Single-cell and multi-omics integration delineates the landscape of gene-wise intratumor heterogeneity and identifies prognostic biomarkers for immunotherapy in non-small cell lung cancer Authors: Jianye Yuan, Zelin Weng, Zhenguo Li, Rui Chen, Chao Cheng, Weixiong Yang, Xiuying Xie, Chang Luo, Tao Wang, Shuishen Zhang, Zihui Tan Published: 2026 ABSTRACT AbstractBackgroundGene-wise intratumor heterogeneity (ITH), defined as spatial variability in the expression of individual genes across tumor regions, remains incompletely characterized in non-small cell lung cancer (NSCLC). Identifying low-ITH genes as predictive biomarkers offers a promising strategy to enhance the reliability of immunotherapy outcome prediction.MethodsWe profiled gene-wise ITH using multi-region scRNA-seq data and a computational framework combining variance and clustering metrics. Prognostic low-ITH genes were screened via immunotherapy RNA-seq datasets and six machine learning algorithms to construct the LITHrisk score. Transcriptomic profiling and in vitro experiments elucidated the biological and functional relevance of key score components.ResultsThe gene-wise ITH landscape in NSCLC was delineated, with several critical immune checkpoint genes, including PD-L1, LAG3, and TIM-3, being identified as low-ITH genes. A low-ITH-derived signature, the LITHrisk score (PD-L1, CDK11B, CXCL13, RPS4Y1, AKR1B10, CRLF1), effectively predicted immunotherapy prognosis and response in NSCLC and other cancers. Tumors with a low LITHrisk score were characterized by enrichment of immune response-related pathways and heightened infiltration of immune cells, such as T cells, B cells, and myeloid cells. Furthermore, we found that component genes of the score were linked to key tumor characteristics: AKR1B10 was associated with tumor immune interactions, functioning via SPP1 regulation to suppress anti-tumor immunity, while CRLF1 was implicated in promoting a fibroblast-enriched microenvironment and modulating tumor cell motility.ConclusionOur study provides a practical transcriptome-based tool for advancing precision immunotherapy and nominates potential therapeutic targets for enhancing treatment efficacy.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s00262-026-04454-4.Keywords: Gene-wise intratumor heterogeneity, Immunotherapy, Prognostic biomarkers, Non-small cell lung cancer Introduction Lung cancer continues to impose a significant burden as a primary cause of cancer mortality worldwide, with non-small cell lung cancer (NSCLC) accounting for the majority of cases as the predominant histological subtype [1, 2]. Immune checkpoint blockade (ICB), particularly targeting PD-1/PD-L1, has revolutionized treatment paradigms, as increasing clinical trials have shown that ICB enhances the prognosis for both locally advanced and metastatic NSCLC patients [3–6]. However, a notable fraction of individuals treated with ICB do not experience favorable outcomes owing to the development of primary or secondary resistance [7, 8], underscoring the critical significance of predicting the therapeutic efficacy of ICB for the precise treatment of patients. Recent multi-omics studies have unveiled extensive spatial and molecular intratumor heterogeneity (ITH) in lung cancer, encompassing a broad spectrum of dimensions, including the genome, transcriptome, and proteome [9–11]. Recently, there has been growing recognition of the clinical significance of ITH. Studies have shown that higher levels of ITH are associated with poorer survival outcomes [11]. Furthermore, ITH has been identified as a critical factor contributing to therapy resistance [12]. Despite these advances, gene-wise ITH—the spatial variability in the expression of individual genes across distinct tumor regions—remains poorly characterized [13]. Most studies focus on cell-type heterogeneity or bulk genomic alterations, overlooking how expression patterns of immunotherapy-related genes (e.g., immune checkpoints, cytotoxic effectors) vary spatially within tumors. This gap impedes immunotherapy biomarker development because existing gene expression signatures are susceptible to tumor sampling bias, leading to inaccurate predictions of ICB response. A feasible strategy is to exclude high-ITH genes beforehand during prediction signature development [14]. In addition, subclones may have differential expression levels of immune checkpoints, which may lead to heterogeneity in the response to ICB. Our research systematically assessed gene-wise ITH in NSCLC and identified immunotherapy efficacy-related genes characterized by low heterogeneity. Consequently, we developed an LITHrisk score using these genes and further elucidated its associated immune microenvironment. This work bridges spatial genomics, biomarker discovery, and immunotherapy optimization, offering a comprehensive framework for precision oncology in NSCLC. Materials and methods Data origins Single-cell RNA sequencing (scRNA-seq) data from multiple regions of NSCLC were obtained from the ArrayExpress database under accession numbers E-MTAB-6149 and E-MTAB-6653 [15]. This dataset comprised eight NSCLC patients, each with three distinct regions sampled. Bulk RNA-seq data were acquired from The Cancer Genome Atlas (TCGA) database (TCGA-LUAD and TCGA-LUSC) and Gene Expression Omnibus (GEO) database under accessions GSE190265, GSE207422, and GSE78220. Additionally, the IMvigor210 cohort was accessed via http://research-pub.gene.com/IMvigor210CoreBiologies [16]. Further bulk RNA-seq datasets were retrieved from the European Nucleotide Archive under accessions PRJEB25780 (Kim dataset) [17] and PRJEB23709 (Gide dataset) [18]. Spatial transcriptomics data generated using the 10 × Genomics Visium platform of 8 NSCLCs were downloaded from BioStudies (https://www.ebi.ac.uk/biostudies/) under accession number E-MTAB-13530 [19]. Detailed information regarding all aforementioned cohorts is systematically summarized in Supplementary Table 1. scRNA-seq data processing Quality filtering and downstream analysis of scRNA-seq data were carried out using the Seurat R package (version 5.1.0) [20]. Cells containing unique feature counts below 200 or above 10,000, or with mitochondrial content exceeding 15%, were excluded. Normalization of the gene expression matrices was processed with the NormalizeData function under default parameters. Subsequently, a