Gold‐Functionalized Multilayer Heterojunction Microarchitectures Enable High‐Fidelity Serum Metabolite Profiling for Skin Cancer Subtype Classification

ABSTRACT

Skin cancer is among the most prevalent malignancies worldwide, with non‐melanoma types ranking among the top five and melanoma characterized by high lethality. Psoriasis, although non‐malignant, imposes a substantial physical and psychological burden. Accurate and timely detection is crucial for improving clinical outcomes. Serum metabolites, as sensitive indicators of systemic physiology, represent promising noninvasive biomarkers. Here, we developed gold‐modified rose‐like multilayer heterojunctions (G‐RMHJ) as an efficient matrix for laser desorption/ionization mass spectrometry (LDI‐MS). The hierarchical multilayer architecture and heterojunction interfaces, combined with gold functionalization, synergistically enhance light absorption, interfacial energy transfer, and ionization efficiency, enabling sensitive and reproducible serum metabolite detection. In a cohort of 130 skin cancer and 23 psoriasis patients together with 218 healthy controls (HC), G‐RMHJ‐assisted LDI‐MS yielded robust serum metabolic fingerprints, and machine learning models built on these data achieved 100% test‐set accuracy for classifying patients vs. HC. Eight discriminative metabolites were identified that reliably differentiated four types of diseases from HC, with area under the curve values ranging from 0.933 to 1.000. This work demonstrates how hierarchical microstructured heterojunctions can directly translate materials‐level design into enhanced bioanalytical performance, providing a generalizable and noninvasive framework for serum metabolomics–based disease classification and skin cancer subtype identification.

We developed a gold‐functionalized rose‐like multilayer heterojunction (G‐RMHJ) platform for LDI‐MS‐based serum metabolic fingerprinting. Integration with machine learning and a panel of eight key metabolites enables multiclass classification of skin diseases and subtype identification of skin cancer, highlighting its potential for accurate diagnosis and improved clinical outcomes.

1. Introduction

Skin, the largest organ of the human body, provides essential protection while regulating diverse physiological functions [1, 2]. Among skin disorders, skin cancer represents one of the most aggressive malignancies in recent decades and currently ranks as the fifth most common cancer worldwide [3, 4]. Although the incidence rate of melanoma is lower than that of non‐melanoma skin cancer, its mortality is substantially higher, with more than 330 000 new cases and nearly 60 000 deaths reported worldwide in 2022 [5]. The rising incidence of skin diseases, particularly basal cell carcinoma (BCC), squamous cell carcinoma (SCC), melanoma (MM) and psoriasis (PSO), poses a growing threat to human health and markedly reduces quality of life [6, 7]. While tissue biopsy remains the diagnostic gold standard for skin cancer [8] by providing precise histological information, it is invasive, increases the risk of infection, and often causes psychological stress. Dermoscopy offers a non‐invasive [9] alternative but relies heavily on physician expertise and is prone to subjective bias. Artificial intelligence−assisted approaches improve diagnostic efficiency, yet they require extensive high‐quality data and are vulnerable to algorithmic errors [10, 11]. Multi‐molecular imaging methods provide high specificity and rapid analysis [12], but their limited penetration depth compromises diagnostic accuracy for deeper lesions. Taken together, these limitations highlight the urgent need for a reliable, accurate, and non‐invasive strategy capable of improving the classification of skin diseases and enabling precise differentiation of skin cancer subtypes.

Diagnosis based on serum metabolites has attracted increasing attention because of its non‐invasive nature, high patient acceptance, and high accuracy. Subtle changes in the tumor microenvironment can be reflected in serum metabolites, offering important clues to disease onset and progression [13]. Among existing analytical techniques, laser desorption/ionization mass spectrometry (LDI MS) has become a preferred tool for serum metabolite detection owing to its rapid analysis, high sensitivity, and high throughput [14]. Despite these advantages, the ionization of metabolites in complex serum samples remains challenging. Conventional organic matrices such as 2,5‐dihydroxybenzoic acid (DHB) and α‐cyano‐4‐hydroxycinnamic acid (CHCA) generate strong background signals that interfere with detection in the low‐mass range [15, 16, 17]. To overcome these limitations, more attention has been directed toward inorganics [18], metal‐organic framework materials [19], and metal oxides [20], which exhibit high UV absorption, efficient ionization, and low cost. However, single metal oxides generally possess a simple electronic structure and limited adsorption capacity for small metabolites. This limitation may result in the loss of valuable small molecules and makes it difficult to achieve the performance required for clinical application [21]. In contrast, bimetallic oxides exhibit electronic synergistic effects that can effectively promote charge transfer and accelerate the ionization of serum metabolites [22]. For instance, In₂O₃‐Co₃O₄ heterojunctions have been reported to overcome the poor conductivity of Co₃O₄, while the construction of a p‐n junction significantly enhances overall thermal stability [23]. In addition to bimetallic systems, noble metals have also been explored due to their ability to generate high‐energy hot electrons and holes under irradiation. When integrated with metal oxides, they not only strengthen UV absorption but also accelerate electron transfer, thereby further improving ionization efficiency. A representative example is Pt‐modified CeO₂‐ZrO₂ with a core‐shell architecture, where strong interfacial electronic interactions enhance the synergistic effects between the two components and greatly improve material stability [24].

