glucose (Glc), and sucrose (Suc)] typically found in human blood. Signal intensities initially increased for FM-350-450 and then declined with rising temperature for FM-500 (Fig. 3A and SI Appendix, Fig. S4A), which can be attributed to heterojunction effects and variations in defect levels. Notably, FM-450 afforded the significant enhancement of ca. two- to 130-folds over other materials, making it the best candidate among five materials for metabolite detection. Performance of FM-450 in LDI MS analysis was further validated in high salt and protein environments by mixing metabolites with 5 mg mL⁻¹ BSA or 0.5 M NaCl. The Na-adducted signals from five metabolites were all clearly detected when using FM-450 as the matrix (Fig. 3B). In contrast, few characteristic peaks were identified without FM-450 (SI Appendix, Fig. S4B). This performance arises from the mesoporous structure of FM-450, which selectively adsorbs metabolites via uniform pores (2 nm, SI Appendix, Fig. S4C) while dispersing salts to suppress self-crystallization interference (27, 28).

Fig. 3.

Detection of metabolites using FM-450-assisted LDI MS. (A) Mean intensities of sodium adducted metabolite signals ([M+Na]⁺) from five independent LDI MS experiments using Fe-MOF and FM-X (X = 350, 400, 450, 500) as matrices in position ion mode. Alanine (Ala), lysine (Lys), arginine (Arg), glucose (Glc), and sucrose (Suc) were prepared at a concentration of 1 mg mL⁻¹ each. (B) Typical mass spectra of five metabolites in the presence of NaCl (0.5 M) and bovine serum albumin (5 mg mL⁻¹) using FM-450 as the matrix. (C) Intensities of sodium adducted metabolite signals ([M+Na]⁺) for 1 mg mL⁻¹ Ala, Lys, and Suc across 10 experiments using FM-450, CHCA, and DHB as the matrices. (D) Coefficients of variation (CVs) of Na-adducted signals from Ala, Lys, and Suc across 10 experiments using FM-450, CHCA, and DHB as the matrices. (E) Intrabatch and interbatch CVs of Na-adducted signals from five metabolites using FM-450 as the matrix.

In addition, we systematically compared the dual-component FM-450 with single-component MOFs (Fe-MIL-101 and Fe-MOF-74) to revalidate the advantages of MOF-derived heterojunctions. As shown in SI Appendix, Fig. S5 in the Supporting Information, FM-450 exhibits a remarkable 10- to 50-fold signal enhancement compared to its monocomponent counterparts, which could be attributed to the spatially separated charge carriers within the heterojunction architecture that synergistically promoted metabolite ionization and detection efficiency through enhanced charge transfer mechanisms.

Then, we benchmarked FM-450 against commercial organic matrices (CHCA, DHB) across two metrics: sensitivity and reproducibility. For sensitivity, FM-450 yielded a significant increase in signal intensities up to 1,000-fold than that of CHCA and DHB (Fig. 3C). Furthermore, FM-450 produced the lowest limit of detection, reaching down to approximately pmol levels (SI Appendix, Fig. S6 and Table S2). For reproducibility, the CVs for three molecules (Ala, Lys, Suc) were calculated and ranged from 10.7 to 13.2% using FM-450 as the matrix, lower than that for CHCA (29.6 to 97.0%) and DHB (31.3 to 86.3%) (Fig. 3D). Additionally, we investigated the reproducibility of FM-450 in both intrabatch (five replicates of five molecules) and interbatch (five replicates of five molecules) analyses, with CVs of 4.1 to 15.0% and 5.7 to 12.8%, respectively (Fig. 3E), achieving highly repeatable spectra.

Conventional organic matrices (e.g., CHCA, DHB) suffer from high background noise and poor cocrystallization uniformity, limiting their sensitivity and reproducibility (29–31). The favorable reproducibility of FM-450 is attributed to its uniform dispersion, which prevents uneven cocrystallization with analytes (SI Appendix, Fig. S7). Although many monocomponent MOFs have been developed as matrices for MALDI MS enabling highly reproducible mass spectra, a universal matrix with high sensitivity remains elusive due to their limited charge separation efficiency (32, 33). In comparison, FM-450 demonstrates superior sensitivity and selectivity for metabolic analysis directly from complex biosamples. We attributed the enhanced performance of the FM-450 matrix to the synergistic effects of its dual components, which distinguish it from monocomponent materials. Therefore, a deeper investigation into the underlying enhancement mechanisms will be the focus of our next phase of research.

LDI Mechanism of FM-450 for Metabolite Detection.