set of 2000 genes exhibiting the highest variation was selected by employing the FindVariableGenes method, also using default settings. Subsequently, a linear transformation was applied with ScaleData, and principal component analysis was conducted using RunPCA. Following correction of batch effects with the Harmony algorithm, a graph-based clustering procedure using the Louvain algorithm was applied on the first 30 principal components with the FindNeighbors and FindClusters functions. For visualization, Uniform Manifold Approximation and Projection (UMAP) embeddings were generated using RunUMAP. Cell types were then annotated based on classical marker genes: epithelial cells (EPCAM, CDH1, KRT19), T cells (CD3D, CD3E), natural killer (NK) cells (NCAM1, NKG7, GNLY), myeloid cells (CD68, LYZ, CD14, CD163), B cells and plasma cells (CD79A, MS4A1, MZB1), mast cells (KIT, MS4A2, TPSB2), fibroblasts (DCN, COL1A1, COL1A2), and endothelial cells (PECAM1, CLDN5, RAMP2). To distinguish malignant cells from the epithelial compartment, we applied copy number variation (CNV) analysis using the inferCNV R package (version 1.20.0) from the Trinity CTAT Project (https://github.com/broadinstitute/inferCNV). To quantitatively evaluate the integration performance and the preservation of regional heterogeneity, the Local Inverse Simpson’s Index (LISI) was calculated using the lisi R package (version 1.0). Gene-wise ITH quantification To quantify gene-wise ITH, regional pseudo-bulk gene expression profiles were extracted exclusively from the unintegrated, log-normalized expression matrix using the AverageExpression function in Seurat (focusing strictly on the data layer). This approach was intentionally designed to ensure that the Harmony batch correction applied during cell clustering did not artificially attenuate the authentic spatial variance among different regions. We calculated the integrated heterogeneity score (IHS) following the mathematical framework established by Yang et al. [13]. The global IHS was computed for each gene by synergistically combining a variance-based approach and a clustering-based approach. First, for the variance-based approach, intratumor variance (W, representing the variance among different regions within the same tumor) and intertumor variance (B, representing the variance between different patients) were computed utilizing linear mixed-effects models via the nlme R package (version 3.1–166). The intratumor variability score (ITVS) was formulated as: A lower ITVS indicates decreased intra-tumor heterogeneity. Second, for the clustering-based approach, hierarchical clustering was performed to group spatial samples into clusters with sequentially increasing numbers (from k = 1 to Npatients). At each clustering step, the patient group overall ratio (PGOR) was calculated. The PGORk is defined mathematically as the proportion of patients whose multi-region samples remained grouped within the same cluster: After computing the area under the PGOR curve AUCPGOR, the clustering concordance score (CCS) was defined as: Finally, the IHS was determined by calculating the geometric mean of ITVS and CCS: Consistent with the established classification criteria by Yang et al., genes were categorized into four discrete tiers based on their global IHS (ranging from 0 to 1): low ITH (0–0.25), medium ITH (0.25–0.50), high ITH (0.50–0.75), and very high ITH (0.75–1.00). Furthermore, to validate the robustness of the ITH metric against sampling variation, a leave-one-out sensitivity analysis was performed. Specifically, the IHS for all genes was iteratively recalculated by sequentially excluding all spatial regions associated with a single patient. The stability of the ITH metric was then quantitatively evaluated by calculating the Spearman rank correlation (ρ) between the IHS derived from each leave-one-out subset and the baseline IHS of the full cohort. Spatial transcriptomics data processing Spatial transcriptomics data were processed using the Seurat R package (version 5.1.0). The SCTransform function was employed to normalize the data. We then identified genes exhibiting significant spatial patterns by calculating Moran’s I statistic, a measure of spatial autocorrelation, using the FindSpatiallyVariableFeatures function with the ‘moransi’ method. The integrated scRNA-seq dataset served as a reference to deconvolute and map cell types onto the spatial transcriptomics slide using the Robust Cell Type Decomposition (RCTD) method, implemented via the spacexr R package (version 2.2.1) [21], thereby generating spatially resolved cell-type distributions. Development of the LITHrisk score A candidate gene set was derived from the intersection of genes with low ITH identified by scRNA-seq analysis and prognostic genes identified from immunotherapy RNA-seq datasets. Subsequently, we performed further gene screening using six algorithms, including Multivariate Cox Regression, Lasso, random survival forest, xgboost, CoxBoost, and Elastic Net. These analyses were implemented using the following R packages: survival (version 3.6–4), glmnet (version 4.1–8), randomForestSRC (version 3.3.2), xgboost (version 1.7.11.1), and CoxBoost (version 1.5). The six genes obtained from the screening were used to calculate a risk score via Multivariate Cox Regression. Gene set enrichment analysis The clusterProfiler R package (version 4.2.2) was used to conduct Gene set enrichment analysis (GSEA) of RNA-seq data [22]. For scRNA-seq data analysis, GSEA was conducted with the SCP R package (version 0.5.6). The reference gene sets utilized in this study were acquired from the Molecular Signatures Database (MSigDB), specifically the Hallmark and Gene Ontology (GO) collections. The false discovery rate (FDR) correction was conducted using the Benjamini–Hochberg procedure, and enrichment results with an FDR < 0.05 were deemed statistically significant. Characterization of tumor immune infiltrate from RNA-seq data The abundance of tumor-infiltrating immune cells was estimated from RNA-seq data using the MCP-counter algorithm, implemented with the MCPcounter R package (version 1.2.0) [23]. Prediction of chemotherapy sensitivity The oncoPredict R package (version 1.2) was employed to evaluate the half-maximal inhibitory concentration (IC50) of common chemotherapeutic agents in the NSCLC samples [24]. Cell communication analysis