Single-cell and machine learning-based neural regulation signature for prognosis prediction and immunotherapy response in lung adenocarcinoma

Highlights

• •Developed a novel neural regulation-based signature (NR.Sig) using advanced machine learning algorithms for lung adenocarcinoma (LUAD).
• •NR.Sig accurately predicts patient prognosis and immunotherapy efficacy in LUAD, outperforming existing models.
• •Single-cell RNA sequencing identified functionally distinct neural regulation-associated epithelial cell subtypes, with CRABP2-positive epithelial cells linked to unfavorable outcomes and immune evasion.
• •NR.Sig effectively stratifies the tumor immune microenvironment, showing an "immune-hot" phenotype in low-risk patients, suggesting higher responsiveness to immunotherapy.
• •This robust and precise NR.Sig holds significant translational potential for clinical risk stratification and personalized therapeutic guidance in LUAD.

Abstract

Objective

Lung adenocarcinoma (LUAD) molecular heterogeneity limits traditional prognostic models. Given the emerging role of neural regulation (NR) in tumor progression, we aimed to delineate NR-associated cellular phenotypes via single-cell RNA sequencing (scRNA-seq) and develop a robust machine-learning-derived signature (NR.Sig) to precisely assess prognosis and guide personalized immunotherapy.

Methods

We integrated three LUAD scRNA-seq cohorts and ten transcriptomic cohorts with immunotherapy records. Single-cell analyses (clustering, cell-cell communication, pseudotime trajectory) identified NR-enriched epithelial subpopulations. Using their prognostic marker genes, we evaluated 101 combinations from 10 machine learning algorithms via leave-one-out cross-validation. The combination yielding the highest C-index formed the NR.Sig model. Its prognostic accuracy, stability, and clinical utility in characterizing the tumor immune microenvironment (TME) and forecasting immunotherapy efficacy were comprehensively validated across multiple independent cohorts.

Results

"CRABP2-positive epithelial cells" were identified as a stem-like, NR-enriched malignant subpopulation correlating strongly with immune exhaustion. The random survival forest (RSF)-based NR.Sig achieved optimal modeling performance. Validation confirmed that NR.Sig high-risk patients had significantly shorter overall and progression-free survival. NR.Sig outperformed conventional clinical indicators and existing prognostic models, with FAM83A identified as the core hub gene. Crucially, high-risk scores inversely correlated with immune infiltration. Conversely, the low-risk group exhibited an "immune-hot" phenotype with enhanced cancer-immunity cycle activity and elevated checkpoint expression, translating to significantly higher immunotherapy response rates in independent clinical cohorts.

Conclusion

By integrating scRNA-seq with an optimized machine learning framework, we developed and validated NR.Sig. This robust signature holds significant clinical translational value, serving as a precise molecular tool for LUAD risk stratification, prognostic assessment, and the guidance of personalized immunotherapy strategies.

Graphical abstract

The overall experimental workflow.

Introduction

Lung cancer represents the most common malignancy of the respiratory system and remains the leading cause of cancer-related mortality globally in both men and women [1]. Based on the World Health Organization (WHO) classification, lung cancers are broadly divided into non‑small cell lung cancer (NSCLC) and small cell lung cancer (SCLC), with NSCLC accounting for approximately 85% of cases and SCLC comprising about 15% [2]. Among NSCLC subtypes, adenocarcinoma and squamous cell carcinoma predominate, of which lung adenocarcinoma (LUAD) is the most frequently diagnosed histological type, representing around 40% of all lung cancer cases [3]. Although surgical resection remains the standard of care for early‑stage LUAD, its invasive nature often hinders complete tumor removal. The advent of immunotherapy has transformed LUAD management, particularly through immune checkpoint inhibitors, which have significantly improved patient outcomes and emerged as a neoadjuvant option for early‑stage resectable disease. Nevertheless, despite continuous therapeutic progress, the five‑year survival rate for LUAD patients remains unsatisfactory [4]. These observations highlight the urgent need to elucidate the molecular mechanisms driving LUAD and to develop reliable molecular stratification models for accurate prognosis prediction and personalized treatment strategies.

The interaction between neural signaling and cancer cells is increasingly recognized as a critical regulator of tumor proliferation, invasion, angiogenesis, immune exhaustion, and immune evasion [5]. Recent studies have highlighted the significance of neural regulation (NR) in tumor progression and poor clinical outcomes, noting that tumor cells can induce neuronal reprogramming to facilitate the recruitment of new nerve fibers [6]. Such neural–tumor crosstalk may reactivate neural‑dependent biological pathways that further drive tumor advancement [7]. Elucidating the dialogue between nerves and tumors is therefore essential to deepening our understanding of the mechanisms underlying tumor initiation, malignant progression, and metastatic dissemination. Numerous studies in the past decade have established critical contributions of NR to the development and progression of central nervous system (CNS) tumors. Specifically, neurons facilitate glioma growth predominantly via secretion of paracrine mitogens, which activate key proliferative signaling cascades such as the PI3K-mTOR, WNT, and β-catenin pathways [8]. Furthermore, a range of neural signals exhibit direct or indirect associations with peripheral solid tumors, influencing both tumor pathogenesis and progression. For instance, breast cancer cells can stimulate neuronal calcium signaling activity, leading to the release of the neuropeptide substance P, which in turn enhances tumor growth and invasive potential [9]. Although the critical involvement of neural signaling in cancer initiation and progression has been established across multiple tumor types, the precise molecular and cellular mechanisms underlying this modulation in LUAD remain to be fully elucidated.