The exceptional metabolite detection performance of FM-450 stems from its optimized ionization and desorption dynamics, mainly driven by photoelectric and photothermal conversion capabilities. For the photoelectric ability, ultraviolet-visible (UV–Vis) spectroscopy first revealed that FM-X (X = 350 to 500) materials exhibit strong absorption within the 300 to 400 nm range (Fig. 4A), aligning with the 355 nm wavelength of the Nd:YAG laser used in LDI MS. Subsequently, photocurrent measurements were conducted to investigate the disparities in photoelectric properties among various materials. As a result, FM-450 exhibited significantly enhanced photoelectric conversion efficiency. This was reflected in its photocurrent density, which was twice that of pristine Fe-MOF (Fig. 4B). We attributed it to interfacial heterojunctions that accelerated charge migration (34, 35). For the photothermal ability, an infrared thermometer was used to measure temperature alterations in Fe-MOF and FM-X (X = 350 to 500) under irradiation with an ultraviolet lamp. FM-450 achieved the highest rate of temperature rise (Fig. 4C and SI Appendix, Fig. S8), suggesting defects and Fe₃O₄ NPs synergistically enhanced localized thermal effects. This property promotes hot electron emission from FM-450, driving thermal desorption of adjacent metabolites. Although Fe-MOF reached a higher absolute temperature, its slower heating kinetics limited its utility in LDI MS, where rapid energy transfer is critical for efficient analyte release (36).

Fig. 4.

Mechanistic investigation of FM-450 in metabolite detection. (A) Ultraviolet-visible absorption spectra and (B) photocurrent-time curves of Fe-MOF and FM-X (X = 350, 400, 450, 500). (C) Temperature variations of Fe-MOF and FM-X (X = 350, 400, 450, 500) in air under the irradiation of a 365 nm ultraviolet lamp. (D) Schematic illustration of charge transfer route in FM-450 via the S-scheme heterojunction. (E) LDI MS results of Suc in 50% juglone solutions using Fe-MOF and FM-X (X = 350, 400, 450, 500) as the matrices.

Besides photoelectric and photothermal conversion capabilities, the potency of charge separation and transfer in the LDI process was also of considerable importance. To verify the charge separation capability of FM-450, we analyzed the energy band structures of Fe-MOF and FM-500 (i.e., Fe₃O₄) using UV–Vis diffuse reflectance spectroscopy and Mott–Schottky measurements. Both materials exhibited consistent broad absorption in the range of 300 to 400 nm (SI Appendix, Fig. S9 A and B), with energy gaps (E_g) determined as 2.24 and 1.68 eV according to the corresponding Tauc plots (SI Appendix, Fig. S9 C and D). Mott–Schottky analysis confirmed their n-type semiconductor characteristics (37), revealing conduction band potentials (E_CB) of −1.15 and −0.35 V vs. Ag/AgCl electrode (SI Appendix, Fig. S10). Combining these results, valence band potentials (E_VB) were calculated as 1.29 V for Fe-MOF and 1.53 V for FM-500. Work function analysis (SI Appendix, Fig. S11) showed a higher Fermi level (E_F) for Fe-MOF (6.14 eV) than FM-500 (8.16 eV), driving electron transfer from Fe-MOF to Fe₃O₄ NPs during in situ formation of the two-phase system (SI Appendix, Table S3, see detailed calculation in Materials and Methods). This created a built-in electric field (Fe-MOF → Fe₃O₄, Fig. 4 D, ii). In particular, the interfacial energy band of Fe-MOF bent upward, while that of Fe₃O₄ NPs bent downward, ultimately forming the step-scheme heterojunction (S-scheme heterojunction) (38). Under LDI MS laser irradiation, photogenerated electrons in Fe₃O₄ conduction band recombined with holes in Fe-MOF valence band (Fig. 4 D, iii), while the retained holes in Fe₃O₄ and electrons in Fe-MOF facilitated efficient charge separation.

To further identify the dominant charge carrier in the LDI process, we evaluated FM-450's performance using an electron scavenger (juglone) and a hole scavenger [ethylenediaminetetraacetic acid (EDTA)]. As shown in SI Appendix, Fig. S12, metabolite (Ala, Lys, Arg, Glc, Suc) signals increased with rising juglone concentrations but decreased as the concentration of EDTA increased. This confirms that holes act as critical drivers in the positive-ion mode through promoting the ionization of cationic analytes (39, 40). Furthermore, an equivalent quantity of juglone was introduced to benchmark the efficacy of Fe-MOF and FM-X (X = 350 to 500). Notably, FM-450 demonstrated superior signal intensity over other materials under equivalent juglone conditions (Fig. 4E). This finding suggested its enhanced capacity to generate active holes for metabolite detection.

Collectively, the integration of Fe₃O₄ NPs into the Fe-MOF achieves three advancements: 1) extended light absorption matching the laser wavelength, 2) defect-mediated photothermal effects enabling rapid desorption, and 3) S-scheme charge separation prolonging carrier lifetimes. These attributes address the longstanding limitations of monocomponent MOF matrices, where photogenerated electrons and holes exhibit extremely short lifetimes ranging from picoseconds to nanoseconds due to rapid recombination, which significantly hampers the efficiency of metabolite detection (41, 42).