The CellChat R package (version 1.6.1) was applied to decipher the cell–cell communication networks present in the NSCLC microenvironment [25]. The normalized gene expression matrix was processed through the default CellChat workflow, which quantifies interaction probabilities by integrating gene expression with a comprehensive repository of validated ligand-receptor pairs. We utilized the CellChatDB.human database for inference, considering all interaction categories. Inferred interactions involving fewer than 10 cells were excluded from subsequent analysis. CytoTRACE analysis The developmental potential of individual tumor cells was assessed using CytoTRACE R package (version 0.3.3), an algorithm that infers cellular differentiation states based on transcriptional diversity [26]. CytoTRACE scores, ranging from 0 (most differentiated) to 1 (least differentiated), were computed for each cell to reconstruct differentiation trajectories. CNV score calculation We utilized the CopyKAT algorithm (implemented via the copykat R package, version 1.1.0) to evaluate CNVs within individual tumor cells [27]. To quantify the extent of genomic alterations, the CNV score for each cell was defined as the mean absolute value of CNVs predicted by CopyKAT [28]. Cell culture The human lung adenocarcinoma A549 cell line and the human lung squamous carcinoma NCI-H226 cell line were obtained from the Type Culture Collection of the Chinese Academy of Sciences, Shanghai, China. All cells were cultured in Dulbecco’s modified Eagle’s medium (DMEM, Gibco, USA) supplemented with 10% fetal bovine serum (FBS, Gibco) under standard incubator conditions (37 °C, 5% CO₂, saturating humidity). Transfection assay AKR1B10 siRNA, CRLF1 siRNA, and control siRNA (siNC) were obtained from RiboBio (Guangzhou, China). The CRLF1 overexpression plasmid was generated by inserting the CRLF1 open reading frame (ORF) into a pReceiver-M02 vector (GeneCopoeia, USA), followed by sequence verification. The NSCLC cells were transfected with siRNA using Lipofectamine RNAiMAX Transfection Reagent (Invitrogen, USA) and with plasmid using Lipofectamine 3000 Transfection Reagent (Invitrogen) according to the manufacturer’s instructions. The cells were collected 48 h post-transfection for subsequent experiments. Real-time quantitative PCR (RT-qPCR) Total RNA was extracted from the cultured cells using TRIzol reagent (Invitrogen), and subsequently, reverse transcription was performed using the RevertAid RT Reverse Transcription Kit (Invitrogen) following the manufacturer’s protocols. qPCR was then carried out on a LightCycler 480 II system (Roche, Switzerland). Gene-specific primers were designed and synthesized, with their sequences listed as follows: AKR1B10 forward primer: 5′-CGCAGCCCAGGTTCTGAT-3′, and reverse primer: 5′-TGGGACATGAGTGGAGGTAGTC-3′; CRLF1 forward primer: 5′-AGTGTGGACTGGAAGGTGGT-3′, and reverse primer: 5′-GCCATAGATGCCAAAGGGGT-3′. Western blot Protein lysates from cell lines were extracted using RIPA buffer with protease inhibitor cocktail (Roche). Protein concentration was determined with a BCA assay kit (Thermo Fisher, USA). Protein lysates were then separated by SDS-PAGE and electrophoretically transferred onto polyvinylidene fluoride membranes (Millipore, USA). The membranes were blocked in 5% skim milk prepared in TBST for 1 h at room temperature and then incubated overnight at 4 °C with specific primary antibodies against AKR1B10 (18252–1-AP, 1:4000, Proteintech, USA), SPP1 (22952–1-AP, 1:2000, Proteintech), CRLF1 (ab211438, 1:1000, Abcam, USA), Vinculin (13901, 1:1000, CST, USA), and GAPDH (97166, 1:1000, CST). The membranes were subsequently incubated with secondary antibodies (7074 or 7076, CST) at room temperature for 1 h. Protein bands were visualized using a ChemiDoc Imaging System (Bio-Rad). Transwell migration assay Cell migration was assessed using 24-well Transwell chambers with 8 µm pore size inserts (353097, Corning, USA). Briefly, cells were resuspended in serum-free medium and seeded into the upper chamber. The lower chamber was filled with 700 µL of complete medium containing 10% FBS. After incubation for 24 h at 37 °C, non-migratory cells on the upper surface of the membrane were removed with a cotton swab. Cells that had migrated to the lower surface were fixed with methanol and stained with crystal violet. Migrated cells were visualized and imaged under an Olympus microscope. Quantification was performed by counting cells in five randomly selected fields per chamber. Statistics and software Overall survival (OS) and progression-free survival (PFS) were analyzed using the Kaplan–Meier method, and comparisons between groups were performed with the log-rank test. Hazard ratios (HR) and 95% confidence intervals (CI) were estimated using univariate Cox proportional hazards regression models. The optimal threshold for grouping was computationally derived through the maxstat R package (version 0.7–25). For continuous variables, comparisons between two groups were conducted using Student’s t tests or Wilcoxon rank-sum tests. To rigorously quantify the magnitude of differences for nonparametric comparisons, Cliff’s δ was calculated as the measure of effect size. Correlations between two continuous variables were assessed using Pearson or Spearman correlation analysis. In this study, all statistical analyses were carried out with R software (version 4.4). Some of the figures were generated using BioGDP.com [29]. Statistical significance was defined as a two-tailed P value below 0.05. Results Gene-wise and cell-type infiltration ITH in NSCLC To systematically investigate the landscape of gene-wise ITH and translate these findings into actionable biomarkers for NSCLC immunotherapy, we designed a comprehensive, multi-stage analytical framework (Fig. 1). As a foundational step, single-cell resolution enables gene expression profiling within defined cell types, allowing for an in-depth dissection of gene-wise ITH in NSCLC. Using single-cell transcriptomic data from a multi-region NSCLC cohort (n = 8 patients, three regions per tumor), we applied Harmony batch correction across patients, while the spatial heterogeneity among distinct regions was faithfully preserved (Supplementary Figs. 1 and 2). Building upon this, we first profiled the expression of key immune checkpoint genes across the annotated cell types (Fig. 