Proteogenomic characterization of pulmonary large-cell neuroendocrine carcinoma identifies molecular features and therapeutic strategy Authors: Lele Zhang, Liangdong Sun, Shengnan Duan, Yilv Yan, Jinyi Song, Linghui Xia, Huansha Yu, Jijun Sun, Junjie Hu, Di Wang, Lu Han, Jue Wang, Yan Chen, Hu Zhou, Chen Wang, Haiyang Hu, Ming Ding, Peng Zhang Published: 2026 ABSTRACT Pulmonary large-cell neuroendocrine carcinoma (LCNEC) is a rare yet highly aggressive subtype of non-small cell lung carcinoma (NSCLC) with limited therapeutic options. We conduct a comprehensive proteogenomic analysis of LCNEC using tumors and paired normal adjacent tissues from 107 patients (81 pure LCNEC and 26 combined LCNEC). APOBEC mutational signatures strongly correlate processes of tumor initiation and immune suppression, and KEAP1 mutations correlate with metabolic reprogramming in LCNEC combined with NSCLC. A conflicting relationship is observed between neuroendocrine and immune phenotypes. We identify three LCNEC subtypes with unique prognosis features, microenvironment dysregulation, genetic alterations, and potential therapeutic targets. Interleukin-33 (IL-33) emerges as a critical therapeutic biomarker associated with enhanced T cell infiltration and antitumor activity. We further optimize recombinant IL-33 (rIL33) with site-directed mutagenesis and develop PEGylated rIL33, demonstrating its prolonged circulation time in cynomolgus monkeys and superior immune agonist activity in mouse models. Overall, this study offers insights into LCNEC biology and provides promising innovative immunoagonist therapy strategies for lung cancer. INTRODUCTION Lung cancer remains the leading cause of cancer-related death globally (1). Pulmonary large-cell neuroendocrine carcinoma (LCNEC) represents a rare yet highly aggressive lung tumor, accounting for 2 to 3% of all resected lung cancers (2, 3). Annual incident cases of LCNECs are estimated to range from 44,000 to 66,000, with its incidence rising from 0.26 per 100,000 individuals in 2004 to 0.39 per 100,000 individuals in 2015 (4, 5). LCNEC, along with typical carcinoid (TC), atypical carcinoid (AC), and small cell lung cancer (SCLC), falls under the category of neuroendocrine lung tumors (2, 3). Unlike TC and AC, LCNEC exhibits a clinically aggressive phenotype, characterized by widely extensive metastases and a poor prognosis, with 5-year survival rates ranging from 15 to 25% (2, 3, 6, 7). The main therapeutic approaches for LCNEC include surgery and chemotherapy; however, the efficacy of chemotherapy remains limited, and no established standard treatment regimen has been established for this tumor type (8). Besides pure LCNEC, LCNEC frequently coexists with non-small cell lung carcinoma (NSCLC) and occasionally with SCLC components in clinical practice, which are referred to as combined large-cell neuroendocrine carcinomas with NSCLC (C-LCNEC with NSCLC) and combined large-cell neuroendocrine carcinomas with SCLC (C-LCNEC with SCLC), respectively (9). These combined large-cell neuroendocrine carcinomas (C-LCNEC) constitute approximately 20 to 25% of pulmonary LCNEC cases (10). Despite this prevalence, there is limited knowledge regarding the molecular differences between pure LCNEC and C-LCNEC. The insufficient knowledge of the molecular underpinnings of LCNEC poses a major barrier to the development of targeted therapies for this distinctive and aggressive subtype of lung cancer. Recent advances in genomic and transcriptomic sequencing analyses have begun to elucidate the molecular landscape of LCNEC (9–12). Despite these initial insights, our comprehension of the molecular landscape and heterogeneity in LCNEC, particularly for the development of targeted therapies, remains incomplete. A major challenge in LCNEC research is the scarcity of high-quality tumor specimens, which hampers both clinical and basic research efforts, due to the rarity of this disease. Comprehensive “omics” strategies, encompassing proteomic and phosphoproteomic profiling in conjunction with genomic analysis, have identified previously unrecognized disease subtypes, signaling pathways, and potential therapeutic targets in NSCLC (13–16). Nonetheless, a systematic proteomic study of LCNEC has yet to be conducted, despite the pivotal role of proteins in determining tumor phenotype. In this study, we conducted an integrated proteogenomic characterization of 107 LCNEC tumors and paired normal adjacent tissues (NATs) by leveraging mass spectrometry (MS)–based quantitative proteomics and phosphoproteomics, along with genomic and transcriptomic sequencing. This comprehensive proteogenomic analysis provided critical insights into genomic aberrations, LCNEC-associated molecular characteristics, and therapeutic biomarkers, thereby improving our understanding of patient stratification for targeted treatment strategies. We also proposed that interleukin-33 (IL-33) and engineered polyethylene glycol–conjugated interleukin-33 (PEG–IL33) emerge as promising therapeutic strategies with immune-enhancing properties. Furthermore, the resulting dataset also serves as a valuable resource for ongoing and future researches in the field of LCNEC. RESULTS The proteogenomic landscape of LCNEC To characterize the proteogenomic landscape of LCNEC in China, we collected tumor and paired NAT samples from 107 patients with surgically resected LCNEC [Tongji University (TU)-LCNEC cohort] (fig. S1A). All samples were cryo-pulverized and aliquoted for molecular profiling using whole-exome sequencing (WES), RNA sequencing (RNA-seq), and isobaric tandem mass tag (TMT) labeling–based global proteomics and phosphoproteomics (Fig. 1, A and B). Clinical information, including gender, age at diagnosis, tumor stage, survival, histology, etc., is summarized in table S1 (Fig. 1C). Four molecular profiling strategies were used to characterize the TU-LCNEC cohort (Fig. 1A and fig. S1A). The liquid chromatography–tandem mass spectrometry (LC-MS/MS) identified 9580 proteins, with 5683 quantified in 103 sample pairs and 8142 detected in more than 50% of the 103 paired samples. TMT-based phosphoproteomics identified 30,009 highly reliable phosphorylation sites (phosphosites) from 7476 phosphoproteins, among which 8932 phosphosites from 3137 phosphoproteins were quantified in more than 50% of the 103 paired samples (table S1, Supplementary Materials). The TMT-based proteomics and phosphoproteomics data were of high quality as evaluated (fig. S1, B to E). Principal components analysis (PCA) showed no overall apparent batch effect in the global proteome and phosphoproteome data (fig. S1F). On the basis of transcriptomic and proteomic data from matched samples, 5393 mRNA-protein pairs were identified, with 88.5% exhibiting significant positive correlation [false discovery rate (FDR) < 0.05, fig. S1G]. The median correlation between mRNA and protein was 0.48 (fig. S1G), indicating a discrepancy in expression between mRNA and protein, which has also been observed in other cancer types (17, 18). Multi-omics landscape of LCNEC. (A) Workflow of LCNEC proteogenomic profiling [created in BioRender. Yan, Y. (2026) https://BioRender.com/gbvfvds]. (B) Summary of data and metadata generated in this study. (C) Stacked histograms indicating the distribution of patient clinical phenotypes. (D) Genetic profile and associated clinicopathologic features of all 107 patients with LCNEC. (E) Significantly mutated genes in TU-LCNEC. False discovery rate (FDR) was calculated by MutSigCV. (F) Co-occurrence and mutual exclusiveness of cancer-associated mutated genes (Fisher’s exact test). (G) Boxplot showing the TMB (tumor mutation burden) differences between PI3K pathway mutated and PI3K pathway wild patients (Wilcoxon ranked-sum test). (H) Kaplan-Meier curves comparing overall survival for high-TMB and low-TMB patients (Log-rank test). (I) Two LCNEC specific mutation signatures. The smoking signature and APOBEC signature are dominated by C>A mutations (blue) and C>T mutations (red), respectively. (J) Boxplot showing the APOBEC signature differences between KEAP1 mutated and KEAP1 wild patients (Wilcoxon ranked-sum test). (K) Boxplot showing the APOBEC signature differences between PIK3CA mutated and PIK3CA wild patients (Wilcoxon ranked-sum test). Using WES data with a mean depth of more than 100× (fig. S1G), we detected 29,439 mutations, including 27,885 single-nucleotide variants (SNVs) and 1554 small insertion-deletions (indels) (table S1, Supplementary Materials). The genomic landscape reaffirmed previously reported somatic alterations in LCNEC (10, 12) (Fig. 1D). We identified six genes exhibiting significant recurrent somatic mutation (q < 0.1), all of which were well-known cancer-related genes: TP53 (74%), RB1 (52%), KEAP1 (13%), MAP3K (9.9%), STK11 (8%), and POU4F1 (7%) (Fig. 1, D and E). Compared to WES or targeted sequencing results of previous LCNEC cohorts, we observed differences in the distribution of three significantly mutated genes (RB1, KEAP1, and STK11) across these cohorts (10, 12, 19, 20) (fig. S1I). The distribution differences might be due to unequal proportions of subgroups in different LCNEC cohorts, including LCNEC transcriptomic type I associated with STK11/KEAP1 mutations and LCNEC type II associated with RB1 mutations (12). By further evaluating the interactions among cancer-associated driver mutations (21), we found a significant comutation relationship between TP53 and RB1 in LCNEC (Fig. 1F). Genetic alterations in the PI3K/AKT/mTOR pathway have been identified as potential therapeutic targets (9, 22). We found that LCNECs with PI3K pathway mutations exhibited a higher tumor mutation burden (TMB) (Fig. 1G), which was also observed in the George et al. LCNEC cohort (12) (fig. S1J). In addition, patients with a higher TMB tended to have more prolonged overall survival (OS) (fig. 1H), similar to a previous observation (23). Through mutational signature analysis, we identified two dominant mutational signatures in the TU-LCNEC cohort, namely, S1 (APOBEC signature) and S2 (Smoking signature) (Fig. 1I). Further analysis revealed that KEAP1 and PIK3CA mutations were significantly associated with a high APOBEC signature (Fig. 1, J and K), consistent with previous reports in lung cancer (24, 25). Moreover, the identified APOBEC signature and Smoking signature, as well as a high APOBEC signature in KEAP1 mutation, were confirmed in the George et al. LCNEC cohort (12) (fig. S1, K and L). These results suggested that the APOBEC signature might contribute to the carcinogenicity of LCNEC. In addition, we found that tumors harboring a high APOBEC enrichment score showed increased immune scores (fig. S1M) and elevated expression of immune checkpoint proteins, such as IDO1 and LGALS9 (fig. S1, N and O), which are critical to tumor immune mechanisms and immune checkpoint therapy (26, 27), indicating a possible role of APOBEC mutational signature in the context of immune checkpoint therapy in LCNEC. Effects of copy number alterations in LCNEC We identified 25,884 somatic copy number alterations (CNAs) using WES and examined their regulatory effects on mRNA, protein, and phosphoprotein expression (Fig. 1B). CNAs affect mRNA, protein, and phosphoprotein abundance in either “cis” or “trans” modes, corresponding to the diagonal and vertical patterns, respectively (fig. S2A). The cis-regulatory effects of CNA were more prominent on mRNAs and proteins than on phosphoproteins, with 6873, 2114, and 628 significant cis-effects observed at mRNA, protein, and phosphoprotein levels, respectively (table S2). We obtained 347 significant cis-effects shared at all three omics levels (Fig. 2A), which were significantly correlated with positive regulation of nucleocytoplasmic transport, base excision repair, and ErbB signaling pathways (Fig. 2B). To further nominate critical genes within CNA regions, we focused on the 593 cancer-associated genes (CAGs) (28), of which 555 CAGs were identified in the TU-LCNEC cohort (Fig. 2C). We identified 25 CAGs associated with significant cis-effects across all three omics levels (Fig. 2C and table S2), which were enriched in the neurophin signaling pathway, EGFR tyrosine kinase inhibitor resistance, FoxO signaling pathway, and ErbB signaling pathway (Fig. 2D). Effects of copy number alterations on multi-omics data. (A) Venn diagram depicting the cascading effects of CNAs. It shows the overlap between significant cis events across the transcriptome, proteome, and phosphoproteome. (B) Heatmap of significant pathways based on significant cis-events observed at mRNA, protein, and phosphoprotein levels. (C) The Venn diagram shows the significant cis-events restricted to cancer-associated genes (CAGs) across multiple data types. (D) Enriched pathway for 25 cis CAG-related CNV events (mRNA-Protein-Phospho). (E) Genome-wide arm amplification and deletions. (F) Genome-wide focal amplification and deletions. Chromosomal locations of peaks of significantly recurring focal amplifications (red) and deletions (blue) were filtered by FDRs. (G) Arm-level alterations associated with prognosis (overall survival). Volcano plot showing log2-based hazard ratio for each alternative chromosome arm. (H) Focal-level alterations associated with prognosis (overall survival). Volcano plot showing log2-based hazard ratio for each alternative chromosome focal. (I) Volcano plot showing log2-based hazard ratio (overall survival) for genes on prognostic focal-level CNA. (J) Survival Kaplan-Meier curves of patients with LCNEC with high (red) or low (blue) LNPK protein abundance in the TU-LCNEC cohort (Log-rank test). (K) Survival Kaplan-Meier curves of patients with LCNEC with high (red) or low (blue) LNPK mRNA expression in the George et al. cohort (Log-rank test). (L) Volcano plot showing the correlation of KEGG pathway enrichment score with LNPK based on protein level. (M) Heatmap and survival analysis of proteins involved in the N-glycan biosynthesis pathway. The heatmap (left) displays protein abundances of LNPK and pathway members significantly positively correlated with LNPK. The forest plot (right) shows the hazard ratios (HRs) and 95% confidence intervals derived from univariate Cox regression analysis for overall survival. For (B), (D), (E), (I), and (L), P values were adjusted using the Benjamini-Hochberg correction. CNA analysis identified the most frequent gains in chromosomes 1p, 1q, 3q, 5p, 8q, and 12p, as well as losses in chromosomes 3p, 5q, 8p, 10q, 15q, 16p, 16q, 17p, 19p, 21p, 21q, and 22p in the TU-LCNEC cohort (Fig. 2E and table S2). In addition, we identified amplifications in driver oncogenes, including MYCL (1p34.3, 22 cases), FGFR1 (8p11.23, 27 cases), MYC (8q24.21, 45 cases), and CEBPA (19q13.1, 16 cases) and deletions in key tumor suppressors, such as NSD1 (5q35.3, 45 cases) and FSTL3 (19p13.3, 49 cases) (Fig. 2F and table S2). Survival analysis of arm-level and focal arm-level somatic CNA events revealed that amplifications of 21q11.2 and 2q31.1, as well as deletions of 10q, 16q, 10q26.2, and 14q32.13, were significantly associated with LCNEC prognosis (Fig. 2, G and H). The 2q31.1 amplification was observed in 23 out of 107 (21.5%) patients with LCNEC (fig. S2B). LNPK, an endoplasmic reticulum junction-forming protein located at 2p31.1, is essential for stabilizing the ER network (29, 30). We identified LNPK as the only focal amplification that remained significantly associated with poor prognosis after multiple-testing correction (Fig. 2I). Survival analysis showed that the high protein abundance of LNPK was significantly associated with poor OS (Fig. 2J), which was also observed based on transcriptomic data of the George et al. cohort (12) (Fig. 2K). Pathway enrichment analysis of proteins significantly positively correlated with LNPK expression showed strong enrichments for N-glycan biosynthesis and GPI-anchor biosynthesis pathways (Fig. 2L), indicating that elevated LNPK abundance was accompanied by increased activity of these biosynthetic processes. Key enzymes involved in the N-glycan biosynthesis pathway, such as MGAT5 and MAN1A2, were positively correlated with LNPK and were independently identified as significant risk factors for poor survival (Fig. 2M). Proteomic alterations associated with tumorigenesis and prognosis Proteomic profiles of both LCNEC tumors and paired NATs provided us with a unique opportunity to explore the molecular events of tumor progression. PCA demonstrated that PC1 discriminated tumors and NATs and explained a much higher expression variance between tumors and NATs at the protein and phosphoprotein levels than at the mRNA level (fig. S3A). We identified 974 differentially expressed proteins (Wilcoxon rank-sum test, FDR < 0.05, |log2 fold change| > 1), with 246 proteins up-regulated and 728 proteins down-regulated in LCNEC tumors compared with NATs (fig. S3B and table S3). Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis revealed that DNA replication, cell cycle, and multiple DNA repair–related pathways were up-regulated in tumors, while complement and coagulation cascades, ECM-receptor interaction, and focal adhesion–related pathways were down-regulated (fig. S3C). After applying more stringent criteria for differential expression and adjusting for stromal and immune content, we further singled out 40 up-regulated proteins in tumors, which we termed LCNEC-associated proteins (fig. S3D). The potential clinical utility of these proteins was annotated using the Human Protein Atlas, revealing 21 plasma proteins, 22 enzymes, 13 known cancer-related genes, and 10 approved drug targets among them (fig. S3D). PLIN2 was the only LCNEC-associated protein with opposite mRNA and protein differential expression (fig. S3E), which is a ubiquitously expressed protein that regulates lipid droplets, particularly in nonadipose tissues such as the liver (31). Two LCNEC-associated proteins, RAB3B and MTHFD2 (32), were associated with poor survival (fig. S3, F and G). We obtained 1892 differentially expressed phosphosites between LCNECs tumor and NAT (Wilcoxon rank-sum test, FDR < 0.05, |LFC| > 1; Fig. 3H and table S3). In general, the differential expression pattern was largely concordant between phosphosites and cognate proteins (fig. S3I and table S3; Pearson correlation, R = 0.77, P < 0.001). Notably, 573 phosphosites increased much stronger (>2-fold) than corresponding proteins in tumors (fig. S3I). For instance, PDHA1, whose dysregulation promotes tumor glycolysis by down-regulating the pyruvate dehydrogenase complex activity, exhibited markedly greater increases in phosphorylation than in protein abundance (fig. S4A) (33). Specifically, three phosphorylation sites on PDHA1 (Ser232, Ser293, and Ser300) showed increases exceeding twofold, whereas the change in protein abundance of PDHA1 was comparatively modest (fig. S4A). Quantitative analysis of immunohistochemical (IHC) staining of PDHA1 and phosphor-PDHA1 (Ser293) further confirmed a significant elevation of phosphorylation of PDHA1 at Ser293 in tumors relative to paired NATs (fig. S4, B and C). Although phosphosites and proteins exhibited comparable prognostic values (fig. S4D and table S3), specific phosphorylation sites exhibited more pronounced prognostic values than their corresponding proteins. For example, LCNECs with high phosphorylation at Ser293 on PDHA1 instead of PDHA1 protein abundance showed significantly worse OS (fig. S4, E and F). Similar observations were identified for the phosphorylation at Ser353 on INSM1 (fig. S4, G and H) and additional 445 phosphosites (table S3). Comparative proteogenomic characteristics of C-LCNEC with NSCLC, C-LCNEC with SCLC, and pure LCNEC. (A) Heatmap showing the signature proteins and phosphosites for patients with C-LCNEC with NSCLC, C-LCNEC with SCLC, and pure LCNEC (C-LCNEC, combined LCNEC). Representative pathways were annotated on the right. (B) Significant pathways enriched in C-LCNEC with NSCLC compared with pure LCNEC via gene set enrichment analysis (C-LCNEC). (C) Boxplot showing the comparison of protein differences for pentose phosphate pathway key enzymes among three histological subtypes (one-way ANOVA test). (D) Comparison of KEAP1 mutation rate among three histological subtypes (chi-square test). (E) Scatterplot showing the correlation between differential regulation of protein log2 fold change in tumors with KEAP1 mutation versus KEAP1 wild tumors versus protein log2 fold change in tumors with C-LCNEC with NSCLC versus pure LCNEC tumors. FC, fold change. (F) Volcano plot showing differentially regulated proteins in KEAP1 mutated versus wild samples (Wilcoxon ranked-sum test). (G) Boxplot showing the comparison of protein differences for pentose phosphate pathway key enzymes between KEAP1 mutated and wild tumor samples (Wilcoxon ranked-sum test). (H) Significantly enriched pathways in KEAP1 mutated tumors compared with KEAP1 wild tumors via gene set enrichment analysis. For (C) and (G), *P < 0.05, **P < 0.01, and ***P < 0.001. Kinase-substrate enrichment analysis (KSEA) identified 23 kinases with significantly increased activity in LCNEC tumors (fig. S5A and table S3). WEE1 and CDK1 were also supported by elevated expression of their protein abundance and their corresponding substrates (fig. S5, B and C). Targeted inhibitor of the WEE1 kinase represents a potential therapeutic strategy in neuroendocrine neoplasms (34, 35), potentially informing the treatment strategy in LCNEC. CDK1-Tyr15 served as a substrate for WEE1, and we further found evidence of the high activity of CDK1 in LCNEC tumors (fig. S5, C and D). The supervised analysis identified 46 prognostic biomarkers after stringent filtering, including 35 proteins associated with good prognosis and 11 proteins associated with poor prognosis (fig. S5E). Among them, FBP1 was significantly down-regulated in LCNEC tumors, which was validated by IHC staining and quantitative analysis (fig. S5F). Patients with higher FBP1 expression showed better survival in both TU-LCNEC and George et al. LCNEC cohorts (12) (fig. S5, G and H). Moreover, patients with high FBP1 protein expression were associated with the activation of multiple immune pathways, such as interferon response and the Janus kinase–signal transducer and activator of transcription (JAK–STAT) signaling pathways (fig. S5I). To verify the tumor suppressor function of FBP1, we overexpressed it in the H727 cell line, which exhibited a high neuroendocrine signature comparable to SCLC cell lines H524 and H82 (table S3). The results demonstrated that overexpression of FBP1 significantly inhibited tumor growth in the H727 cell (fig. S5, J and K) and H727 cell-derived xenograft (CDX) model (fig. S5, L to N). Together, these results suggested that FBP1 might serve as both a prognostic biomarker and a tumor suppressor in LCNEC. KEAP1 mutation drives metabolic reprogramming in C-LCNEC with NSCLC In addition to 81 pure LCNEC, the TU-LCNEC cohort comprised 26 C-LNCEC, including 21 LCNEC combined with NSCLC components (C-LCNEC with NSCLC) and 5 LCNEC combined with SCLC components (C-LCNEC with SCLC), and the histology information and representative IHC staining of pure LCNEC, LCNEC combined with NSCLC, and LCNEC combined with SCLC are shown in table S1 and fig. S6. We compared molecular features and clinical characteristics among C-LCNEC with NSCLC, C-LCNEC with SCLC, and pure LCNEC to identify biological characteristics and potential drug targets specific to C-LCNEC (Fig. 3A). Gene set enrichment analysis (GSEA) demonstrated a unique activation of pathways related to antigen presentation and processing, MTOR1 signaling, NRF2, and the pentose phosphate pathway in C-LCNEC with NSCLC (Fig. 3B, fig. S7A, and table S4). Consistently, key enzymes in the pentose phosphate pathway, including NQO1, G6PD, TKT, PGD, TALDO1, and ME1, were specifically up-regulated in C-LCNEC with NSCLC (Fig. 3C). These results indicated an active metabolic alteration in C-LCNEC with NSCLC. The mTORC1 signaling-induced metabolic reprogramming enables tumor cells to be glycolysis-independent via the pentose phosphate pathway (36), implying that co-inhibition of mTORC1 and glycolysis may hold potential as a treatment strategy for C-LCNEC with NSCLC. We then investigated the genomic characteristics associated with C-LCNEC. We found a high frequency of mutational events involving KEAP1 in C-LCNEC with NSCLC (chi-square test, P = 0.04; Fig. 3D). Notably, the differential protein expression caused by KEAP1 mutation largely recapitulated that between C-LCNEC with NSCLC and pure NSCLC (Fig. 3E), providing evidence of a possible link between LCNEC driver mutation KEAP1 and the histological subtype of LCNEC. Compared to wild tumors, KEAP1 mutation caused strong induction of key enzymes of the pentose phosphate pathway, especially G6PD (Fig. 3, F and G; fig. S7B; and table S4). Moreover, KEAP1 mutated tumors exhibited up-regulation of NRF2-regulated proteins and NRF2-regulated pathways (fig. S7C and table S4), including NR0B1, AKR1B10, and CYP4F11 (fig. S7D). GSEA further revealed a significant enrichment of pathways related to glutamine metabolism in KEAP1-mutated tumors (Fig. 3H and fig. S5E). These results suggested a critical role of KEAP1 mutation in driving metabolic reprogramming in LCNEC combined with NSCLC. Given that KEAP1-mutated cells display a robust sensitivity to glutaminase inhibitors (37), it is therefore appealing to speculate that a combination of glutaminase and G6PD inhibition might be a promising therapeutic approach for patients with C-LCNEC with NSCLC harboring KEAP1 mutation. Immune landscape and regulation in LCNEC Consensus clustering based on the xCell (38) signatures identified two sets of tumors representing immunologically hot and cold clusters, as well as a NAT-enriched cluster (Fig. 4A and table S5). The hot tumor cluster was characterized by an increased abundance of immune cells (macrophages M1, B cells, CD8+ T cells, and aDC), higher immune score, better survival, higher TMB, lower tumor immune dysfunction and exclusion (TIDE) score (39), and decreased abundance of neuron cells compared to the cold tumor cluster (Fig. 4, A to E). GSEA found that the hot tumor cluster was chiefly enriched for T cell receptor/B cell