SMF for TET Diagnosis.

Serum samples from 163 TET patients and 80 BCs (including 14 thymic cysts and 66 healthy individuals) were collected. Based on a pilot study dataset of 16 samples (8/8, TETs/BCs), the minimum sample size of 10 (5/5, TETs/BCs) was determined to achieve a predicted statistical power > 0.8 at FDR of 0.1 for meaningful machine learning (SI Appendix, Fig. S13A), which can be at sufficient confidence level to conclude the statistical meaningful results according to previous references (43, 44). Notably, our cohort represents one of the largest metabolic studies of TETs published to date (SI Appendix, Fig. S13B) (45, 46). TET classification strictly adhered to the 2021 World Health Organization Classification of Tumors of the Thymus and Mediastinum (47), categorizing tumors into types A, AB, B1, B2, B3, and type C (thymic carcinoma). Based on prognostic risk stratification, types A, AB, and B1 were classified as LR, while B2, B3, and C constituted the HR group (SI Appendix, Table S4) (48). Demographic, clinical, and medical history data were retrieved from electronic health records (Fig. 5A).

Fig. 5.

Characterization of TET-associated SMFs. (A) Demographic profile of 80 benign controls (BC, including 14 thymic cyst patients and 66 healthy individuals) and 163 patients (70 LR TETs and 93 HR TETs). SMFs with 698 features were generated for each sample (Created in https://BioRender.com). (B) 3D principal component analysis (PCA) score plot of 698 SMFs by different TET risks showed overlap among BC (benign control, green), LR (orange), and HR (red) groups. The cumulative explained variance corresponds to the first three components, which explain 66.2%, 12.7%, and 5.3% of the total variance, respectively. (C) Performance evaluation of different machine learning algorithms in distinguishing between BC and TET patients. Receiver operating characteristic (ROC) showing areas under curves for the (D) discovery and (E) validation cohort. (F) Differentiation between BC and TET patient groups by the LASSO model. (G) Venn diagram of 10 m/z features selected as diagnosis biomarker panel following multiple screening rounds. (H) Diagnostic model constructed using Panel-10 M for discovery and validation cohorts. (I) Diagnostic performance of differential metabolites in BC and TETs [including pyruvic acid (PA), deoxyinosine (D-ino), glycerol (Gly), linoleic acid (Lin-A), pyroglutamic acid (P-Glu), lactic acid (LA), octanedloylcarnitine (O-ine), uracil (Ura), glycerol 3-phosphate (G 3-P), and nutriacholic acid (NA)]. (J) Diagnostic performance of the Panel-10 M in differentiating LR and HR-TET patient groups.

We employed the controlled pyrolysis-engineered FM-450 to extract SMFs for thymoma screening, which enabled rapid (30 s/sample), high-throughput analysis with minimal serum consumption (1 μL) (49). Raw spectrum for each sample with ~210,000 data points were recorded (Fig. 5A) and further processed to extract 698 features in the range of 80 to 1,000 Da (defined as SMF). Notably, the lower molecular weight range (80 to 500 Da) accounted for over 90% of the total spectral intensity, primarily consisting of end products of metabolic activity, such as amino acids, organic acids, and monosaccharides. High similarity scores (scores > 0.8) for SMFs within both groups confirmed the data stability and reproducibility (SI Appendix, Fig. S14A). Unsupervised PCA and t-distributed stochastic neighbor embedding map revealed partial cohort separation but persistent spectral overlap (Fig. 5B and SI Appendix, Fig. S14 B and C, analysis of similarities R = 0.203, P = 0.001; permutational multivariate ANOVA R² = 0.155, P = 0.001) (3, 50).

To optimize clinical utility, the constructed cohort was divided into a discovery (n = 172, 56 BCs and 116 TETs) and a validation cohort (n = 71, 24 BCs and 47 TETs). Notably, the age and gender distribution between the BC and TET patients were matched in the discovery cohort (P = 0.146 for age by Student's t test and P = 0.395 for gender by the chi-square test), to prevent potential biases during the training and subsequent panel screening phase (SI Appendix, Tables S4 and S5). Four machine learning algorithms [least absolute shrinkage and selection operator (LASSO), ridge regression (RR), neural network (NN), support vector machine (SVM)] were evaluated (Fig. 5C). While all algorithms achieved high performance, LASSO regression outperformed others in the discovery cohort (AUC = 0.975 vs. RR = 0.962, NN = 0.937, SVM = 0.917; Fig. 5D). Validation confirmed LASSO's robustness (AUC = 0.930, 95% CI: 0.894 to 1.000; Fig. 5E), with high sensitivity (87.2%), specificity (87.5%), and accuracy (87.3%). In addition, we also visualized the diagnostic performance of the LASSO model, which showed good diagnostic performance in 243 serum samples and different groups (Fig. 5F). Among them, the diagnostic effect was significantly different between BC and LR/HR patients (P < 0.05, SI Appendix, Table S6). Heatmap analysis further distinguished metabolic profiles between BC, LR, and HR groups (SI Appendix, Fig. S15).