2A–B). PD-L1 was heterogeneously expressed in malignant cells and also detectable in immune lineages (T, myeloid, and B cells). PD-1, CTLA-4, LAG3, TIGIT, HAVCR1, TIM-3, and CD39 were predominantly expressed in T cell subsets, with partial expression in myeloid and B cells. Schematic overview of the study design and analytical workflow. First, multi-region scRNA-seq and spatial transcriptomics were utilized to quantify gene-wise intratumor heterogeneity (ITH) and intra-sectional spatial patterns of NSCLC. Next, by intersecting low-ITH genes with immunotherapy-prognostic signatures, we applied integrative machine learning pipelines for feature selection to construct a six-gene LITHrisk model, which was subsequently validated across multiple independent cohorts. Finally, comprehensive biological insights were delineated at both the overall risk-score level (including pathway enrichment, tumor microenvironment profiling, and drug sensitivity) and the individual core gene level (including pathway enrichment, cell–cell communication, and in vitro experimental validation). Part of the figures were created with BioGDP.com Multi-region scRNA-seq revealed gene-wise ITH in NSCLC. A UMAP visualization of all cell subsets. The major cell subsets are shown in the circle. B UMAP visualization showing the expression of 8 immune checkpoint genes. C The integrated heterogeneity score (IHS) of 8 immune checkpoint genes in whole tissue, immune cells, and tumor cells, and the IHS of cell subtype infiltration Gene-wise ITH was quantified in whole tissue, immune, and tumor cells using the integrated heterogeneity score (IHS), a comprehensive metric based on intra-tumor variability and clustering concordance [13]. Based on the IHS, genes were categorized into four groups using predefined intervals adopted from the aforementioned established framework: low ITH (0–0.25), medium ITH (0.25–0.50), high ITH (0.50–0.75), and very high ITH (0.75–1.00). At the whole-tissue level, PD-L1, LAG3, and TIM-3 were low-ITH genes (Fig. 2C). PD-L1 showed low ITH in both tumor and immune cells, while LAG3 and TIM-3 had low ITH in immune cells. Targeting these low-ITH genes may achieve more uniform therapeutic efficacy across regions. In contrast, PD-1, CTLA-4, TIGIT, HAVCR1, and CD39 were medium-ITH genes at the whole-tissue level and showed medium-to-high ITH in immune cells, suggesting heterogeneous responses across regions if targeted, with regions of poor response potentially contributing to treatment resistance. These findings suggest that for NSCLC immunotherapy selection, low-ITH targets PD-L1, LAG3, and TIM-3 may be more suitable for ICB targeting, whereas relatively higher-ITH molecules like HAVCR1 and TIGIT may be less optimal. We also evaluated ITH in cell infiltration using the same methodology to assess the reliability of cell infiltration as a predictive biomarker. Previously associated with better immunotherapy response [30–33], plasma cells, atypical memory B cells (Bm_atypical), and precursors of exhausted CD8 + T cells (CD8_Tpex) exhibited medium or high ITH in infiltration patterns (Fig. 2C), indicating substantial spatial heterogeneity across NSCLC tumor regions. Spatial heterogeneity in gene expression within tissue sections of NSCLC Spatial transcriptomics enables the characterization of spatial heterogeneity in gene expression within tissue sections, prompting our exploration of its association with gene-wise ITH using an independent spatial transcriptomics cohort (n = 20 tissue sections from 8 NSCLC tumors; Fig. 3A). Moran’s I was used to quantify spatial heterogeneity in gene expression, where values near 1 indicate strong spatial clustering and values near 0 reflect a random, low-heterogeneity distribution. Low Moran’s I was observed for LAG3, PD-L1, CTLA-4, PD-1, TIM-3, and TIGIT, suggesting limited spatial heterogeneity within sections (Fig. 3B). Conversely, CD39 exhibited a relatively high Moran’s I, reflecting elevated spatial heterogeneity. HAVCR1 was not evaluated due to undetectable expression levels. Notably, Moran’s I did not correlate with IHS (Fig. 3C), suggesting that spatial heterogeneity within a single section does not adequately reflect gene-wise ITH across multiple tumor regions. However, this lack of correlation should be interpreted with caution, as the analysis might be underpowered due to the limited number of spatial transcriptomic samples. Spatial transcriptomics assessed spatial heterogeneity in gene expression within tissue sections of NSCLC. A H&E staining of the representative tissue section and the corresponding spatial transcriptome atlas depicting the expression of immune checkpoint genes. B Moran’s I of 7 immune checkpoint genes assessed by spatial transcriptomics. C A scatter plot shows the correlation between the integrated heterogeneity score (IHS) of genes evaluated by scRNA-seq and Moran’s I of genes assessed by spatial transcriptomics. Pearson correlation analysis LITHrisk score predicts prognosis of immunotherapy in NSCLC We subsequently focused on low-ITH genes identified at the whole-tissue level as candidate biomarkers for predicting immunotherapy outcomes. Whole-tissue gene expression data were selected due to their greater feasibility in routine detection. We first performed a leave-one-out sensitivity analysis to rigorously ensure the reliability of the scRNA-seq-derived IHS against sampling variation. Iterative recalculation of the IHS demonstrated remarkable stability, maintaining a highly consistent rank order across permutations compared to the full cohort baseline (median Spearman’s ρ = 0.93; Fig. 4A). Using bulk RNA-seq data from immunotherapy-treated NSCLC samples (GSE190265, anti-PD-1, n = 43), we identified genes associated with immunotherapy prognosis. Of the 1,958 low-ITH genes, 19 were found to be immunotherapy-prognostic genes (Fig. 4B). Six machine learning methods refined these to six core genes (Fig. 4C–D), which were integrated into a multivariate Cox model to construct a LITHrisk score: Low-ITH genes for immunotherapy efficacy prediction in NSCLC. A Leave-one-out robustness evaluation of the continuous integrated heterogeneity score (IHS). The high Spearman rank correlation (ρ) between each leave-one-out subset and the full cohort demonstrates the high stability of the metric against sampling variation. B Venn diagram identifying common genes