The Overlooked Autoantibody Repertoire: Exploring the Biomarker Potential of Downregulated Autoantibodies in NSCLC

ABSTRACT

Autoantibodies (AAbs) represent promising biomarkers in cancer. While most AAbs are elevated in cancer, a substantial subset is downregulated, and their diagnostic and prognostic potential remains largely unexplored. Here we used the HuProt protein microarray to identify downregulated AAbs in non‐small cell lung cancer (NSCLC) serum. Indirect ELISA quantified serum levels in 781 samples. Ten machine learning algorithms were used to construct diagnostic models. An independent cohort of 353 NSCLC patients was used to assess prognostic value and develop a prognostic model. Six downregulated AAbs were identified, among which five AAbs (anti‐HIST1H1B, anti‐HIST1H1C, anti‐DYDC2, anti‐CAMKK2, and anti‐GRPEL1) were significantly reduced in NSCLC. The gradient boosting machine (GBM) model showed the best performance for NSCLC and BPNs, with AUCs of 0.869 (95% CI: 0.833–0.905) in the training set and 0.813 (95% CI: 0.745–0.880) in the validation set. For early‐stage NSCLC, the model achieved an AUC of 0.809 (95% CI: 0.729–0.890) in the validation set, with a sensitivity of 74.0% and specificity of 81.3%. Multivariate Cox regression identified four AAbs significantly associated with patient prognosis. A prognostic model integrating age and AAb levels demonstrated robust predictive performance for long‐term survival (7‐year AUC = 0.79). Bioinformatics analyses further supported the relevance of the corresponding genes/proteins of these AAbs to NSCLC outcomes. Overall, our findings demonstrate that downregulated AAbs possess significant diagnostic and prognostic value in NSCLC and may contribute to improved patient management and survival prediction.

This study identifies downregulated autoantibodies in NSCLC and demonstrates their diagnostic and prognostic utility. Integrating proteome‐wide screening and large‐cohort validation, we highlight a novel biomarker class that enhances early detection and patient stratification.

Abbreviations

AAbs

autoantibodies

AUC

area under curve

BPN

benign pulmonary nodules

CEA

carcinoembryonic antigen

CI

confidence interval

CNV

Copy number variation

CPTAC

Clinical Proteomic Tumor Analysis Consortium

DM

distant metastasis

ELISA

indirect enzyme‐linked immunosorbent assay

GBM

gradient boosting machine

GEO

gene expression omnibus

GO

gene ontology

IgG

immunoglobulin G

LASSO

least absolute shrinkage and selection operator

LC

lung cancer

LM

lymph node metastasis

LUAD

lung adenocarcinoma

LUSC

lung squamous cell carcinoma

NCs

normal controls

NSCLC

non‐small cell lung cancer

OD

optical density

OR

odds ratio

OS

overall survival

PPI

protein–protein interaction

SBI

specific binding index

SNR

signal‐to‐noise ratio

TCGA

The Cancer Genome Atlas

1. Introduction

Lung cancer (LC) remains a leading cause of cancer incidence and mortality worldwide. In 2022, there were approximately 2.5 million new cases and over 1.8 million deaths globally, accounting for nearly one‐eighth (12.4%) of all cancer incidences and nearly one‐fifth (18.7%) of all cancer‐related deaths [1, 2]. Although LC survival and mortality have improved in recent years due to early detection and treatment, it still maintains an unfavorable prognosis, with 5‐year survival rates typically below 20% [2, 3]. Studies have shown that NSCLC is the major subtype of LC, mainly including two types: lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC) [4]. Therefore, in‐depth investigation of various aspects of NSCLC is of great importance for improving the overall prevention and treatment of LC.

Autoantibodies (AAbs), which are immunoglobulins produced by the immune system, can be detected in peripheral blood and remain stable over time, and are not limited to tumor tissues or cells but may also be present in normal tissues [5, 6]. AAbs, as crucial factors influencing human health and cancer patients, are largely underexplored. “Science” has pointed out that differences in the “autoantibody(AAb) response groups” (pathologic and protective) between LC patients and healthy individuals have become particularly critical for the diagnosis and prognosis of LC [7]. There is a bias in current studies on AAbs. Most studies focus on AAbs with elevated expression in tumors [5, 8, 9, 10]. The exploration of downregulated AAbs in tumors is still insufficient. Notably, a clinical study demonstrated that serum anti‐KL6/MUC1 AAbs levels were significantly reduced in NSCLC, and that patients with high concentrations exhibited better one‐year survival, highlighting the clinical importance of downregulated AAbs [11]. Therefore, identifying novel downregulated AAb biomarkers is crucial for improving the diagnosis and prognostic evaluation of NSCLC.

In this study, we identified a set AAbs by applying HuProt protein chip. Compared with normal control serum, several AAbs were generally down‐regulated in the serum of NSCLC patients. We evaluated the diagnostic and prognostic value of these AAbs. In summary, our aim was to identify AAbs of significant value for the diagnosis and prognostic assessment of NSCLC, providing new strategies to improve the survival outcome of NSCLC patients.