Previous studies have shown that noble metal–functionalized metal oxide heterojunctions can significantly improve charge transfer, suppress electron–hole recombination, and enhance light absorption. Gold is one of the most widely used noble metals because of its resistance to oxidation, stable interfacial properties, and excellent electrical conductivity [25, 26]. In this work, we further modified these composite gold nanoparticles, which not only enhanced the structural stability of the heterojunction but also strengthened the synergistic interactions between Ag₂O and SnO₂. As a result, the material exhibited more efficient charge transfer, reduced electron–hole recombination, and improved light absorption. In addition, the rose‐like multilayer structure provided a large surface area and abundant active sites, facilitating the binding of small metabolites with ions such as sodium and potassium. Building on these advantages, we constructed gold‐modified rose‐like multilayer heterojunctions (G‐RMHJ) as an efficient matrix to assist the LDI‐MS platform, which enabled serum metabolic fingerprints (SMFs) combined with machine learning for skin disease classification and skin cancer subtype diagnosis (Scheme 1). A neural network (NN) model achieved 100% accuracy in distinguishing patients with skin diseases from healthy controls in the test set and yielded area under the curve (AUC) values above 0.9 for binary classification of skin cancer subtypes. Furthermore, analysis of eight metabolites revealed their critical roles in the development of PSO, BCC, SCC, and MM. The results demonstrated the potential of the G‐RMHJ‐based platform for highly sensitive detection of small‐molecule metabolites below 1000 Da and highlight its promise for early diagnosis and large‐scale clinical applications.

SCHEME 1.

Schematic of the application of G‐RMHJ‐assisted LDI MS for dermatologic surveillance. (a) Construction and application of G‐RMHJ for serum metabolite profiling. (b) Multiclass classification and subtype differentiation of SD based on SMFs by a machine learning algorithm.

2. Results and Discussion

2.1. Synthesis and Characterization of G‐RMHJ

The facile preparation method of G‐RMHJ was refined based on previous reports [27], as shown in Scheme 1a. Briefly, smooth 2D nanosheets were first synthesized via a hydrothermal method, during which layer‐by‐layer interactions facilitated the formation of rose‐like multilayer hetero‐scaffold (RMHS). The as‐prepared RMHS were subsequently calcined in air at high temperature, which preserved the original precursor morphology while increasing the surface roughness of the 2D nanosheets. Gold nanoparticles were then deposited onto the surface of the resulting rose‐like multilayer heterojunctions (RMHJ) by the chemical reduction method. The incorporation of gold nanoparticles markedly enhanced the ionization efficiency of the material, ultimately yielding gold‐functionalized RMHJ (G‐RMHJ) with significantly improved surface roughness.

The morphologies of G‐RMHJ and its precursors were characterized by scanning electron microscopy (SEM) and field emission transmission electron microscopy (FETEM). As shown in Figure 1a,b, the tin sulfide precursor exhibits an interlocked rose‐like architecture composed of thin, well‐defined nanosheet “petals” with smooth and flat surfaces. After calcination in air, the resulting tin oxide precursor (Figure 1c,d) retained the rose‐like structure, but the nanosheet surfaces became rougher, and raised nodules appeared. Following gold nanoparticle deposition, the final G‐RMHJ product (Figure 1e,f) maintained the flower‐like morphology but exhibited a higher density of raised nodules and interparticle linkers, which are expected to enhance the adsorption of small molecules on the material surface. The TEM images of RMHS and RMHJ (Figure 1g,h) reveal that both structures are assembled from interlocked 2D nanosheets. Elemental mapping of G‐RMHJ (Figure 1i) confirms the uniform distribution of Sn, O, Ag, and Au throughout the hierarchical structure, indicating that the well‐dispersed metal/metal oxide active sites likely contribute to the observed synergistic effects in the rose‐like architecture. Detailed information is shown in Table S1. The elemental distribution profiles further show that Sn and O are homogeneously distributed within the SnO₂ microscale roses, while Ag and Au are relatively evenly anchored, with Ag and Au signals appearing more intense. High‐resolution FETEM (Figure 1j) reveals lattice spacing of 0.338 nm corresponding to the (101) plane of tetragonal rutile‐phase SnO₂, 0.268 nm for the (111) crystal face of Ag₂O, and 0.212 nm for the (111) plane of Au [28].

FIGURE 1.

Characterization of G‐RMHJ. FESEM images of (a, b) RMHS, (c, d) RMHJ, and (e, f) G‐RMHJ. FETEM images of (g) RMHS, (h) RMHJ, and (i) G‐RMHJ, and elemental mapping images of G‐RMHJ showing Sn in red, O in green, Ag in purple, Au in yellow, and an overlap image. (j) High‐resolution FETEM of G‐RMHJ. (k) Sn 3d, (l) Ag 3d, (m) O 1s, and (n) Au 4f XPS spectra of G‐RMHJ.