To develop a streamlined diagnostic model, we constructed a 10-metabolite panel (Panel-10 M; Fig. 5G) from biomarkers significantly altered in the TET group. The 10 metabolites (SI Appendix, Table S7) were selected using stringent criteria (P < 0.001, fold change > 1.5, LASSO score > 0.3, mean intensity > 500; see details of adduct verification in Materials and Methods). Notably, all metabolites were downregulated in TETs vs. BC (P < 0.05, SI Appendix, Fig. S16 and Table S8). The Panel-10 M displayed comparable diagnostic potential of TETs with an AUC of 0.960 (95% CI: 0.931 to 0.988, Fig. 5H). The LASSO model on Panel-10 M demonstrated optimal diagnostic performance in discovery and validation cohorts, with sensitivity, specificity, and accuracy values of 94.0%, 82.1%, and 90.1%, respectively, in the discovery cohort. Validation cohort yielded consistent results with sensitivity of 87.2%, specificity of 87.5%, and accuracy of 87.3% (SI Appendix, Table S8), outperforming any single metabolite (Fig. 5I). Furthermore, we validated the robust performance of Panel-10 M in differentiating the BC with not only HR-TETs but also LR-TETs, while the HR's AUC is higher than LR’s (Fig. 5J). The model’s performance revealed differences in the efficacy of Panel-10 M among patients stratified by thymoma risk. This finding suggests that metabolic heterogeneity underlying the risk groups may account for the disparate outcomes.

To further validate the specificity of the metabolite signature, we have constructed a new control cohort (including 7 TETs, 6 germ cell tumors, 7 mediastinal lymphoma patients, 7 diabetes, and 7 mediastinal inflammation patients). As a result, our analysis confirmed that the metabolic signature was highly specific to TETs, with a specificity of 88.2% (SI Appendix, Fig. S17A). Notably, PCA demonstrated clear separation between TETs and all comparator groups, including other mediastinal masses (germ cell tumors, lymphomas, and thymic cysts, SI Appendix, Fig. S17B) and systemic conditions (diabetes and inflammation, SI Appendix, Fig. S17C). This showed that neither malignancies in the mediastinal compartment nor metabolic disorders replicate the TET-specific profile.

Pathway topological analysis of the 10 differentially abundant metabolites revealed significant change of six metabolic pathways in TET patients: glycerolipid metabolism, linoleic acid metabolism, citrate cycle (TCA cycle), pantothenate and CoA biosynthesis, beta-alanine metabolism, and pyruvate metabolism (SI Appendix, Table S9). Notably, pyruvate—a key hub metabolite during TET proliferation—demonstrated mechanistically coordinated dysregulation across both the TCA and pyruvate metabolic cycles, consistent with prior reports of cancer-associated rewiring of energy flux (51). Core lipidomic mediators, including glycerol 3-phosphate, glycerol, and octanoylcarnitine, exhibited perturbations within glycerolipid and fatty acid oxidation pathways. These alterations may underpin paraneoplastic lipid dysregulation in TETs, suggesting therapeutic entry points for metabolic reprogramming and prognostic stratification (52).

Risk Stratification of TETs.

After using this platform for TET diagnosis, we sought to differentiate TET patients with different risks without the aid of imaging techniques. Certain metabolic disparities were observed among BC, LR, and HR patients (SI Appendix, Figs. S14 and S15A). In addition, the PCA plot displayed the regular altered trajectory from HC, LR, to HR (Fig. 6A), suggesting the potential metabolic evolution during disease progression. However, only two out of the 10 metabolites exhibited significant differences between the LR and HR groups (Student’s t test, SI Appendix, Table S10). When applied to risk stratification (LR-TET vs. HR-TET), the previously established Panel-10 M achieved moderate performance AUC of 0.700 (95% CI: 0.603 to 0.797, Fig. 6B) for differentiating LR and HR. This observation suggested potential limitations in resolving disease severity or progression, underscoring the need to prioritize metabolites with mechanistic relevance to TET subclassification.

Fig. 6.