2. Materials and Methods

2.1. Study Participants

This study comprises two independent cohorts: a diagnostic cohort and a prognostic cohort. The diagnosis cohort was recruited at the First Affiliated Hospital of Zhengzhou Universityfrom December 2017 to June 2023. The cohort included 306 newly diagnosed NSCLC patients, 215 patients with benign pulmonary nodules (BPN), and 260 normal controls (NC). Serum levels of carcinoembryonic antigen (CEA) were measured and recorded for all participants. Thereinto, 15 plasma samples were also collected from NSCLC patients before surgery and approximately 2 weeks after surgery. Patients admitted with suspicious pulmonary nodules detected on chest CT underwent fasting blood sampling on the first day of admission before treatment. NSCLC was histopathologically confirmed in all patients. BPN was confirmed using clinical and pathological evaluation. NC had no history of lung or autoimmune diseases. The prognostic cohort included 353 NSCLC patients diagnosed by radiological or pathological methods and treated according to standard protocols, with complete follow‐up over a 7‐year period (December 2017–August 2024). Clinical data regarding survival status were collected regularly by telephone follow‐up or clinical report review. Detailed clinical characteristics are summarized in Table 1. All serum samples were collected and stored according to standard protocols [12].

TABLE 1.

Clinical information of diagnosis cohort and prognostic cohort.

Characteristic	Diagnosis cohort				Prognostic cohort
	NSCLC (n = 306)	BPN (n = 215)	Normal (n = 260)		NSCLC (n = 353)
Age (years)
≤ 55	113 (36.9%)	85 (39.5%)	92 (35.3%)		108 (30.6%)
> 55	193 (63.1%)	130 (60.5%)	168 (64.6%)		245 (69.4%)
Mean ± SD	58.8 ± 9.9	57.1 ± 9.7	58.4 ± 10.1	p > 0.05
Sex
Male	177 (57.8%)	135 (62.8%)	149 (57.3%)		224 (63.5%)
Female	129 (42.2%)	80 (37.2%)	111 (42.7%)	p > 0.05	129 (36.5%)
T
T1	190 (62.1%)				145 (41.1%)
T2	52 (17.0%)				76 (21.5%)
T3	15 (4.9%)				31 (8.8%)
T4	23 (7.5%)				85 (24.1%)
Unknown	26 (8.5%)				16 (4.5%)
Stage of tumor
Early	164 (53.6%)				147 (41.6%)
Advanced	101 (33.0%)				191 (54.1%)
Unknown	41 (13.4%)				15 (4.3%)
LM
Yes	113 (36.9%)				183 (51.8%)
No	165 (53.9%)				150 (42.5%)
Unknown	28 (9.2%)				20 (5.7%)
DM
Yes	46 (15.0%)				122 (34.6%)
No	211 (69.0%)				201 (56.9%)
Unknown	49 (16.0%)				30 (8.5%)
History of smoking
Yes	95 (31.0%)	71 (33.0%)			154 (43.6%)
No	205 (67.0%)	139 (64.7%)			194 (55.0%)
Unknown	6 (2.0%)	5 (2.3%)			5 (1.4%)
Histological subtypes
LUAD	256 (83.7%)
LUSC	50 (16.3%)
Inflammatory/infectious nodules		182 (84.7%)
Benign neoplastic		14 (6.5%)
Other		19 (8.8%)

Note: Inflammatory/infectious nodules: non‐neoplastic pulmonary nodules caused by infectious or inflammatory processes, including pulmonary tuberculosis, granulomatous inflammation, inflammatory pseudotumors, and other infection‐related or non‐specific inflammatory nodules.

Abbreviations: Advanced, StageIII and Stage IV; BPN, benign pulmonary nodule; DM, distant metastasis; Early, StageI and StageII; LM, lymph node metastasis.

Comprehensive multi‐omics data and clinical characteristics were obtained from three public databases to further support the study conclusions. From The Cancer Genome Atlas (TCGA), RNA‐seq data were downloaded for 528 LUAD tumors with 59 matched normal tissues and 501 LUSC tumors with 51 matched normals. Clinical Proteomic Tumor Analysis Consortium (CPTAC) included 111 LUAD tumor‐normal pairs and 110 LUSC tumors with 102 matched normals. Immunohistochemistry (IHC) data were obtained from the Human Protein Atlas (HPA) database [13] (https://www.proteinatlas.org/). NSCLC single‐cell transcriptomic datasets were retrieved from the Gene Expression Omnibus (GEO) database, including GSE148071 (n = 42) and GSE117570 (n = 4).

2.2. HuProt Array Preparation and Analysis

This study is considered to be a reanalysis of our previous data [8]. The HuProt protein chip was described and detailed in our previous paper. Plasma samples were incubated on the arrays, and IgG AAbs were detected using isotype‐specific fluorescent secondary antibodies. The “limma” package of R was used to screen down‐regulated AAbs in NSCLC. For each AAb, a positive threshold was defined based on normalized signal‐to‐noise ratio (SNR), and positive rates were calculated for the NSCLC and NC groups. A protein–protein interaction (PPI) network of genes corresponding to significantly downregulated AAbs was constructed using the STRING database (http://string‐db.org/). Interaction data were analyzed in Cytoscape (v3.10.2), and key hub proteins were identified using CytoHubba based on MCC and Degree scores.

2.3. Enzyme‐Linked Immunosorbent Assay (ELISA)

An indirect ELISA was performed to detect the expression levels of candidate downregulated AAbs in serum. Six recombinant proteins used in this study included HIST1H1A, HIST1H1B, HIST1H1C, and CAMKK2 (from Cloud‐Clone Corp), as well as GRPEL1 and DYDC2 (from Cusabio Technology LLC). Six recombinant proteins were coated onto 96‐well plates at a concentration of 0.125 μg/mL and incubated overnight at 4°C, followed by blocking with 2% BSA. Serum samples (1:100) served as the primary antibody, and HRP‐conjugated goat anti‐human IgG (1:5000) as the secondary antibody. Each plate included a blank control (without serum) and a quality control (pooled serum from 100 healthy controls).