Current prognostic frameworks, including the TNM staging system, depend largely on clinicopathological characteristics. Although these approaches have been useful for stratifying patients into general risk groups, they frequently overlook the substantial molecular heterogeneity present in LUAD. The field of medical research is now experiencing a profound shift, propelled by sophisticated bioinformatics techniques. Advances in RNA sequencing (RNA‑seq) analysis, investigations of genetic variation, and high‑resolution single‑cell RNA sequencing (scRNA‑seq) are reshaping our understanding of disease pathogenesis [10]. When leveraged for prognosis prediction and immunotherapy response assessment in LUAD, these methods offer fresh perspectives on potential therapeutic avenues. To date, a number of molecular markers have been identified as significant predictors of immunotherapy outcome and prognostic indicators in LUAD [11]. Predictive markers such as programmed death‑ligand 1 (PD‑L1) expression, tumor mutational burden (TMB), and hematologic biomarkers serve as important determinants of therapeutic response and are strongly associated with clinical outcomes in LUAD. Beyond genetic factors, environmental and lifestyle influences also contribute to LUAD pathogenesis; for example, tobacco‑derived carcinogens can promote bronchial epithelial cell differentiation and thereby foster tumor initiation [12]. Nevertheless, conventional biomarkers often exhibit limited predictive power, highlighting the need for more advanced prognostic tools to improve risk stratification and immunotherapy guidance in LUAD.

To overcome these limitations, we have developed a fundamentally novel approach that distinguishes itself through three key innovations. First, we employ high‑resolution scRNA‑seq to delineate NR‑associated cellular phenotypes with unprecedented cellular resolution, moving beyond bulk tissue analysis that obscures critical intra-tumor heterogeneity. Second, we integrate this single‑cell perspective with comprehensive multi‑omics data within a carefully constructed multi‑algorithm machine learning framework—leveraging complementary algorithmic strengths to achieve robust and generalizable predictions that surpass the performance of conventional single‑method approaches. Our NR‑associated signature provides enhanced precision for immunotherapy response prediction and patient risk stratification, offering a paradigm shift toward truly personalized, molecularly informed therapeutic guidance in LUAD.

Materials and methods

Transcriptomic cohorts and NR related genes

A total of ten transcriptomic datasets of LUAD were included in this study: GSE13213, GSE26939, GSE29016, GSE30219, GSE31210, and GSE42127 in GEO, TCGA-LUAD in TCGA, and three independent cohorts containing immunotherapy clinical records (OAK [13], POPLAR [14], and Jung [15]). Transcriptomic reads were normalized using the log₂(x+1) transformation, and batch effects were mitigated using the ComBat method implemented in the “sva” R package [16]. Additionally, three scRNA-seq cohorts of LUAD (GSE189357, GSE207422 and GSE131907) were obtained from GEO. NR associated genes were extracted from the Gene Ontology (GO) database (https://www.geneontology.org/) and categorized into 14 distinct GO terms (Supplementary Table 1).

Single-cell analysis of NR associated genes

The scRNA-seq data analysis was conducted using the Seurat R package [17], starting with the conversion of raw mRNA matrices into Seurat objects. Quality control was performed by filtering out cells exhibiting over 40,000 UMIs, expressing fewer than 500 or more than 5,000 genes, and having a mitochondrial gene fraction exceeding 20%. Following the removal of potential doublets via the DoubletFinder tool [18], batch effects across different cohorts were mitigated utilizing Harmony [19]. We utilized FindVariableFeatures algorithm to find the top variant 2000 genes [20], at which PCA was performed. We used FindClusters algorithm to identify clusters at resolution 0.3. We performed RunUMAP algorithm to reduce dimensions and manually annotated cell subpopulations by canonical markers [21]. Finding marker genes of epithelial subpopulations was employed by criteria of log2FC >0.25 and p value <0.05 by FindAllMarkers algorithm. We performed scRNA-seq scoring analysis by various algorithms (AUCell in “AUCell” R package, Ucell in “Ucell” R package, ssGSEA and GSVA in “GSVA” R package, singscore in “singscore” R package, AddModuleScore and PercentageFeatureSet in “Seurat” R package) [17,[22], [23], [24]]. Function enriching analyses within GO and KEGG was conducted with “ClusterGVis” R package (https://github.com/junjunlab/ClusterGVis). A continuous differentiation score, reflecting cellular differentiation status from least differentiated (1.0) to most differentiated (0.0), was computed from mRNA expression matrix using “CytoTRACE” R package [25]. Cell lineage trajectories and pseudotemporal ordering were inferred with “slingshot”, “Monocle”, and “Monocle3” R packages [[26], [27], [28]]. To identify cellular subpopulations associated with prognostic differences, we applied “Scissor” R algorithm, which integrates single‑cell sequencing data with bulk‑sample phenotype information [29]. Copy number variations (CNVs) in epithelial cells were delineated using “InferCNV” R package, with immune cells serving as the reference [30]. Intercellular communication networks were explored with “CellChat” R package [31].

