Machine learning based on metabolomics unveils neutrophil extracellular trap-related metabolic signatures in non-small cell lung cancer patients undergoing chemoimmunotherapy

Abstract

BACKGROUND

Non-small cell lung cancer (NSCLC) is the primary form of lung cancer, and the combination of chemotherapy with immunotherapy offers promising treatment options for patients suffering from this disease. However, the emergence of drug resistance significantly limits the effectiveness of these therapeutic strategies. Consequently, it is imperative to devise methods for accurately detecting and evaluating the efficacy of these treatments.

AIM

To identify the metabolic signatures associated with neutrophil extracellular traps (NETs) and chemoimmunotherapy efficacy in NSCLC patients.

METHODS

In total, 159 NSCLC patients undergoing first-line chemoimmunotherapy were enrolled. We first investigated the characteristics influencing clinical efficacy. Circulating levels of NETs and cytokines were measured by commercial kits. Liquid chromatography tandem mass spectrometry quantified plasma metabolites, and differential metabolites were identified. Least absolute shrinkage and selection operator, support vector machine-recursive feature elimination, and random forest algorithms were employed. By using plasma metabolic profiles and machine learning algorithms, predictive metabolic signatures were established.

RESULTS

First, the levels of circulating interleukin-8, neutrophil-to-lymphocyte ratio, and NETs were closely related to poor efficacy of first-line chemoimmunotherapy. Patients were classed into a low NET group or a high NET group. A total of 54 differential plasma metabolites were identified. These metabolites were primarily involved in arachidonic acid and purine metabolism. Three key metabolites were identified as crucial variables, including 8,9-epoxyeicosatrienoic acid, L-malate, and bis(monoacylglycerol)phosphate (18:1/16:0). Using metabolomic sequencing data and machine learning methods, key metabolic signatures were screened to predict NET level as well as chemoimmunotherapy efficacy.

CONCLUSION

The identified metabolic signatures may effectively distinguish NET levels and predict clinical benefit from chemoimmunotherapy in NSCLC patients.

Key Words: Non-small cell lung cancer, Chemoimmunotherapy, Neutrophil extracellular traps, Metabolomics, Machine learning

Core Tip: This study found that high neutrophil extracellular trap (NETs) levels in patients with non-small cell lung cancer were associated with poor chemotherapy immunotherapy outcomes. Further study found that 54 different metabolites existed between the high NET and low NET groups. Through screening by a machine learning algorithm, three metabolites, 8,9-epoxy-eicosatrienoic acid, L-malate, and bis(monoacylglycerol)phosphate (18:1/16:0), were found to be biomarkers to predict the effect of chemoimmunotherapy in non-small cell lung cancer patients. This study provided important clues to understanding the relationship between NET levels and the effect of chemoimmunotherapy.

INTRODUCTION

Lung cancer is the leading cause of cancer-related deaths worldwide, accounting for about 18.4%[1]. Non-small cell lung cancer (NSCLC) accounts for 80% of lung cancer[2], with a low 5-year survival rate[3]. Platinum-based doublet chemotherapy combined with immunotherapy is recommended as the first-line treatment for NSCLC patients without driver gene mutation[4]. Despite the benefits of immunotherapy, primary and acquired resistance remains a huge clinical challenge. The response rate of immunotherapy is estimated to be approximately 32%[5]. Hence, the identification of biomarkers in patients with an effective response to immunotherapy is of great clinical significance.

Evidence suggests that the nature of the interaction between tumor and immune cells in the tumor microenvironment (TME) determines the anti-tumor response. Among them, neutrophils (the most abundant cells in the immune system) serve a pivotal role in the TME[6]. Neutrophil extracellular traps (NETs) also affect the efficacy of immunotherapy. NETs can impede anti-tumor immune responses via impairing the function of CD8+ T cells in NSCLC[7]. NETs also physically obstruct tumor cell contact with CD8+ T cells and natural killer cells, protecting the tumor cell from immune toxicity[8]. NETs interact with macrophages, which might enhance tumor immunosuppression[9]. In addition, NETs may also participate in the regulation of tumor immune response by acting on B cells[10]. Thus, NETs might be used to predict the efficacy of immunotherapy.

With in-depth exploration of immune metabolism, it has been recognized that metabolic alterations have a significant impact on tumor growth and metastasis[11]. Studies have shown that neutrophils are dependent on multiple sources such as carbohydrates, amino acids, and fatty acids for fuel selection and metabolism, indicating that neutrophil subsets exhibit metabolic flexibility[12]. Metabolic flexibility refers to the switch between metabolic programs in response to nutrient supply or metabolic demand[13]. In addition, neutrophils undergo metabolic switching and carry out their functions through glycolysis and fatty acid metabolism[14]. These metabolic changes are crucial for the process of chromatin decondensation and DNA release, facilitating NET formation[15].