As shown in Figure S1, the nitrogen adsorption−desorption isotherms of G‐RMHJ exhibit a type IV curve with an H3 hysteresis loop, which is characteristic of mesoporous structures. Specifically, the specific surface area of G‐RMHJ is 59.3734 m²/g. Furthermore, based on the data in Figure S2, the BJH adsorption average pore size is approximately 18.50 nm. The pore size distribution reveals two dominant regions: mesopores in the range of 3–20 nm and a broader distribution centered around ∼40 nm within the 30–60 nm range, the latter showing relatively higher intensity. Micropores are negligible, while a minor fraction of pores larger than 50 nm can be attributed to macropores. In summary, the prepared porous G‐RMHJ has a high specific surface area and mesoporous structure, providing abundant active sites for surface interactions and facilitating the enrichment and detection of metabolites and other biomolecules in complex biological samples.

The XRD patterns (Figure S3a) of RMHS display eight distinct diffraction peaks at 14.9°, 28.3°, 32.1°, 41.8°, 50.0°, 52.4°, 54.8°, and 60.6°, which correspond to the (001), (100), (002), (102), (110), (111), (103), and (201) planes of hexagonal SnS₂ (JCPDS No. 23–0677) [29]. Upon calcination, RMHS is converted to SnO₂, as evidenced by the diffraction peaks in Figure S3b at 2θ = 26.5°, 33.9°, 38.1°, 51.8°, and 64.7°, indexed to the (110), (101), (200), (211), and (112) planes of tetragonal rutile SnO₂ (JCPDS No. 77–0447) [30]. For G‐RMHJ (Figure S3c), the XRD pattern shows the characteristic peaks of SnO₂ alongside those of Au. Notably, the Au (200) reflection appears at 44.2°, confirming the successful deposition of Au nanoparticles (JCPDS No. 99‐0056) [31]. The Au (111) peak at 38.1° and Au (200) peak at 64.7°overlap with the corresponding SnO₂ peaks. No distinct diffraction peaks of Ag₂O or Ag were detected, likely due to their low content and the overlap of their main reflections with those of SnO₂ [32].

X‐ray photoelectron spectroscopy (XPS) analysis was conducted to further elucidate the elemental composition and valence states of G‐RMHJ. As shown in Figure S4, the full survey spectrum reveals characteristic peaks of Sn 3d, O 1s, Ag 3d, and Au 4f. All binding energies were calibrated against the C 1s peak at 284.8 eV. The high‐resolution Sn 3d spectrum (Figure 1k) exhibits peaks at 486.7 and 495.1 eV, corresponding to Sn⁴⁺3d_5/2 and Sn⁴⁺3d_3/2, respectively [33]. In the Ag 3d region (Figure 1l), deconvolution reveals binding energies at 374.2, 368.2, 373.6, and 367.5 eV, assigned to Ag⁰ 3d_3/2, Ag⁰ 3d_5/2, Ag⁺ 3d_3/2, and Ag⁺ 3d_5/2, respectively [34]. The O 1s spectrum (Figure 1m) is resolved into three components at 530.5, 531.3, and 532.5 eV, corresponding to lattice oxygen (O_L), vacancy oxygen (O_V), and adsorbed oxygen (O_C) bonded to Sn⁴⁺, respectively [35]. The Au 4f spectrum (Figure 1n) presents peaks at 87.1 eV (Au 4f_5/2) and 83.3 eV (Au 4f_7/2), confirming the presence of Au [36].

To evaluate the optical and surface charge properties relevant to laser desorption/ionization, UV–vis absorption spectra were recorded. A 355 nm UV laser, which is one of the most commonly used wavelengths in MALDI [37], matches well with the absorption of the matrix and enables efficient photon‐to‐analyte energy transfer. As shown in Figure S5, G‐RMHJ and its precursors exhibit pronounced absorption at 355 nm at the same concentration, with G‐RMHJ showing the strongest absorbance. This enhanced optical response suggests superior energy‐transfer capability to analytes during laser irradiation. In addition, the negative surface charge of G‐RMHJ is expected to promote the formation of cationic adducts in LDI MS. Zeta potential measurements (Figure S6) reveal that G‐RMHJ possesses the most negative potential among the tested samples [38], with an average value of −27.3 mV. This evident negative charge favors the acquisition of protons or the coordination with alkali metal ions (Na⁺, K⁺) in positive ion mode, leading to enhanced generation of [M + H]⁺, [M + Na]⁺, and [M + K]⁺ species.

2.2. Probing the Performance of G‐RMHJ With Standard Metabolites and Optimization of Analytical Conditions

We compared the capture efficiency of standard metabolites when using RMHS, RMHJ, and G‐RMHJ as matrices. As shown in Figure S7, G‐RMHJ exhibited the strongest mass spectrometry signal intensity, demonstrating superior capture efficiency compared to the other two materials. To demonstrate that the rose‐like structure enhances ionization performance, we grounded G‐RMHJ to disrupt the rose‐like structure (Figure S8). Three standard metabolites were selected to test the ionization performance of both materials. Figure S9a shows that the material retaining the rose‐like structure exhibits high signal intensity and low background interference. As depicted in Figure S9b, after the rose‐like structure was disrupted, the overall mass spectrometry signal intensity decreased, and more critically, background interference increased significantly. Furthermore, we compared this material with other inorganic materials, as shown in Figure S10. Under identical mass spectrometry conditions and matrix concentrations, although gold nanoparticles exhibited stronger mass spectrometry signals, background interference was also notably pronounced. Overall, our material G‐RMHJ demonstrated superior ionization capability.