Establishment of NR.Sig with machine learning frameworks

The TCGA‑LUAD cohort was designated as the training set for prognostic model construction. The remaining six cohorts served as independent validation sets and were subsequently pooled to generate a consolidated test cohort. Utilizing prognostic marker genes derived from NR-associated epithelial subtypes, we developed an NR‑related signature (NR.Sig) to predict survival outcomes in patients with LUAD. Based on previous studies, the prognostic modeling integrated ten ML algorithms, including random survival forest (RSF), elastic net (Enet), Lasso, Ridge, stepwise Cox regression, CoxBoost, partial least squares Cox regression (plsRcox), supervised principal components (SuperPC), generalized boosted regression modeling (GBM), and survival support vector machine (survival-SVM) [32,33]. To optimize model performance, we constructed the NR.Sig in the training cohort through grid parameter tuning and a leave-one-out cross-validation (LOOCV) framework. Specifically, a total of 101 ML prognostic models were trained. Models containing fewer than five genes were excluded from further consideration. For each ML combination, the average concordance index (C-index) was calculated across the validation datasets. The prognostic ML framework achieving the highest mean C-index was selected as the optimal signature. Risk scores for each prognostic ML model were derived via a linear combination of the expression values of its constituent genes.

Verification of NR.Sig

Comprehensive validation was performed to assess the accuracy, stability, and discriminative power of the NR.Sig. LUAD patients were stratified into high‑ and low‑risk groups based on the median NR.Sig risk score derived from the training cohort. Survival differences between the two groups were evaluated using Kaplan–Meier (KM) analysis and the log‑rank test, implemented via the “survival” and “survminer” R packages. Receiver operating characteristic (ROC) curves, calibration plots, and decision curve analysis (DCA) were used to quantify the predictive accuracy, discriminative performance, and clinical utility of the signature. Time‑dependent area under the curve (AUC) values were further calculated to compare the predictive capability of NR.Sig against conventional clinical parameters. Finally, univariate and multivariate Cox regression analyses were conducted to determine the independent prognostic significance of the NR.Sig score.

Function enrichment analysis

Differentially expressed genes (DEGs) were discovered between two risk groups divided by NR.Sig. DEGs were defined by "limma" R package, with tcriteria of False-discovery rate (FDR) <0.05 and absolute log2fold change (FC) >1. The functional enrichment of DEGs was employed in Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) terms within the "clusterprofiler" R package [34].

Delineating tumor microenvironment

Immune infiltration levels across risk groups were quantified using multiple immune deconvolution algorithms implemented in “IOBR” R package [[35], [36], [37], [38], [39], [40], [41], [42], [43]], along with single‑sample gene set enrichment analysis (ssGSEA) based on established marker gene sets [44]. Subsequently, Spearman correlation analysis was applied to examine associations among risk scores, model gene expression, and immune cell abundances. Additionally, ssGSEA was employed to evaluate immune functional activity using predefined immune‑function gene sets [45] and to characterize the seven steps of the cancer‑immunity cycle with corresponding gene signatures from the Tracking Tumor Immunophenotype (TIP) database (http://biocc.hrbmu.edu.cn/TIP/) [46]. Expression levels of immune checkpoint genes were compared between the two risk groups.

Statistical Analysis

All statistical analyses were conducted in R software (version 4.2.1; R Foundation for Statistical Computing, Vienna, Austria). Continuous variables are expressed as mean ± standard deviation when normally distributed, or as median with interquartile range (IQR) otherwise. Normality was evaluated using the Shapiro–Wilk test. Categorical variables are summarized as frequencies (percentages). Between‑group comparisons for continuous data were performed using the independent two‑sample t‑test (for normally distributed data) or the Mann–Whitney U test (for non‑normal distributions). Comparisons across three or more groups were conducted by one‑way analysis of variance (ANOVA) followed by Tukey’s post‑hoc test. Associations between categorical variables were examined using the chi‑square test or Fisher’s exact test, as appropriate. A two‑tailed p‑value of less than 0.05 was considered statistically significant.

Results

Single-cell transcriptomic overview in LUAD

After integrating sequencing data from three scRNA‑seq cohorts and applying Harmony‑based batch correction, batch effects were effectively reduced (Supplementary Fig. S1A). Quality control showed well sequencing quality of these three scRNA‑seq cohorts (Supplementary Fig. S1B). Using established canonical markers, we manually annotated twelve major cell populations: epithelial cells, T cells, B cells, natural killer (NK) cells, endothelial cells, macrophages, monocytes, plasma cells, fibroblasts, mast cells, neutrophils, and proliferating cells (Fig. 1A), which originated from three scRNA-seq cohorts (Fig. 1B). The top five up-regulated and down-regulated marker genes for each major cell type are displayed (Fig. 1C). Functional annotation via GO and KEGG enrichment analyses further supported the cellular classifications (Fig. 1D). Representative marker genes for each cell type are illustrated in a dot plot (Fig. 1E). Cell‑type proportions varied substantially across patients, reflecting the pronounced tumor heterogeneity in LUAD (Fig. 1F). Additionally, expression patterns of partial NR‑related genes were visualized in a heatmap, revealing their molecular landscape within the LUAD microenvironment (Supplementary Figure S1C).

FIG. 1.

Overview of the single-cell transcriptomic landscape within the LUAD microenvironment. (A) UMAP dimensional reduction plot depicting the spatial distribution and clustering of the primary cellular lineages within the LUAD tumor microenvironment (TME) following quality control and batch correction. (B) UMAP visualization colored by dataset origin, demonstrating the robust integration and distribution of cells pooled from three independent scRNA-seq cohorts. (C) Dot plot summarizing the expression profiles of the top five differentially expressed marker genes specific to each major cell cluster. (D) Comprehensive heatmap detailing the expression signatures of cluster-specific marker genes. The adjacent panels display Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment results, revealing the core biological pathways and functional roles distinct to each cell subpopulation. (E) Expression landscape of the established canonical reference genes relied upon to manually identify and annotate the main cell clusters. (F) Stacked bar chart illustrating pronounced inter-patient cellular heterogeneity by quantifying the relative proportions of each major cell subset across individual LUAD patients.