Under nutrient-limited conditions, immature low-density neutrophils support persistent NET formation through mitochondria-dependent catabolism, thereby facilitating the development of liver metastasis of breast cancer[16]. Thus, enhanced neutrophil metabolic flexibility is essential for tumor growth and metastasis. Unfortunately, the specific link between NETs and metabolism in lung cancer has not been deciphered. Moreover, previous studies have commonly focused on NET-related long noncoding RNA or a gene signature for NSCLC[17,18]. NET-related metabolic signatures have not yet been developed to explore novel biomarkers for predicting NET levels or chemoimmunotherapy efficacy of NSCLC.

In our study, NSCLC patients who received chemotherapy combined with immunotherapy as the first-line treatment were analyzed retrospectively. Further, using metabolomics sequencing data and machine learning methods, key metabolic signatures were screened to help predict the NET level as well as efficacy of chemoimmunotherapy. Thus, the identified metabolic signatures may effectively distinguish NET levels and the clinical benefit of chemoimmunotherapy.

MATERIALS AND METHODS

Participants recruitment

Between June 2021 and December 2022, 159 NSCLC patients receiving first-line chemoimmunotherapy from Hunan Cancer Hospital, China, were analyzed retrospectively. NSCLC was diagnosed by pathological examination. Information on clinical characteristics, including age, sex, pathological type (squamous carcinoma and adenocarcinoma), PD-L1, smoking status, performance status score, and TNM stage were collected. The clinical characteristics of the patients are shown in Table 1. The TNM classification was according to the National Comprehensive Cancer Network Classification Standard Eighth Edition. All patients received first-line chemotherapy plus immune checkpoint inhibitors (ICIs). Treatment regimens for lung squamous carcinoma was intravenous (IV) nab-paclitaxel (260 mg/m²) plus IV carboplatin area under the curve (AUC) 5 d once every 3 wk plus ICIs. Treatment regimens for lung adenocarcinoma was IV pemetrexed (500 mg/m²) plus carboplatin AUC 5 d once every 3 wk plus ICIs. The measurable tumor was evaluated once every 6 wk in the first 12 mo and once every 9 wk in year 2 and beyond using Response Evaluation Criteria in Solid Tumors version 1.1. The efficacy of chemoimmunotherapy were classified into complete response (CR), partial response (PR), stable disease (SD), and progressive disease (PD). All protocols were approved by the Institutional Ethics Committee of Hunan Cancer Hospital, and the subjects provided written informed consent.

Table 1 The basic clinical information of enrolled non-small cell lung cancer patients.

Variables	PD/SD, n = 62	CR/PR, n = 97	t/χ2/Z	P value
Age group			0.832	0.362
< 60 yr	34 (54.8)	46 (47.4)
≥ 60 yr	28 (45.2)	51 (52.6)
Sex			0.313	0.576
Female	11 (17.7)	14 (14.4)
Male	51 (82.3)	83 (85.6)
BMI	22 (16.02, 27.7)	22.39 (16.33, 29.69)	0.242	0.842
Histologic subtypes			2.586	0.108
Squamous carcinoma	33 (53.2)	64 (66.0)
Adenocarcinoma	29 (46.8)	33 (34.0)
Smoking index			1.451	0.237
< 400	20 (32.3)	23 (23.7)
≥ 400	42 (67.7)	74 (76.3)
PS scoring			0.477	0.788
0	10 (16.1)	19 (19.6)
1	50 (80.6)	76 (78.4)
2	2 (3.2)	2 (2.1)
T stage			1.685	0.641
1	2 (3.2)	8 (8.2)
2	18 (29.1)	25 (25.8)
3	11 (17.7)	17 (17.5)
4	31 (50)	47 (48.5)
N stage			0.319	0.956
0	2 (3.2)	4 (4.1)
1	12 (19.4)	21 (21.6)
2	26 (41.9)	41 (42.3)
3	22 (35.5)	31 (32.0)
M stage			1.994	0.425
0	13 (21.0)	18 (18.6)
1	49 (89.0)	45 (81.4)

Data are n (%). BMI: Body mass index; CR: Complete response; PD: Progressive disease; PR: Partial response; PS: Performance status; SD: Stable disease.

Measurement of biochemical indexes

Venous blood samples were collected from the patients and placed in EDTA and serum tubes. The total and sorted white blood cells in venous blood samples were counted using an automatic blood cell analyzer. The central laboratory of Hunan Cancer Hospital measured the percentage of CD3, CD4, CD8, natural killer cells, B cells, and lactate dehydrogenase concentration. Interleukin (IL)-6 and IL-8 levels were assayed using enzyme-linked immunosorbent assay kits (Cusabio Biotech Co., Ltd., Wuhan, China), according to the manufacturer’s instructions.