The analytic value was expressed as the specific binding index (SBI), which was calculated using the following formula:

$$\begin{array}{cc}\mathrm{SBI}\hfill & =\left(\mathrm{OD}\text{value of study subjects}-\mathrm{OD}\text{value of blank control}\right)\hfill \\ & /\left(\mathrm{OD}\text{value of quality control sample}-\mathrm{OD}\text{value of blank control}\right).\hfill \end{array}$$

2.4. Single‐Cell Analysis of Genes Corresponding to AAbs in NSCLC

Single‐cell RNA sequencing data processing was conducted using the “Seurat” package (version 5.3.0). Cells with fewer than 300 or more than 5000 expressed genes (nFeature_RNA), total UMIs (nCount_RNA) below 300, mitochondrial gene content exceeding 20%, or ribosomal gene content below 3% were excluded. Highly variable genes (n = 3000) were selected for PCA. Other processes followed the standard Seurat workflow [14]. To eliminate batch effects, the Harmony method was used to correct datasets. Clustering results were visualized using the t‐distributed stochastic neighbor embedding (t‐SNE) algorithm. By integrating marker genes from the CellMarker 2.0 (http://117.50.127.228/CellMarker/) and published literature [15, 16], seven major cell types were identified: fibroblasts, endothelial cells, mast cells, B cells, epithelial cells, NK/T cells, and myeloid cells. Copy number variation (CNV) analysis was performed using the “inferCNV” R package [17], with fibroblasts as reference cells to identify malignant epithelial cells. The “CytoTRACE” R package was employed for inferring differentiation potential across epithelial subclusters [18].

2.5. Statistical Analysis

Data analysis and visualization was performed using R‐4.4.0. The Wilcoxon test and paired t‐test were used to analyze differences in AAbs levels and molecular expression among different groups. Optimal cutoffs for positive classification were determined by the maximum Youden's index (YI) on receiver operating characteristic (ROC) curves. Ten machine learning algorithms were implemented using the “caret” package in R, including support vector machine (SVM), random forest (RF), generalized linear model (GLM), gradient boosting machine (GBM), C5.0 decision tree (C5.0), linear discriminant analysis (LDA), naïve Bayes (NB), neural network (Nnet), k‐nearest neighbors (KNN), and adaptive boosting (AdaBoost). Serum samples were randomly divided into training and testing sets at a 7:3 ratio. To ensure robustness, 10‐fold cross‐validation was repeated five times. Model performance was evaluated by accuracy, sensitivity, specificity, YI, positive predictive value (PPV), negative predictive value (NPV), Cohen's kappa, F1 score, and area under the curve (AUC). Decision curve analysis (DCA) was performed using “dcurves” package, and AUCs were compared using the DeLong test. Kaplan–Meier (K‐M) survival analysis was performed using the “survival” R package and differences between K‐M survival curves were compared using the log‐rank test. The optimal cutoff value for each AAb was determined with “survminer” to stratify patients into high and low expression groups. Variables of AAbs and age were analyzed using least absolute shrinkage and selection operator (LASSO) Cox regression with the “glmnet” R package, and the optimal penalty parameter (λ) for risk model construction was determined via 10‐fold cross‐validation. Based on the regression coefficients of key factors, a risk score was calculated for each patient. The “rms” were used to construct a nomogram and calibration curves. Time‐dependent ROC curve analysis was performed using the “timeROC” package. Univariate and multivariate Cox proportional hazards regression models were applied to evaluate prognostic factors. The “shiny” package was used to develop and deploy a web application on the shinyapps.io platform. p < 0.05 (two‐sided) was considered statistically significant for all analyses.

3. Results

3.1. Identification of Downregulated AAbs in NSCLC Based on HuProt Protein Chip

The overall workflow of this study is illustrated in Figure 1. To correct for inconsistent background and systematic errors within the chip, we applied normalization to reduce batch effects (Figure 2A). PCA demonstrated the clustering results between NSCLC and NC, which were distinguishable overall (Figure 2B). We applied the HuProt protein chip, utilizing p < 0.05 and log₂FC < −0.5 as a threshold for differential analysis between NSCLC and NCs, 84 AAbs were identified (Figure 2C). The expression heatmap of each AAb is shown in Figure 2D. GO and KEGG enrichment analyses indicated that the associated genes were mainly involved in mRNA splicing, protein folding, amino acid metabolism (Figure S1). Using the NSCLC as a reference, we set the cutoff threshold to calculate the positivity rate of 84 AAbs, 12 positive AAbs were identified (Table S1). Based on criteria of positivity rate > 0.6 in NC, log₂FC < −1, and p < 0.05, four AAbs were selected: anti‐CAMKK2, anti‐GRPEL1, anti‐DYDC2, and anti‐HIST1H1B. The expression of these four candidate AAbs on the chip are shown in Figure S2.

FIGURE 1.

The overall workflow of this study. Figure created with support from Figdraw.

FIGURE 2.