SMF-based risk stratification of TETs. (A) Visualization of SMFs for healthy controls (HC, green), LR (orange), and HR (red) TET samples using the Panel-10 M. (B) The model using the Panel-10 M for discriminating LR vs. HR TETs in discovery and validation cohorts. (C) ROC curves for discriminating LR vs. HR TETs using all 698 m/z features. (D) Venn diagram identifying the 6 m/z features selected for the Panel-6 M based on significance (P < 0.001), fold change (FC > 1.5), LASSO score (> 0.3), and signal-to-noise ratio (S/N > 10). LASSO scores for (E) BC vs. LR and (F) BC vs. HR [including pyruvic acid (PA), deoxyinosine (D-ino), glycerol (Gly), linoleic acid (Lin-A), pyroglutamic acid (P-Glu), lactic acid (LA), octanedloylcarnitine (O-ine), uracil (Ura), glycerol 3-phosphate (G 3-P), and nutriacholic acid (NA)]. (G) Schematic workflow for classifier construction. Feature selection (green): Stages I identifying key metabolites to form the Panel-6 M that met statistical criteria of P < 0.001, FC > 1.5, and LASSO score > 0.3 in the comparison of LR-TET vs. HR-TET; and II identifying metabolites from the Panel-10 M to form the 6 score metabolites panel (Panel-6SM) with a ratio (LASSO score from BC vs. LR-TET divided by LASSO score from BC vs. HR-TET) greater than 4. Classifier building (blue): The 12 metabolites from stages I and II are integrated into the final classifier of Panel-12 M using a random forest algorithm. (H) Superiority of the optimal Panel-12 M classifier over imaging model (mediastinal mass size, MMS) and alternative hierarchical models (Panel-6 M; Panel-6SM).

To address this, we randomized 163 serum samples (LR/HR: 70/93) into discovery (LR/HR: 49/67) and validation cohorts (LR/HR: 21/26), with balanced demographics (Fig. 6C). Our analytical pipeline began with performing LASSO on the discovery cohort, implemented with fivefold cross-validation and stratified by disease risk. This analysis identified six metabolic biomarkers with higher sensitivity to form a refined panel (Panel-6 M), according to the criteria of P < 0.001, fold change > 1.5, LASSO score > 0.3 and signal-to-noise ratio (S/N) > 10 (Fig. 6D). The Panel-6 M consisted of stearoylcarnitine, palmitic acid, squalene, oleic acid, erucic acid, and deoxyadenosine (SI Appendix, Table S11).

To further enhance stratification accuracy, we utilized a previously established Panel-10 M to differentiate risk populations, as metabolic profiles differed significantly between risk groups (metabolite score fold change range: 0.87 to 4.67; Fig. 6 E and F). From this panel, we selected the six most significantly regulated metabolites to create a new panel (Panel-6SM). Selection was based on a LASSO score ratio > 4, comparing BC vs. LR-TET to BC vs. HR-TET, reflecting distinct metabolic changes during thymoma progression across patient subtypes. The final Panel-6SM consisted of pyruvic acid, lactic acid, glycerol, octanedloylcarnitine, deoxyinosine, and nutriacholic acid (SI Appendix, Table S12).

Concerning the restricted diagnostic performances of Panel-6 M [AUC = 0.825 (0.750 to 0.900)] and Panel-6SM [AUC = 0.768 (0.681 to 0.855), SI Appendix, Table S13], we further combined these panels via random forest algorithm, forming a 12-hierarchical metabolite panel (Panel-12 M, Fig. 6G). This model synergized dual-phase feature selection with metabolomic variance across risk strata to maximize discriminative power (SI Appendix, Fig. S18). The Panel-12 M achieved an AUC of 0.856 (95% CI: 0.780 to 0.931) in discovery cohorts, with validation cohorts confirming the model generalizability (AUC = 0.864, 95% CI: 0.762 to 0.967). Notably, the Panel-12 M outperformed existing panels, including the Panle-6 M (AUC = 0.825), Panle-6SM (AUC = 0.768), and mediastinal mass size (AUC = 0.627, Fig. 6H).

We have performed the pathway topology analysis to explore in-silico interrogation of potentially altered metabolic pathways (SI Appendix, Fig. S19A). The observed metabolic alterations associated with the metabolites of Panel-12 M point to specific dysregulated pathways (e.g., purine metabolism, pyruvate metabolism, and glycerolipid metabolism) that derived TET pathogenesis in HR compared to LR (SI Appendix, Fig. S19B and Table S14). Specifically, the changes in pyruvic acid, pyroglutamic acid, and lactic acid suggested differential regulation of pyruvate metabolism, gluconeogenesis, and Warburg effect during tumor growth (53, 54). The differential expression of uracil (the nucleotide metabolism intermediate metabolites) reflected the metabolic adaptation to the increased transcriptional activity and regulation of purine and pyrimidine metabolism due to the proliferation of cancer cells (55, 56). Additionally, abnormal fatty acid (FA) expression corresponds with impaired FA degradation in tumor cells (57, 58), which was considered as a key aspect of metabolic reprogramming of malignant tumors. These progression-linked metabolic shifts mapped dynamic trajectory deviations, facilitating high-precision longitudinal monitoring of disease evolution (59). The optimized Panel-12 M balanced analytical simplicity with biological granularity (SI Appendix, Fig. S20), capturing critical metabolic network reprogramming during tumor progression and ensuring clinical feasibility (60). This hierarchical integration of multifeatures selection established a paradigm for precision risk stratification in TETs.