receptor (TCR/BCR) signaling and interferon-related pathways (fig. S8A), while certain neuronal pathways were especially enriched in the cold tumor cluster (fig. S8B). The immune repertoire analysis further showed higher TCR/BCR reads, clonality, and unique complementarity-determining region 3 (CDR3) in hot tumors (40) (fig. S8C), which were consistent with the TCR/BCR pathways’ enrichment in hot tumor cluster (fig. S8A). The JAK-STAT signaling pathway was also enriched in the hot tumor cluster along with higher protein expression and TF activity of STAT3 (fig. S8, A and D). In addition, the hot tumor cluster showed low EPCAM but high VIM protein expression accompanied by epithelial–mesenchymal transition (EMT) pathway enrichment (fig. S8, A and E). The immune landscape of LCNEC. (A) Heatmaps illustrate cell-type compositions and activities of selected individual genes/proteins and pathways across three immune clusters: hot tumor cluster, cold tumor cluster, and NAT-enriched cluster. Successive heatmaps illustrate xCell immune signatures, mRNA, and protein expression of key immune-related markers, and ssGSEA pathway scores based on global proteomic data for biological pathways that were differentially regulated in immune groups based on both mRNA and global protein abundance. (B) Kaplan-Meier curves comparing overall survival for hot tumor and cold tumor (Log-rank test). (C to E) Hot tumor cluster showed high immunescore (C), TMB (tumor mutation burden) (D), and low TIDE (tumor immune dysfunction and exclusion) score (E) (Wilcoxon ranked-sum test). (F) Higher APOBEC enrichment score in hot tumor cluster (Wilcoxon ranked-sum test). (G) APOBEC enrichment score was significantly positively associated with IFN-α activity, IFN-γ activity, immune scores, and multiple APOBEC family members’ expression (Pearson’s test). (H) LRP1B mutation frequency between hot and cold tumor cluster (Fisher’s exact test). (I and J) LRP1B mutated tumor samples showed high immunescore (I) and TMB (J) (Wilcoxon ranked-sum test). (K and L) High activity (K) and mRNA expression (L) of INSM1, SALL2, and ONECUT2 in cold tumor cluster (Wilcoxon ranked-sum test). *P < 0.05, **P < 0.01, and ***P < 0.001. (M) Negative correlation between immunescore and neuroendocrine-related characteristics (Pearson’s test). The APOBEC mutational signature was crucial in both the TU-LCNEC cohort and the George et al. LCNEC cohort (12) (Fig. 1I and fig. S1K). We observed a higher APOBEC enrichment score in the hot tumor cluster (Fig. 4F and table S5). The positive correlation between APOBEC enrichment score and interferon-gamma (IFN-γ) signaling has also been observed in thoracic tumors (41). We found the enrichment of interferon-alpha (IFN-α) and IFN-γ pathways in the hot tumor cluster (fig. S8A). Moreover, the evaluation of the activities of IFN-α and IFN-γ pathways, using proteomic data, showed that the APOBEC enrichment score was positively associated with IFN-α activity, IFN-γ activity, immune score, and the expression of multiple APOBEC family members (Fig. 4G and table S5), suggesting an intriguing connection between the APOBEC mutational signature and expression of APOBEC members, IFN-related pathways, and immune activity in the tumor microenvironment of LCNEC. Consistently, higher expression of multiple APOBEC members was observed in the hot tumor cluster at both the protein level (fig. S8F) and the mRNA level (fig. S8G). Furthermore, patients with higher expression of APOBEC3A mRNA, APOBEC3H mRNA, APOBEC3C protein, and APOBEC3G protein expression showed better survival (fig. S8H). In addition to APOBEC, LRP1B mutations were also significantly enriched in the hot tumor cluster (Fisher’s exact test, P < 0.01; Fig. 4H). LCNECs with LRP1B mutations showed higher immune score and TMB than those wild-type tumors (Fig. 4, I and J), indicating an association between LRP1B mutation and increased tumor immunogenicity in LCNEC. Both APOBEC and LRP1B mutagenesis are associated with the good immune checkpoint blockade (ICB) response in lung cancer (42, 43). Consistently, several important immunosuppressive mediators, including CTLA4, PD-1 (PDCD1), PD-L1(CD274), and LAG3, were up-regulated in the hot tumors (Fig. 4A and table S5). These results suggested the potential application of ICB therapy in the hot tumor of LCNEC. In the cold tumor cluster, neuronal cell activity and neuronal regulatory pathways were specifically recognized (Fig. 4A and fig. S8B). In addition, we observed higher expression and activity of several critical neuroendocrine-related transcriptional factors (TFs), including INSM1, SALL2, and ONECUT2 (Fig. 4, K and L). These neuroendocrine-related TFs were negatively associated with the immune score but positively associated with neuroendocrine (NE) score (Fig. 4M and table S5). A strong inverse correlation between immune score and NE score indicated a conflicting role between neuroendocrine phenotype and immune phenotype in LCNEC (Fig. 4M, Pearson’s correlation, R = −0.54, P < 0.01). Proteomic subtypes with distinct biological and clinical features Using the non–negative matrix factorization (NMF)–based unsupervised clustering, we identified three distinct proteomic subtypes exhibiting diverse clinical, immunologic, epigenetic, and genomic features (Fig. 5A; fig. S9, A and B; and table S6). These three proteomic subtypes were involved in distinct pathways and significantly differed in OS, with the C1 subtype showing the best and the C3 subtype showing the worst OS (overall Log-rank P = 0.0035; Fig. 5, B and C). Patient stratification using phosphoproteomic data similarly identified three phosphoproteomic subgroups (S1, n = 24; S2, n = 47; S3, n = 32), which were largely congruent with proteomic subtypes (chi-square test, P < 0.01; Fig. 5D). Intriguingly, patients with C-LCNEC with NSCLCs were significantly enriched in the C1 subtype, whereas patients with pure LCNECs were dominant in the C2 and C3 subtypes (chi-square test, P < 0.01; Fig. 5E). Differentially expressed genes in the C1 subtype and C-LCNEC with NSCLC were highly concordant at the protein level (fig. S9C; Pearson correlation, R = 0.80, P < 0.001), indicating a link between the histological subtype and proteomic subtype. The comparison between proteomic subtypes, hot tumor clusters, and cold tumor clusters showed the dominance of hot tumors in the C1 subtype, while cold tumors were more prevalent in the C2 and C3 subtypes (chi-square test, P < 0.01; fig. S9D). Proteomic stratification of TU-LCNEC cohort and the corresponding molecular and pathway features. (A) Proteomic profiles of the signature proteins and the enriched biological processes of the three proteomic subtypes. (B) Kaplan-Meier curves for overall survival based on the proteomic subgroups in the LCNEC cohort (Log-rank test). (C) Heatmap representing significantly enriched pathways (MSigDB Hallmark) in three proteomic subtypes via gene set enrichment analysis. *P < 0.05 and **P < 0.01. (D) Stacked bar plot showing the proportions of phosphoproteomic subtypes distribution in each proteomic subtype (chi-square test). (E) Stacked bar plot showing the proportions of LCNEC histological subtypes distribution in each proteomic subtype (chi-square test). (F) Cluster-wise up- and down-regulation of transcription factors in three proteomic subtypes. (G) Cluster-wise up- and down-regulation of human kinases in three proteomic subtypes. (H) Boxplot showing high HDAC inhibition sensitive score for protein level in C2 subtype (one-way ANOVA followed by Tukey’s multiple comparisons test). (I) Significant enrichment of E2F targets and RB loss pathways in C2 subtype. (J) Boxplot showing high EZH2 protein expression in C2 subtype (one-way ANOVA followed by Tukey’s multiple comparisons test). (K) Significant down-regulation of EZH2 targets in C2 subtype. In line with the enrichment of hot tumors in the C1 subtype, we also found the enrichment of IFN-α/IFN-γ and JAK-STAT pathways in the C1 subtype, indicating an immune activity activation state in the C1 subtype (Fig. 5C). Consistently, the C1 subtype exhibited higher immune score (fig. S9E) and tertiary lymphoid structure (TLS) activity (fig. S9F). Notably, transcription factors (TFs) of the STAT family were strongly expressed in the C1 subtype (Fig. 5F). Protein-protein interaction (PPI) analysis of TFs with C1-biased expression further indicated a central role of the STAT family (fig. S9G). In addition, the C1 subtype exhibited the highest TF activity of STAT3 (fig. S9H) and phosphorylation levels of Ser727 on STAT3 (fig. S9I). Moreover, the STAT3 TF activity also exhibited a significant positive association with the abundance of STAT3 protein and phosphorylation sites (fig. S9J). STAT3 phosphorylation at Ser727 