Screening of candidate AAbs based on HuProt protein chip. (A) Removing batch effects by chip normalization. (B) Principal component analysis (PCA). (C) Volcano plot of the detected AAbs in HuProt protein chip. (D) The heatmap of 84 down‐regulated AAbs expression HuProt protein chip of LC patients and normal volunteers. (E) PPI network diagram of proteins corresponding to 84 AAbs (p < 0.001). (F) PPI visualization using cytoscape. (G) Top 10 Hubba proteins of PPI based on degree algorithm. (H) Top 10 Hub proteins of PPI based on MCC algorithm. (I) The Venn diagram of candidate AAbs between MCC and degree and positive AAbs. p < 0.05 was considered significant.

Subsequently, a PPI network was constructed based on the corresponding proteins of 84 AAbs using the STRING database and Cytoscape (Figure 2E,F). The top 10 core proteins were identified using the MCC (Figure 2G) and Degree (Figure 2H) algorithms. Three AAbs (anti‐HIST1H1A, anti‐HIST1H1B, and anti‐HIST1H1C) were identified using a Venn diagram (Figure 2I).

In summary, a total of six candidate AAbs (anti‐CAMKK2, anti‐GRPEL1, anti‐DYDC2, anti‐HIST1H1A, anti‐HIST1H1B, and anti‐HIST1H1C) were ultimately identified.

3.2. Validation and Diagnostic Assessment of Downregulated AAbs in NSCLC by ELISA Approach

We measured the expression of six candidate AAbs in the diagnosiscohort (NSCLC & NC) by indirect ELISA. No significant differences were observed among NSCLC, BPN, and NCs groups in terms of age, sex (p > 0.05). Compared to NC group, the levels of AAbs against GRPEL1, CAMKK2, DYDC2, HIST1H1B, and HIST1H1C were significantly lower in NSCLC patients (p < 0.05) (Figure 3A). Subsequently, we measured the expression of five AAbs (anti‐GRPEL1, anti‐CAMKK2, anti‐DYDC2, anti‐HIST1H1B, and anti‐HIST1H1C) between NSCLC and BPN. Compared with BPN groups, the levels of AAbs against CAMKK2, DYDC2, HIST1H1B, and HIST1H1C were significantly lower in NSCLC patients (p < 0.05) (Figure 3B). Subtype‐specific analysis showed that anti‐HIST1H1B AAb were significantly higher in LUSC than in LUAD, while no differences were observed for the other AAbs (Figure S3). To access the diagnostic value, ROC curve analysis was performed for each candidate AAb, with AUC values ranging from 0.53 to 0.70 (Figure 3C and Figure S4A).

FIGURE 3.

Validation of six candidate AAbs and construction of machine learning–based diagnostic models in the diagnosis cohort. (A) Differential expression of six AAbs in NSCLC and NC. (B) Differential expression of six AAbs in NSCLC and BPN. (C) ROC analysis of six AAbs in NSCLC and BPN. (D, E) ROC curve of 10 models in the train set (D) and test set (E). (F, G) ROC curves, confusion matrices, and DCA of the GBM model in the training set (F) and test set (G). (H, I) Diagnostic performance of the GBM model in differentiating early and advanced ESCC from BPN in the training set (H) and test set (I). (J) SHAP analysis revealed the contribution of each feature in the GBM model. GBM, gradient boosting machines; GLM, generalized linear model; KNN, k‐nearest neighbors; LDA, linear discriminant analysis; NB, naive Bayes; NNET, neural network; RF, random forest; SVM, support vector machine. Adaboost, adaptive boosting; ***p < 0.001; **p < 0.01; *p < 0.05. p < 0.05 was considered significant.

3.3. Diagnostic Performance of a Machine Learning Algorithms Using Downregulated Autoantibodies

To improve discrimination of NSCLC, we developed machine learning algorithms combining five downregulated AAbs with serum CEA. The machine learning algorithms exhibited modest performance in distinguishing NSCLC from NC, with the AUC ranging from 0.64 to 1.00 in the training set and from 0.60 to 0.72 in the validation set (Figure S4B,C). Given the complexity of differentiating between benign and malignant pulmonary nodules [19, 20], it is noteworthy that the models achieved better performance in distinguishing NSCLC from BPN. The AUCs of the ten algorithms ranged from 0.73 to 1.00 in the training set and from 0.71 to 0.81 in the validation set, while the corresponding accuracy values ranged from 0.68 to 1.00 and from 0.65 to 0.76 (Figure 3D,E). Other diagnostic metrics were presented in Table 2. DeLong's test indicated no significant differences in AUC values between the training and validation sets for the GLM, GBM, C5.0, LDA, NB, and Nnet models. GBM model demonstrated the best overall performance. In the training set, the AUC was 0.869 (95% CI: 0.833–0.905), with sensitivity of 75.8%, and specificity of 82.8% (Figure 3F). In the validation set, the AUC was 0.813 (95% CI: 0.745–0.880), with sensitivity of 78.8%, and specificity of 73.4% (Figure 3G). DCA confirmed the clinical utility of the GBM model. We further assessed the model's performance in distinguishing early and advanced NSCLC from BPN. In the training set, the AUCs were 0.855 (95% CI: 0.810–0.901) for early NSCLC and 0.887 (95% CI: 0.837–0.937) for advanced NSCLC, with sensitivities of 80.7% and 84.2% and specificities of 74.1% and 81.5%, respectively (Figure 3H). In the validation set, the AUCs were 0.809 (95% CI: 0.729–0.890) for early NSCLC and 0.818 (95% CI: 0.724–0.912) for advanced NSCLC, with sensitivities of 74.0% and 61.3% and specificities of 81.3% and 90.6%, respectively (Figure 3I). SHAP analysis revealed the contribution of each feature in the model, with anti‐DYDC2, anti‐GRPEL1, and anti‐HIST1H1B AAb identified as key contributors to the GBM model predictions (Figure 3J).

