The blood samples were collected from the First People’s Hospital of Yunnan Province from January to June 2023. The diagnosis of all patients was confirmed. Lung adenocarcinoma was diagnosed through histopathology. All patients with lung adenocarcinoma underwent contrast-enhanced computed tomography (CT) scans of the brain, chest, and abdomen, with female patients also receiving contrast-enhanced CT scans of the pelvis. All lung adenocarcinoma patients underwent single-photon emission computed tomography (SPECT) bone imaging. The TNM staging of lung cancer was determined according to the 8th edition of the TNM classification issued by the International Association for the Study of Lung Cancer (IASLC). Patients with stage TNM I and stage TNM II underwent video-assisted thoracoscopic surgery (VATS), with postoperative histopathology confirming adenocarcinoma. Patients with stage TNM III and stage TNM IV underwent tissue biopsy (with the exception of five cases who underwent medical thoracoscopic pleural biopsy, the rest underwent transbronchial lung biopsy and endobronchial ultrasound-guided transbronchial needle aspiration), with histopathology confirming adenocarcinoma. Pulmonary cryptococcosis, pulmonary sclerosing pneumocytoma, hamartoma, organizing pneumonia, and chronic suppurative inflammation were all diagnosed through surgical lung biopsy. Pulmonary aspergillosis was diagnosed through tissue samples obtained via transbronchial lung biopsy. The remaining non-cancerous lung diseases were diagnosed clinically. Stage TNM I and stage TNM II were set as early-stage lung cancer, and stage TNM III and stage TNM IV as advanced-stage lung cancer.

The blood samples of all patients were taken before treatment, invasive medical tests, and surgical intervention. To avoid the effects of food, all of the samples were taken in the morning before breakfast (fasting for at least 8 hours). EDTA anticoagulation tubes were used to collect blood. The collected blood was immediately sent to the laboratory and centrifuged at 4°C and 3,000 g for 10 min. The separated plasma was stored in a −80°C refrigerator.

The samples (100 μL plasma) were resuspended with 400 μL prechilled 80% methanol. The mixtures were incubated on ice for 5 min and then centrifuged at 4°C and 15,000 g for 20 min. The supernatant was collected and diluted with liquid chromatography–mass spectrometry (LC–MS)-grade water to a methanol concentration of 53%. The samples were centrifuged at 4°C and 15,000 g for 20 min. Finally, the supernatant was injected into the LC–MS system analysis. A quality control (QC) sample used for all LC–MS runs was prepared by combining equal volumes of diluted plasma from each participant.

An Orbitrap Q Exactive™ HF-X mass spectrometer (Thermo Fisher, Bremen, Germany) coupled with a Vanquish UHPLC system (Thermo Fisher, Germany) was used to perform LC–MS analyses. The mass spectrometer was operated in positive and negative polarity modes. The m/z acquisition range was 100–1,500. Samples were injected onto a Hypersil Gold column (100 × 2.1 mm, 1.9 μm) using a 12-min linear gradient at a flow rate of 0.2 mL/min. The eluents for the positive polarity mode were eluent A (0.1% formic acid (FA) in water) and eluent B (methanol). The eluents for the negative polarity mode were eluent A (5 mM ammonium acetate, pH 9.0) and eluent B (methanol). The solvent gradient was set as follows: 2% B, 1.5 min; 2%–85% B, 3 min; 85%–100% B, 10 min; 100%–2% B, 10.1 min; and 2% B, 12 min. The parameter settings for the mass spectrometer were as follows: MS/MS secondary scanning, data-dependent scans; aux gas heater temperature, 350°C; spray voltage, 3.5 kV; aux gas flow rate, 10 L/min; S-lens radio frequency (RF) level, 60; sheath gas flow rate, 35 psi; and capillary temperature, 320°C.

To perform peak picking, peak alignment, and quantitation for each metabolite, Compound Discoverer 3.3 (CD3.3; Thermo Fisher) was used to process the raw data files generated by LC–MS. The main parameters were set as follows: signal intensity tolerance, 30%; peak area was corrected with the first QC; actual mass tolerance, 5 ppm; and minimum intensity.