contributes to EMT pathway activity (44). Consistently, the C1 subtype was strongly enriched in the EMT pathway (Fig. 5C), featuring the highest abundance of mesenchymal marker with VIM, ZEB1, and ZEB2, and the lowest abundance of EPACM (fig. S9K). The kinase protein abundance significantly differed among three proteomic subtypes, and more than half of differentially expressed kinases were up-regulated in the C1 subtype (Fig. 5G). We observed strong up-regulation of receptor tyrosine kinases (RTKs), such as PDGFRA and PDGFRB, mainly in the C1 subtype (fig. S9K). STAT3 phosphorylation at Ser727, the substrate of PDGFRB, showed the highest level in the C1 subtype (fig. S9L). AXL exhibited elevated expression in the C1 subtype at both mRNA and protein levels (fig. S9M). Evaluation of immune checkpoints based on inhibitory receptors (IRs) and corresponding ligands (45, 46) indicated that the expression levels of IRs generally correlated with immune infiltration and mRNA expression of AXL (fig. S9N; Pearson correlation, R = 0.89, P < 0.001). The previous observation that targeting AXL favors ICB response indicated that the combination of AXL inhibitor (AXLi) and ICB might represent a possible treatment option for the C1 subtype (47). The aberrate epigenetic activity and SCLC-like feature in the C2 subtype After comparison among three proteomic subtypes of the TU-LCNEC cohort and previously reported LCNEC tumor clusters using genomic or transcriptomic data (10, 12), we found a more preferred SCLC-like feature for the C2 subtype of the TU-LCNEC cohort (chi-square test, both P < 0.05, fig. S10, A to C). Consistently, the C2 subtype exhibited more frequent comutation of TP53 and RB1 (chi-square test, P < 0.05, Fig. 5A) and enriched pathways relating to cell proliferation (Fig. 5C). Differentially expressed genes in the C2 subtype and SCLC-like subtype were largely concordant at the protein level (fig. S10D; Pearson correlation, R = 0.53, P < 0.001), indicating a link between SCLC-like tumors and the C2 subtype. Notably, two SCLC-associated proteins, STMN1 and STMN2 (48), showed specific up-regulation in both the C2 subtype and SCLC-like tumors (fig. S10D), indicating a link between SCLC-like features and the C2 subtype. Aberrate epigenetic reprogramming is emerging as an important driver in tumorigenesis (49). Intriguingly, multiple histone deacetylase (HDAC) family members were most abundant at the protein level in the C2 subtype (fig. S10E). The query for compounds that reversed the C2 signature using the Library of Integrated Network-Based Cellular Signatures (LINCS) further presented with enriched HDAC inhibitor (HDACi) in the C2 subtype (fig. S10F). Furthermore, we further confirmed that the C2 subtype harbored a significantly higher sensitivity score to HDACi than other tumors using the HDAC sensitivity score proposed previously (50) (Fig. 5H). Meanwhile, C2 showed higher MYCN mRNA expression (fig. S10G) and stemness-like signature enrichment (fig. S10H), likely due to aberrant epigenetic activity (51). GSEA further revealed significant enrichment for E2F targets and RB-loss pathways in the C2 subtype compared with other subtypes (Fig. 5I). EZH2, a critical epigenetic repressor downstream of the pRB-E2F pathway (52), was also the most abundant in the C2 subtype at both mRNA and protein levels (Fig. 5J and fig. S10I). Moreover, SCLC-like tumors also showed significantly higher EZH2 expression (fig. S10J). In line with the higher expression of EZH2 in the C2 subtype, the target genes of EZH2 were significantly down-regulated in the C2 subtype (Fig. 5K). KCTD12, a target gene of EZH2 that contributes to the stemness of melanoma and colorectal cancer (53, 54), was down-regulated explicitly in the C2 subtype (fig. S10K). Intriguingly, KCTD12 was down-regulated at both mRNA and protein levels in LCNEC tumors compared to NATs (fig. S10L), and patients with lower KCTD12 protein expression showed worse survival (fig. S10M), suggesting an oncogenic role of EZH2 possibly via increasing tumor cell stemness by down-regulating KCTD12 in the C2 subtype. Considering the synergistic role of HDACi and EZH2 inhibitor (EZH2i) in NSCLC (55), our results suggested that the combination treatment of HDACi and EZH2i might be an option for the C2 subtype. Elevated oxidative phosphorylation and neuroendocrine features in the C3 subtype The C3 subtype, displaying the worst OS, was the most dominant (89.7%) with cold tumors (fig. S9D). A notable feature of this proteomic subtype was the elevated oxidative phosphorylation activity (Fig. 5C). The higher abundance of mitochondrial proteins, including NDUFS2, UQCRC1, and COX4I1 of respiratory complexes I, III, and IV, but not SDHB of respiratory complex II, was observed in the C3 subtype (Fig. 6, A to C). GSEA consistently demonstrated a strong enrichment of respiratory complexes I, III, and IV, except for complex II in the up-regulated genes of the C3 subtype (Fig. 6D). These results indicated a dependency of the C3 subtype on NADH oxidation, making them vulnerable to complex I inhibitors (complex Ii). The m6A reader protein IGF2BP3 is critical to promote oxidative phosphorylation in NSCLC (56). We found the up-regulation of IGF2BP3 in the C3 subtype (fig. S11A). By categorizing LCNEC tumors into “IGF2BP3 high” (n = 29), “IGF2BP3 medium” (n = 32), and “IGF2BP3 low” (n = 42) groups based on IGF2BP3 abundance, we found that the high IGF2BP3 protein expression was associated with poor prognosis (fig. S11B). Moreover, the significantly up-regulated genes in the “IGF2BP3 high” group exhibited strong enrichment of known IGF2BP3 target genes (Fisher’s exact test, one-sided P = 0.054, fig. S11C) and were increased more evidently in the C3 subtype (fig. S11D). Modular-based network analysis of the genes up-regulated in the C3 subtype further identified four significant modules (modules 1, 2, 3, and 7) exhibiting enrichment of IGF2BP3 targets (Fisher’s exact test, P < 0.05; fig. S11E and table S6). Notably, these four modules were significantly enriched in mitochondrial-related processes, including mitochondrial translational transport, respiratory chain complex I assembly, and oxidative phosphorylation (fig. S11F). These results indicated a role of IGF2BP3 in regulating the elevated oxidative phosphorylation activity in the C3 subtype. OXPHOS dependence for C3 subtype. (A) Expression of mitochondrial proteins stratified by proteomic subtype. Proteomic subtype C3 has significantly higher expression of mitochondrial proteins than C1 and C2 (one-way ANOVA followed by Tukey’s multiple comparisons test). (B) Protein log2 fold changes between C3 and other tumor samples grouped by mitochondria or nonmitochondria proteins. (C) Protein expression of representative markers in OXPHOS complex I to IV subunits stratified by proteomic subtype (one-way ANOVA followed by Tukey’s multiple comparisons test). (D) Four plots showing the gene set enrichment score for the four mitochondrial complexes (I, II, III, and IV, respectively) with C3 subtype. Complexes I, III, and IV were enriched in C3 expect for complex II. (E) Boxplot showing higher NE score in the C1, C2, and C3 subtype (one-way ANOVA followed by Tukey’s multiple comparisons test). (F) PI3K-NEU mutations in the C1, C2 and C3 subtype. (G) PPI and comutation interactions among 17 PI3K-NEU mutation genes. (H and I) Boxplots showing high PI3K-NEU activity for C1, C2, and C3 in both protein (H) and mRNA levels (I) (one-way ANOVA followed by Tukey’s multiple comparisons test). (J and K) Kaplan-Meier curves for overall survival based on PI3K-NEU activity in the TU-LCNEC cohort (J) and George et al. LCNEC cohort (K). The neuroendocrine feature was another hallmark of the C3 subtype. We also observed the highest neuroendocrine activity measured using the NE score in the C3 subtype (Fig. 6E). The neuroendocrine markers, including ASCL1 and INSM1, showed the highest mRNA expression in the C3 subtype (fig. S12A). The Notch signaling pathway controls NE differentiation and plasticity through up-regulating REST and down-regulating ASCL1 and DLL3 (57). In line with high NE activity, we observed the lowest activity of Notch signaling (fig. S12B) and REST (fig. S12C), as well as the highest level of DLL3 in the C3 subtype (fig. S12A). DLL3 has been identified as a promising target of lung cancers, the abundance of which correlates strongly with the therapeutic efficacy of DLL3-targeted drugs (58), suggesting that C3 subtype patients with high expression of DLL3 might benefit from DLL3-targeted therapy. We next examined the specific mutations of the C3 subtype. A total of 30 genes with markedly high mutation frequency specific to the C3 subtype were identified (Fisher’s exact test, C3 versus other