Transcriptomic data of LUAD were obtained from The Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov), including gene expression profiles and corresponding clinical information. An independent cohort from the International Cancer Genome Consortium (ICGC) database (https://dcc.icgc.org), specifically the ICGC-LUAD-US cohort, was used for external validation of the prognostic model. Cisplatin-resistant transcriptomic data were retrieved from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo), specifically dataset GSE213102 (42). Single-cell RNA sequencing (RNA-seq) data were obtained from the GEO dataset GSE229253 (43). All datasets were processed and normalized according to their respective preprocessing pipelines prior to downstream analyses.

Cisplatin resistance-associated genes were retrieved from the GeneCards database (https://www.genecards.org/). The keyword ‘cisplatin-resistant’ was used to perform a systematic search and genes with relevance scores provided by GeneCards were extracted. The resulting gene list was downloaded and curated to obtain a cisplatin resistance gene set for subsequent analyses.

Unsupervised clustering was performed using the ‘ConsensusClusterPlus’ R package (44) (version 1.68.0) based on the expression profiles of treatment group samples from the TCGA-LUAD cohort. The expression matrix was subjected to consensus clustering with the k-means algorithm and Euclidean distance. Clustering stability was evaluated across 50 resampling iterations, with 80% of the samples and all the features randomly selected in each iteration. The optimal number of clusters was determined by inspecting the cumulative distribution function curves and consensus heatmaps, supplemented by item-consensus and cluster-consensus scores derived from the integrated consensus likelihood analysis (44). The samples were subsequently assigned to their respective consensus clusters and the resulting subtype information was exported for downstream analyses.

To quantify pathway activity at the individual sample level, ssGSEA was performed using gene expression data from the TCGA-LUAD cohort. Gene expression matrices were preprocessed by converting expression values to numeric formats, removing duplicated genes through average expression consolidation and genes with low overall expression were filtered. Curated gene sets were imported from a gene matrix transposed (GMT) file using the GSEABase R package (version 1.70.0; DOI, 10.18129/B9.bioc.GSEABase). ssGSEA scores for each gene set were then computed using the ‘ssgsea’ method and Gaussian kernel estimation and the genes were ranked by absolute expression. To facilitate cross-sample comparison, raw ssGSEA enrichment scores were normalized using a min-max scaling transformation. The resulting normalized enrichment matrix was used for subsequent statistical analyses.

GSEA was performed using gene expression data from the TCGA-LUAD cohort to identify pathway differences between the two molecular subgroups. The analysis was conducted using GSEA software (https://www.gsea-msigdb.org/gsea/index.jsp) with the Kyoto Encyclopedia of Genes and Genomes (KEGG) and REACTOME gene sets downloaded from the Molecular Signatures Database (MSigDB; http://www.gsea-msigdb.org/gsea/msigdb/). Genes were ranked according to the signal-to-noise ratio between the two subgroups and enrichment was assessed using 1,000 permutations. Pathways with a nominal P<0.05 were considered to be significantly enriched.

Immune cell infiltration was quantified using the CIBERSORT algorithm (45) based on bulk RNA sequencing data from the TCGA-LUAD cohort, which applied ν-support vector regression (ν-SVR) to deconvolute the relative proportions of 22 immune cell subsets. The LM22 leukocyte signature matrix was used as the reference gene expression profile and the mixed expression matrix was quantile-normalized prior to deconvolution. To improve robustness, CIBERSORT fitted multiple ν-SVR models with different ν parameters and selected the model with the minimal root-mean-square error. For each sample, CIBERSORT returned the estimated immune cell fractions along with correlation coefficients and root-mean-square error values reflecting deconvolution accuracy. Permutation testing was performed when needed to compute empirical P-values and samples with P<0.05 were considered to have reliable immune infiltration estimates.