Epithelial heterogeneity and exploration of NR related subtype

To delineate the molecular landscape of neural regulation (NR) within the integrated single‑cell atlas of LUAD, we performed single‑cell enrichment analysis using an NR‑specific gene set to quantify NR enrichment scores across major cell populations (Fig. 2A). The analysis revealed that epithelial cells and endothelial cells exhibited the highest NR enrichment scores, indicating a prominent involvement of NR‑associated biological processes in these two cell types. Given the central role of epithelial cells in LUAD oncogenesis, we isolated the epithelial compartment for deeper characterization and performed refined sub‑clustering. Batch effects among three scRNA‑seq cohorts were mitigated in epithelial cells using the Harmony algorithm (Supplementary Figure S1D). At a clustering resolution of 0.3, we identified 12 distinct epithelial subclusters in LUAD (Fig. 2B). Single‑cell enrichment analysis revealed that cluster 0 had the highest NR enrichment scores (Fig. 2C-D). We next applied InferCNV analysis to delineate putative malignant epithelial populations, which highlighted that clusters 0, 2, and 10 harbored the most extensive copy‑number alterations among epithelial cells (Fig. 2E). Integrating these mutational patterns with established marker genes from prior literature [21] (Supplementary Fig. S2A-D), we manually annotated the epithelial cells into eleven subgroups: AT2‑like epithelial cells, CRABP2‑positive epithelial cells, AT1 cells, lymphoid‑like epithelial cells, Clara‑like epithelial cells, myeloid‑like epithelial cells, VIM‑positive epithelial cells, proliferating epithelial cells, FOS‑positive epithelial cells, mitochondria epithelial cells, and ribosomal epithelial cells (Fig. 2F). Among these, AT2‑like epithelial cells and CRABP2‑positive epithelial cells were identified as likely malignant epithelial populations. Following the identification of epithelial subgroups, we aimed to pinpoint NR‑associated cellular subpopulations. To this end, Cox regression analysis was performed to identify NR‑related genes with risk‑promoting or protective prognostic effects in LUAD (Fig. 2G). Subsequently, we applied AddModuleScore, PercentageFeatureSet, AUCell, and ssGSEA‑based single‑cell scoring methods to epithelial cells. These analyses consistently identified the CRABP2‑positive epithelial subpopulation as the most prominently NR‑enriched subset within the epithelial compartment (Fig. 2H). Furthermore, a heatmap visualization of NR‑related gene expression confirmed their pronounced expression in CRABP2‑positive epithelial cells (Supplementary Figure S2E). We next employed the CytoTRACE algorithm to assess differentiation heterogeneity among LUAD epithelial cells. The results indicated that CRABP2‑positive malignant cells exhibited the highest CytoTRACE scores, suggesting a stem‑like phenotype, whereas AT2‑like epithelial cells appeared to occupy a more differentiated state (Fig. 2I, Supplementary Figure S2F). Pseudotemporal trajectory reconstruction was performed using Monocle3 (Fig. 2J) and slingshot (Fig. 2K) analyses, with CRABP2‑positive malignant cells set as the trajectory origin. These analyses delineated distinct differentiation lineages and computed pseudotime scores across epithelial subpopulations, highlighting the complexity and heterogeneity of epithelial differentiation in LUAD.

FIG. 2.

Epithelial cell heterogeneity and identification of neural regulation (NR)‑associated epithelial cells in LUAD. (A) Single‑cell NR enrichment scores across major cell types, computed using the AddModuleScore algorithm based on an NR‑specific gene set in three scRNA‑seq cohorts. (B) UMAP visualization of epithelial cells following sub‑clustering via the Seurat pipeline. (C‑D) NR enrichment scores within epithelial cells derived from the AddModuleScore (C) and PercentageFeatureSet (D) algorithms applied to the NR gene set. (E) Copy‑number variation (CNV) landscapes of epithelial cells and aggregated CNV levels per epithelial cluster, inferred using InferCNV. (F) Annotation of epithelial subgroups displayed on UMAP, with color‑coding reflecting subtype identity. (G) Forest plot of univariate Cox regression analysis evaluating the prognostic association of NR‑related genes in the TCGA‑LUAD cohort. (H) Single‑cell scoring of epithelial subtypes based on the NR gene set, computed by AddModuleScore, PercentageFeatureSet, AUCell, and ssGSEA algorithms. (I) Differentiation potential of epithelial subpopulations assessed by CytoTRACE, with higher scores indicating less‑differentiated (stem‑like) states. (J‑K) Pseudotemporal trajectories of epithelial differentiation reconstructed using Monocle3 (J) and Slingshot (K) analyses.