subtypes, OR > 5, P < 0.05), including 3 genes of the PI3K pathway (STK11, KRAS, and RELN) and another 14 genes related to neuronal diseases (CAMTA1, KCHN5, ATP13A1, KIF1A, NELL2, SOX5, TSHZ3, ADAM11, PAX1, ZNF804A, POU3F3, ZFPM2, PXDN, and TANC1) based on literature checking and the annotation of GeneCards database (Fig. 6F). We also found that these 30 genes were preferentially involved in the PI3K pathway and localized in the cellular component of neuronal cells (table S6). Furthermore, focusing on 17 genes involved in the PI3K pathway and neuronal diseases (PI3K-NEU genes), we identified close links among them through comutation and PPI analysis (Fig. 6G), suggesting an intrinsic connection between PI3K-NEU genes. By annotating LCNEC tumors into “PI3K-NEU mutated” (n = 48) or “PI3K-NEU wild” (n = 59) according to the mutation status of PI3K-NEU genes, we found that PI3K-NEU mutated tumors exhibited a highly concordant differential protein expression pattern with the C3 subtype (fig. S12D; Pearson correlation, R = 0.88, P < 0.001). By further defining a PI3K-NEU signature using PI3K-NEU genes with significant up-regulation at both mRNA and protein levels in PI3K-NEU mutated tumors, we found that the C3 subtype exhibited specific high activity of PI3K-NEU signature (PI3K-NEU activity) at both mRNA and protein levels (Fig. 6, H and I, and table S6). Patients with high PI3K-NEU activity showed poor survival in the TU-LCNEC cohort and George et al. LCNEC cohort (12) (Fig. 6, J and K). In addition, GSEA and traditional correlation analysis revealed a positive association between PI3K-NEU activity and respiratory complexes I, III, and IV, but not respiratory complex II (fig. S12, E and F), indicating a link between PI3K-NEU signature and elevated oxidative phosphorylation feature in the C3 subtype. IL-33 preactivation enhances the antitumor efficacy of expanded γδT cells Compared with paired NATs, differential protein expression analysis identified nearly three times more down-regulated proteins than up-regulated proteins in LCNEC tumors (fig. S3B). These down-regulated proteins markedly enriched in immune-related biological processes, such as adaptive immune response, leukocyte migration, immune response, and innate immune response (fig. S13A), indicating a global immune suppression in the LCNEC tumors. By analyzing the expression alterations of immune-related genes categorized by ImmProt (59), we found that interleukins exhibited the most down-regulation in LCNEC tumors (fig. S13B). Notably, IL-33 was decreased in almost all LCNEC tumors, exhibiting an approximate sixfold reduction on average (fig. S13C). We then investigated the potential role of IL-33 in tumorigenesis via in vivo mice models. We constructed Lewis lung carcinoma (LLC) CDX models and randomly assigned them to recombinant IL-33 (rIL33) or saline treatment groups when tumor volume reached 100 mm3 (Fig. 7A). As shown, rIL33 treatment greatly inhibited the growth of LLC tumors (Fig. 7, B to D). IHC staining and quantitative analysis identified higher CD45 expression in rIL33-treated tumors, indicating more immune cell infiltration (Fig. 7E and fig. S14A). We sorted the CD45+ cells of lung tumors in LLC mice treated with rIL33 (n = 5) and saline (n = 4), then single-cell RNA sequencing (scRNA-seq) was performed to examine the immune cell infiltration difference in lung tumors more specifically. Cell clustering analysis identified 23 clusters corresponding to 10 major cell types: B cell, cDC1, fibrocyte, macrophage, MAST cell, MoDC, Mreg DC, neutrophil, natural killer (NK) cell, and T cell (Fig. 7F and fig. S13, D to F). Compared to the saline-treated group, rIL33 treatment increased the total number of T cells in the tumor tissue of LLC mice (Fig. 7, G and H). IL-33 preactivation enhances the antitumor efficacy of expanded γδT cells. (A) Schematic of the experimental design: Lewis lung carcinoma (LLC) cell-derived xenograft (CDX) models were established, and mice were randomly divided into recombinant IL-33 (rIL33, 0.25 mg/kg, intraperitoneal injection, qd for 10 days) or saline groups when tumor volume reached ∼100 mm3 [created in BioRender. Yan, Y. (2026) https://BioRender.com/a88xaz7]. Tumor growth was monitored, and tumor tissues were collected for IHC, CD45+ cell sorting, and single-cell RNA sequencing analyses. (B) Tumor growth in LLC CDX models treated with recombinant murine IL-33 (rIL33) or saline. Data are represented as mean ± SEM (n = 8 mice per group, two-way ANOVA test). (C) Statistical analyses of tumor weight from LLC CDX models treated with IL-33 or saline. Data are represented as mean ± SEM (Wilcoxon ranked-sum test). (D) The tumors from LLC CDX models treated with IL-33 or saline. (E) LLC CDX models treated with rIL33 or saline. Tumor tissues were performing IHC and stained with anti-CD45 antibody. Representative IHC images are derived from different mice in the rIL33 and saline treatment groups. Scale bar, 25 μm. (F) A UMAP plot of all scRNA-seq data, showing a total of 10 distinct cell types that were identified. (G) Cell-type distribution in IL-33 and saline groups. (H) Boxplot showing high T cell infiltration after IL-33 treatment (Student’s t test). (I) Boxplots showing high CD4T cells and γδT cells after IL-33 treatment (Student’s t test). (J) Boxplot showing high TCR clone numbers in CD4T cells (Wilcoxon ranked-sum test). (K) LLC CDX models treated with rIL33 or saline. Tumor tissues were subjected to IHC and stained with anti-RORγt antibody. Representative IHC images are derived from different mice in the rIL33 and saline treatment groups. Scale bar, 25 μm. Further analysis for the NK/T population classified CD4+ T cells, CD8+ T cells, γδT cells, regulatory T cells, and NK cells (fig. S13G). Higher infiltrations of CD4+ T cells, CD8+ T cells, and γδT cells were observed after rIL33 treatment (Fig. 7I). Moreover, significantly higher TCR clones were also observed in CD4+ T cells after rIL33 treatment, suggesting the clonal expansion of CD4+ T cells after IL-33 preactivation (Fig. 7J), which may render better ICB responses (60). IHC staining and quantitative analysis also verified higher infiltration of RORγt+ cells in rIL33-treated tumors (Fig. 7K and fig. S14B). We observed a strong interaction between γδT cells and CD8+ T cells using cell-cell communication analysis via CellChat (fig. S13H). These results suggested that IL-33 preactivation might promote γδT cell proliferation, cytokine production, and activation of endogenous CD8+ T cells in a cell-cell interaction–based manner (61). We also observed high cytotoxic markers, such as Gzmb, in CD4+ T and CD8+ T cells after rIL33 treatment (fig. S13, I and J), indicating the positive role of IL-33 in promoting the function of CD4+ T and CD8+ T cells. rIL33 treatment resulted in higher expression of Prf1 in CD8+ T cells and NK cells (fig. S13, J and K). These results collectively demonstrate that IL-33 preactivated murine γδT cells could enhance tumor-killing capacity and may exert more effective antitumor activity upon adoptive transfer. Cysteine mutation and PEGylation prominently enhanced the IL-33 immune agonist activity The short half-life of cytokine limits its therapeutically effective concentration (62). rIL33 contains four cysteine residues, the oxidization of which in the extracellular environment reduces its protein activity. To address this, we introduced cysteine mutation and PEGylation modification on IL-33 to enhance its protein activity and half-life period. We engineered PEGylated rIL33 using maleimide PEG on the terminal cysteine residue of rIL33 and then sequentially mutated four cysteine residues and studied the impact on PEG modification activity (Fig. 8A). Among various cysteine mutation combinations (SSSS, CSSC, FSSC, YSSC, TSSC, GSSC, and ASSC), the GSSC mutant exhibited stable and enhanced activity (Fig. 8B). We then generated mono-PEG-5K and mono-PEG-20K modified rIL33 variants based on the GSSC mutant. In both CT26 and LLC CDX models, tumors treated with PEG5k-GSSC and PEG20k-GSSC exhibited inhibited growth compared to those treated with saline (Fig. 8, C to H, and fig. S14, C to E). Notably, PEG20k-GSSC demonstrated superior tumor growth inhibition compared to PEG5k-GSSC in both models (Fig. 8, C to H, and fig. S14, C to E). To further evaluate the circulation time of PEGylated rIL33 in nonhuman primates, two cynomolgus monkeys were administered the PEGylated rIL33, and serum samples were collected from 12 to 240 hours posttreatment. DIA (data-independent acquisition) analysis revealed that PEGylated rIL33, along with downstream protein markers IL18BP and IL1RL1, maintained stable intensity levels up to 72 hours (Fig. 8I). Collectively, these results indicate that PEGylation