with both low intratumor heterogeneity (ITH) and association with immunotherapy prognosis. C UpSet plot displaying immunotherapy-prognostic genes identified by six different filtering methods. D Ranking of immunotherapy-prognostic genes based on the six filtering methods. E The distribution of LITHrisk score and progression-free survival (PFS) in the GSE190265 dataset (n = 43). F Survival curves for PFS of patients grouped by LITHrisk score in the GSE190265 dataset. G Comparisons of LITHrisk scores in the GSE207422 dataset between patients with non-major pathological response (NMPR, n = 15) and major pathological response (MPR, n = 9). Wilcoxon rank-sum test Clinically, the LITHrisk score successfully stratified patient survival and showed a concordant trend with therapeutic responses. Patients were stratified into high and low LITHrisk groups according to an optimal threshold. The scatter plot of LITHrisk score indicated that as the LITHrisk score increased, PFS decreased, and the progression rate increased (Fig. 4E). The high LITHrisk group showed significantly shorter PFS compared with the low-risk group (P < 0.001, HR = 7.83 [95%CI: 3.52–17.45], Fig. 4F). In an independent bulk RNA-seq NSCLC cohort (GSE207422, anti-PD-1, n = 24), lower LITHrisk scores trended with major pathological response (MPR) (P = 0.123, Cliff’s δ = 0.39, Fig. 4G), suggesting its association with pathological treatment response. We further validated the capacity of the LITHrisk score to predict immunotherapeutic efficacy using independent bulk RNA-seq cohorts across multiple cancer types. In urothelial cancer (IMvigor210, anti-PD-L1, n = 348) and melanoma (GSE78220, anti-PD-1, n = 26), low LITHrisk was associated with improved survival and response (Fig. 5A–D). Consistent results were observed in the Gide dataset (anti-PD-1, n = 41; anti-PD-1&CTLA-4, n = 32), where melanoma patients with a low LITHrisk score showed improved survival outcomes and tended to have better responses to both anti-PD-1 monotherapy and combination therapy with anti-PD-1 and CTLA-4 inhibitors (Fig. 5E–J). In addition, in the Kim dataset (anti-PD-1, n = 45), gastric cancer (GC) patients with a low LITHrisk score exhibited enhanced responses to anti-PD-1 immunotherapy (Fig. 5K). The LITHrisk score for immunotherapy efficacy prediction in multi-cancer cohorts. A Survival curves for overall survival (OS) of patients grouped by LITHrisk score in the IMvigor dataset (n = 348). B Comparisons of LITHrisk scores in the IMvigor dataset between patients with stable disease/progressive disease (SD/PD) (n = 230) and those with complete response/partial response (CR/PR) (n = 68). Wilcoxon rank-sum test. C Survival curves for OS of patients grouped by LITHrisk score in the GSE78220 dataset (n = 25). D Comparisons of LITHrisk scores in the GSE78220 dataset between non-responders (NR, n = 12) and responders (R, n = 14). Wilcoxon rank-sum test. E–H Survival curves for OS and progression-free survival (PFS) of patients treated with anti-PD-1 monotherapy (n = 41) or anti-PD-1 combined with anti-CTLA-4 (n = 32) grouped by LITHrisk score in the Gide dataset. I–J Comparisons of LITHrisk score between patients with SD/PD and those with CR/PR in the Gide dataset among patients treated with anti-PD-1 monotherapy or anti-PD-1 combined with anti-CTLA-4. Wilcoxon rank-sum test. K Comparisons of LITHrisk score in the Kim dataset between patients with SD/PD (n = 33) and those with CR/PR (n = 12). Wilcoxon rank-sum test. Note: Sample sizes for specific analyses may vary slightly due to missing clinical data in the original cohorts. Specifically, for patients treated with anti-PD-1 combined with anti-CTLA-4 in the Gide cohort (G-H), the hazard ratios for both OS and PFS were not estimable due to the absence of death or progression events in the low LITHrisk group Uncovering biological insights of the LITHrisk score Tumors with a low LITHrisk score are characterized by a highly inflamed, immune-active microenvironment. GSEA showed that the low LITHrisk group in GSE190265 was enriched for immune and inflammatory response pathways (e.g., ALLOGRAFT_REJECTION, INTERFERON_GAMMA_RESPONSE) and cytokine signaling pathways (INTERFERON_GAMMA_RESPONSE, IL6_JAK_STAT3_SIGNALING, TNFA_SIGNALING_VIA_NFKB) (Fig. 6A), suggesting a more robust anti-tumor immune response. In contrast, the high LITHrisk group was linked to oncogenic pathways like MYC_TARGETS_V1, indicating a greater degree of tumor malignancy. Consistently, immune deconvolution revealed significantly greater infiltration of B cells, T cells, CD8 + T cells, myeloid cells, and dendritic cells in the low LITHrisk group (Fig. 6B), further supporting the presence of enhanced anti-tumor immunity. Biological insights of the LITHrisk score. A Barplot of normalized enrichment score (NES) and false discovery rate (FDR) from gene set enrichment analysis (GSEA) for HALLMARK gene sets enriched in the high (n = 20) and low (n = 23) LITHrisk score groups of the GSE190265 dataset (n = 43). B Box plots showing comparisons of the infiltration of 10 cell types quantified by MCP-counter between high and low LITHrisk score groups in the GSE190265 dataset. Wilcoxon rank-sum test. C Box plots showing comparisons of the predicted IC50 of drugs between high and low LITHrisk score groups in the GSE190265 dataset. Wilcoxon rank-sum test. *P < 0.05; **P < 0.01; ***P < 0.001; and ns, not significant Alongside these distinct transcriptomic and immune profiles, we also observed divergent pharmacological vulnerabilities. Drug sensitivity analysis predicted that the low LITHrisk group was more sensitive to various agents, including chemotherapeutics (Decitabine, Carboplatin), pro-apoptotic compounds (PAC-1, O-6-benzylguanine), targeted inhibitors (IC-87114 for PI3K, Trametinib for MEK1/2, TPCA-1 for IKKβ, NVP-BSK805 for JAK2, ML-210 for GPX4, RAF265 for RAF, Fedratinib for JAK2, UNC0638 for G9a, PF-3758309 for PAK4), and statins with reported anti-tumor effects [34, 35]. In contrast, the high LITHrisk group showed higher sensitivity only to Ifosfamide (Fig. 6C). Biological insights into the functional roles of LITHrisk component genes Single-cell transcriptome analysis was conducted to further investigate the biological insights underlying the LITHrisk score. Single-cell transcriptomic mapping confirmed that the six LITHrisk component genes exhibit distinct