Tumor mutation burden analysis was performed using somatic mutation data downloaded from TCGA-LUAD cohort. Masked somatic mutation (MAF) files were retrieved from the National Cancer Institute Genomic Data Commons database (https://portal.gdc.cancer.gov/) and merged to generate a unified mutation profile. The MAF data were processed and annotated using the ‘maftools’ R package (version 2.18.0) (46) and TMB was calculated as the total number of non-synonymous somatic mutations per megabase for each sample. To compare mutational landscapes between molecular subgroups, sample identifiers were matched to cluster assignments and subgroup-specific MAF objects (for example, C1 and C2) were generated. Mutation profiles were visualized using the oncoplots function implemented in maftools, which displays the most frequently mutated genes across molecular subgroups.

Differential expression analysis was performed on TCGA-LUAD transcriptomic data using the ‘limma’ R package (version 3.58.1) (47). Expression matrices were log₂-transformed prior to analysis. A design matrix was constructed based on control and treatment groups, and linear models were fitted using the lmFit function implemented in limma, followed by empirical Bayes moderation. Differentially expressed genes were identified using a contrast comparison between the C1 and C2 groups. In parallel, the GSE213102 dataset, which includes cisplatin-resistant and cisplatin-sensitive cell lines, was analyzed using the same ‘limma’-based differential expression pipeline to identify resistance-associated genes.

WGCNA was performed using the ‘WGCNA’ R package (version 1.73) (48) based on gene expression data from the TCGA-LUAD cohort. After low-variance genes were removed, the top 25% of the most variable genes were selected for network construction. The sample quality was assessed by hierarchical clustering and outlier samples were excluded. Soft-thresholding power was determined using the scale-free topology criterion (48) and an adjacency matrix was constructed and transformed into a topological overlap matrix (TOM). Genes were clustered on the basis of TOM dissimilarity and coexpression modules were identified using dynamic tree cutting. Module eigengenes were calculated and correlated with cluster membership (C1 and C2 subgroups) to identify trait-associated modules. Gene significance and module membership were computed to screen for hub genes within key modules.

Gene Ontology (GO) (49) and KEGG (50) enrichment analyses were performed using the ‘clusterProfiler’ R package (version 4.10.1) (51). Gene sets were first converted from gene symbols to Entrez IDs using the org.Hs.eg.db annotation database (https://bioconductor.org/packages/org.Hs.eg.db/) and genes without valid Entrez IDs were removed. GO enrichment analysis [biological processes (BPs), cellular components and molecular functions] and KEGG pathway enrichment analysis were subsequently performed, with the organism set to ‘hsa’. Enriched terms were filtered using P<0.05 as the significance threshold. The results were visualized using the ‘enrichplot’ (version 1.22.0) (https://bioconductor.org/packages/enrichplot/) and ‘ggplot2’ R packages (version 3.5.0) (https://cran.r-project.org/package=ggplot2).

Gene set were imported into the Search Tool for the Retrieval of Interacting Genes/Proteins (STRING) database (https://string-db.org/) to construct a protein-protein interaction network using a confidence score threshold of 0.4. The resulting interaction network was downloaded and further visualized in Cytoscape (https://cytoscape.org/). Topological parameters were calculated and key hub genes were identified using the built-in ‘NetworkAnalyzer’ tool or the ‘CytoHubba’ plugin.