Background noise is an important parameter for evaluating the performance of a matrix. To compare the performance of G‐RMHJ with conventional matrices, α‐cyano‐4‐hydroxycinnamic acid (CHCA), N‐naphthalene‐1,2‐diamine dihydrochloride (NEDC), and 2,5‐dihydroxybenzoic acid (DHB) were tested under identical experimental conditions. As shown in Figure S11, CHCA, NEDC, and DHB both produced strong background noise, which could interfere with analyte identification. On the contrary, G‐RMHJ exhibited extremely low background noise, indicating that its use as a matrix would not compromise analytical accuracy. The matrix concentration was then optimized using three standard metabolic small molecules, namely arginine (Arg), phenylalanine (Phe), and L‐asparagine (Asp), to within the range of 0.3–4.0 mg/mL. Results from three independent experiments (Figure S12) consistently showed that 2.0 mg/mL yielded the highest signal intensities, confirming it as the optimal concentration for subsequent analysis.

The reproducibility and ionization properties of G‐RMHJ were further examined. Ion‐adduct signals for Arg, Asp, and histidine (His) were collected from multiple positions within the same spot and across different spots to assess within‐spot and between‐spot stability. As shown in Figure S13a,b, the ion‐adduct peak intensities remained highly stable, with coefficients of variation (CVs) ranging from 6.1% to 10.9% for within‐spot measurements and from 2.5% to 9.6% for between‐spot tests in ten replicate experiments. We then collected ion‐adduct signals for Arg, Asp, and His from different points, and the ionic adduct peak intensities of all metabolites remained stable with coefficients of variation ranging in ten replicate experiments, which demonstrated excellent reproducibility of G‐RMHJ. The proton affinity of G‐RMHJ was further validated using Arg, Asp, glutamate (Glu), and His. Strong [M + H]⁺ and [M + Na]⁺ signals were observed (Figure S14), confirming the high efficiency of G‐RMHJ in promoting protonation and cationization processes.

To further evaluate the performance of G‐RMHJ under complex sample conditions, standard metabolite mixtures containing Asp, Phe, Glu, Arg, and His were analyzed, and ten distinct adduct ion signals were detected (Figure S15). The salt and protein tolerance of G‐RMHJ was then assessed by adding high concentrations of NaCl or bovine serum albumin (BSA) to standard metabolite mixtures to simulate real biological matrices. As shown in Figure S16a–d, G‐RMHJ maintained strong detection capability, producing seven adduct ion signals in the presence of NaCl and nine in the presence of BSA. These results suggest that G‐RMHJ possesses high sensitivity, excellent reproducibility, and strong tolerance to complex sample environments, highlighting its potential as a robust matrix for LDI‐based metabolite analysis in real biological samples.

2.3. Serum Metabolic Fingerprints for SD Multiclass Analysis and Skin Cancer Subtype Identification

Here, G‐RMHJ‐assisted LDI MS was employed to obtain serum metabolic fingerprints (SMFs) from a total of 371 samples, comprising 153 cases of skin diseases (SD) and 218 healthy controls (HC). The SD cohort included 66 cases of basal cell carcinoma (BCC), 34 cases of squamous cell carcinoma (SCC), 30 cases of malignant melanoma (MM), and 23 cases of psoriasis (PSO). Demographic and clinical information, including age, gender, and category, is presented in Figure 2a. Statistical analysis using the Mann‐Whitney U test indicated no significant differences in age or sex (P > 0.05) between SD and HC groups (Tables S2 and S3). Representative MS spectra from HC, SCC, BCC, MM, and PSO obtained under the same analytical conditions are shown in Figure 2b, revealing considerable differences in multiple signal peaks. Data preprocessing in RStudio yielded 410 m/z signals for further machine learning analysis. The heatmap in Figure 2c illustrates substantial differential expression across these features between groups.

FIGURE 2.

Comprehensive analysis of all features from bio‐samples. (a) Demographic data of 371 clinical samples. (b) Typical MS spectra of HC, SCC, BCC, MM, and PSO. (c) The heatmap of all m/z features. (d) PCA scores plot of SD and HC. (e) The radar chart illustrating comprehensive evaluation of AUC, accuracy, F1 score, MCC, CA, and recall for distinguishing HC from SD based on all features. (f)The confusion matrices for distinguishing SD from HC based on all features using the NN model in the training and test sets. (g) The sample‐level plots for the classification of SD and HC in the training and test sets. The ROC curves of (h) BCC vs. PSO, (i) BCC vs. SCC, and (j) MM vs. SCC in the training and test sets, respectively. The sample scatter plots of (k) BCC vs. PSO, (l) BCC vs. SCC, and (m) MM vs. SCC in the training and test sets, respectively.