Discussion

Precise diagnosis and differentiated risk stratification have emerged as a strategy for optimizing therapeutic outcomes in TETs. The management of HR thymoma presents dual clinical challenges: Systemic organ failure from metastatic dissemination complicates complete tumor eradication, while chemotherapy-induced tumor suppression paradoxically fosters chemoresistance and escalates treatment costs (61). These realities underscore the critical need for rapid, precise risk stratification tools to guide precision interventions and arrest disease progression. While biopsy and CT imaging remain current gold-standard diagnostic modalities, they are limited for population-level screening due to the invasiveness or poor scalability (62).

Our study addressed these challenges through facile acquisition of SMFs by advanced materials-assisted LDI MS and machine learning-enabled analysis. In particular, the engineered heterojunction in FM-450 substrate synergized UV absorption of Fe-MOF with photothermal conversion efficiency of Fe₃O₄, while interfacial charge separation effects collectively enhance LDI MS detection sensitivity by 1,000-fold compared to conventional matrices. Concurrently, the constrained thermal decomposition within the MOF imposes spatial confinement on metal ion migration, enabling the nucleation of ultrafine metal NPs (63, 64).

On the other hand, Fe₃O₄ NPs have garnered attention as substrates for LDI MS owing to their low thermal conductivity and superior photothermal conversion efficiency (16, 65, 66). Nevertheless, their inherent magnetic dipole interactions and high surface energy predispose them to agglomeration (67), which induces heterogeneous crystallization and stochastic hot-spot effect, resulting in weak analyte ionization and poor reproducibility. A spatially confined in situ growth protocol to anchor Fe₃O₄ NPs on MOF is an effective strategy to mitigate NPs aggregation. Moreover, this synergistic architecture enhances photogenerated charge separation and leverages exceptional photothermal conversion efficiency of Fe₃O₄ NPs.

For clinical applications, the FM-450 matrix exhibits a favorable profile in terms of user safety, reagent stability, and cost-effectiveness for high-volume diagnostic testing. For reagent stability, the particles show no aggregation for at least 6 mo when stored as an aqueous suspension (1 mg/mL) at room temperature and maintain consistent signal intensity without degradation over 10,000 laser irradiations (SI Appendix, Fig. S21). For user safety, the matrix consisted of Fe₃O₄ within MOFs, both of which are considered biocompatible for clinical use (68, 69). For cost, the FM-450 matrix offers economic advantages for scaled diagnostics with costs of $0.03 per test, based on the usage of 1 mg of matrix to process 384 samples. This represents a substantial ~100-fold reduction compared to the estimated $4 per test cost associated with conventional LC–MS/MS methodologies (70).

To ensure sample handling does not introduce error or compromise reliability in hospital labs, we have established two key protocols: 1) prefabrication of single-use MS chips with preloaded nanoparticles during in vitro diagnostic manufacturing, and 2) integration of automated instrumentation for clinical sample loading. The manufacturing-controlled nanoparticle prespotting eliminates manual manipulation steps at the hospital, significantly reducing operational complexity for clinical staff without MS expertise while shortening processing time (SI Appendix, Fig. S22 A and B). This is complemented by automated sample loading systems that standardize liquid handling, thereby minimizing variability from manual pipetting (SI Appendix, Fig. S22C). Together, these protocols ensure consistent chip preparation and sample preparation, which are critical for reliability control. To validate this approach, we conducted robustness testing with four inexperienced operators performing full assays. The resulting CVs for all standard analytes remained below 10% (SI Appendix, Fig. S22D), demonstrating operational resilience suitable for clinical environments (71, 72).

In total, the developed platform demonstrates three advances: 1) noninvasive detection with trace sample consumption (1 μL of serum) (73), 2) clinically validated diagnostic performance (AUC 0.960) on larger cohorts (243 cases) across diverse cohorts (encompassing healthy individuals, benign lesions, and various thymoma subtypes), surpassing conventional liquid chromatography–tandem mass spectrometry (MS) approaches (74, 75), and 3) biologically interpretable risk stratification through the identified metabolic biomarkers.

Future directions should focus on multicenter validation across broader geographical populations and integration with multimodal clinical data. Prospective studies incorporating longitudinal monitoring could further validate the prognostic value of the metabolic panel, while technical extensions to other biofluids (e.g., pleural effusions) may expand diagnostic applicability.

Materials and Methods

Chemicals and Reagents.