To construct a robust prognostic signature, an integrative machine learning framework encompassing 101 different algorithmic combinations was employed. Gene expression data from the TCGA served as the training set, while ICGC data were used for independent validation. Candidate genes were preselected based on univariate Cox regression analysis (P<0.05). The machine learning framework incorporated 10 algorithms, including Random Survival Forest (RSF), Elastic Net (Enet), Stepwise Cox regression (StepCox), CoxBoost, partial least squares regression for Cox (plsRcox), supervised principal components (superpc), gradient boosting machine (GBM), survival support vector machine (survival-SVM), Ridge regression and Lasso regression, as well as their specified pairwise combinations, resulting in 101 candidate models. Model development and selection were performed using ML.Dev.Prog. The Sig pipeline in R (with the following packages: ‘GSEABase’ (version 1.70.0; DOI, 10.18129/B9.bioc.GSEABase), Gene Set Variation Analysis (‘GSVA’; version 1.50.5) (52), ‘cancerclass’ (version 1.46.0) (53), ‘mixOmics’ (version 6.26.0) (54), ‘sparrow’ (version 1.8.5; DOI, 10.18129/B9.bioc.sparrow), ‘sva’ (version 3.50.0) (55), ‘ComplexHeatmap’ (version 2.18.0) (56), ‘CoxBoost’ (version 1.5) (57), ‘fastAdaboost’ [version 1.0.0; http://github.com/souravc83/fastAdaboost) and ‘Mime1’ (version 0.0.0.9) (58)] was used. Model performance was evaluated by the concordance index (C-index) in both the training and validation cohorts. Visualization of C-index distributions across all the models was performed using the ‘cindex_dis_all’ function, with the top-performing models further selected for downstream analyses.

Single-cell RNA-seq data from the GSE229253 dataset were processed using the ‘Seurat’ R package (version 5.3) (59). Quality control was performed by calculating the mitochondrial gene ratio (percent.mt) and hemoglobin gene ratio (percent.HB), followed by filtering cells with nFeature_RNA >200, nFeature_RNA <7,500, nCount_RNA >1,000 and percent.mt <20%. After normalization, highly variable genes were identified and scaling and dimensionality reduction were conducted using principal component analysis. Batch effects were corrected using the Harmony algorithm. Clustering was carried out with ‘FindNeighbors’ and ‘FindClusters’ and the results were visualized using Uniform Manifold Approximation and Projection. Cluster-specific marker genes were identified using ‘FindAllMarkers’ with the Wilcoxon test. Cell-type annotation was performed with ‘SingleR’ (version 2.4.1) (60) using the Human Primary Cell Atlas as the reference and further validated by cross-referencing significant marker genes with the CellMarker database (http://bio-bigdata.hrbmu.edu.cn/CellMarker/).

Cell-cell communication was analyzed using the ‘CellChat’ R package (version 1.6.1) (61). Normalized count matrices and cell-type annotations were used to construct the CellChat object and the human ligand-receptor database was used to identify overexpressed signaling genes and ligand-receptor pairs. The communication probabilities and pathway-level interactions were computed, filtered and visualized to characterize the intercellular signaling networks.

To evaluate the activity of a custom cisplatin resistance-associated gene set at the single-cell level, tumor epithelial cells were extracted from the GSE229253 single-cell RNA-sequencing dataset and analyzed using the immune-related GSEA (‘irGSEA’) R package (version 3.3.2; http://github.com/chuiqin/irGSEA). The custom gene set was generated from cisplatin resistance-associated genes identified in the present study and converted into GMT format for enrichment analysis. The expression matrix was normalized; highly variable genes were identified and data were scaled using the ‘Seurat’ R package. Gene set activity in individual cells was calculated using six scoring methods implemented in ‘irGSEA’, including AUCell, UCell, Variance-adjusted Mahalanobis, Single-Cell Signature Explorer, ssGSEA and singscore. The resulting enrichment scores are visualized as violin plots to compare the activity of the custom cisplatin resistance-associated gene set across different tumor epithelial cell subpopulations.

Single-cell CNVs were inferred using the ‘inferCNV’ R package (version 1.18.1) (62). Raw unique molecular identifier count matrices were extracted from the Seurat object and gene genomic coordinates were annotated and ordered using the AnnoProbe R package (version 0.1.7; http://github.com/shengqh/AnnoProbe), excluding genes on chromosome (chr)-M, chrX and chrY. Normal cells were used as the reference group. The inferCNV object was constructed from the expression matrix, cell annotations and gene order file and CNV inference was performed with a cut-off of 0.1, denoising enabled and hierarchical clustering using the ward.D2 method. The resulting CNV profiles were used to identify malignant epithelial cell populations.