Characterizing the molecular profiles of epithelial subgroups

To systematically delineate the heterogeneity within the epithelial compartment, we performed a series of single‑cell RNA‑seq analyses aimed at elucidating the biological functions of NR‑associated subpopulations. The top five marker genes for each epithelial subtype are displayed (Fig. 3A). Functional enrichment analysis based on GO and KEGG terms was employed to annotate the biological roles of each epithelial subtype, providing mechanistic insight into their respective functions (Fig. 3B). The variation in epithelial subtype proportions across patients further illustrated the substantial heterogeneity present in LUAD (Fig. 3C). Additionally, Scissor analysis was applied to assess the prognostic relevance of epithelial subpopulations, revealing that “Mitochondria epithelial cells” and “CRABP2‑positive epithelial cells” were identified as Scissor‑positive populations associated with unfavorable outcomes (Fig. 3D). Furthermore, we explored the tumor‑promoting role of CRABP2‑positive epithelial cells in the TCGA‑LUAD cohort. Although survival analysis based solely on the median expression of CD8A failed to stratify patient risk, we observed that a combined stratification using both median CD8A expression and median CRABP2‑positive epithelial cell score could partition TCGA‑LUAD patients into four distinct subgroups. Among these, patients exhibiting high CD8A expression together with a high CRABP2‑positive epithelial cell score demonstrated the poorest prognosis, thereby reinforcing the malignant phenotype associated with this epithelial subpopulation (Fig. 3E). Survival analysis of TCGA‑LUAD patients stratified by the median CRABP2‑positive epithelial cell score further indicated that patients with higher scores experienced significantly poorer outcomes, supporting the malignant phenotype of this subpopulation (Fig. 3F). Spearman correlation analysis between immune dysfunction scores, exhaustion scores and CRABP2‑positive epithelial cell scores revealed a strong positive association, suggesting a potential link between CRABP2‑positive epithelial cell abundance and immune dysfunction, as well as immune exhaustion, which may contribute to tumor immune evasion in LUAD (Fig. 3G). Cell‑cell communication analysis across major cell populations illustrated extensive interaction networks, with nearly all clusters showing pronounced communication signals directed toward CRABP2‑positive epithelial cells (Fig. 3H). Analysis of signaling contributions revealed that distinct cell subtypes differentially influenced total, incoming, and outgoing signals, with CRABP2‑positive epithelial cells and fibroblasts displaying particularly prominent roles in interaction number and strength (Fig. 3I). Interaction landscapes were mapped to visualize over‑expressed ligand–receptor pairs and communication profiles between CRABP2‑positive epithelial cells and other cellular clusters (Fig. 3J).

FIG. 3.

Molecular profiling of epithelial subpopulations. (A) Circle plot displaying the top five marker genes for each major cell type. (B) Heatmap visualization of marker‑gene expression across cell types, accompanied by Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis highlighting the most enriched functional terms per cell type. (C) Stacked bar plots illustrating the proportion of each epithelial subgroup across individual LUAD patients, reflecting inter‑tumor heterogeneity. (D) Dimplot depicting epithelial cells with unfavorable (Scissor‑positive) or favorable (Scissor‑negative) prognostic associations identified through Scissor analysis. (E) Kaplan–Meier curves compare survival of LUAD patients stratified by high vs. low CRABP2‑positive epithelial cell scores and high vs. low CD8A mRNA expressions. (F) Kaplan–Meier curves compare survival of LUAD patients stratified by high vs. low CRABP2‑positive epithelial cell scores. (G) Correlation analysis between ssGSEA‑derived CRABP2‑positive epithelial cell enrichment scores and immune dysfunction scores, as well as immune exhaustion scores. (H) Circle plots summarizing interaction strength and number of ligand–receptor pairs between epithelial subpopulations, immune cells, and stromal cells. (I) Analysis of signaling contributions revealed that distinct cell subtypes differentially influenced total, incoming, and outgoing signals in interaction number and strength. (J) Interaction landscapes were mapped to visualize over‑expressed ligand–receptor pairs and communication profiles.

Development and validation of NR.Sig

Prognostically relevant marker genes derived from the CRABP2‑positive epithelial subpopulation were employed for feature selection and model training. Through the leave‑one‑out cross‑validation (LOOCV) framework, we evaluated 101 ML prognostic combinations to identify the optimal model for constructing the neural regulation‑related signature (NR.Sig) The most robust prognostic combination, which achieved the highest mean C‑index (0.7) across all seven cohorts, was obtained through random survival forest (RSF)‑based feature selection and model construction (Fig. 4A). Survival analysis revealed that patients in the low‑risk group exhibited significantly better overall survival (OS) than those in the high‑risk group, both in the training set and the combined test dataset (Fig. 4B-C). ROC curves of the train and the combined test cohorts indicated high specificity of NR.Sig (Fig. 4D-E). Calibration curves revealed that NR.Sig provided accurate and stable probability estimates in both cohorts, confirming good model calibration (Fig. 4F-G). Comparison of 3‑year OS AUC values showed that NR.Sig, as well as a Cox regression model incorporating NR.Sig and clinical variables, outperformed conventional clinical indicators in prognostic discrimination (Fig. 4H). Time‑dependent AUC analysis further verified that NR.Sig and the combined Cox model possessed superior discriminative ability compared to several clinical parameters (Fig. 4I). Multivariate Cox regression confirmed that the NR.Sig risk score served as an independent prognostic predictor in the TCGA‑LUAD cohort (P < 0.001) (Fig. 4J). DCA indicated that using NR.Sig for prognosis prediction would yield clinical net benefit across a range of threshold probabilities, supporting its utility in guiding decisions such as treatment intensification or enhanced monitoring (Fig. 4K). Collectively, these evaluation metrics demonstrate that NR.Sig exhibits precision, robustness, and clinical relevance in prognostic stratification.