Detection of NETs

NETs were detected using the Quant-iTTM PicoGreen dsDNA kit. Briefly, the anti-MPO antibody was diluted 1:2000 in a coating buffer (pH 9.6), and 100 μL of antibody was added to each well for incubation at 4 °C overnight. After removing the coating solution from the wells, the plates were washed thrice with PBS. Cells were incubated with 200 μL blocking buffer for 120 min at room temperature. Next, 100 μL of tested samples and standards were added to a 96-well plate and incubated for 2 h at room temperature. In addition, 100 μL of PicoGreen solution was added to the samples and incubated in the dark for 5 min. Fluorescence intensity was determined using a fluorescence microplate reader with excitation and emission wavelengths of 490 and 540 nm, respectively.

Immunofluorescence staining

To assess the NET levels in patients with different treatment responses, we performed immunofluorescence staining for NETs and neutrophils. Staining was performed as described previously[19]. Briefly, the NETs were stained with the following primary antibodies: anti-MPO rabbit monoclonal antibody (1:1500; Invitrogen, Carlsbad, CA, United States) and cit-histone 3 rabbit polyclonal antibody (1:2500; Invitrogen). DNA in tissue sections was stained with 4',6-diamidino-2-phenylindole. Images were observed and photographed using a fluorescence microscope.

Liquid chromatography tandem mass spectrometry analysis

Based on NET scores, patients were divided into the low levels of NETs (L_NET) group and the high levels of NETs (H_NET) group. Sixty samples (30 per group) were selected for liquid chromatography tandem mass spectrometry (LC-MS/MS) analysis. LC-MS/MS analysis was performed using a Shimadzu Nexera X2 LC-3ad system equipped with an Acquity UPLC HSS T3 column (1.8 μm, 2.1 mm × 50 mm column) and a triple quadruplex mass spectrometer (5500 QTRAP; AB Sciex, Framingham, MA, United States). The solvent system used was a 0.1 formic acid aqueous solution (A) and 100% acetonitrile (B). The gradients were as follows: 100% A for 2.5 min, 100%-70% A for 9.0 min, 70%-0% A for 1.0 min, and 0% A continued for 5.4 min. The column temperature was maintained at 40 °C at a flow rate of 200 μL/min. This work was carried out in positive and negative ion modes. The transition was detected using the multiple reaction monitoring mode. Raw MRM data of the MT1000 KIT metabolites were extracted using MultiQuant software (Version 3.0.2; AB Sciex) to obtain the peak areas of each metabolite.

Metabolomic profiling of L_NET and H_NET groups

Three methods, principal component analysis (PCA), partial least square discriminant analysis (PLS-DA), and orthogonal projection to latent structure-discriminant analysis (OPLS-DA), were utilized to perform a detailed unsupervised or supervised analysis of the obtained metabolomic data. PCA was used to reduce the dimensions of the data and reflect the overall distribution of the samples. In addition, PLS-DA and OPLS-DA methods employ partial least squares regression algorithms to develop a correlation model between metabolite expression and sample type to distinguish and classify samples.

Identification of significantly altered metabolites between the L_NET and H_NET groups

According to the OPLS-DA model, the variable importance in projection (VIP) was calculated to measure the contribution of metabolites to the model and mine meaningful metabolites. In this study, metabolites with VIP > 1 and P < 0.05 were considered differential metabolites of the L_NET and H_NET groups. Metabolites were visualized using a heatmap. The classification of differential metabolites was statistically analyzed based on their structure and function. To understand the biological functions of these metabolites, pathway enrichment analysis was performed using Kyoto Encyclopedia of Genes and Genomes[20], with a P value of < 0.05.

Machine learning approach

The distinct metabolites underwent initial filtration and screening through receiver operating characteristic (ROC) analysis. Metabolites demonstrating robust predictive performance (AUC > 0.7) were subsequently analyzed by machine learning. Machine learning is an emerging field in medicine, and a large number of resources are applied to integrate computer science and statistics into medical problems[21,22]. Recently, with the progress of machine learning algorithms and the increase in computational power, numerous diagnostic biomarkers were identified in some diseases. In this study, three algorithms were used to further evaluate the significance of metabolites: Least absolute shrinkage and selection operator regression; support vector machine (SVM)-recursive feature elimination (RFE), and random forest (RF). LASSO regression is a statistical technique that can be used to study the effects of clinical variables in outcome prediction, which can reduce overfitting by constructing a penalty function[23]. A 10-fold cross-validation was performed using the glmnet package in R software to screen for characteristic metabolites with minimal error[24] with the following parameters: Family = “binomial,” type.measure = “class.”