Human LUAD cell line A549 and NSCLC cell line H1299 were obtained from the American Type Culture Collection. A549 cells are derived from human LUAD epithelial cells, while H1299 cells are derived from human non-small cell lung carcinoma lacking p53 expression. A549 cells were cultured in F-12K medium (cat. no. PM150910; Procell Life Science & Technology Co., Ltd.) and H1299 cells were maintained in RPMI-1640 medium (cat. no. PM150110; Procell Life Science & Technology Co., Ltd.), both supplemented with 10% FBS (cat. no. FSD500; Shanghai ExCell Biology, Inc.) and 1% penicillin-streptomycin. All cells were cultured at 37°C in a humidified incubator with 5% CO₂.

Gene knockdown was performed using short hairpin (sh)-RNA expression vectors targeting human CCT2. A total of two shRNA sequences targeting human CCT2 were newly designed and cloned into the pGPU6-GFP-Neo vector backbone by Shanghai GenePharma Co., Ltd. A non-targeting shRNA vector was used as the negative control. Cells were seeded in 6-well plates at 50–60% confluence and transfected with 2.5 µg shRNA plasmid per well using Hieff Trans™ Booster DNA/RNA Transfection Reagent (Lipo3000 alternative; cat. no. 40801ES04; Shanghai Yeasen Biotechnology Co., Ltd.) according to the manufacturer's instructions. After 48–72 h, the knockdown efficiency was assessed by reverse transcription-quantitative PCR (RT-qPCR) and western blotting. For RT-qPCR, total RNA was extracted using TRNzol Universal Reagent (cat. no. DP424; Tiangen Biotech Co., Ltd.), reverse transcribed according to the manufacturer's instructions with a Hifair^® III 1st Strand cDNA Synthesis SuperMix for qPCR (gDNA digester plus; cat. no. 11141ES10; Shanghai Yeasen Biotechnology Co., Ltd.) and amplified using Hieff UNICON™ Universal Blue qPCR Master Mix (SYBR; cat. no. 11184ES03; Shanghai Yeasen Biotechnology Co., Ltd.). qPCR amplification was performed on a QuantStudio™ 3 Real-Time PCR System with an initial denaturation at 95°C for 30 sec, followed by 40 cycles of 95°C for 3 sec and 60°C for 20 sec. Relative gene expression was normalized to GAPDH and calculated using the 2^−ΔΔCq method (63). All shRNA and primer sequences are listed in the Table SI.

Total protein was extracted from A549 and H1299 cells using RIPA lysis buffer (Yeasen Biotechnology Co., Ltd.; cat. no. 20115ES60) and quantified using a BCA assay. Equal amounts of protein (20 µg per lane) were separated by 12% SDS-PAGE and transferred onto PVDF membranes. After blocking with 5% skim milk at room temperature for 2 h, membranes were incubated overnight at 4°C with primary antibodies against CCT2 (cat. no. HA722240; 1:1,000; HUABIO) and GAPDH (cat. no. ET1601-4; 1:50,000; HUABIO), diluted in 5% BSA. After washing, membranes were incubated with HRP-conjugated secondary antibodies (cat. no. HA1001; 1:50,000; HUABIO) at room temperature for 1 h. Protein bands were visualized using an enhanced chemiluminescence system (Enhanced Chemiluminescence Detection Kit; Yeasen Biotechnology Co., Ltd.; cat. no. 36208ES) and quantified using ImageJ software (version 1.8.0; National Institutes of Health).

Cell proliferation was evaluated using a Cell Counting Kit-8 (CCK-8; cat. no. AC11L054; Heyuan Liji (Shanghai) Biotechnology Co., Ltd.) assay. Cells were seeded in 96-well plates at 2×10³ cells per well and incubated at 37°C for the indicated durations. At each time point, 10 µl CCK-8 solution was added to each well and incubated for 2 h. The absorbance at 450 nm was measured using a microplate reader. All experiments were performed in triplicate.