Ferrous chloride tetrahydrate (FeCl₂·4H₂O, 99.0%), ferric chloride hexahydrate (FeCl₃·6H₂O, 99.0%), ferrous acetate (Fe(CH₃COO)₂, 98.0%), terephthalic acid (99.0%), 2,5-dihydroxyterephthalic acid (97.0%), N, N-dimethyl formamide (DMF, 99.5%), ethanol, triethylamine (TEA, 98.0%), sodium chloride (NaCl, 99.5%), acetonitrile (ACN, 99.0%), trifluoroacetic acid (TFA, 99.0%), α-cyano-4-hydroxy-cinnamic acid (CHCA, 99.0%), and 2,5-dihydroxybenzoic acid (DHB, 99.0%) were purchased from Aladdin (Shanghai, China). L-alanine (Ala, 98.0%), L-arginine (Arg, 99.5%), L-lysine (Lys, 98.0%), D-glucose (Glc, 99.5%), sucrose (Suc, 99.0%), ethylenediaminetetraacetic acid (EDTA, 99.0%), juglone (97.0%), and bovine serum albumin (BSA, 98.0%) were purchased from Sigma-Aldrich (St. Louis, MO). The deionized water (DIW, 18.2 MΩ·cm) was produced through a Millipore Milli-Q system (Milli-Q, Millipore, GmbH).

Synthesis of Materials.

Synthesis of Fe-MOF.

The Fe-MOF was synthesized by the ultrasonic method according to previous literature with a minor modification (11). First, 32 mL of DMF, 2 mL of ethanol, and 2 mL of water were mixed in a 50 mL polytetrafluoroethylene tube. Next, 0.75 mmol of terephthalic acid was dissolved into the mixed solution under ultrasonication. Subsequently, 0.75 mmol of FeCl₂·4H₂O was added. 0.8 mL of TEA was quickly injected into the solution after Fe²⁺ salt was dissolved. Then, the solution was stirred for 5 min to obtain a uniform colloidal suspension. Afterward, the colloidal solution was treated with continuous ultrasonication for 8 h (40 kHz) at 4 °C. Finally, the product was obtained via centrifugation, washed with deionized water and ethanol (three to five times), and dried in a vacuum oven at 60 °C.

Synthesis of FM-X.

In order to synthesize a series of FM-X (X = 350, 400, 450, 500), the as-synthetized Fe-MOF was transferred into a ceramic boat and placed in a tube furnace. The sample was heated to 350, 400, 450, and 500 °C for 1 h under flowing argon gas and then naturally cooled to room temperature. The obtained materials were directly used without further treatment.

Synthesis of Fe-MIL-101.

The Fe-MIL-101 was synthesized according to previous literature (76). First, 2.45 mmol FeCl₃·6H₂O and 1.24 mmol terephthalic acid were dissolved in 15 mL of DMF. Next, the mixture was heated at 110 °C for 20 h in a 50 mL Teflon-lined stainless-steel autoclave. Finally, the product was obtained via centrifugation, washed with ethanol three times, and dried in a vacuum oven at 60 °C.

Synthesis of Fe-MIL-74.

The Fe-MOF-74 was synthesized according to previous literature (77). First, 0.45 mmol 2,5-dihydroxyterephthalic acid and 1.44 mmol Fe(CH₃COO)₂ were dissolved in a mixed solvent containing 30 mL of DMF, 1.8 mL of ethanol, and 1.8 mL of water. Then, the mixture was heated at 120 °C for 24 h in a 50 mL Teflon-lined stainless-steel autoclave. After cooling down to room temperature, the products were obtained via centrifugation, washed with ethanol 3 times, and dried in a vacuum oven at 60 °C.

Characterization Methods.

PXRD patterns were collected by a Rigaku Miniflex600 PXRD (Cu Kα). TEM and HRTEM images were collected by FEI Tecnai G2 F30. Atomic force microscopy (AFM) image was characterized by SPI-3800. XPS measurements were performed on an ESCALAB 250 spectrometer (Thermo-VG Scientific). Thermogravimetry analysis (TGA) was performed on a TG209F1 Libra R analyzer. Scanning electron microscopy (SEM) images were recorded by a SU8010 system. UV–Vis spectra and UV–Vis diffuse reflectance spectra (UV–Vis DRS) were obtained by a Hitachi UH5300 spectrometer. Both the photocurrent measurements and the Mott–Schottky tests were carried out in a standard three-electrode system with materials as the working electrode, platinum plate as a counter electrode, Ag/AgCl as a reference electrode, and 0.5 M Na₂SO₄ as electrolyte solution in a CHI 660E electrochemistry workstation (ChenHua Instruments Co., China). Thermographic images were recorded by an IR thermal camera (Fotric 246 M Thermal Imaging Camera).

Density Function Theory Calculation.