Peak intensities were normalized to the total spectral intensity, and then the molecular formula with the normalized data was predicted. To obtain accurate qualitative and relative quantitative results, the peaks were matched with the mzVault, mzCloud (https://www.mzcloud.org/), and MassList database.

Statistical analyses were performed using the software R (R version R-3.4.3), CentOS (CentOS release 6.6), and Python (Python 2.7.6 version). Data that were not normally distributed were standardized using the following equation to obtain relative peak areas:

rPA = r Q V Q V s / Q V q c . rPA is the relative peak area. rQV is the sample raw quantitation value, which refers to the result obtained using the CD3.3 software to perform peak integration and calculate the area under the curve for the mass spectrometry peaks corresponding to each metabolite. QVs is the sum of sample metabolite quantitation value. QVqc is the sum of the QC1 sample metabolite quantitation value.

Only the compounds whose coefficient of variation (CV) of relative peak areas in QC samples was less than or equal to 30% were retained, and the metabolite identification and relative quantification results were finally obtained.

E-LUAD was compared with the other three groups to identify significantly different plasma metabolite levels. A Venn diagram was constructed to identify overlapping differential metabolites of the class comparisons. The data were randomly divided into a training set and a validation set. The overlapping differential metabolites were subjected to logistic regression analysis; then, the diagnostic model was constructed. The receiver operating characteristic (ROC) curve was plotted to evaluate the diagnostic performance of the model. The area under the ROC curve (AUC) was used to evaluate diagnostic accuracy: high accuracy (0.9 < AUC < 1) and moderate accuracy (0.7 < AUC ≤ 0.9). An AUC metabolite score meant the diagnostic accuracy for E-LUAD.

The metabolites were annotated using the Human Metabolome Database (HMDB) database (https://hmdb.ca/metabolites), Kyoto Encyclopedia of Genes and Genomes (KEGG) database (https://www.genome.jp/kegg/pathway.html), and LIPIDMaps database (http://www.lipidmaps.org/). The metabolites with variable importance in the projection (VIP) score >1.0, fold change (FC) > 1.5 or FC < 0.667, and p-value < 0.05 were considered to be differential metabolites. Partial least squares discriminant analysis (PLS-DA) and principal component analysis (PCA) were performed at metaX. The goodness of fit for the PLS-DA models was evaluated using three quantitative parameters: R²X, R²Y, and Q². Univariate analysis (t-test) was applied to calculate the statistical significance (p-value).

For clustering heatmaps, z-scores of the intensity areas of differential metabolites were used to normalize the data. The functions of metabolites and metabolic pathways were studied using the KEGG database. When the ratio was satisfied by x/n > y/N, the metabolic pathway was considered enriched (x, the number of differential metabolites associated with a specific metabolic pathway; y, the total number of all metabolites associated with a specific metabolic pathway; and n, the number of differential metabolites annotated by KEGG). The metabolic pathway whose p < 0.05 was considered statistically significantly enriched.

GraphPad Prism 8.0 was used for the statistical analysis of baseline data from the four groups of participants. The Kruskal–Wallis statistic was used for the comparison of non-normally distributed data. The chi-square test was used for the statistical analysis of counting data. p-Value less than 0.05 was considered statistically significant.

The study included four groups: E-LUAD (n = 21), HC (n = 17), NCC (n = 17), and A-LUAD (n = 25).

The NCC group included eight cases of pulmonary infectious diseases (five cases of bacterial pneumonia, one case of active pulmonary tuberculosis, one case of invasive pulmonary aspergillosis, and one case of pulmonary cryptococcosis), three cases of pulmonary sclerosing pneumocytoma, two cases of hamartoma, one case of chronic suppurative inflammation, one case of chronic obstructive pulmonary disease, one case of organic pneumonia, and one case of bronchiectasis.

There were no significant differences in age, smoking status, tumor history, or comorbidity among the four groups (p > 0.05); see Table 1 for details.