TABLE 2.

Evaluation of ten machine learning algorithms for diagnostic performance in the training and validation cohorts.

	SVM	RF	GLM	GBM	C5.0	LDA	NB	NNET	KNN	Adaboost
Train data
AUC	0.871	1.00	0.758	0.869	0.765	0.729	0.806	0.815	0.834	1.00
Sensitivity	0.647	1.00	0.586	0.758	0.809	0.749	0.665	0.735	0.856	1.00
Specificity	0.940	1.00	0.821	0.828	0.622	0.609	0.815	0.722	0.642	1.00
PPV	0.939	1.00	0.823	0.862	0.753	0.732	0.836	0.790	0.773	1.00
NPV	0.651	1.00	0.582	0.706	0.696	0.630	0.631	0.657	0.758	1.00
F 1	0.765	1.00	0.685	0.807	0.780	0.740	0.740	0.761	0.812	1.00
Kappa	0.551	1.00	0.384	0.571	0.439	0.360	0.460	0.450	0.510	1.00
Accuracy	0.768	1.00	0.683	0.787	0.732	0.691	0.727	0.729	0.768	1.00
Test data
AUC	0.773	0.772	0.798	0.813	0.689	0.769	0.775	0.785	0.713	0.768
Sensitivity	0.549	0.846	0.670	0.780	0.791	0.791	0.692	0.758	0.780	0.824
Specificity	0.859	0.453	0.797	0.734	0.516	0.547	0.719	0.672	0.484	0.562
PPV	0.847	0.687	0.824	0.807	0.699	0.712	0.778	0.767	0.683	0.728
NPV	0.573	0.674	0.630	0.702	0.635	0.648	0.622	0.661	0.608	0.692
F 1	0.666	0.756	0.739	0.794	0.742	0.749	0.732	0.762	0.728	0.773
Kappa	0.381	0.315	0.449	0.511	0.316	0.346	0.402	0.429	0.273	0.398
Accuracy	0.677	0.684	0.723	0.761	0.677	0.690	0.703	0.723	0.658	0.716
p‐value	0.02	< 0.001	0.353	0.149	0.118	0.384	0.485	0.479	< 0.001	< 0.001

Abbreviations: Adaboost, adaptive boosting; AUC, area under the ROC curve; F ₁, F ₁ score; GBM, gradient boosting machines; GLM, generalized linear model; KNN, k‐nearest neighbors; LDA, linear discriminant analysis; NB, naive Bayes; NNET, neural network; NPV, negative predictive value; PPV, positive predictive value; p‐value, AUCs were compared using the DeLong test; RF, random forest; SVM, support vector machine.

3.4. Survival Analysis of Downregulated AAbs in NSCLC

To evaluate the prognostic value of the above five downregulated AAbs, we followed up a cohort of 353 NSCLC patients and recorded OS. Notably, a more favorable prognosis and relatively higher OS among patients in this cohort was observed in the anti‐DYDC2_high group (p = 0.038), and a poor prognosis in the high‐expression groups of anti‐CAMKK2, anti‐GRPEL1 or anti‐HISTIH1C (p < 0.01) (Figure 4A–E and Table S2). After adjusting for sex, age, tumor stage and other relevant risk factors, high expression of anti‐DYDC2 (HR = 0.71, 95% CI: 0.51–0.98, p = 0.038) was still associated with improved prognosis, while high expression of anti‐CAMKK2 (HR = 1.47, 95% CI: 1.06–2.05, p = 0.022), anti‐GRPEL1 (HR = 1.49, 95% CI: 1.07–2.06, p = 0.017) or anti‐HIST1H1C (HR = 2.05, 95% CI: 1.39–3.04, p < 0.001) was associated with poor prognosis (Figure 4F–J). The survival curves of five downregulated AAbs in different clinical features of NSCLC are presented in Figure S5.

FIGURE 4.

Kaplan–Meier survival curves and multivariate Cox regression analyses. (A–E) Kaplan–Meier survival curves of five downregulated AAbs, using the optimal cutoff value. (F–J) Forest plot showed the results of multivariate cox regression analysis for five downregulated AAbs. CI, confidence interval; DM, distant metastasis; HR, hazard ratio; LM, lymph node metastasis. p < 0.05 was considered significant.

Fifteen NSCLC patients' plasma samples were collected before and 1–2 weeks after surgery, and the AAbs levels were measured. As shown in Figure S6, four AAbs significantly increased postoperatively (p < 0.05), while anti‐CAMKK2 AAb showed a trend toward increase (p = 0.069), suggesting that the downregulation of these AAbs was likely associated with the tumor presence.