A549 and H1299 cells were first transfected with shRNA expression vectors targeting CCT2 or control shRNA vectors. After transfection, cells were treated with cisplatin (cat. no. P4394; Sigma-Aldrich; Merck KGaA) at final concentrations of 20 µM for A549 cells and 15 µM for H1299 cells at 37°C for 48 h. Cells were then incubated for 48 h at 37°C in a humidified atmosphere with 5% CO₂. A total of four experimental groups were included: i) Control; ii) sh-CCT2; iii) cisplatin alone and; iv) cisplatin combined with sh-CCT2. After treatment, cell viability was assessed using the CCK-8 assay according to the manufacturer's instructions.

Molecular docking of CCT2 was performed using the Molecular Operating Environment (MOE; http://www.chemcomp.com). The crystal structure of human CCT2 was prepared by removing water molecules, adding hydrogens, optimizing protonation states and minimizing the protein with the AMBER10:EHT force field. Natural-product ligands were energy-minimized and docked into the predicted binding pocket using the Triangle Matcher placement method (64) and London dG scoring, followed by refinement with Generalized Born Volume Integral/Weighted Surface Area (GBVI/WSA) dG scoring function to obtain the optimal binding poses. The top docking complexes were subsequently subjected to molecular dynamics simulations using GROMACS (https://www.gromacs.org). Protein topologies were generated with the AMBER ff14SB force field and ligand parameters with General Amber Force Field (GAFF). Each complex was placed in a Transferable Intermolecular Potential 3-Point (TIP3P) water box with 0.15 M sodium chloride, energy-minimized, equilibrated under NVT and NPT conditions and simulated for 100 ns at 310 K and 1 bar. Trajectories were analyzed for RMSD, radius of gyration and solvent accessible surface area to evaluate the dynamic behavior and binding stability of each ligand-CCT2 complex.

Statistical analyses were performed using GraphPad Prism (version 10.0; Dotmatics). Data are presented as the mean ± SD from at least three independent biological replicates. Comparisons among multiple groups were performed using one-way ANOVA tests followed by Dunnett's or Tukey's multiple comparisons test, as appropriate. For experiments involving two independent variables, statistical significance was determined using two-way ANOVA tests followed by Šidák's multiple comparisons test. P<0.05 was considered to indicate a statistically significant difference.

First, the transcriptome data of LUAD from the TCGA-LUAD dataset was downloaded, which included 541 tumor samples and 59 adjacent normal lung tissue samples. Differential expression analysis [log₂ fold change (FC) >1; P<0.05] revealed 3,234 differentially expressed genes (DEGs), among which 1,229 were upregulated and 2,005 were downregulated in tumor tissues (Fig. 1A). To further identify cisplatin resistance-associated genes, the GSE213102 dataset was retrieved from the GEO database, which contains the transcriptomic profiles of cisplatin-resistant cell lines (A549, H1299, 3B1A and H1573). GEO2R analysis with the same thresholds yielded 1,976 shared DEGs across all the resistant cell lines (Fig. 1B). Univariate Cox regression was then performed to screen prognostic genes among these candidates and obtained cisplatin resistance-associated genes from the GeneCards database. The intersection of upregulated genes, prognostic risk genes, GeneCards-derived cisplatin resistance genes and cisplatin resistance-associated DEGs in LUAD resulted in the identification of a final set of nine cisplatin resistance-associated genes (Fig. 1C). Gene correlation analysis using GeneMANIA (https://genemania.org/) revealed extensive coexpression, physical interactions and colocalization among these genes, which were enriched primarily in BPs such as ‘cell cycle G₂/M phase transition’, ‘chromosome separation’ and ‘regulation of nuclear division’ (Fig. 1D). Unsupervised consensus clustering based on these nine genes revealed two robust molecular subtypes, designated C1 and C2 (Fig. 1E). ssGSEA demonstrated that the cisplatin resistance score was significantly greater in the C1 subtype compared with the C2 subtype (Fig. 1F). Survival analysis further revealed that patients in the C1 cluster exhibited significantly worse overall survival compared with those in the C2 cluster (log-rank test, P<0.001; Fig. 1G).