A total of 527 metabolites were tentatively identified in positive ion mode ( Supplementary Table 1 ) and 286 metabolites in negative ion mode (Supplementary Table 2 ).

Based on the relative quantification values of metabolites, Pearson’s correlation coefficient between QC samples was calculated. All R² values were greater than 0.98 in positive ion mode, and all R² values were greater than or equal to 0.99 in negative ion mode ( Supplementary Figure 1 ).

We performed PCA on the peaks extracted from all experimental and QC samples. The PCA score plots ( Figures 1A, B ) demonstrated good clustering of pooled QC samples, indicating good stability of the experimental method and high data quality.

The KEGG pathway analysis showed that the identified metabolites were involved in metabolic pathways including cellular processes, environmental information processing, genetic information processing, human diseases, metabolism, and organic systems. The main biological metabolic pathway involved was metabolism in positive/negative ion mode ( Figures 2A, B ).

Compared with the HC group, 82 metabolites were upregulated and eight metabolites were downregulated in the E-LUAD group in positive ion mode ( Supplementary Table 3 ). In negative ion mode, 30 metabolites were upregulated and one metabolite was downregulated in the E-LUAD group (Supplementary Table 4 ).

The PLS-DA model showed that the samples in the E-LUAD and HC groups were distributed in different quadrants ( Figures 3A, D ), indicating significant metabolic differences between the two groups. The permutation test demonstrated that the PLS-DA model was not overfitting: in positive and negative modes, the Q² regression line had a negative intercept, and all blue Q² points were below the original blue Q² point ( Figures 3B, E ). The volcano plot of the overall characteristics of the metabolite peaks is shown in Figures 3C, F .

The metabolic pathways are visually displayed in Supplementary Figures 2 , 3 . The details of metabolic pathways are presented in Supplementary Table 9 . The KEGG pathway enrichment analysis of differential metabolites showed six metabolic pathways with a p-value < 0.05 in positive ion mode, including steroid hormone biosynthesis, ovarian steroidogenesis, aldosterone synthesis and secretion, cortisol synthesis and secretion, Cushing’s syndrome, and the prolactin signaling pathway.

Plotting the ROC curves of these metabolites, the AUC values of the differential metabolites are shown in Table 2 ; Supplementary Tables 3 , 4 . Most metabolites had a moderate discrimination ability.

Compared with the NCC group, there were 33 differentially expressed metabolites (25 upregulated and eight downregulated) in the E-LUAD group in positive ion mode ( Supplementary Table 5 ). In negative ion mode, 33 metabolites were upregulated and one metabolite was downregulated in the E-LUAD group ( Supplementary Table 6 ).

Significant metabolic differences between the two groups were displayed in the PLS-DA model ( Figures 4A, D ). The metabolites of the two groups were distributed in different quadrants. The permutation test demonstrated that the PLS-DA model was not overfitting ( Figures 4B, E ). The volcano plot visually displayed differential metabolites between the two groups ( Figures 4C, F ).

The metabolic pathways are visually displayed in Supplementary Figures 4 , Supplementary Tables 5 . The details of metabolic pathways are presented in Supplementary Table 9 . The KEGG pathway enrichment analysis of differential metabolites showed five metabolic pathways (aldosterone synthesis and secretion, cortisol synthesis and secretion, porphyrin and chlorophyll metabolism, steroid hormone biosynthesis, and Cushing’s syndrome) with a p-value < 0.05 in positive ion mode and one metabolic pathway (biosynthesis of unsaturated fatty acids) in negative ion mode.

Most differential metabolites had moderate discrimination ability ( Table 3 ). Detailed information on these metabolites can be found in Supplementary Tables 5 , 6 .

Compared with the A-LUAD group, we identified nine metabolites with upregulated expression and 31 metabolites with downregulated expression in the E-LUAD group in positive ion mode ( Supplementary Table 7 ). In negative ion mode, 13 metabolites were upregulated and one metabolite was downregulated in the E-LUAD group ( Supplementary Table 8 ).