cell-type-specific expression profiles (Fig. 7A). As previously described, PD-L1 expression was observed in tumor cells and immune cells, including T cells, myeloid cells, and B cells. CDK11B and RPS4Y1 expressions were observed in multiple cell types, including tumor, immune, and stromal cells. CXCL13 expression was predominantly observed in T cells. Notably, AKR1B10 and CRLF1 expressions were primarily observed in tumor cells. Of these genes, PD-L1 is an established immune checkpoint and predictive biomarker for immunotherapy [36], while CXCL13—expressed by T cell subsets such as exhausted CD8 + T cells—also correlates with treatment efficacy [32, 37]. The inclusion of these key immunotherapy-related molecules supports the validity of the LITHrisk score derived from unsupervised screening. Biological insights into the LITHrisk score component genes. A UMAP visualization of the expression of 6 component genes of the LITHrisk score. B UMAP visualization of tumor cells stratified by AKR1B10 expression (positive vs. negative). C Barplot of normalized enrichment score (NES) and false discovery rate (FDR) from gene set enrichment analysis (GSEA) for Gene Ontology-Biological Process (GO-BP) gene sets enriched in the AKR1B10 positive and negative tumor cells. D Bubble heatmap showing ligand–receptor pairs involved in cell–cell communication between AKR1B10 positive or negative tumor cells and T cell subsets. Dots were colored by communication probability. E UMAP visualization of tumor cells stratified by CRLF1 expression (positive vs. negative) and their CytoTRACE score reflecting differentiation order. F Violin plot comparing the CytoTRACE scores between CRLF1-positive and negative tumor cells. Higher CytoTRACE scores indicate a less differentiated state. Wilcoxon rank-sum test. G Violin plot comparing CopyKAT-inferred copy number variation (CNV) scores between CRLF1-positive and negative tumor cells. Wilcoxon rank-sum test. H Barplot of NES and FDR from GSEA for GO-BP gene sets enriched in the CRLF1-positive and negative tumor cells. I Cell communication analysis based on scRNA-seq data. The heatmap depicts the predicted interaction numbers between cell subsets and communication probability of COLLAGEN, LAMININ, and FN1 pathways in cell communication between CRLF1-positive or negative tumor cells and other cell subsets We next characterized AKR1B10 and CRLF1—two tumor-enriched genes whose elevated expression correlates with higher LITHrisk and poorer immunotherapy prognosis—to elucidate the specific mechanisms through which they drive a pro-tumorigenic microenvironment. AKR1B10 expression delineated a tumor subpopulation characterized by both elevated metabolic activity and an immunosuppressive interaction profile. Specifically, AKR1B10-positive tumor cells were enriched for metabolic pathways and exhibited blunted lymphocyte activation responses (Fig. 7B–C). Cell communication analysis indicated that AKR1B10-positive tumor cells had stronger SPP1-CD44 interactions with T cells (Fig. 7D), a known axis of T cell inhibition [38], which may contribute to immunosuppression. Similarly, CRLF1 expression identified a distinct tumor cell state characterized by a less differentiated and genomically unstable phenotype, motility-regulatory transcriptomic programs, and profound stroma-remodeling capabilities. CytoTRACE analysis revealed that CRLF1-positive tumor cells exhibit a less differentiated state (Fig. 7E–F). Notably, these cells also exhibited higher CNV scores inferred by CopyKAT, indicating stronger genomic alterations (Fig. 7G). Collectively, this dedifferentiated and genomically unstable phenotype highlights the enhanced malignant potential of the CRLF1-positive subpopulation. GSEA showed that these cells were enriched in motility-related pathways (e.g., cilium movement, microtubule-based movement; Fig. 7H). Furthermore, cell communication analysis demonstrated that CRLF1-positive tumor cells exhibited stronger interactions with fibroblasts via canonical fibroblast-associated pathways like COLLAGEN, LAMININ, and FN1 (Fig. 7I). This implies a link between CRLF1-positive tumor cells and fibroblast activity, which is often associated with reduced immune cell recruitment and impaired anti-tumor immunity [39]. To empirically corroborate specific mechanistic hypotheses derived from our computational analyses, we focused on validating the AKR1B10-SPP1 regulatory axis, alongside the spatial stromal interactions and functional impact of CRLF1. First, we experimentally confirmed that AKR1B10 positively regulates SPP1 expression in NSCLC. Notably, AKR1B10 and SPP1 expression was positively correlated in the TCGA NSCLC RNA-seq dataset (TCGA-LUAD, n = 528; TCGA-LUSC, n = 501; Fig. 8A). We further knocked down AKR1B10 in NSCLC cells (Fig. 8B) and found that its knockdown led to downregulation of SPP1 protein expression (Fig. 8C), further confirming that AKR1B10 upregulates SPP1 and modulates anti-tumor immunity in NSCLC. Functional exploration of AKR1B10 and CRLF1 in NSCLC. A Scatter plot showing the correlation between AKR1B10 and SPP1 expression in the TCGA NSCLC RNA-seq dataset (TCGA-LUAD, n = 528, and TCGA-LUSC, n = 501). Pearson correlation analysis. B qPCR quantification of AKR1B10 mRNA in the AKR1B10 knockdown group (si-AKR1B10) and control group (siNC) of NSCLC cells. Student’s t test. Results are shown as mean ± SD from three independent experiments (n = 3). C Western blot showing changes in protein expression after AKR1B10 knockdown in NSCLC cells. D Spatially resolved cell communication analysis. The heatmap shows the predicted interaction numbers between cell subsets and the communication probability of COLLAGEN, LAMININ, and FN1 pathways in cell communication between cell subsets. E Representative spatial visualization of cell communication of the classical fibroblast-related pathway (COLLAGEN). Nodes (cell subsets) are connected by edges whose width denotes the probability of predicted ligand-receptor interactions. F Box plots showing comparisons of the CRLF1 expression between the pathological node-negative (pN0) and node-positive (pN +) groups in the TCGA NSCLC datasets. Wilcoxon rank-sum test. G Survival curves for overall survival (OS) of patients grouped by CRLF1 expression in the TCGA-LUAD dataset (n = 507). H qPCR quantification of CRLF1 mRNA in the CRLF1 knockdown group (si-CRLF1) and