To elucidate the differences in gene expression between the two subtypes, TMB analysis was first performed. The highly mutated genes included CUB and sushi multiple domains 3 (CSMD3), erythrocytic spectrin α 1 and xin actin binding repeat containing 2 (Fig. 2A). Notably, compared with the C2 subtype, the C1 subtype exhibited significantly higher mutation frequencies for TP53, titin and CSMD3 (Fig. 2B). The present study subsequently investigated differences in pathway activity between subtypes using GSEA. KEGG-based GSEA revealed that ‘KEGG cell cycle’, ‘KEGG p53 signaling pathway’ and ‘KEGG mismatch repair’ were markedly enriched in the C1 subtype (Fig. 2C). Consistent with these findings, a reactome GSEA revealed significant activation of ‘REACTOME DNA repair’, ‘REACTOME transcriptional regulation by TP53’ and ‘REACTOME Mitotic G₂/M phases’ in the C1 subtype (Fig. 2D). These findings suggested that enhanced genomic instability and checkpoint activation were key molecular hallmarks of the high-resistance C1 cluster.

Immunological heterogeneity between the two cisplatin resistance subtypes was first characterized by comparing the expression levels of immune checkpoint-associated genes. The majority of immune checkpoint-associated genes, including glutamate-cysteine ligase catalytic subunit, twist family BHLH transcription factor 1 and phorbol-12-myristate-13-acetate-induced protein 1, were significantly upregulated in the C1 subtype relative to the C2 subtype (Fig. 3A). Immune cell infiltration was then assessed using the CIBERSORT algorithm. A number of immune cell populations, such as naïve B cells, plasma cells and resting CD4 memory T cells, were found to be more abundant in the C2 subtype, whereas CD8 T cells were significantly more abundant in the C1 subtype (Fig. 3B). Numerous computational algorithms were subsequently applied to evaluate the correlation between immune infiltration patterns and cisplatin resistance scores. The resistance score was found to be positively correlated with that of resting natural killer cells and follicular helper T cells, indicating an association between the immunological activation state and the cisplatin resistance phenotype (Fig. 3C).

To investigate the molecular disparities between the two subtypes, differential expression analysis was performed (log₂ FC >0.5; P<0.05) to identify 2,375 DEGs, including 1,224 upregulated genes and 1,151 downregulated genes, in the C2 relative to C1 subtype (Fig. 4A). WGCNA was then conducted, with a soft-thresholding power of 3 (Fig. 4B), resulting in a number of coexpression modules (Fig. 4C). Module-trait correlation analysis revealed numerous clinically relevant modules (Fig. 4D), which were selected for downstream analyses. Given the key role of protein palmitoylation in chemoresistance, palmitoylation-associated genes were incorporated into the analysis. An intersection of DEGs, WGCNA-derived module genes, prognostic risk genes and palmitoylation-associated genes yielded 28 shared genes (Fig. 4E). To elucidate the biological functions associated with these genes, GO and KEGG enrichment analyses were performed. GO enrichment revealed involvement in ‘de novo protein folding’, ‘regulation of protein localization to Cajal body’, ‘chaperonin-containing T-complex’ assembly, ‘microtubule’ and ‘protein folding chaperone’ activity (Fig. 4F). KEGG analysis revealed significant enrichment of the ‘HIF-1 signaling pathway’ and ‘glycolysis/gluconeogenesis’ (Fig. 4G). PPI analysis using the STRING database revealed GAPDH, CCT2 and CCT3 as central hub genes within the network (Fig. 4H). Finally, CMAP (https://clue.io/cmap) analysis identified several compounds with negative or low connectivity scores, including ropivacaine, epoxycholesterol, naltrindole and beclometasone in A549 cells; SR-59230A, dihydrosamidin, talniflumate, GW-5074, DMBI, cosmosiin and isoeugenol in HCC515 cells; and BRD-K40269473 in NCI-H2110 cells (Fig. 4I).