of rIL33 prolongs its circulation time in cynomolgus monkeys and enhances its immune agonist activity in mouse models. Cysteine mutation and PEGylation prominently enhanced the IL-33 immune agonist activity. (A) Schematic diagram of IL-33 modification and functional evaluation [created in BioRender. Yan, Y. (2026) https://BioRender.com/k1kkv3n]. (B) Cysteine sites were mutated to generate single maleimide reaction sites. The mutant IL-33 (recombinant Homo sapiens IL-33) and PEGylated IL-33 activity were measured with NF-κB luciferase reporter assay in HEK293 cells. (C) Tumor growth in LLC CDX models treated with IL33-PEGylation or saline. Data are represented as mean ± SEM (n = 6 mice per group, two-way ANOVA followed by Tukey’s multiple comparisons test). (D) Statistical analyses of tumor weight from LLC CDX models treated with IL33-PEGylation or saline. Data are represented as mean ± SEM (Wilcoxon ranked-sum test). (E) The tumors from LLC CDX models treated with IL33-PEGylation or saline. (F) Tumor growth in CT26 CDX models treated with IL33-PEGylation or saline. Data are represented as mean ± SEM (n = 6 mice per group, two-way ANOVA followed by Tukey’s multiple comparisons test). (G) Statistical analyses of tumor weight from CT26 CDX models treated with IL33-PEGylation or saline. Data are represented as mean ± SEM (Wilcoxon ranked-sum test). (H) The tumors from CT26 CDX models treated with IL33-PEGylation or saline. (I) Data-independent acquisition (DIA) analysis was used for IL-33 half-life study in nonhuman primate serum. PEGylated IL-33 and downstream protein markers, IL18BP and IL1RL1, were plotted with different time points. DISCUSSION In this study, we present the first comprehensive proteogenomic profiles of primary LCNEC tumors and paired NATs by leveraging WES, RNA-seq, proteomic, and phosphoproteomic analyses. From the WES data, we identified the APOBEC mutational signature as the crucial mutational pattern potentially contributing to LCNEC pathogenesis and immunosuppression. Recent studies demonstrated that APOBEC mutational signatures are the primary drivers in early-onset squamous cell carcinomas and are notably prevalent in younger female patients with lung adenocarcinoma (16, 63). Furthermore, the association between a higher APOBEC mutational signature and prolonged progression-free survival in the ICB-treated NSCLC cohort suggests the potential utility of the APOBEC signature in identifying patients with LCNEC who may benefit from immunotherapy (64). In our study, as well as in the George et al. study, KEAP1 mutation was identified as a significantly mutated gene in LCNEC (12). The proteogenomic analysis provided deeper insights into the downstream signaling effects of KEAP1 alterations, unveiling potential regulatory mechanisms and therapeutic targets. Key enzymes in the pentose phosphate pathway, such as G6PD, along with NRF2-regulated proteins like NR0B1, exhibited high expression levels in LCNECs with KEAP1 mutations. In addition, we observed significant enrichment of glutamine metabolism–related pathways in KEAP1 mutated tumors, and these KEAP1-mutant cells showed strong sensitivity to a glutaminase inhibitor (37). Coincidentally, a combined glutaminase and G6PD inhibitor has been proposed as a promising therapeutic approach for patients with KEAP1-mutated lung cancer (65). NRF2-regulated proteins, including NR0B1, harbor covalently druggable cysteines that contribute to the growth of KEAP1-mutant cancer cells, making them potential druggable transcriptional regulators (66). These findings may open promising therapeutic opportunities for the treatment of KEAP1-mutated LCNECs. Despite the promise of immunotherapy in LCNEC treatment, its efficacy remained limited (67–69). Through a thorough analysis of the tumor microenvironment, we characterized the immune landscape of LCNEC and identified three immune clusters. We identified the LRP1B mutation as a potential predictive biomarker for patients with LCNEC who could benefit from immunotherapy by linking it to the hot tumor cluster (43). Furthermore, we found that immune-hot tumors exhibited heightened APOBEC mutational signature activity, potentially contributing to immune infiltration by activating IFN-related pathways. Given the APOBEC signature’s association with a better ICB response (70), the immune-hot cluster may represent a viable candidate for ICB therapy in LCNEC. Unsupervised NMF-based proteomic clustering categorized LCNEC tumors into three subtypes, each characterized by distinct prognosis, biological differences, and therapeutic vulnerabilities. The C1 subtype displayed elevated immune activity and RTK expression, suggesting a potential responsiveness to AXL inhibition and ICB (47). The C2 subtype exhibited high abundance of HDAC family members and EZH2, indicating a potential response to the combination inhibition of HDAC and EZH2 (55). The C3 subtype was characterized by high OXPHS, NE features, and dependence on respiratory complex I, suggesting a potential benefit from mitochondrial complex I inhibition (71). These findings underscore the potential utility of these proteomic subtypes in guiding personalized treatment selection for LCNEC. IL-33, an unconventional cytokine, acts as a danger-associated molecular signal released upon necrotic cell death (72). A significant down-regulation of IL-33 was observed in LCNEC tumor samples. Treatment of LLC mouse models with rIL33 exhibited significantly enhanced antitumor activity. scRNA-seq analysis of rIL33-treated and control mice tumor samples revealed the clonal expansion of CD4+ T cells and the interaction between γδT cells and CD8+ T cells, potentially contributing to enhanced antitumor activity following rIL33 treatment. Therefore, our findings indicate that IL-33 plays a critical and complicated role in antitumor immunity and may serve as a robust therapeutic biomarker for patients with LCNEC. Most cytokines have a short lifetime due to their low molecular weight and because they tend to be degraded by plasma protease. We engineered the rIL33 with mutant combination and PEG modification, and the final version of GSSC rIL33 with 20 kDa PEG showed a marked enhancement on its immune-agonist activity in the mouse model. This version of rIL33 exhibits an extended half-life in nonhuman primates (cynomolgus monkeys), remaining detectable for up to 1 week. The animals tolerated a dosage of 1 mg/kg without apparent adverse effects. These preclinical findings suggest that the GSSC-PEG20K protein holds promising potential as an effective therapeutic agent. In summary, the integrating of comprehensive genomic, transcriptomic, proteomic, and phosphoproteomic data unveiled the underlying molecular mechanisms and disease subtypes of LCNEC that were not fully captured by genomic and transcriptomic analysis alone. These findings, analyses, and underlying data are expected to contribute to more effective clinical treatment for LCNEC and provide a valuable resource for in-depth investigation into the carcinogenesis, progression, and therapeutic strategies of this disease. MATERIALS AND METHODS Clinical specimens All LCNEC tumor and NAT samples for this study were collected from Shanghai Pulmonary Hospital (Tongji University, Shanghai, China) between November 2012 and January 2021. This study was approved by the Ethics Committee of Shanghai Pulmonary Hospital (K23-309) and written informed consents of all participating patients were obtained. Paired NATs were collected from the same patient at tumor resection. Tissues were obtained and stored in liquid nitrogen less than 30 min after resection. For accurate tumor cellularity analysis, hematoxylin and eosin staining of the middle section of each tissue was performed. Pathological features of all tumor samples were confirmed independently by two certified and experienced pathologists. High-quality tumors and paired NATs from 107 patients were chosen for WES, RNA-seq, proteomic, and phosphoproteomic profiling. The clinicopathological parameters were collected and summarized in table S1, including gender, age, smoking status, tumor site, histologic type, tumor size, date of surgical resection, tumor-node-metastasis (TNM) staging [American Joint Committee on Cancer (AJCC) cancer staging system 8th edition], and so on. This cohort included 95 males and 12 females with a median age at diagnosis of 65 (range, 31 to 83). In total, 49, 22, and 36 patients were classified as TNM stages I, II, and III-IV, respectively. Histologically, 77% (n = 82) of the tumors were pure LCNEC, and 23% (n = 25) were combined LCNEC, which had an additional component of NSCLC or SCLC. OS was defined as the interval between surgery and death or last date of follow-up. Animals All animal protocols