FIG. 4.

Construction and validation of NR.Sig for prognostic prediction in LUAD. (A) Performance evaluation of 101 prognostic models trained under a leave‑one‑out cross‑validation (LOOCV) framework, summarized by concordance index (C‑index) values. (B‑C) Kaplan–Meier survival curves comparing overall survival (OS) between high‑risk and low‑risk groups stratified by NR.Sig in the training (B) and combined test (C) cohorts. (D‑E) Receiver operating characteristic (ROC) curves for 1‑, 3‑, and 5‑year OS predicted by NR.Sig in the training cohort (D) and the combined test cohort (E). (F‑G) Calibration curves for 1‑, 3‑, and 5‑year OS probabilities derived from NR.Sig in the training (F) and combined test (G) cohorts. (H) Area under the curve (AUC) values for 3‑year OS prediction by NR.Sig, a Cox regression model integrating NR.Sig and clinical variables, and clinical variables alone in the training cohort. (I) Time‑dependent AUC curves comparing the discrimination performance of NR.Sig, the combined Cox model, and clinical variables over follow‑up time. (J) Forest plot of multivariate Cox regression analysis evaluating the independent prognostic value of NR.Sig alongside clinical covariates in the training cohort. (K) Decision curve analysis (DCA) comparing the clinical net benefit of NR.Sig, the integrated Cox model, and clinical variables across a range of risk thresholds.

Model comparison and model evaluation in immunotherapy prediction

To assess the prognostic performance of NR.Sig, we obtained model‑gene coefficients from several previously established LUAD prognostic signatures and compared their concordance indices (C‑indices) with NR.Sig across multiple LUAD RNA‑seq cohorts. The results demonstrated that NR.Sig outperformed most existing models in prognostic stratification (Fig. 5A), indicating its robust predictive capacity in LUAD. Feature‑importance analysis based on the RSF algorithm identified FAM83A as the most influential variable within the NR.Sig model (Fig. 5B). For immunotherapy prediction, survival analysis was conducted in three independent LUAD cohorts with available immunotherapy data. Patients with higher NR.Sig risk scores exhibited significantly worse PFS (Fig. 5C) and OS (Fig. 5D) compared to those with lower scores. Furthermore, immunotherapy responders displayed significantly lower NR.Sig risk scores than non‑responders, supporting the utility of NR.Sig in predicting treatment response (Fig. 5E).

FIG. 5.

Model comparison and performance in predicting immunotherapy response. (A) Concordance index (C‑index) comparison between NR.Sig and previously published prognostic signatures across eight independent LUAD cohorts (TCGA‑LUAD, GSE31210, GSE42127, GSE13213, GSE26939, GSE29016, GSE30219, and the meta‑cohort). Significance levels: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. (B) Feature selection and relative importance of variables included in the NR.Sig model, derived from the random survival forest (RSF) algorithm. (C) Kaplan–Meier curves for progression‑free survival (PFS) stratified by NR.Sig risk groups in three independent immunotherapy‑treated cohorts. (D) Kaplan–Meier curves for overall survival (OS) stratified by NR.Sig risk groups in two immunotherapy‑treated cohorts. (E) Distribution of NR.Sig risk scores in immunotherapy responders versus non‑responders across three immunotherapy cohorts.

Functional enrichment and immunogenomic applications of NR.Sig

To elucidate the molecular mechanisms associated with the NR.Sig model genes, we performed principal component analysis (PCA), which revealed clear separation between high‑ and low‑risk groups stratified by NR.Sig risk scores in the TCGA‑LUAD cohort (Fig. 6A). Differentially expressed genes (DEGs) identified between the two risk groups were subjected to functional enrichment analysis using GO and KEGG databases. GO analysis highlighted enrichment in biological processes such as regulation of establishment of protein localization telomere, while KEGG pathway analysis indicated significant involvement of proteasome (Figs. 6B-C). The expression patterns of NR.Sig model genes and clinicopathological characteristics differed markedly between the two risk groups (Fig. 6D). To assess the discriminative capacity of NR.Sig within the immune microenvironment, we quantified immune‑cell infiltration levels in the TCGA‑LUAD cohort using eight deconvolution algorithms. The analysis demonstrated significantly reduced immune‑cell infiltration in high‑risk cases (Fig. 6E). Furthermore, ssGSEA of gene sets representing six key steps of the cancer‑immunity cycle revealed enhanced immune‑cycle activity in low‑risk patients (Fig. 6F). Besides, based on immune‑function signatures indicated that the low‑risk group displayed a more active immune microenvironment (Fig. 6G). Notably, low‑risk patients exhibited elevated expression of immune‑checkpoint genes, suggesting a greater likelihood of response to immunotherapy (Fig. 6H). Spearman correlation analysis confirmed a significant inverse relationship between immune cell abundance and NR.Sig risk scores (Fig. 6I). Collectively, these results underscore the ability of NR.Sig to effectively stratify the tumor immune microenvironment in LUAD.

FIG. 6.