A total of 101 machine learning algorithms were integrated to refine the expression of cisplatin resistance-associated core genes. The TCGA cohort was used for model training and the ICGC cohort was used for external validation. The performance heatmaps of all the algorithmic models are presented in Fig. 5A. Among all combinations, the StepCox (forward) + CoxBoost model demonstrated the highest predictive accuracy. Survival curves generated from both the training and validation datasets, on the basis of the model-derived risk scores, consistently revealed significant prognostic stratification (Fig. 5B). This optimized model identified five key genes, CCT2, CCT3, CCT5, CCT6A and centromere protein K (CENPK), as the most robust predictors. All five genes were markedly upregulated in LUAD tissues compared with normal lung tissues and patients with high expression of these genes exhibited significantly poorer overall survival compared with those with low expression levels (Fig. 5C). These findings suggested that the five-gene signature served a key role in mediating cisplatin resistance and may serve as a potent biomarker panel for risk assessment. Notably, numerous members of the CCT family (CCT2, CCT3, CCT5 and CCT6A) were consistently identified, suggesting that the CCT family may serve a broader and coordinated role in cisplatin resistance. Among these, CCT2 exhibited the most significant prognostic association, as it exhibited the smallest log-rank P-value in Kaplan-Meier survival analysis (Fig. 5C) and patients with high CCT2 expression demonstrated significantly worse overall survival, which led it to be prioritized for further investigation.

Single-cell RNA-seq data from lung cancer samples retrieved from the GEO database (GSE229253) were used to explore the cellular origin and functional heterogeneity of the core genes at single-cell resolution. Dimensional reduction and clustering revealed key cell populations, including epithelial cells, macrophages, endothelial cells, T cells and B cells (Fig. 6A), with representative marker genes for each lineage displayed in Fig. 6B. Cell-cell communication networks inferred using CellChat revealed extensive intercellular interactions across these cell types (Fig. 6C). Visualization of core gene expression across cell populations demonstrated that CCT2, CCT3, CCT5 and CCT6A were predominantly enriched in epithelial cells (Fig. 6D). Further reclustering of epithelial cells revealed six distinct epithelial cell subpopulations, each characterized by unique marker gene signatures (Fig. 6F). According to the results of the IRGSEA for resistance gene-set enrichment, subclusters 0 and 5 exhibited strong enrichment for cisplatin resistance signatures across numerous algorithms (Fig. 6G). Finally, mapping the five core genes onto epithelial subclusters demonstrated that CCT2, CCT3, CCT5 and CCT6A were significantly enriched in both cisplatin-resistant subclusters 0 and 5, whereas CENPK expression was predominantly enriched in subcluster 5 (Fig. 6H). These findings highlighted specific epithelial cell subpopulations as the principal cellular reservoirs of cisplatin resistance-associated gene activation.