control group (siNC) of NSCLC cells. Student’s t test. Results are shown as mean ± SD from three independent experiments (n = 3). I Western blot showing changes in protein expression after CRLF1 knockdown in NSCLC cells. J Transwell assay images and bar graph of CRLF1 knockdown and control NSCLC cells. Student’s t test. Results are shown as mean ± SD from three independent experiments (n = 3). K Western blot showing changes in protein expression after CRLF1 overexpression in NSCLC cells. L Transwell assay images and bar graph of CRLF1-overexpressing and control NSCLC cells. Student’s t test. Results are shown as mean ± SD from three independent experiments (n = 3). *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001; and ns, not significant. Note: Sample sizes for specific analyses may vary slightly due to missing clinical annotations in the original TCGA NSCLC cohorts Furthermore, we corroborated the spatial association of CRLF1 with fibroblasts and experimentally delineated its metastasis-suppressive role. To validate the interactions initially predicted by scRNA-seq, spatial transcriptomics confirmed intimate cross talk—prominently featuring canonical fibroblast-related pathways—between CRLF1-positive tumor cells and local fibroblasts (Fig. 8D–E). This spatial resolved interaction directly substantiates the link between CRLF1 expression and a fibroblast-enriched microenvironment. While the prior GSEA implied a potential role for CRLF1 in regulating cell motility, the specific impact remained unclear, prompting us to conduct subsequent investigations. Transcriptome analysis of TCGA NSCLC RNA-seq cohort revealed an association between CRLF1 expression and an earlier N stage in NSCLC (Fig. 8F). Additionally, high CRLF1 expression correlated with more favorable survival in the TCGA-LUAD dataset (Fig. 8G), suggesting a metastasis-suppressive role for CRLF1 in NSCLC. We further knocked down CRLF1 in NSCLC cells (Fig. 8H–I). Subsequent Transwell assays demonstrated that CRLF1 depletion enhanced the migratory ability of NSCLC cells (Fig. 8J). Correspondingly, overexpression of CRLF1 in NSCLC cells resulted in decreased cell migration (Fig. 8K–L). Taken together, our data imply a context-dependent role for CRLF1: It may improve prognosis in non-immunotherapy settings by inhibiting metastasis, yet potentially worsen outcomes under immunotherapy by fostering a fibroblast-dominated, immunosuppressive microenvironment. Discussion The landscape of gene-wise ITH in NSCLC remains incompletely defined. The identification of low-ITH genes as biomarkers facilitates more reliable prediction of immunotherapy outcomes. Furthermore, elucidating the ITH of immune checkpoints also aids in selecting targets that can elicit a homogeneous therapeutic effect across different tumor regions for immunotherapy. In this study, we systematically profiled gene-wise ITH in NSCLC and constructed a prognostic signature, the LITHrisk score, based on low-ITH genes, to predict immunotherapy efficacy. Patients with low LITHrisk scores exhibited stronger anti-tumor immunity, and the score’s component genes were functionally linked to immune regulation or tumor–microenvironment cross talk. Among common immune checkpoint targets, PD-L1, LAG3, and TIM-3 showed low ITH, leading us to hypothesize that targeting them might yield more consistent efficacy across regions. In contrast, targeting higher-ITH genes carries a theoretical risk of suboptimal response in low-expression regions, which could potentially contribute to resistance. Clinically, anti-PD-L1—a low-ITH target—is widely used with good efficacy in NSCLC [5, 6]. Although PD-1 displays moderate ITH, anti-PD-1 therapy remains effective [3, 4], likely because it blocks the PD-1/PD-L1 axis and PD-L1 exhibits low ITH. Similarly, anti-LAG3 (relatlimab) and anti-TIM-3 (cobolimab) have shown promising clinical activity in combination regimens with anti-PD-1 therapy [40, 41]. Thus, targeting low-ITH ICB genes represents a theoretically advantageous strategy in NSCLC, which warrants further experimental and prospective clinical validation. Since biopsy samples are routinely used for predictive testing, biomarkers with low ITH offer more consistent detection. Based on low-ITH genes linked to immunotherapy prognosis in NSCLC, we developed the LITHrisk score. Low LITHrisk was consistently associated with improved immunotherapy outcomes not only in NSCLC but also across multiple other cancers. Further investigation into the biological insights of the LITHrisk score helps reveal key mechanisms underlying resistance to immunotherapy and illuminate new directions for enhancing therapeutic efficacy. Biologically, the LITHrisk score correlated with anti-tumor immune pathways and MYC signaling. Accordingly, immune profiling revealed significantly stronger anti-tumor immune infiltration in the low LITHrisk group. The gene components of the LITHrisk score include PD-L1 and CXCL13, both of which have been previously linked to immunotherapy efficacy and play a role in T cell modulation in the tumor microenvironment, thereby influencing anti-tumor immunity [32, 36, 42]. Regarding other genes, CDK11B belongs to the cyclin-dependent kinases family and has been shown to affect tumor cell proliferation [43]. RPS4Y1, a ribosomal protein encoded by the Y chromosome, has been investigated for its potential role in shaping the immune landscape of ulcerative colitis [44]. AKR1B10, a member of the human aldo–keto reductase superfamily, has been shown to promote the proliferation and migration of tumor cells [45]. Targeting AKR1B10 has been reported to overcome chemoresistance in NSCLC patient-derived tumor organoids [46]. CRLF1, a type I cytokine receptor family member, has been implicated in conferring cisplatin resistance in ovarian cancer [47]. In our study of NSCLC, AKR1B10 was found to upregulate SPP1, which contributes to the suppression of anti-tumor immunity, while CRLF1 promotes a fibroblast-enriched microenvironment and modulates tumor cell motility. These findings nominate AKR1B10 and CRLF1 as promising therapeutic targets in NSCLC. Furthermore, the precise molecular mechanisms underlying the core LITHrisk genes warrant deeper investigation. As an aldo–keto reductase involved in lipid metabolism, AKR1B10 can activate the PI3K/AKT/NF-κB signaling cascade [45]. Because NF-κB is a known