Functional enrichment and transcriptomic profiling of model genes, and comprehensive characterization of the TME across risk groups in TCGA‑LUAD. (A) Principal component analysis (PCA) plot distinguishing high‑risk and low‑risk patients. (B‑C) Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses of differentially expressed genes (DEGs) between risk groups. (D) Heatmap displaying differential expression of model genes and clinicopathological variables between the two risk groups. (E) Immune‑cell infiltration levels estimated by eight deconvolution algorithms, compared between high‑ and low‑risk groups. (F) Single‑sample gene‑set enrichment analysis (ssGSEA) derived scores for key steps of the cancer‑immunity cycle, compared between risk groups. (G) The ssGSEA derived scores for immune‑function signatures across risk groups. (H) Differential expression of immune‑checkpoint‑related genes between risk groups. (I) Correlation heatmap illustrating associations among immune‑cell abundances and NR.Sig risk scores.

Discussion

Lung adenocarcinoma (LUAD) is a highly heterogeneous and complex malignancy, posing substantial challenges for accurate prognosis prediction [47]. Prognostic biomarkers play a critical role in anticipating disease progression, guiding treatment selection, and estimating clinical outcomes. Among key prognostic determinants, histological subtype and tumor grade are of particular importance, with higher grades generally associated with more aggressive disease, shorter PFS, and poorer OS. Molecular biomarkers have increasingly been recognized as powerful prognostic tools in LUAD. Patient age at diagnosis also significantly influences outcomes, with younger patients typically exhibiting longer OS compared to older individuals. Tumor size and anatomical location further contribute to prognostic stratification, as smaller lesions and tumors in more favorable sites correlate with better prognosis. Additionally, the presence of residual disease after surgical resection represents another major factor affecting long‑term survival. As our understanding of LUAD biology continues to evolve, novel prognostic factors and emerging molecular signatures are being investigated to refine risk stratification and enable more personalized therapeutic approaches.

Recent studies have revealed associations between neural regulation (NR) and the pathogenesis of LUAD. However, the potential role of NR‑mediated cellular modulation in tumorigenesis remains inadequately characterized. Meanwhile, ML approaches are increasingly employed in cancer prognosis research [48], providing a powerful framework for handling complex, high‑dimensional datasets through sophisticated algorithmic pipelines. These methods are particularly adept at predictive tasks, capable of identifying subtle patterns and extracting biologically meaningful insights from large‑scale data [49]. Nevertheless, translating ML‑based prognostic models into clinical practice with robust accuracy and reliability continues to pose substantial challenges. Two critical considerations in this process are the rationale for selecting a specific ML algorithm and the determination of an optimal modeling strategy. Notably, algorithm choice in research settings is often influenced by investigator preference and prior experience, which may introduce bias and affect generalizability. In this study, we integrated transcriptomic profiles from seven international multi‑center LUAD cohorts and developed a neural regulation‑based signature (NR.Sig) using a novel computational framework. NR.Sig is constructed from the mRNA expression of NR‑associated genes that demonstrate strong prognostic utility. Functional enrichment analysis revealed that the high‑risk group was predominantly enriched in processes related to cellular functions, environmental response, and organismal systems. Notably, we observed a significant association between high‑risk status and biological pathways involved in proliferation‑related signaling, which may partially explain the unfavorable prognosis in this subgroup. Model validation further indicated that NR.Sig can serve as a valuable clinical tool to inform therapeutic decisions and potentially improve patient management. Importantly, NR.Sig effectively stratified prognosis in independent immunotherapy‑treated cohorts, providing clinicians with a reliable means of identifying patients who are more likely to benefit from immune‑checkpoint inhibition. Besides, feature‑importance analysis based on the RSF algorithm identified FAM83A as the most influential variable within the NR.Sig model. FAM83A, the smallest protein molecular weight member of the FAM83 family [50], was initially identified as BJ-TSA-9 in peripheral blood mononuclear cells of lung cancer patients, located on chromosome 8q24 [51]. Growing evidence underscores FAM83A's crucial role in tumor regulation. Numerous studies consistently report its upregulation in various malignant tumors, including pancreatic, lung, breast, and cervical cancers. Furthermore, FAM83A contributes to key physiological phenomena such as drug resistance, invasion, migration, and tumor growth [52]. These findings further demonstrate the robustness of NR.Sig.

We established an integrated ML framework to construct and validate the NR.Sig using mRNA expression profiles of marker genes derived from NR‑associated epithelial cells. This signature demonstrated robust performance in both prognostic prediction and immunotherapy‑response forecasting. Through LOOCV pipeline, 101 distinct ML prognostic combinations were evaluated across training and test datasets. Systematic assessment across seven independent LUAD cohorts identified the RSF algorithm as the optimal approach for both feature selection and model construction. The strength of this integrative framework lies in its capacity to systematically compare multiple ML algorithms, thereby enabling the selection of a model with superior predictive accuracy after exhaustive combinatorial evaluation. This integrated framework simplifies the model for functional and translational application by reducing dimensionality through systematic feature selection. The superior performance of NR.Sig was subsequently validated using time‑dependent AUC analysis, ROC curves with corresponding AUC values, Kaplan–Meier survival analysis, Cox regression, calibration plots, and DCA. These evaluations consistently demonstrated that NR.Sig outperformed existing models and conventional clinical variables in both accuracy and stability. Furthermore, C‑index comparisons across seven independent LUAD cohorts, along with a meta‑analysis of pooled C‑indices, confirmed that NR.Sig achieved the highest predictive precision and robustness among all externally validated signatures. These findings underscore the translational potential of NR.Sig for clinical risk stratification and treatment‑decision support.