3.5. Prognostic Risk Assessment and Model Construction Based‐On Downregulated AAbs

To identify key variables for the prognostic model, we employed LASSO Cox regression with 10‐fold cross‐validation to identify risk variables and determine their coefficients (Figure S7A,B). Age and all downregulated AAbs were included as candidate variables, because age had a significant impact on patient prognosis in our analysis. Anti‐HIST1H1B was also retained, due to its strong biological relevance within the PPI network. Five risk variables were included in the final risk model. According to the median risk score (risk score = 0.2117 × anti‐GRPEL1_exp − 0.2054 × anti‐DYDC2_exp − 0.3236 × anti‐HIST1H1B_exp + 0.3633 × anti‐HIST1H1C_exp + 0.04022 × age), patients were stratified into high‐ and low‐risk groups, and the high‐risk group had poorer prognosis (Figure 5A,B). Figure 4C illustrates the correlations among the key variables in the model. Compared to the individual risk variables, the risk score demonstrated superior predictive power for mortal events (DeLong test p < 0.05) (Figure 5D). The model achieved the maximal AUC at 7 years (AUC = 0.79), and its performance was further evaluated in early‐stage NSCLC patients (AUC = 0.80) (Figure 5E,F). Calibration curves at different time points showed a good fit of prediction to actual survival (Figure S7C). The distribution of risk score across clinical subgroups and its prognostic value within each subgroup are provided (Figure S8). Multivariate Cox regression analysis demonstrated that the risk score remained an independent prognostic factor (HR = 2.04, 95% CI: 1.38–3.02, p < 0.001) (Figure 5G). Finally, a nomogram was constructed using bidirectional stepwise regression to optimize the model (Figure 5H). The calibration curve demonstrated highly consistent outcomes between the predicted and actual outcomes, indicating the high reliability of the nomogram model (Figure 5I).

FIGURE 5.

Prognostic risk assessment and model construction. (A) Risk factor plot: Distribution of risk score stratified by median risk score and the survival status and time. (B) Kaplan–Meier survival curves for patients with high (red) and low (blue) risk scores. (C) Correlation analysis between five risk variables. (D) ROC curves show the risk score predictive ability for death events compared to individual variables. (E, F) Time‐dependent ROC curves of risk score at 1, 2, 5, and 7 years in the total cohort (E) and early cohort (F). (G) Multivariate Cox analysis for risk score and clinical relevant characteristics. (H) Construction of nomogram using risk score and clinical characteristics. (I) Calibration curves of the nomogram model over 1, 2, 5, and 7 years.

3.6. Bioinformatics Analysis of Five Genes to Downregulated AAbs

To investigate the roles of genes corresponding to downregulated AAbs in tumor, we analyzed transcriptomic and proteomic data. In LUAD and LUSC, CAMKK2 and DYDC2 were consistently downregulated, whereas GRPEL1 was upregulated. HIST1H1B showed increased mRNA but unchanged protein levels, while HIST1H1C displayed mRNA upregulation and protein downregulation (Figure 6A,B), consistent with TCGA paired data (Figure S9A,B). Immunohistochemistry data indicated that DYDC2 was undetectable in both normal lung and NSCLC tissues, whereas the remaining proteins were mediumly/highly expressed in tumors, with some also showing relatively high expression in normal lung samples (Figure S10). Survival analyses showed that high CAMKK2, GRPEL1, or HIST1H1C predicted poor OS in TCGA‐LUAD, whereas high DYDC2 indicated better prognosis. In TCGA‐LUSC, elevated CAMKK2 and HIST1H1B were associated with worse survival, while high HIST1H1C was protective. CPTAC data similarly identified DYDC2 and HIST1H1C proteins as protective and HIST1H1B protein as a risk factor (Figure S11).

FIGURE 6.

Bioinformatics analysis of five downregulated AAbs corresponding genes/proteins (A, B). Differential expression of the genes of candidate AAbs in LUAD and LUSC based on the TCGA (A) and CPTAC (B) database. (C) Clustering of 56,008 cells from GSE117570 and GSE148071 identified 7 cell types in NSCLC. (D) Marker gene expression profiles in different cell types. (E) The expressional profiles of the genes of candidate AAbs in different cell types. (F) Clustering of 36,760 cells from Epithelial cells identified 7 subclusters. (G) The hierarchical heatmap showing large‐scale CNVs in NSCLC epithelial cells. (H) The expressional profiles of the genes of candidate AAbs in different epithelial subclusters.

To investigate the expression and distribution of AAb‐associated genes at single‐cell resolution, we analyzed scRNA‐seq data from NSCLC samples (GSE148071 and GSE117570). After stringent quality control, a total of 59,438 tumor‐derived cells were clustered into 16 clusters (Figure S9C), which were subsequently annotated into 7 major cell types (Figure 6C,D and Figure S9D). Notably, GRPEL1 and HIST1H1B were broadly expressed across all cell types, CAMKK2 was lowly expressed in all clusters and DYDC2 exhibited specific expression within some epithelial cell subsets (Figure 6E). To further assess gene expression heterogeneity, 36,760 epithelial cells were extracted and classified into seven subclusters (Figure S9E). To infer CNVs of epithelial cells, we applied inferCNV. According to the inferCNV results, epithelial subclusters 0, 1, 2, 3, and 4 exhibited obvious CNVs (Figure 6G). The t‐SNE map showed epithelial cells clustered into seven preliminary subgroups (Figure 6F and Figure S9F). CytoTRACE indicated that the Cancer1 subcluster had the highest differentiation potential and expressed proliferation markers such as MKI67 and TOP2A, suggesting a key role in tumor progression (Figure S9G). Next, we explored the expression of genes in different epithelial cell subclusters. Notably, DYDC2 was specifically expressed in the ciliated epithelial subcluster; GRPEL1 was expressed in all subclusters except ciliated cells and HIST1H1C was expressed in all epithelial subclusters (Figure 6H). The heterogeneity of gene expression may, to some extent, help explain the impact of AAbs on the prognostic outcome of NSCLC patients.