transcriptional activator of SPP1 [48], future studies should integrate lipidomics with NF-κB ChIP-seq and pathway inhibitors to definitively map this regulatory axis, alongside immune co-cultures to identify the specific immune subsets it suppresses. Similarly, the predicted interactions between CRLF1-positive tumor cells and fibroblasts require rigorous functional validation. Future 3D tumor-fibroblast organoid co-cultures using CRLF1-manipulated cells, combined with transcriptomic profiling and multiplexed cytokine arrays, are needed to precisely delineate how CRLF1 mediates fibroblast activation and shapes the immunosuppressive microenvironment. To translate the LITHrisk score into routine clinical practice, its cost-effectiveness and clinical utility metrics must be carefully evaluated. While whole-transcriptome sequencing is powerful, a more cost-effective and clinically feasible implementation for this concise six-gene signature would involve developing a targeted multiplex RT-qPCR assay or a multiplex immunohistochemistry scoring system applicable to standard formalin-fixed paraffin-embedded biopsy specimens. Furthermore, to establish true clinical utility and integrate this score into current decision-making frameworks, the LITHrisk signature must be rigorously evaluated in prospective, randomized trials. Future studies should directly compare LITHrisk-guided immunotherapy selection against current standard-of-care biomarkers, primarily the PD-L1 tumor proportion score and tumor mutational burden (TMB) [36], to determine whether LITHrisk offers superior or complementary predictive accuracy. This study has several limitations. First, our gene-wise ITH landscape was derived from multi-region scRNA-seq data of a relatively small cohort (n = 8), with three regions sampled per tumor. While our sensitivity analyses and previous statistical models [13] suggest that ITH metrics largely stabilize at this sampling depth, we acknowledge that this limited sample size may not fully capture the extensive heterogeneity across diverse NSCLC molecular subtypes and clinical stages. Additionally, the spatial transcriptomics cohort used to compare 2D intra-sectional patterns (Moran’s I) with macroscopic gene-wise ITH was similarly constrained, which may have limited the statistical power of this specific comparative analysis. Consequently, the identified gene-wise ITH patterns and the predictive performance of the LITHrisk score may exhibit variability across unrepresented patient subgroups. Future multi-center studies utilizing larger spatial multi-omics cohorts with denser regional sampling are warranted to further delineate these findings. Conclusions Through multi-region single-cell transcriptomic analysis of NSCLC samples, we identified a set of genes with low ITH in NSCLC and developed the LITHrisk score to predict the efficacy of immunotherapy in NSCLC patients. Further bioinformatic analyses and experiments demonstrated that the LITHrisk score is strongly correlated with anti-tumor immune responses and key malignant features of tumor cells. Collectively, the LITHrisk score not only provides a practical tool for advancing precision immunotherapy in NSCLC but also reveals promising targets for enhancing treatment efficacy. Supplementary Information Below is the link to the electronic supplementary material. Supplementary file1 (DOCX 2333 kb) Acknowledgements Not applicable. Abbreviations Copy number variation False discovery rate Gene expression omnibus Gene set enrichment analysis Immune checkpoint blockade Integrated heterogeneity score Intratumor heterogeneity Local inverse Simpson’s index Lung adenocarcinoma Lung squamous cell carcinoma Non-small cell lung cancer Overall survival Progression-free survival Single-cell RNA sequencing The cancer genome atlas Uniform manifold approximation and projection Authors contributions Jianye Yuan contributed to conceptualization, methodology, formal analysis, data curation, writing—original draft, visualization, and project administration. Zelin Weng was involved in methodology, validation, investigation, resources, and data curation and provided software. Zhenguo Li contributed to methodology, validation, formal analysis, and visualization. Rui Chen was involved in conceptualization, software, investigation, and data curation. Chao Cheng contributed to conceptualization and resources. Weixiong Yang was involved in visualization and formal analysis. Xiuying Xie contributed to methodology and investigation. Chang Luo was involved in investigation and data curation. Tao Wang contributed to investigation and formal analysis. Shuishen Zhang was involved in writing—review and editing, supervision, and project administration. Zihui Tan contributed to conceptualization, writing—review and editing, supervision, project administration, and funding acquisition. Funding This work was supported by the Medical Science and Technology Foundation of Guangdong Province (No. A2024053) and the Guangdong Esophageal Cancer Institute Science and Technology Program (No. Q201909). Data availability Public datasets analyzed in this study are available from ArrayExpress (E-MTAB-6149, E-MTAB-6653) for single-cell RNA-seq and BioStudies (E-MTAB-13530) for spatial transcriptomics. Bulk RNA-seq data are available from TCGA (LUAD, LUSC), GEO (GSE190265, GSE207422, GSE78220), the European Nucleotide Archive (PRJEB25780, PRJEB23709), and the IMvigor210 cohort. Declarations Conflict of interest The authors declare that they have no conflict of interest. Ethics approval Not applicable. Consent for publication Not applicable. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Jianye Yuan, Zelin Weng, Zhenguo Li, and Rui Chen have contributed equally to this work. Contributor Information Jianye Yuan, Email: yuanjy55@mail.sysu.edu.cn. Shuishen Zhang, Email: zhangshsh9@mail.sysu.edu.cn. Zihui Tan, Email: tanzh@sysucc.org.cn. References Associated Data Supplementary Materials Supplementary file1 (DOCX 2333 kb) Data Availability Statement Public datasets analyzed in this study are available from ArrayExpress (E-MTAB-6149, E-MTAB-6653) for single-cell RNA-seq and BioStudies (E-MTAB-13530) for spatial transcriptomics. Bulk RNA-seq data are available from TCGA (LUAD, LUSC), GEO (GSE190265, GSE207422, GSE78220), the European Nucleotide Archive (PRJEB25780, PRJEB23709), and the IMvigor210 cohort.