First-principles calculations were performed by using the Cambridge sequential total energy package module of Materials Studio software (version 5.5) to investigate the work functions of Fe-MOF and FM-500 (i.e., Fe₃O₄) surfaces. Exchange-correlation functions were depicted by generalized gradient approximation with Perdew–Burke–Ernzerhof pseudopotentials (78). To minimize interactions from neighboring slabs, a vacuum of 20 Å was added along the z direction. The convergence criterion of geometry relaxation was set to 0.01 eV Å⁻¹ in force on each atom. The energy cutoff for plane wave-basis was set as 500 eV. The K points were sampled with 1 × 1 × 1 by the Monkhorst–Pack method (79).

Sample Preparation.

The standard metabolite solution (Ala, Lys, Arg, Glc, and Suc) was prepared with deionized water at a concentration of 1 mg mL⁻¹. The serum samples were diluted 10-fold with deionized water at room temperature. CHCA (4 mg mL⁻¹) and DHB (10 mg mL⁻¹) matrices were dissolved in a mixed solution of ACN and water (3:7, v/v) with 0.1% TFA. The matrices of Fe-MOF and FM-X (X = 350, 400, 450, 500) were prepared as aqueous suspensions at a concentration of 1 mg mL⁻¹. To evaluate protein and salt tolerance, BSA (5 mg mL⁻¹) and NaCl (0.5 M) were mixed with a mixture of typical small metabolites including Ala, Lys, Arg, Glc, and Suc, each at a concentration of 1 mg mL⁻¹. Blood samples were collected from Shanghai Chest Hospital. This study was approved by the Medical Ethics Committee of Shanghai Chest Hospital (approval ks25049). We obtained written informed consent from all participants prior to blood collection. This study enrolled 80 benign controls (BC) and 163 participants with TETs. Inclusion criteria for patients were as follows: 1) all patients were clinically validated by medical professionals using radiological imaging and pathology, while BC showed no detectable lesions; and 2) samples were free from hemolysis during handling. All plasma samples included (n = 243) were stored at −80 °C. Written consent was received from all chosen patients for the utilization of their samples for biomedical research.

MS Analysis.

During MS analysis, 1 μL of matrix suspension was added to the sample spot and dried at room temperature. Then, 1 μL of serum was dropped onto the target plate and dried naturally. This drying process took ~60 min for each batch of 243 samples (equivalent to 14 s per sample). We used ultrafleXtreme MALDI-TOF MS (Bruker, Germany) instruments with a 355 nm Nd:YAG laser to detect the samples on the target plate. Mass spectra were acquired in positive ion reflector mode. The optimized parameters are configured as follows: positive ion reflector mode, 60% laser intensity, acceleration voltage of 20 KV, and repetition rate of 200 Hz. All MS results were directly exported for further analysis. Scanning range is mass-to-charge (m/z) 80 to 500 in positive ion mode. The MS instrument was tuned and calibrated with Bruker peptide II standards in a matrix mix (CHCA) in MALDI source. The focus of the laser beam was set as “small” level. The detector plate was set to 210 V, with a plate offset of 80 V, laser power at 20 to 30%, frequency at 1 kHz, 2,000 laser shots, and a time of light of 1.1 ms. Final metabolite identification was achieved by cross-referencing accurate MS from the Human Metabolome Database (HMDB) and FT-ICR-MS (Bruker Daltonics, Germany). Metabolite assignments were based on exact m/z ratios (accurate within 0.01 Da). To confirm identities, molecular weights were calculated using four common adduct formation patterns (+H, +Na, +K, +2Na-H) and matched against the reference.

Data Processing.

The data processing was carried out using Python (version 3.8). Baseline correction and spectral smoothing were performed to eliminate noise in raw mass spectra. The preprocessing steps included peak detection, peak alignment, peak filtration, standardization, and peak filtration. The discrete wavelet transform was applied for spectrum smoothing and baseline correction. Peaks with signal-to-noise greater than 3 and present in ≥2/3 of the samples were extracted from each sample for subsequent analyses. Before data analysis, we used Orange (version 3.36.2) to normalize the sample data after preliminary feature extraction. Then, all the extracted peak information was utilized to construct several predictive models using LASSO, RR, NN, SVM, and random forest algorithms by Orange (version 3.36.2). Each model was developed to evaluate the significance of different features in achieving effective classification. Data normalization was employed using SPSSPRO (version 1.1.28, https://www.spsspro.com/) to remove less significant peaks and identify the optimal combination of features for reliable classification. Chi-square test, two-sided Student’s t test, and unpaired t tests were conducted using IBM SPSS Statistics (version 24). Data with P < 0.05 were considered as statistically significant.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

MS data have been deposited in GitHub (https://github.com/Haochen-Han/PNAS-TETS.SMFs.git) (80). All other data are included in the manuscript and/or SI Appendix.