Appropriate machine learning model selection is critical to ensure both sensitivity and specificity in disease classification. To preliminarily assess the separability of the groups, principal component analysis (PCA) was performed. As shown in Figure 2d, the PCA score plot indicates that simple 2D projection does not provide sufficient distinction between SD and HC, implying the need for more sophisticated machine learning algorithms. Accordingly, the dataset was randomly divided into a training set (70%; 107/153, SD/HC) for model development and a test set (30%; 46/65, SD/HC) for independent validation. For the SD multiclass and skin cancer (SC) subtypes, the same stratified randomization method was used. This resulted in 21 MM, 46 BCC, 24 SCC and 16 PSO cases in the training set and 9 MM, 20 BCC, 10 SCC, and 7 PSO cases in the test set (Scheme 1b). Six machine learning algorithms, including k‐nearest neighbor (kNN), three models, neural network (NN), Naive Bayes (NB), gradient boosting (GB), and AdaBoost (AB), were constructed to differentiate SD from HC based on all the features. Ten‐fold cross‐validation was applied during the training process to ensure that model performance evaluation was reliable.

We compared the performance of the six models using six evaluation metrics, which were AUC value, Matthews correlation coefficient (MCC), recall, precision, F1 score, and classification accuracy (CA). As shown in Figure 2e and detailed in Table S4, the NN consistently outperformed the other models and achieved values greater than 0.98 for all six metrics in the training set. The NN model correctly classified all SD cases and achieved 0.98 accuracy for HC in the training set, as shown in Figure 2f. In the independent test set, the NN model achieved 100% classification accuracy for both SD and HC. To better visualize the classification results, scatter plots were generated to display the separation between the two groups. Figure 2g shows that SD and HC are clearly separated in both training and test sets, which suggests that G‐RMHJ‐assisted LDI MS effectively captures the metabolic differences between SD patients and HC.

Based on the initial findings, the NN model was further optimized for multiclass analysis. Using all m/z features derived from SMFs, the NN model differentiated BCC, SCC, MM, and PSO from HC. In the binary classification setting, the differentiation between BCC and HC achieved an accuracy of 100% with an area under the curve (AUC) value of 1.000 in both the training and independent test sets (Figure S17a). The remaining three disease groups also reached AUC values up to 0.99 in both sets, as shown in Figure S17b–d. Confusion matrices presented in Figures S18a–d and S19a–d indicate that the classification accuracy for each disease type against HC exceeded 90%. Scatter plots in Figures S18e–h and S19e–h further demonstrate a clear separation between each disease group and HC in both training and test sets. These results highlight the strong capability of the NN model in distinguishing different SD from HC.

To explore the robustness of the NN model, more complex multiclass analysis of SD was performed. We applied the NN model based on SMFs to distinguish BCC, MM, PSO, SCC, and HC, with the corresponding AUC values for HC, BCC, MM, PSO, and SCC being 1.000, 0.993, 0.918, 0.973, and 0.982, respectively, in the independent test set (Figure S20). These results demonstrate the high credibility of the NN model in handling diverse classification scenarios. Encouraged by these findings, we extended the application of the NN model to more demanding conditions, such as distinguishing PSO, MM, BCC, and SCC. As shown in Figure S21a, the training set yielded AUC values of 0.964, 0.899, 0.861, and 0.983, respectively, while the AUC values in the test set achieved 0.977, 0.841, 0.974, and 0.994 (Figure S21b). The corresponding confusion matrices (Figure S21c,d) indicated that misclassifications were mainly concentrated in MM, which reflects the inherent difficulty of detecting melanoma at early stages due to its aggressive and concealed clinical nature. Further multiclass analyses were conducted on grouped combinations including PSO vs. BCC vs. MM, PSO vs. BCC vs. SCC, and PSO vs. SCC vs. MM. The ROC curves in Figure S22a–c show that the AUC values for these groups ranged between 0.918 and 0.993 in the test set. The confusion matrices (Figure S22d–f) also confirmed that the groups could be reliably distinguished with high accuracy. To evaluate diagnostic specificity more precisely, pairwise comparisons between the four skin diseases were performed. Notably, the ROC curve of BCC vs. PSO achieved an AUC value of 1.000 in the test set (Figure 2h). For PSO vs. MM, the AUC values were 0.908 in the training set and 0.889 in the test set (Figure S23a), while the AUC values for PSO vs. SCC reached 0.849 and 0.971, respectively (Figure S23b). Scatter plot visualization exhibited clear separation for BCC vs. PSO, while PSO vs. MM and PSO vs. SCC also presented visible separation (Figure 2k; Figure S23d,e). Consistently, the confusion matrices (Figure S24a,d) indicated excellent classification accuracy for PSO vs. BCC, whereas minor misclassifications were observed for PSO vs. MM and PSO vs. SCC in the training set (Figure S24b,c,e,f), although performance improved in the test set.