SVM-RFE consists of the classification algorithm and the feature selection algorithm wrapped around, strategized to select or remove some features from the high-dimensional feature set and obtain the optimum feature subset from various candidate subsets generated[25]. We employed the SVM algorithm with the RFE method in R package “e1071” (Version 1.7-12) to rank the above metabolites and then select the lowest error rate as the best combination.

RF is a regression tree technique that uses bootstrap aggregation and randomization of predictors to achieve a high degree of predictive accuracy, which is widely used in biomedical research[26]. The metabolites were identified by the R package randomForest (Version 4.7-1.1), and the characteristic obtained metabolites were sorted using the “Mean DecreaseAccuracy” and “Mean Decrease Gini” methods. Then the top 10 metabolites were selected as characteristic biomarkers. Ultimately, the essential metabolomic signatures were derived by intersecting the feature metabolites identified through the three algorithms.

Correlation analysis between clinical features and metabolic signatures

We also evaluated the association between the metabolic markers and clinical characteristics. Correlation analyses were performed using Spearman’s correlation coefficient, and the results were presented using a heatmap.

Statistical analysis

The R language (Version 4.1.1) was used for data analysis and plot generation, and SPSS (Version 22.0; IBM Corp., Armonk, NY, United States) was used for statistical analysis. Clinical data were expressed as mean ± standard deviation. Univariate analysis was used to calculate the relationship between clinical characteristics and efficacy. Spearman’s correlation coefficient was applied for correlation analysis among clinical characteristics, and the Mann-Whitney U test was used to assess the differences between groups. ROC curve was analyzed for predictive value of key metabolites in chemoimmunotherapy. Differences were considered statistically significant at P < 0.05.

RESULTS

Correlation between clinical variables and chemoimmunotherapy efficacy

First, 159 NSCLC patients receiving immunotherapy were enrolled, with 62 in the SD/PD group and 97 in the CR/PR group. Detailed basic clinical information is summarized in Table 1. The basic characteristics of the two groups were similar. Furthermore, we compared clinical characteristics between chemoimmunotherapy responders and non-responders. As depicted in Table 2, the levels of PD-L1 (P < 0.05), circulating IL-8 (P < 0.05), circulating NETs (P < 0.05), and neutrophil-to-lymphocyte ratio (NLR) (P < 0.05) were obviously different between the PD/SD and PR/CR groups. Patients in the PD/SD group had lower PD-L1 levels and higher IL-8, NET, and NLR levels than those in the CR/PR group. Univariate analysis further confirmed the clinical importance of NET and PD-L1 in the prediction of chemoimmunotherapy efficacy (Table 3). Among the chemoimmunotherapy effective group, the proportion of L_NET was significantly higher (Figure 1A). Neutrophil infiltration and NET formation were found in chemoimmunotherapy ineffective samples (Figure 1B). Furthermore, correlation analysis confirmed that NET was positively correlated with NLR (r = 0.288, P < 0.001) (Figure 1C) and IL-8 (r = 0.188, P = 0.018) (Figure 1D). Thus, increased NET formation and release might play a key role in IL-8 and neutrophil-related chemoimmunotherapy resistance.

Figure 1 Correlation between neutrophil extracellular traps and chemoimmunotherapy efficacy. A: Portion of high neutrophil extracellular traps (NETs) and low NETs in different efficacy group; B: NET formation in clinical samples. 4',6-diamidino-2-phenylindole (blue); MPO (red); cit-histone 3 (green); C: Levels of NETs positively correlated with the neutrophil-to-lymphocyte ratio; D: Levels of NETs positively correlated with the interleukin-8.

Table 2 Potential clinical factors correlated with chemoimmunotherapy efficacy in non-small cell lung cancer patients.

Factors	PD/SD, n = 62	CR/PR, n = 97	t/χ2/Z	P value
IL-6 in pg/mL	16.38 (0.14, 92.85)	17.93 (0.05, 219.79)	0.616	0.538
IL-8 in pg/mL	146.29 (19.87, 1038.52)	112.19 (15.76, 915.15)	1.986	0.047
PD-L1			2.413	0.035
< 1%	20 (32.3)	35 (36.1)
1%-49%	38 (61.3)	43 (44.3)
> 50%	4 (6.4)	19 (19.6)
NETs in pg/mL	740.63 (28.05, 3128.41)	467.54 (11.36, 2036.62)	2.564	0.011
Neutrophil percent	5.88 (1.93, 10.85)	5.09 (1.35, 13.5)	2.078	0.058
CD3 cell percent	63.03 (28.22, 86.55)	64.30 (29.63, 85.35)	0.533	0.594
CD4 cell percent	35.64 (17.56, 53.86)	37.88 (16.13, 62.02)	1.199	0.231
CD8 cell percent	25.07 (10.25, 64.67)	24.13 (10.28, 46.54)	0.215	0.829
B cell percent	10.88 (0.77, 36.50)	8.96 (0.98, 23.31)	1.831	0.067
NK cell percent	23.89 (5.90, 65.47)	24.79 (6.76, 53.68)	0.496	0.622
NLR	4.14 (2.43, 5.81)	3.13 (2.3, 4.2)	2.551	0.011
NLR grouping			6.947	0.008
≤ 4	32 (51.6)	70 (72.2)
> 4	30 (48.4)	27 (27.8)
Lymphocyte percent	1.53 (0.38, 3.04)	1.56 (0.67, 3.32)	0.457	0.647
Hepatic metastases			1.428	0.232
No	47 (75.8)	81 (83.5)
Yes	15 (24.2)	16 (16.5)
Bone metastasis			2.109	0.146
No	49 (79.0)	85 (87.6)
Yes	13 (21.0)	12 (12.4)
Metastatic sites			1.067	0.302
< 3	42 (67.7)	73 (75.3)
≥ 3	20 (32.3)	24 (24.7)