GSVA across six epithelial subclusters revealed distinct biological functions of CCT2, with subcluster 0 exhibiting significant enrichment of genes associated with oncogenic and proliferative pathways, including genes associated with ‘KRAS signaling up’, ‘Wnt beta-catenin signaling’ and ‘cell cycle’ (Figs. 7A and S1). RNA velocity analysis further demonstrated a directional flow from subcluster 0 toward other epithelial subclusters, suggesting that subcluster 0 may represent a transcriptionally active and developmental origin state (Fig. 7B). On the basis of the IRGSEA-derived drug resistance signatures, the epithelial cells were stratified into high- and low-resistance groups. Cell cycle phase scoring indicated that, compared with the low-resistance group, the high-resistance group exhibited a markedly higher proportion of proliferating cells (S and G₂/M phases; Fig. 7C). Cell-cell communication analysis revealed that the high-resistance epithelial cell population exhibited stronger interactions with endothelial cells and mast cells compared with the low-resistance population, as indicated by the increased interaction strength in the communication network (Fig. 7D). Consistently, ligand-receptor analysis identified a greater number of significant interactions between these cell populations, including representative ligand-receptor pairs such as semaphorin-3C (SEMA3C)-[neuropilin (NRP)-2 + plexin-A2 (PLXNA2)], SEMA3C-(NRP1 + PLXNA2), SEMA3C-(NRP1 + NRP2), cell adhesion molecule 1 (CADM1)-CADM1 and amphiregulin (AREG)-EGFR, among others (Fig. S2). CNV profiling further demonstrated increased chromosomal instability in the resistant epithelial population (Fig. 7E). To clarify the contribution of CCT2 expression, epithelial cells were separately analyzed according to their CCT2 expression level. High-CCT2 epithelial cells exhibited stronger interactions with endothelial cells and mast cells compared with low-CCT2 epithelial cells, as indicated by increased interaction strength in the communication network (Fig. 7F). Consistently, ligand-receptor analysis revealed a greater number of significant interactions between high-CCT2 epithelial cells and these cell populations, including representative ligand-receptor pairs such as SEMA3C-(NRP1 + PLXNA2), SEMA3C-(NRP1 + NRP2), secretoglobin family 3A member 2-macrophage receptor with collagenous structure, CADM1-CADM1 and AREG-EGFR, among others (Fig. S3). Consistent with the resistance-associated cell population, CCT2-high epithelial cells also exhibited a higher proportion of proliferating cells (S and G₂/M phases; Fig. 7G). In addition, CNV analysis showed greater genomic alterations in the high-CCT2 subgroup (Fig. 7H), collectively indicating that CCT2 upregulation was associated with proliferative activity, enhanced cellular communication and genomic instability within epithelial tumor cells.

Functional roles of CCT2 were next validated in the cell lines A549 and NCI-H1299, respectively. A total of two shRNA expression constructs targeting CCT2 were transfected into both cell lines and qPCR determined the efficient knockdown by sh-CCT2-1 and sh-CCT2-2 (Fig. 8A; n=3; P<0.05). Western blotting analysis demonstrated that CCT2-targeting shRNA vectors effectively reduced CCT2 protein expression in A549 and NCI-H1299 cells (Fig. 8B; n=3; P<0.05). Cell proliferation assessed by CCK-8 assays at 24 and 48 h revealed that CCT2 depletion significantly reduced the proliferative capacity of both A549 and NCI-H1299 cells (Fig. 8C; n=5; P<0.05). To further evaluate whether CCT2 influences the cellular response to cisplatin, cells were treated with cisplatin following CCT2 knockdown. Cisplatin alone significantly suppressed cell viability, whereas combined treatment with cisplatin and CCT2 silencing resulted in a notably greater reduction in cell viability (Fig. 8D; n=5; P<0.05). Compared with cisplatin treatment alone, CCT2 knockdown further enhanced the inhibitory effect of cisplatin on cell viability, indicating that depletion of CCT2 increased the sensitivity of lung cancer cells to cisplatin. These findings indicated that CCT2 promoted lung cancer cell proliferation and modulated the sensitivity of lung cancer cells to cisplatin.

Molecular docking was performed using a natural-product compound library within the MOE platform to identify potential inhibitors of CCT2. A number of compounds, including deslanoside (−11.9 kcal/mol; Fig. 9A), digitoxin (−10.7 kcal/mol; Fig. 9B), lanatoside C (−11.8 kcal/mol; Fig. 9C), lecithin (−11.7 kcal/mol; Fig. 9D) and thymopentin (−10.9 kcal/mol; Fig. 9E), demonstrated strong binding affinities. Subsequent molecular dynamics simulations revealed stable RMSD, radius of gyration and solvent-accessible surface area profiles for each ligand-protein complex, indicating stable binding conformations. These results suggested that the identified compounds may serve as promising candidates for targeting CCT2 in future therapeutic development.