A separate subtype analysis of skin cancer was subsequently carried out to further evaluate the diagnostic performance of the NN model. Using ternary classification of BCC, SCC, and MM, the test set yielded AUC values of 0.993, 0.918, and 0.982, respectively (Figure S25a). The confusion matrix in Figure S25b revealed that most errors originated from MM, again suggesting the difficulty of early melanoma detection. In binary subtype comparisons, the AUC values of BCC vs. SCC reached 1.000 in both the training and test sets (Figure 2i), whereas MM vs. SCC and BCC vs. MM achieved AUC values ranging from 0.840 to 0.988 (Figure 2j; Figure S23c). Scatterplot confirmed that BCC and SCC can be completely separated (Figure 2l), while MM vs. SCC and BCC vs. MM also showed clear boundaries (Figure 2m; Figure S23f). In addition, the confusion matrices in Figure S26 demonstrated consistently high classification accuracies for BCC vs MM, BCC vs. SCC, and SCC vs. MM in both the training and test sets, although more errors appeared in BCC vs. MM classification in the training set, which may be related to the relatively small sample size. Taken together, these findings provide strong evidence that an SMF‐based NN model not only achieves accurate differentiation between SD and HC but also demonstrates excellent performance in both multiclass analysis of SD and subtype analysis of SC. The results highlight the potential of NN models as a powerful diagnostic tool, although the challenge of early melanoma detection remains noteworthy.

2.4. Construction and Evaluation of the Biomarker Panel in SD Multiclass Classification and SC Subtype Differentiation

The NN machine learning algorithm based on SMFs demonstrated efficient analytical performance and revealed significant metabolic differences between SD and HC. To further investigate these differences, we visualized the extracted 410 m/z features through a volcano plot. The up‐regulated and down‐regulated features in dermatological patients were screened according to statistical significance thresholds, a t‐test p‐value below 0.05, and a fold change (FC) greater than 1.2 or less than 0.83. As shown in Figure 3a, the number of down‐regulated signals (pink) was slightly higher than the number of up‐regulated features (blue). To accurately identify characteristic metabolites for the multiclass diagnosis of dermatological diseases and for the subtype identification of malignant cancers in clinical settings, we applied a projection variable importance (VIP) threshold to prioritize and retain the most discriminative serum metabolites (Figure 3b). Based on this screening, several candidate metabolites were initially screened. The results of this statistical analysis were combined with information from the Human Metabolome Database (http://www.hmdb.ca/), leading to the final selection of eight key metabolites as a potential biomarker panel (Table S5). As shown in Figure S27, the heatmap visualization of these eight metabolites clearly demonstrates marked differences in pression levels among BCC, MM, SCC, PSO, and HC, which provides a strong basis for subsequent analysis aimed at assessing the contribution of these metabolites to disease detection performance. As shown in Figure S28, the MS/MS spectra of the screened biomarkers further confirm their structural identities. The good agreement between the standard metabolites and the corresponding signals in actual samples supports the reliability of biomarker identification, suggesting their potential for clinical diagnosis.

FIGURE 3.

The comprehensive evaluation of the biomarker panel for the multiclass classification and subtype differentiation. (a) Volcano plots of all features. The blue dots and pink dots indicate the up‐regulated and down‐regulated features in SD patients, respectively. (b) Venn diagram of 410 m/z features selected as a potential biomarker panel with FC > 1.2 or < 0.83, p < 0.05, and VIP >1. (c) The ROC curves and (d) confusion matrices for the NN model distinguishing BCC vs. MM vs. PSO vs. SCC vs. HC based on 8 features in the training set. (e) The correlation heatmap of 8 key metabolites between SD and HC. (f) Pathway analysis of six metabolites. (g–i) The violin plots and box plots of the three key metabolites showing differential expression between SD and HC. ^**** p < 0.0005, ^*** p < 0.005. White circles represent the median, rectangular boxes define the upper and lower quartiles of the distribution, and whiskers indicate 1.5‐fold interquartile spacing.

The diagnostic utility of these eight key features was subsequently evaluated using six machine learning models to classify skin diseases and HC. Encouragingly, all models achieved AUC values greater than 0.90, indicating robust diagnostic capability in the independent test set. Radar plot visualization (Figure S29) and the statistical summary in Table S6 confirmed that all models achieved values of AUC, CA, F1 score, and precision above 0.8, with recall and MCC also exceeding 0.80. Among these models, the NN model demonstrated the highest classification performance, with all six evaluation metrics surpassing 0.95. These results highlight the strong potential of the biomarker panel, particularly when integrated with NN algorithms, for the accurate and efficient diagnosis of SD.