Data are n (%). CR: Complete response; IL: Interleukin; NETs: Neutrophil extracellular traps; NK: Natural killer; NLR: Neutrophil to lymphocyte ratio; PD: Progressive disease; PR: Partial response; SD: Stable disease.

Table 3 Univariate analysis of chemoimmunotherapy efficacy in non-small cell lung cancer patients.

Variables	Univariate analysis
	HR	OR (95%CI)	P value
Age	1.004	0.956-1.061	0.798
Sex	1.047	0.305-2.734	0.871
Histologic subtypes	2.033	0.835-4.817	0.119
IL-6	1.005	0.990-1.023	0.463
IL-8	1.000	0.997-1.002	0.762
NETs	0.999	0.998-1.000	0.006
PS score	0.494	0.193-1.267	0.142
T stage	0.858	0.600-1.352	0.614
N stage	0.863	0.576-1.504	0.769
M stage	0.326	0.276-1.626	0.376
PD-L1 expression	1.029	1.010-1.051	0.004
CD3 percentage	0.985	0.919-1.079	0.769
CD4 percentage	1.068	0.989-1.156	0.094
CD8 percentage	1.008	0.941-1.079	0.827
NK percentage	1.025	0.969-1.093	0.354
B cell percentage	0.933	0.860-1.025	0.162
NLR	1.135	0.739-1.739	0.565
PMN percentage	0.784	0.544-1.141	0.206
Lymphocyte percentage	0.811	0.457-5.365	0.476

CI: Confidence interval; HR: Hazard ratio; IL: Interleukin; NETs: Neutrophil extracellular traps; NK: Natural killer; NLR: Neutrophil-to-lymphocyte ratio; OR: Odds ratio; PMN: Polymorphonuclear neutrophil; PS: Performance status.

Metabolic characteristics of the H_NETs and L_NETs groups

Because NETs are closely related to the efficacy of chemoimmunotherapy, patients were divided into the L_NETs and H_NETs groups and analyzed for metabolic characteristics. A total of 524 plasma metabolites were detected using LC-MS/MS. The identified metabolites were input for the R software for PCA. The outcomes indicated that the two principal components elucidated 6.31% and 29.14% of data variance. Notably, there was partial overlap and separation between the H_NETs and L_NETs groups (Figure 2A). In addition, the PLS-DA and OPLS-DA models clearly distinguished the H_NETs and L_NETs groups (Figure 2B and C). The permutation test of the OPLS-DA model found that the model did not overfit, with R2 = 0.79 and Q2 = -1.04 (Figure 2D).

Figure 2 Metabolic profiling of the low neutrophil extracellular traps and high neutrophil extracellular traps groups. A: Low neutrophil extracellular traps (L_NETs) and high neutrophil extracellular traps (H_NETs) groups based on principal component analysis; B: L_NETs and H_NETs groups based on partial least square discriminant analysis; C: L_NETs and H_NETs groups based on orthogonal partial least square discriminant analysis; D: L_NETs and H_NETs groups based on permutation plots.

Comparison of metabolic profiling between the H_NETs and L_NETs groups

Based on the VIP (calculated using OPLS-DA) and P values, the differential metabolites in the H_NETs and L_NETs comparisons were screened. A total of 54 differential metabolites were identified, including 39 upregulated and 15 downregulated metabolites. All differential metabolites were visualized using a heat map (Figure 3A). These metabolites could be classified into 14 categories, which mainly included arachidonic acid metabolites (22.22%), organic acids and derivatives (18.52%), and nucleosides, nucleotides, and analogs (12.96%) (Figure 3B). Next, ranking by VIP value, the top 30 metabolites are shown in Figure 3C. We observed that metabolites such as LTB4, 5−hydroxyeicosatatraeniocacid, and 8,9-epoxyeicosatrienoic acid were increased, whereas PG (16:0/18:1), SM 40:1; 2, and SM 38:3; 3 were decreased in the H_NETs group.