The samples distributed in different quadrants in the PLS-DA model indicated the significant metabolic differences between the E-LUAD and A-LUAD ( Figures 5A, D ). The permutation test demonstrated that the PLS-DA model was not overfitting ( Figures 5B, E ). The volcano plot visually displayed the overall characteristics of the metabolites ( Figures 5C, F ).

The metabolic pathways are visually displayed in Supplementary Figures 6 , 7 . The details of metabolic pathways are presented in Supplementary Table 9 . The KEGG pathway enrichment analysis of differential metabolites showed only one metabolic pathway (bile secretion) with a p-value < 0.05 in positive ion mode.

The AUCs of 87.5% (35/40) of the metabolites were greater than 0.70 in positive ion mode, but only one (2.5%) metabolite had an AUC greater than 0.9 ( Table 4 ; Supplementary Table 7 ). The AUCs of 71.43% (10/14) of the metabolites were greater than 0.70 in negative ion mode; however, none of the metabolites had an AUC greater than 0.9 ( Table 4 ; Supplementary Table 8 ).

The heatmap based on the differentially expressed metabolites provided intuitive visualizations of the trends in metabolites between the four groups in positive/negative ion mode ( Supplementary Figures 8 , 9 ). As can be seen intuitively from the heatmap, different metabolites increased or decreased in samples of the four groups.

The Venn diagram visualized the overlapping results between the differentially regulated metabolites found at E-LUAD vs. HC, E-LUAD vs. NCC, and E-LUAD vs. A-LUAD comparisons ( Figure 6 ).

In positive ion mode, 16 metabolites were commonly differentially regulated between E-LUAD vs. HC and E-LUAD vs. NCC comparisons, indicating that 16 metabolites can simultaneously distinguish E-LUAD from HC and NCC; 81.25% of these 16 metabolites had AUCs between 0.7 and 0.9, and 12.5% had AUCs greater than 0.9. In negative ion mode, the Venn diagram visualized 13 overlapping differential metabolites between E-LUAD vs. HC and E-LUAD vs. NCC comparisons, indicating that 13 metabolites can simultaneously distinguish E-LUAD from HC and NCC; 76.92% of these 13 metabolites had AUCs between 0.7 and 0.9, and 7.69% had AUCs greater than 0.9.

In positive ion mode, only one metabolite was commonly differentially regulated between E-LUAD vs. HC, E-LUAD vs. NCC, and E-LUAD vs. A-LUAD comparisons, indicating that only one metabolite can simultaneously distinguish E-LUAD from HC, NCC, and E-LUAD. In negative ion mode, the Venn diagram visualized three overlapping differential metabolites, indicating that only three metabolites were specific diagnostic markers for E-LUAD.

Compared with the other three groups, the box plot showed that the four metabolites were significantly upregulated in the E-LUAD group ( Figure 7 , colored green).

The AUCs of the four metabolites are shown in Table 5 . The AUC of one ROC curve was less than 0.70 (0.690), and the AUCs of the remaining ROC curves were all greater than 0.70. The detailed information on these four metabolites is presented in Table 6 .

The data were randomly divided into a training set and a validation set in a 7:3 ratio. The Venn diagram showed that four metabolites can distinguish E-LUAD from HC, NCC, and E-LUAD. Logistic regression analysis showed that except for taurodeoxycholic acid (sodium salt), the p-values of the other three metabolites were all less than 0.05. The three identified metabolites, namely, 12-epileukotriene B4, LPE O-18:0, and 2-[2-(3,4-dichlorophenyl)acetyl]-N-propylhydrazine-1-carbothioamide, were used as diagnostic feature biomarkers for predicting E-LUAD and constructing a diagnostic model.

The ROC curve was used to evaluate the diagnostic performance of the model. The AUC value of the training set was 0.918 ( Figure 8A ), and it was 0.983 in the validation set ( Figure 8B ), indicating that the diagnostic model exhibited a good predictive value for E-LUAD.