Immunotherapy has introduced new therapeutic avenues for patients with LUAD, providing a promising strategy to improve survival outcomes in this challenging disease [53]. By examining the relationship between NR.Sig risk scores and the TME, we observed an inverse correlation between NR.Sig risk scores and the abundance of most immune cells, immunomodulatory molecules, and immune‑checkpoint gene expression. Enrichment analysis further highlighted a stronger representation of immune‑relevant functions in the low‑risk group. Consequently, patients with lower NR.Sig scores exhibited an “immune‑hot” phenotype, characterized by heightened immune‑cell infiltration. Notably, prior studies have reported that increased infiltration of multiple immune‑cell subsets within the TME is associated with favorable prognosis in LUAD, supporting their role in restraining tumor progression [54]. Therefore, these observations may partially account for the poorer clinical outcomes observed in high‑risk patients. Meanwhile, the NR.Sig risk scores of non-responders were higher than those of responders. Collectively, the results underscore the utility of NR.Sig risk stratification as a predictive biomarker for immunotherapy response in LUAD.

Previous research has established that tumors can influence neural structures, which may subsequently modulate tumor biology through direct or indirect signaling pathways [6]. While neural‑tumor crosstalk has been relatively well‑characterized in central nervous system (CNS) malignancies, its functional significance in peripheral solid tumors remains poorly defined. In this study, we systematically examined the role of neural signaling pathways in LUAD epithelial cells, with a specific focus on their potential implications for therapeutic design and immunotherapy responsiveness. Our scRNA-seq analysis of neural‑pathway activity demonstrated that elevated neural signaling occurs not only in CNS tumors but also across a range of peripheral solid malignancies including LUAD. Both intra‑tumor and inter‑tumor heterogeneity in neural activity were substantial. Using hierarchical clustering based on neural‑signal activity profiles, patients were classified into high‑ and low‑ NR activity subgroups. These neural subtypes exhibited significant differences in clinical characteristics, survival outcomes, biological functions, and immune‑microenvironment features. We characterized the TME through the lens of neural signaling activity. Patients exhibiting higher neural‑signal activity (high NR.Sig risk scores) displayed a distinct tumor immune microenvironment (TIME) profile, marked by diminished immune‑cell infiltration, elevated proportions of tumor‑associated stromal and malignant cells, and upregulation of immunosuppressive genes. This pattern suggests a heightened propensity for immune evasion and a lower probability of response to immunotherapy.

The NR.Sig can be readily translated into clinical practice through quantitative reverse‑transcription PCR (qRT‑PCR), supporting its broad applicability. Transcriptomic profiling of biopsy or surgical LUAD specimens could be routinely incorporated into diagnostic workflows to guide therapeutic decisions. By incorporating such data into the NR.Sig framework, clinicians would be able to more precisely stratify patient risk and tailor treatment regimens, particularly for cases with unfavorable prognosis. Nevertheless, several limitations of our study should be acknowledged. First, our analysis relied on retrospective sequencing data and clinical annotations from public databases; therefore, large‑scale, multi‑center prospective validation is required to confirm generalizability. The lack of granular treatment histories, detailed metastatic patterns, and recurrence records may also influence the interpretation of our findings. Second, the functional roles of the two central hub genes identified in LUAD remain to be fully characterized. Further investigation using expanded clinical cohorts, along with in vitro and in vivo experimental models, is warranted to elucidate their biological mechanisms in LUAD pathogenesis. Finally, our current modeling strategy, which is based primarily on transcriptomic sequencing, could be significantly enhanced through the integration of multi‑modal data. Incorporating diverse molecular, imaging, and clinical datasets enables a more comprehensive view of underlying biological mechanisms and pathophysiological processes, thereby improving the robustness and predictive accuracy of prognostic models. The inclusion of multi‑modal features introduces a higher dimensional variable space, which is particularly amenable to advanced artificial intelligence (AI) architectures. Unlike conventional machine‑learning methods that rely on manual feature engineering, deep‑learning approaches can autonomously learn discriminative patterns directly from raw data. Thus, leveraging emerging deep‑learning algorithms in conjunction with multi‑modal data integration offers a powerful and scalable framework for advancing precision oncology and personalized therapeutic strategies in LUAD.

Conclusion

In summary, we successfully developed a neural regulation‑based signature (NR.Sig) that accurately predicts prognosis and immunotherapy response in lung adenocarcinoma (LUAD) using a robust machine‑learning framework. Through comprehensive exploration encompassing model performance, immune microenvironment profiling, immunotherapy response assessment, and single‑cell transcriptomic evaluation, NR.Sig demonstrates consistent robustness and precision in outcome prediction, establishing it as a highly reliable prognostic tool in LUAD.

Funding

No funding was received for this research.

Ethics statement

This study did not involve any animal or human experiments. All analyses were performed using publicly available datasets; therefore, no ethical approval was required.

Clinical trial number

Not applicable.

Consent to Publish declaration

Not applicable.

Consent to participate declaration

Not applicable.

CRediT authorship contribution statement

Yiping Zheng: Writing – review & editing, Writing – original draft, Visualization, Software. Xiaye Miao: Writing – review & editing, Writing – original draft, Visualization, Software. Yumin Wang: Writing – review & editing, Writing – original draft, Visualization, Software. Shuzhen Wei: Writing – review & editing, Conceptualization. Qing Zhang: Writing – review & editing, Conceptualization.

Declaration of competing interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Data availability statement

Publicly available datasets were analyzed in this study. The names of the repositories and accession numbers can be found within the article/Supplementary Materials. The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.