3.7. A Web Tool for Diagnostic and Prognostic Assessment

We developed a Shiny web platform to enhance the usability and accessibility of our models. This platform allows users to input downregulated AAbs identified in this study, along with clinical characteristics, to perform diagnostic prediction and prognostic evaluation for NSCLC. The web can be accessed freely at https://liangyh.shinyapps.io/D_AAb_Pred_v2/.

4. Discussion

NSCLC remains the leading cause of lung cancer–related mortality [1, 3]. AAbs represent a class of crucial yet underexplored immune molecules; in‐depth investigation into their protective roles and pathological mechanisms in the disease process may unlock novel therapeutic strategies [7]. In this study, we developed a machine learning–based diagnostic model and prognostic risk score focusing on downregulated AAbs, which have been largely overlooked in prior research. Our findings highlight the potential of integrating these biomarkers to improve early identification of NSCLC and support patient risk stratification.

The HuProt protein microarray was used as a high‐throughput platform for the discovery of potential AAbs [21]. In this study, five downregulated AAbs against HIST1H1B, HIST1H1C, CAMKK2, DYDC2, and GRPEL1 were identified using a protein microarray and ELISA validation. Notably, these AAbs exhibited better performance in distinguishing NSCLC from BPN than from NC.

Previous research demonstrated the diagnostic potential of downregulated AAbs by developing a six AAbs model to distinguish NCs from early‐stage LC [22]. Zhang et al. constructed an early lung adenocarcinoma diagnostic model by combining autoantibody signatures with carcinoembryonic antigen (CEA), achieving an AUC of 0.86 [23]. However, differentiating BPN and NSCLC remains a major clinical challenge. Machine learning was widely used to construct AAb diagnostic models. Li et al. identified a panel of esophageal squamous cell carcinoma–specific autoantibodies (anti‐CAST, HDAC1, HSF1, PTMS, ZPR1, and FAM131A) and developed a machine learning–based diagnostic model, ESCCPred, with an AUC of 0.83 [24]. In contrast to the previous research, we developed machine learning models combining five downregulated AAbs with CEA to distinguish NSCLC from BPN and NC. Interestingly, this panel demonstrated particularly good performance in differentiating NSCLC from BPN. The GBM showed the best overall performance, with AUCs of 0.869 (95% CI: 0.833–0.905) in the training set and 0.813 (95% CI: 0.745–0.880) in the validation set. The GBM model also showed good early diagnostic performance, achieving an AUC of 0.809 (95% CI: 0.729–0.890) in the validation set, with a sensitivity of 74.0% and specificity of 81.3%. Although machine learning models often behave as black boxes, we used SHAP analysis and deployed a web to enhance the interpretability and accessibility of the GBM model. In summary, our findings suggest that a diagnostic model based on downregulated AAbs holds great potential for effectively discriminating NSCLC patients.

Growing evidence has demonstrated that AAbs are closely associated with patient survival outcomes and have potential clinical applications [25, 26, 27]. We found that these downregulated AAbs may serve as risk indicators for OS in NSCLC patients. Importantly, our results not only demonstrated the protective roles of specific AAbs but also highlight their possible functional heterogeneity across different clinical contexts. A previous study reported that a model consisting of 13 AAb biomarkers could be used for the prediction of NSCLC prognosis [27], in line with our study. Age has been shown to be a risk factor for NSCLC prognosis [28]. In this study, we developed a prognostic risk score, which showed the best predictive performance at 7‐year follow‐up (AUC = 0.79), and with superior predictive ability in early‐stage patients. Multivariate Cox regression confirmed that the risk score is an independent prognostic factor, highlighting its potential for long‐term survival stratification in NSCLC. This model may provide a novel strategy for survival risk stratification in clinical settings.

Finally, we explored the expression patterns and prognostic relevance of the target proteins and corresponding mRNAs of downregulated AAbs using public databases, further supporting our findings. In this study, these AAbs were generally reduced in NSCLC patients, which may reflect differences in antigen abundance, subcellular localization, and humoral immune accessibility. Linker histone H1, a critical component of chromatin, comprises five major subtypes (H1.1–H1.5) [29]. Among them, H1.2 (HIST1H1C) and H1.5 (HIST1H1B) have been linked to multiple cancers [30, 31, 32, 33]. H1.2 has been reported to show high expression in NSCLC and associate with poor prognosis [34]. This is consistent with our findings. H1.5 (HIST1H1B) is the only linker histone involved in the regulation of mRNA splicing [31, 35]. H1.2 and H1.5 AAbs levels were decreased in our cohort, suggesting that the nuclear localization and strong chromatin binding may restrict antigen release and limit exposure to the humoral immune system. Interestingly, our results demonstrated that anti‐HIST1H1B AAbs exhibited opposite prognostic implications in early and advanced patients, which may reflect a different role in NSCLC progression. Its role in NSCLC remains to be explored.