Subsequently, we assessed the performance of the NN model in distinguishing SD from HC based on the biomarker panel. As shown in Figure S30a, the ROC curve for both the training and test sets reached an AUC value of 1.000. The scatter plots (Figure S30b) revealed a nearly perfect separation of SD and HC, while the confusion matrices (Figure S30c,d) indicated classification accuracies ranging from 98.0% to 100.0%. Based on these findings, we further used the NN model to categorize and analyze BCC vs. HC, MM vs. HC, PSO vs. HC, and SCC vs. HC. The ROC curves in Figure S31 showed the AUC values for all four training sets exceeded 0.9, whereas the test sets achieved AUC values greater than 0.99. Consistent with these results, the confusion matrices and scatter plots in Figure S32 demonstrated classification accuracies of 97.8% for BCC, 81.0% for MM, 81.2% for PSO, and 91.7% for SCC in the training set. In the independent test set, the corresponding classification accuracies ranged from 77.8% to 100% (Figure S33). As shown in Figure 3c, the ROC curve distinguishing BCC vs. MM vs. PSO vs. SCC vs. HC based on 8 features from the training set demonstrates the NN model's high‐accuracy differentiation of the four skin conditions, with an area under the ROC curve ranging from 0.863 to 1.000. The corresponding confusion matrix also reveals distinct classification accuracy, with melanoma exhibiting a higher rate of misclassification. This finding aligns with previous research outcomes (Figure 3d).

Beyond binary classification with HC, we explored the ability of the NN model to distinguish among the four skin diseases. Pairwise comparisons revealed variable performance. In the test set, the AUC values were 0.943 for PSO vs. BCC, 0.556 for PSO vs. MM, and 0.700 for PSO vs. SCC (Table S7). The relatively low accuracy in melanoma differentiation was consistent with findings obtained when all features were considered, reflecting the intrinsic challenges of diagnosing this highly malignant and clinically elusive cancer. For skin cancer subtype analysis, the test set yielded AUC values of 0.867 for SCC vs. MM, 0.772 for BCC vs. MM, and 0.965 for BCC vs. SCC, demonstrating generally strong discriminatory performance, though differentiation of melanoma remained less robust. Taken together, these results suggest that the biomarker panel not only reduces model complexity by simplifying the feature set but also maintains classification performance comparable to that of the complete metabolomic profile.

To investigate potential biological underpinnings, we constructed a metabolite‐metabolite interaction network based on these eight biomarkers. As illustrated in Figure S34, the network consisted of 189 nodes and 291 edges, revealing multiple functional associations among core differential metabolites. glutamic acid, serine, phenylalanine, tyrosine, and glucose emerged as central hubs, implying their pivotal roles in metabolic reprogramming associated with BCC, MM, SCC, and PSO. Pearson correlation analysis was further conducted to examine the interrelationships among these metabolites. The correlation heat map in Figure 3e indicated that more than half of the metabolites showed absolute correlation coefficients greater than 0.5, suggesting significantly positive or negative associations. These findings imply that the metabolites are likely to be co‐regulated and may participate in shared or related metabolic pathways. Thereby, providing a possible explanation for their synergistic diagnostic utility. In addition, differential expression analysis was conducted to compare the biomarker levels between SD and HC. As shown in Figure S35a, the violin plots revealed significant up‐regulation in SD, whereas the box plots demonstrated variable levels of up‐regulation across individual dermatological conditions. In Figure S35b, SD also showed significant up‐regulation overall, although SCC and PSO displayed no marked differences compared with HC in the box plots. By contrast, violin plots in Figure S35c–e revealed significant down‐regulation of certain metabolites, with disease‐specific variation observed in the corresponding box plots. These differences between SD and HC were statistically significant (p < 0.005).

To explore the biological significance of these eight metabolites in depth, we examined their combined contribution to metabolic pathways. As shown in Figure 3f, pathway analysis indicated that the eight high‐confidence metabolites were primarily involved in five key pathways, including the biosynthesis of phenylalanine, tyrosine, and tryptophan, the metabolism of phenylalanine, the metabolism of sphingolipids, the metabolism of starch and sucrose, and the metabolism of tyrosine. Correspondingly, violin plots in Figure 3g–i indicated that phenylalanine and L‐tryptophan levels were significantly reduced, whereas ceramide expression was up‐regulated in SD compared with HC. The box plots further revealed that the expression of these metabolites varied across different skin disorders, suggesting disease‐specific metabolic alterations. Enrichment analysis (Figure S36) supported these findings and highlighted the close association of skin diseases with phenylalanine and tyrosine metabolism, sphingolipid/glutathione/glucose/metabolism, phospholipid ethanolamine biosynthesis, and homocysteine/lactose degradation. To gain additional insights into potential mechanisms, we reviewed the literature concerning interactions among the eight metabolites. Previous studies have shown that phosphoethanolamine (pEtn) is significantly elevated in tumor interstitial fluid, whereas its level in normal tissues is extremely low. pEtn suppresses T‐cell receptor signaling by depleting diacylglycerol in T cells [39], thereby promoting immune escape. Given that pEtn shares the same precursor as PGF2α ethanolamine, a competitive or synergistic relationship may exist between these metabolites, influencing tumor development. Abnormal expression of PGF2α ethanolamine in dermatological disorders may therefore serve as a biomarker for disease progression. Ceramide, specifically N‐palmitoyl‐sphing‐4,8‐dienine (cer (d18:1/16:0)), is synthesized from the sphingomyelin backbone and fatty acid chains. It represents a central molecule in sphingolipid metabolism and plays a critical role in skin barrier function. Ceramide also influences cellular processes such as proliferation, inflammation, apoptosis, and autophagy [40]. Abnormal expression of ceramide has been observed in BCC, SCC, MM, and PSO, and is likely to contribute to dysregulation of cell death and proliferation [41].