Figure 3 Identification of differential metabolites between the low neutrophil extracellular traps and high neutrophil extracellular traps groups. A: Heatmap displaying 54 differential metabolites within the low neutrophil extracellular traps and high neutrophil extracellular traps groups; B: Classification, number, and proportion of differential metabolites; C: Variable importance in projection score plot of the top 30 differential metabolites ranked according to importance.

Pathway enrichment analysis was performed to explore the specific functions of these metabolites. As shown in Figure 4A, the differential metabolites were primarily involved in 22 Kyoto Encyclopedia of Genes and Genomes pathways, particularly metabolism-related pathways. Trend analysis was employed to observe changes in these pathways between the two groups. The results revealed that most metabolism-related pathways were upregulated in the H_NETs group, including arachidonic acid and purine metabolism (Figure 4B). Thus, these differential metabolites may influence the final therapeutic outcome of NSCLC by modulating these metabolic pathways.

Figure 4 Kyoto Encyclopedia of Genes and Genomes pathway analysis of differential metabolites between the low neutrophil extracellular traps and high neutrophil extracellular traps groups. A: Bubble map of enriched Kyoto Encyclopedia of Genes and Genomes pathways; B: Differential abundance score of Kyoto Encyclopedia of Genes and Genomes pathways.

Development of metabolite-based diagnostic model via machine learning

To establish the diagnostic model, ROC analysis and the AUC were used to evaluate the diagnostic performance of the differential metabolites. Notably, metabolites with an AUC > 0.7 represented accurate prediction ability. Thus, markers with an AUC ≥ 0.7 were selected as candidate metabolites for subsequent machine learning analysis, and a total of 15 metabolites were obtained. Based on these 15 markers, three diverse algorithms were employed to screen for potential metabolite signatures.

In the LASSO regression analysis, eight metabolites were revealed as diagnostic biomarkers after ten times cross-validation: 13S-hydroxyoctadecadienoic acid; 8,9-epoxyeicosatrienoic acid; 15-hydroxyeicosatetraenoic acid; 5-hydroxyeicosatatraeniocacid; 5-hydroxy-6,8,11,14,17-eicosapentaenoic acid; bis(monoacylglycerol)phosphates (BMP) (18:1/16:0); azelaic acid; and L-malate (Figure 5A and B).

Figure 5 Screening of metabolic biomarkers via three machine learning algorithms. A and B: Least absolute shrinkage and selection operator regression model showed eight metabolites corresponded to the lowest binominal deviation; C: Support vector machine-recursive feature elimination model showed six biomarkers with the highest accuracy; D and E: Top 10 metabolites ranked by mean decrease Gini in the random forest model; F: Venn diagram revealing three shared metabolites in three algorithms.

In SVM-RFE, the accuracy of the model was the highest when there were six characteristic metabolites, including L-malate, 8,9-epoxyeicosatrienoic acid, BMP (18:1/16:0), 14,15-epoxy-5,8,11-eicosatrienoic acid, 15-hydroxyeicosatetraenoic acid, and 8(R)-hydroxy-eicosatetraenoic acid (Figure 5C).

In addition, according to the mean decrease Gini value, ten key characteristic metabolites were obtained by the RF algorithm, namely 14,15-epoxy-5,8,11-eicosatrienoic acid, 8,9-epoxyeicosatrienoic acid, BMP (18:1/16:0), L-malate (MTID74), 13S-hydroxyoctadecadienoic acid, 11-hydroxyeicosatetraenoic acid, 11,12-epoxyeicosatrienoic acid, 8(R)-hydroxy-eicosatetraenoic acid, (9R)-9-hydroxyeicosatetraenoic acid, and azelaic acid (Figure 5D and E).

Next, the metabolites obtained by the three algorithms were integrated to reveal a common signature; 8,9-epoxyeicosatrienoic acid, BMP (18:1/16:0), and L-malate were finally selected (Figure 5F). The AUC of these metabolites were 0.798, 0.754, and 0.702, respectively (Figure 6A). Moreover, we observed the relative abundance of the three metabolites in the different NET groups. Compared to the L_NETs group, 8,9-epoxyeicosatrienoic acid, BMP (18:1/16:0), and L-malate levels were significantly higher in the H_NETs group (P < 0.05, Figure 6B). These data confirmed the potential application of these three metabolites in NSCLC genesis and treatment.

Figure 6 Receiver operating characteristic analysis and relative abundance of three metabolites. A: Receiver operating characteristic analysis of three shared metabolites; B: Relative abundance level of three shared metabolites. ^aP < 0.01; ^bP < 0.001.