Similarly, GRPEL1 and CAMKK2 are intracellular functional proteins localized to the mitochondrial matrix and cytoplasm. They act as a mitochondrial nucleotide exchange factor and a serine/threonine kinase, respectively [36, 37, 38] and have been widely reported in cancers [39, 40, 41]. Anti‐GRPEL1 and anti‐CAMKK2 AAbs were decreased in NSCLC patients. The intracellular sequestration of these antigens may limit antigen presentation and subsequent antibody production. In addition, tumor‐associated immune modulation, including immune tolerance and suppression of humoral immune responses, may further contribute to diminished AAb levels. Higher levels of these AAbs predicted worse outcomes, may be related to tumor immune escape mechanisms. A previous study showed that depletion of the LIV‐1–GRPEL1 axis contributes to tumor cell evasion from mitotic damage‐induced death [39], partially may explaining the tumor‐promoting effects of anti‐GRPEL1 AAbs in NSCLC.

In contrast to the intracellular antigens described above, we found that DYDC2 was highly expressed in ciliated epithelial cells. Ciliated epithelial cells play a pivotal role in maintaining epithelial homeostasis [42]. Consistently, HPA showed no detectable DYDC2 expression in lung samples, while specific expression was observed in bronchial respiratory epithelial and ciliated cells (Figure S12). Lung cancer development may disrupt bronchial epithelial integrity and impair ciliated cell structure and function, potentially affecting DYDC2 expression. Cochrane et al. reported that DYDC2 is a marker of ciliated cells and positively correlated with OS in endometrial cancer [43]. Therefore, the decrease of anti‐DYDC2 AAb levels in NSCLC may be attributed to decreased antigen abundance resulting from the loss or dysfunction of ciliated epithelial cells during tumorigenesis. Despite this overall decrease, elevated levels of DYDC2 and its AAbs were strongly associated with better prognosis. These results support that anti‐DYDC2 AAbs may be a protective AAbs. Interestingly, the prognostic impact of anti‐DYDC2 AAbs differed between EGFR‐mutant and wild‐type patients, indicating that their function may depend on driver mutation context. As current research on DYDC2 in cancer is limited, our findings offer new insights.

However, due to limitations in our study design, all serum samples in this study were collected from a single institution, and we did not establish multicenter cohorts. This may limit the generalizability of our findings across different populations and clinical settings. In addition, antigen expression in clinical tumor tissues was not directly validated using patient‐derived specimens, and supportive evidence was obtained from publicly available resources. These AAbs were detectable not only in NSCLC patients but also in healthy individuals and in individuals with BPNs, likely because their target proteins are essential for normal cellular functions and the AAbs may reflect baseline immune surveillance in the normal state. Future studies should involve independent cohorts from multiple institutions and validation of antigen expression in clinical tissue samples. Despite these limitations, this study provides a foundation for understanding the biological and clinical implications of downregulated AAbs in NSCLC and offers a new perspective for developing AAb‐based diagnostic and therapeutic strategies.

Overall, our findings underscore the diagnostic and prognostic significance of downregulated AAbs in NSCLC and provide novel insights for future research.

Author Contributions

Yihao Liang: conceptualization, methodology, software, data curation, formal analysis, visualization, writing – review and editing, writing – original draft, investigation. Hanke Ma: writing – review and editing, validation, data curation, investigation. Wenke Sun: writing – review and editing, investigation, methodology, validation. Ying Chen: writing – review and editing. Fengqi Chen: writing – review and editing. Yutong Li: writing – review and editing. Songyun Ouyang: resources, data curation, supervision. Liping Dai: conceptualization, funding acquisition, project administration, resources, writing – review and editing, supervision, data curation.

Funding

This study was supported by the National Natural Science Foundation of China (Grant 8167291) and Project of Basic Research Fund of Henan Institute of Medical and Pharmacological Sciences (Grant 2025BP0103).

Ethics Statement

This study was approved by the Ethics Committee of the First Affiliated Hospital of Zhengzhou University (Approval No. 2021‐KY‐1057‐002) and conducted in accordance with the Declaration of Helsinki and relevant institutional guidelines.

Consent

Written informed consent was obtained from all participants prior to their inclusion in the study.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

TABLE S1: Screening results of 12 AAbs by protein microarray analysis.

TABLE S2: Univariate regression analysis of AAbs and clinical variables.

FIGURE S1: Functional enrichment analysis of proteins associated with genes to 84 downregulated AAbs.

FIGURE S2: Four candidate AAbs expression comparing LC patients (colored in red) and normal control (colored in green).

FIGURE S3: Expression comparison of AAbs in LUAD and LUSC.

FIGURE S4: Construction of machine learning–based diagnostic models to evaluate their diagnostic value in distinguishing NC from NSCLC.

FIGURE S5: The survival curves of five down‐regulated AAbs in different clinical features of NSCLC.

FIGURE S6: Changes in circulating AAb levels before and after surgery in NSCLC patients.

FIGURE S7: Development and evaluation of the prognostic model.

FIGURE S8: The distribution of risk score across clinical subgroups and its prognostic value within each subgroup.

FIGURE S9: Transcriptomics analysis of five downregulated AAbs corresponding genes.

FIGURE S10: IHC staining of five autoantibody‐associated proteins in normal lung and NSCLC tissue (HPA database).

FIGURE S11: The prognostic value of the corresponding genes/proteins of AAbs using TCGA and CPTAC database.

FIGURE S12: IHC staining of DYDC2 in normal bronchus tissue (HPA database).

Data Availability Statement

The datasets used for the bioinformatics analysis of five genes/proteins corresponding to downregulated AAbs are available in the database. Data generated and analyzed during this study are available from the corresponding author, Liping Dai, upon reasonable request.