L‐Tryptophan and tyrosine compete with one another. Tyrosinase is a key rate‐limiting enzyme in melanin biosynthesis, catalyzing the conversion of tyrosine into dopamine and melanin precursors [42, 43]. Down‐regulation of tyrosine in melanoma impairs melanin production, which may facilitate tumor growth and metastasis. Consistently, phenylalanine expression levels were significantly decreased in all four skin diseases compared with HC. Since phenylalanine is converted into tyrosine by phenylalanine hydroxylase, reduced levels of both metabolites disturbances that contribute to tumorigenesis and disease progression. These two metabolites together represent potential biomarkers for melanoma diagnosis [44, 45].

In this study, glucose levels were markedly reduced in BCC, SCC, MM, and PSO, whereas ketone bodies such as acetoacetate were significantly increased compared with HC. Under conditions of glucose limitation, ketone bodies provide an alternative energy source to sustain metabolic activity [46]. Acetoacetate, a major ketone body, is closely linked to acetic acid metabolism [47]. Although few studies have directly connected acetoacetate with skin cancer or psoriasis, acetate has been shown to promote lipid synthesis under hypoxic conditions. This suggests that acetoacetate not only compensates for reduced glucose availability but also participates in epigenetic regulation [48, 49]. Specifically, acetic acid promotes lipid synthesis through ACSS2, whereas acetoacetate influences gene expression through Kacac modification, both of which enhance immune escape via c‐Myc activation in the tumor microenvironment [50]. Reduced glucose may also impair T cell function, thereby facilitating tumor progression. 3‐Phosphoglyceric acid (3‐PG), a glycolytic intermediate, serves as a substrate for serine synthesis [51], which provides essential metabolites for cancer cell proliferation. In both skin cancers and psoriasis, serine expression levels were down‐regulated, consistent with the observed reduction in all glucose expression and impairment of glycolysis [52]. Glutamate metabolism was also altered in these conditions. A cyclized derivative of glutamate, pyroglutamate, participates in the γ‐glutamyl cycle that regulates glutathione synthesis. Glutathione, a tripeptide composed of glutamate, cysteine, and glycine, maintains lymphocyte function and modulates inflammatory responses [53, 54]. Abnormal expression of glutamate can therefore influence glutathione levels and contribute to inflammation in disease states. Collectively, our analyses show that eight characterized metabolites both distinguish SD from HC and elucidate pathogenic mechanisms, supporting a diagnostic panel for BCC, MM, SCC, and PSO.

3. Conclusion

In summary, we successfully prepared G‐RMHJ as an efficient matrix for LDI‐MS through simple synthesis and calcination, enabling rapid acquisition of comprehensive serum metabolic fingerprints without the need for complex sample pretreatment. The synergistic effects of SnO₂ and Ag₂O facilitated efficient charge transfer through a well‐defined p–n heterojunction, while gold incorporation further enhanced stability and optimized the electronic structure. This matrix exhibited high sensitivity, throughput, and reproducibility, allowing accurate classification of four skin diseases and reliable identification of skin cancer subtypes when combined with machine learning (AUC: 0.861−0.983). Eight discriminative metabolites were identified, which not only enabled precise disease differentiation but also provided insights into disease‐associated metabolic pathways. This work establishes a noninvasive and patient‐friendly diagnostic strategy with strong potential for clinical translation in the accurate detection and classification of skin diseases.

4. Experimental Section

4.1. Synthesis of RMHS, RMHJ, and G‐RMHJ

The synthesis of RMHS was slightly modified based on previously reported protocols [55]. Initially, 212.01 mg of SnCl₄·5H₂O and 134.50 mg of C₂H₅NS were added to 25 mL of ethanol, followed by sonication for 10 min to form a homogeneous solution. Subsequently, 6.10 mL of AgNO₃ solution was added to the mixture and sonicated until a uniform dispersion was achieved. The resulting solution was then transferred into a 50 mL Teflon‐lined stainless‐steel autoclave and subjected to hydrothermal reaction at 150°C for 12 h. After cooling to room temperature, the precipitate was collected by centrifugation at 8000 rpm, followed by three successive washes with ethanol and three with deionized water to remove residual impurities. The pale‐yellow product (RMHS) was obtained after drying at 50°C overnight and was calcined in a muffle furnace at 550°C in air for 2 h to yield RMHJ. To prepare G‐RMHJ, a simple post‐modification was performed on RMHJ. Specifically, 100 mg of RMHJ was dispersed in 100 mL of HAuCl4 solution (20 mg/mL) and heated to a boil. Then, 10 mL of sodium citrate solution was quickly added, and the reaction mixture was allowed to reflux for an additional 15 min. After completion, the suspension was cooled to room temperature and the resulting precipitate was collected by centrifugation at 8000 rpm and washed with water. Finally, the G‐RMHJ were obtained by drying in an oven at 50°C overnight.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting File: smtd70782‐sup‐0001‐SuppMat.docx.

Data Availability Statement

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