Relationship of diagnostic metabolites and chemoimmunotherapy efficacy

We also observed correlations between the metabolites and clinical features. As shown in Figure 7, all three diagnostic signatures were positively correlated with NETs. Then we compared the key metabolites in the effective and ineffective groups. The AUC of 8,9-epoxyeicosatienoic acid reached 0.74, indicating predictive ability for treatment efficacy (Figure 8A). 8,9-epoxyeicosterogenic acid had a significant difference (P < 0.05) (Figure 8B). Thus, 8,9-epoxyeicosatienoic acid might be used to predict chemoimmunotherapy efficacy through regulation of NETs release.

Figure 7 Heatmap displaying the correlation between differential metabolites and clinical characteristics. ^aP < 0.05.

Figure 8 The predictive value of key metabolites in chemoimmunotherapy efficacy prediction. A and B: Receiver operating characteristic analysis (A) and relative abundance level (B) of three shared metabolites.

DISCUSSION

As immune defense mechanisms that promote tumor metastasis and progression, NETs might be related to chemoimmunotherapy efficacy. In breast cancer with lung metastasis, chemotherapy induced neutrophil infiltration and NET formation[27]. The biological processes of NETs are closely related to cellular metabolism. There is evidence of a cancer-induced shift in neutrophil metabolism toward glycolysis, which causes an increase in NETs[28]. Although NETs are involved in the initiation and maintenance of immune suppression in lung cancer, the NET-related metabolic signatures that predict therapeutic responses have not been fully identified.

We first found clinical indicators clearly associated with efficacy and found that increased NETs led to poor chemoimmunotherapy efficacy. Recent research has confirmed that NETs protect tumor cells from cytotoxicity attacks activated by ICIs through spatial blockade[29]. NETs promote T cell exhaustion and affect the killing activity of immune cells, which negatively impacts the efficacy of immunotherapy[30]. NET inhibitors combined with immunotherapy attenuated tumor growth in animal studies[8].

IL-8 levels were negatively correlated with therapeutic efficacy and positively correlated with NETs. Consistent with our results, high levels of IL-8 predicted cancer in patients who failed to benefit from chemoimmunotherapy[7]. Mechanistically, IL-8 rapidly recruits neutrophils to the TME, triggering the release of NETs and facilitating tumor immune escape[31].

In addition, neutrophils can inhibit or stimulate cytotoxicity T cell responses, and the inhibitor/stimulation balance can be mediated by the NLR[32]. Briefly, an increase in neutrophils is frequently accompanied by a decrease in lymphocytes, representing a decline in the adaptive immune response[33]. Evidence has shown that NLR has prognostic value in multiple tumors and can be used as a predictor of treatment outcomes. Low NLR was associated with a better response and predicted better survival in patients with metastatic colorectal cancer receiving immunotherapy[34]. Moreover, in patients with biliary tract cancer undergoing chemoimmunotherapy, those with a persistently high NLR 1 mo after treatment had a significantly worse prognosis than those with normal or low NLR levels[35]. The results of these studies support our findings. Taken together, these findings provide a valuable reference for NETs in predicting chemoimmunotherapy efficacy.

It is generally accepted that metabolomics can contribute to cancer progression through aberrant glycolysis and fatty acid synthesis[36]. However, studies on the metabolic features associated with chemoimmunotherapy are limited. Hence, in the subsequent analyses, we explored the underlying metabolic alterations based on different NET scores. A total of 54 differential metabolites were screened, and the pathways enriched by these metabolites were analyzed. These results revealed that arachidonic acid metabolism was upregulated in the H_NETs group.

Abnormalities in arachidonic acid metabolism can lead to the development of diseases, including cancer, by activating multiple signaling pathways[37]. Arachidonic acid and its metabolites trigger oxidative stress and stimulate immune responses[38]. In lung cancer, key enzymes involved in arachidonic acid metabolism, such as cyclooxygenase-2 and cytosolic phospholipase A2, may be involved in the occurrence of tumors. Modulation of arachidonic acid metabolism by inhibiting the activities of these enzymes is an effective strategy for cancer treatment[39]. In addition, arachidonic acid metabolism is involved in the immune escape and immunotherapy resistance of cancer cells.

Recent experiments have revealed that CYP1B1 induces resistance to ferroptosis in colorectal cancer cells by activating the protein kinase C signaling pathway and promoting the degradation of ACSL4, thereby affecting the sensitivity to anti-PD-1 therapy[40]. Subsequent animal studies have confirmed that CYP1B1 inhibition enhances the sensitivity of tumor cells to immunotherapy.

We also observed that purine metabolism was enriched in the H_NETs group. A close relationship has been reported between purine metabolism and cancer. Purines are an important raw material for cell proliferation, and its metabolism is regulated by various enzymes[41]. The dysfunction of enzyme function leads to excessive cell proliferation and immune balance, causing tumor progression[42]. Purine anti-metabolizers represent a large portion of the currently available cancer therapies and can either block DNA replication or cause apoptosis of cancer cells through DNA damage[43]. Collectively, these studies support our findings that differential metabolites may influence the final therapeutic outcome by modulating key metabolic pathways.

We utilized three machine learning algorithms to select three specific metabolites with excellent predictive power that could accurately distinguish the L_NETs and H_NETs groups. They included 8, 9-epoxyeicosatrienoic acid, L-malate, and BMP (18:1/16:0). 8, 9-epoxyeicosatrienoic acid belongs to the epoxyeicosatrienoic acids (EETs) and are metabolites of cytochrome P450 synthesized from essential arachidonic acid[44]. EETs are endothelial cell-derived hyperpolarizing factors that may play an important role in the regulation of apoptosis and inflammation[45]. There is also evidence that EETs affect immune response mechanisms.

Gao et al[46] found that 8,9-EETs specifically inhibit the function of B cells, slow cell proliferation, and reduce cell survival, leading to attenuated and shortened cellular immunity. The metabolism of EETs is vasoactive and affects intracellular calcium concentrations[47]. The contribution of EETs to tumorigenesis and metastasis is attributed to the downstream metabolism of COX[48]. Mechanistically, ct-8,9-epoxy-11-hydroxy-eicosatrienoic acid, the major product formed by 8, 9-EETs in the COX pathway, promotes angiogenic tumor products, which in turn enhance the cancer response[48].

L-malate is an essential intermediate in the C4-dicarboxylic and tricarboxylic acid cycles and plays a crucial role in the transport of NADH from the cytoplasm to the mitochondria[49]. Feeding with L-malate can improve energy metabolism and reverse oxidative stress in aged rats, thereby improving mitochondrial function and enhancing performance during forced swimming[50].

BMP is a structural isomers of phosphatidylglycerol that is highly enriched in endolysosomes and are responsible for activating sphingolipid catabolic enzymes[51]. A recent study described the changes in BMP diversity in breast cancer[52]. The investigators found that the total BMP levels were significantly reduced in breast cancer cells, suggesting that BMP was specifically inhibited during malignant transformation. This may result from the disruption of lysosomal membrane stability due to cellular oxidative stress in cancer cells[53].

The above studies confirm the potential application of these three metabolites in NSCLC genesis and treatment, and the metabolic signatures might be highly sensitive and specific biomarkers for NSCLC diagnosis and treatment and effectively distinguish circulating NET levels. However, the specific link between metabolites and the NET mechanism has not been explored.

Targeting tumor metabolism has been found to be effective in clinical practice[54]. Therefore, identifying potential metabolism-based therapeutic targets is important. Our study provided valuable information regarding the metabolic profiles of NETs in patients with lung cancer receiving chemoimmunotherapy. Importantly, the relative abundances of the three metabolites, 8,9-epoxyeicosatrienoic acid, BMP (18:1/16:0), and L-malate, was observed in the different NETs groups. Compared to the L_NETs group, 8,9-epoxyeicosatrienoic acid, BMP (18:1/16:0), and L-malate levels were significantly higher in the H_NETs group. This study also found that all three diagnostic signatures were positively correlated with NETs, and there is evidence that they are involved in tumor progression.

In addition, the key metabolites in the effective and ineffective therapy groups were compared. The AUC of 8,9-epoxyeicosatienoic acid reached 0.74, indicating the predictive ability for treatment efficacy. Therefore, targeting these metabolites may help improve the clinical outcome of NSCLC immunotherapy. This study revealed the correlation between the efficacy of chemoimmunotherapy and NET levels in lung cancer. Importantly, for the first time, we employed machine learning algorithms based on serum metabolic sequencing data to screen for differential metabolites with diagnostic value between the L_NETs and H_NETs groups. These metabolites and their associated pathways provide valuable clues for understanding the mechanisms underlying NET levels and their therapeutic effects.

This study had several limitations. First, the sample size was relatively small and lacked proper validation, limiting the generalization of the results to larger populations. Second, the exact mechanism by which the identified metabolites affect NETs was not elucidated. However, the present findings require further confirmation, particularly in large cohorts.

CONCLUSION

In summary, we found the circulating NET level was closely related to first-line chemoimmunotherapy efficacy in NSCLC. By using plasma metabolic profiles and machine learning algorithms, predictive metabolic signatures were established. The identified metabolic signatures might effectively distinguish circulating NET levels. Using metabolomics sequencing data and machine learning methods, key metabolic signatures were screened to predict NETs level as well as chemoimmunotherapy efficacy.