were approved by the Institutional Animal Care and Use Committee (IACUC) at China Pharmaceutical University (Nanjing, China) in accordance with the standards established by the Laboratory Animal Welfare and Ethics Committee of China and approved by the Ethics Committee of Shanghai Pulmonary Hospital (K23-309). Macaca fascicularis were bred in Triapex Laboratories Ltd. under the protocol AP-24173-740N, which was approved by the IACUC of Triapex Laboratories. Cell lines The murine LLC cell line (TCM47) and CT26 cell line (TCM37) were purchased from the Cell Bank of Shanghai Institute of Biochemistry and Cell Biology (Shanghai, China) and maintained in Dulbecco’s modified Eagle’s medium (DMEM) with 10% fetal bovine serum (FBS). Human embryonic kidney (HEK) 293 (CRL-1573) and HEK293FT (CRL-3249) cell lines were purchased from the American Type Culture Collection (ATCC) and maintained in DMEM with 10% FBS. The H727 (CRL-5815) cell line was purchased from ATCC and maintained in RPMI 1640 medium with 10% FBS. Genomics and transcriptomics profiling experiments DNA extraction and WES Genomic DNA was extracted from tumors and NATs using QIAamp Fast DNA tissue kit (QIAGEN) according to the manufacturer’s protocol. Total DNA was quantified by the Qubit 2.0 Flurometer (Life Technologies) and NanoDrop 2000 (Thermo Fisher Scientific), and the integrity was assessed by TapeStation (Agilent Technologies). For Illumina sequencing library construction, the genomic DNA was fragmented to an average size of 180 to 280 bp using a Covaris focused ultrasonicator. Then, WES libraries were prepared and captured using the Agilent SureSelect Human All Exon V6 kit (Agilent Technologies) following the manufacturer’s instructions. The DNA library with 150-bp paired-end reads was sequenced on the Illumina Novaseq 6000 platform. RNA extraction and RNA-seq Total RNA was extracted and purified from fresh frozen tissues using the TRIzol reagent (Invitrogen). All RNA analytes were assayed for RNA concentration and purity using the Qubit RNA Assay Kit in Qubit 2.0 Flurometer (Life Technologies) and NanoPhotometer spectrophotometer (IMPLEN). RNA integrity was assessed using the RNA Nano 6000 Assay Kit of the Bioanalyzer 2100 system (Agilent Technologies). Samples with RNA integrity number (RIN) > 6.0 were considered high quality and used to prepare the transcriptome library. Total RNA-seq library construction was performed from the RNA samples using the NEBNext UltraTM RNA Library Prep Kit for Illumina (NEB) following the manufacturer’s recommendations and index codes were added to attribute sequences to each sample. Quality control was performed at every step, and the libraries were quantified using the Agilent Bioanalyzer 2100 system (Agilent Technologies). Then, paired-end libraries were sequenced on an Illumina Novaseq 6000 platform and 150-bp paired-end reads were generated. Peptides preparation for MS analysis Protein extraction and digestion For the TMT 11-plex proteomic and phosphoproteomic experiments, cryopulverized human LCNEC and corresponding NAT samples were homogenized in lysis buffer. This lysis buffer consisted of 1% SDS, 10 mM Hepes (pH 7.0), 2 mM MgCl2, Benzonase (diluted at a ratio of 1:1000), and Roche PhosSTOP (diluted at a ratio of 1:50). Following homogenization, the samples were centrifuged at 12,000 rpm for 15 min to separate the supernatant, and the protein concentrations in the supernatant were quantified using the bicinchoninic acid (BCA) assay. Subsequently, each sample (4.5 mg) was reduced with 2 mM dithiothreitol (DTT) (Sigma-Aldrich, D9163) for 20 min and then alkylated with 20 mM iodoacetamide (Sigma-Aldrich, I1149) in the dark at room temperature for 20 min. Protein precipitation was performed using the methanol-chloroform method (1 part lysate:4 parts methanol:1 part chloroform:3 parts H2O). After precipitation, the protein pellet was washed with 1 ml of methanol. The resulting precipitated protein was then reconstituted in an 8 M urea solution, which contained 20 mM tris-HCl and 10 mM EDTA, pH 7.5. The protein digestion was performed in 8 M urea buffer using Lys-C enzyme (Promega, no. VA1170) at a ratio of 1:200 (enzyme:protein) for 3 hours. Subsequently, trypsin enzyme (Promega, no. V5280) was used at a ratio of 1:100 (enzyme:protein) in a 2 M urea diluted solution for overnight digestion at 37°C. The digested samples were acidified with trifluoroacetic acid (TFA) to achieve the desired pH range of 2 to 3. The resulting digested peptides were subjected to desalting, which was performed using C18 SepPak cartridges (60 mg, Waters Oasis HLB, WAT094226). Ten percent of each eluate was dried using Speedvac vacuum concentrators, and then the concentration of peptides for global proteomic analysis was measured using a BCA assay. The remaining eluate samples were supplemented with 2 M lactic acid for subsequent phosphopeptides enrichment. The reference samples used in the TMT-11 experiments were prepared by pooling 10 pairs of tumor and adjacent nontumor lung samples, which were randomly selected and mixed to yield equal amounts of protein. Aliquots of this reference mixture (4.5 mg) were frozen and used consistently for each multiplex experiment, beginning at the protein extraction and digestion stage. Phosphopeptide enrichment Phosphopeptide enrichment was carried out with titanium dioxide (TiO2) beads. The TiO2 beads underwent two washes with washing/binding buffer (comprising 50% acetonitrile, 0.1% TFA, and 2 M lactic acid) and were subsequently resuspended in the same buffer. Each digested sample was incubated with TiO2 beads at a ratio of 1:14 (peptide to beads) overnight, with end-over-end rotation at room temperature. Following incubation, the supernatant was removed through centrifugation, and the beads were subjected to three additional washes using the washing/binding buffer. The peptides captured on the beads were eluted twice with 600 μl of K2HPO4 (100 mM, pH 9.2). These eluates were combined, acidified to a pH of 3.0 using TFA, and desalted using 10-mg C18 SepPak cartridges (Waters Oasis HLB, 186000383). Last, the eluates were dried using a vacuum centrifuge concentrator. TMT 11-plex labeling TMT 11-plex labeling was performed for the analysis of 103 unique tumor samples along with their paired NAT samples, distributed across 21 TMT 11-plex experiments. In this process, individual channels within the 11-plex were used for labeling, with one channel assigned to the reference sample and the remaining 10 channels used for labeling individual samples. For the proteome analysis, desalted peptides were reconstituted in a 200 mM Hepes (pH 8.0) solution into a final concentration of 1 mg/ml. Each 50-μg aliquot of peptides from an individual sample was labeled using 150 μg of TMT 11-plex reagents. In the case of phosphoproteome analysis, dried phosphopeptides were reconstituted in a 200 mM Hepes (pH 8.0) solution into a concentration of 0.5 mg/ml. Subsequently, 25 μg of phosphopeptides was labeled with 120 μg of TMT 11-plex reagents, resulting in a final acetonitrile concentration of 40%. After 1-hour incubation at room temperature, the reaction was quenched using hydroxylamine for an additional 30 min. The labeled peptides were then pooled, acidified with 20 μl of 20% TFA, then lyophilized and subjected to desalting using a 10-mg C18 SepPak column (Waters Oasis HLB, 186000383). Offline HPLC fractionation To enhance protein identification depth and reduce sample complexity, peptide fractionation was performed using basic high pH reverse-phase high-pressure liquid chromatography (HPLC). The fractionation process used solvent A (10 mM ammonium formate, pH 9.9) and solvent B [10 mM ammonium formic acid (FA) in 80% acetonitrile, pH 9.9] as the mobile phases. TMT-labeled peptides were initially solubilized in solvent A and centrifuged at 13,000 rpm for 10 min to remove insoluble peptides. The resulting soluble peptides were then loaded onto a ZORBAX 300 Extend-C18 column (2.1 mm by 150 mm, 3.5 μm particle, Agilent). Subsequent peptide separation was achieved through basic reverse-phase chromatography using a Thermo Fisher Scientific Ultimate 3000 HPLC instrument. For the proteome analysis, we used a 126-min separation gradient with a flow rate of 0.15 ml/min, using the following gradient settings: (min: %B) 0:4; 101:65; 106:100; 124:100; and 126:0. Eluents were collected to a total of 96 fractions spanning from 5 to 100 min. To reduce the number of fractions, a concatenation strategy was used, where these 96 fractions were grouped into 24 pooled fractions. Subsequently, these pooled fractions were acidified with TFA. For phosphoproteome analysis, a 100-min separation gradient was applied with the following settings: (min: %B) 0:4; 75.5:65; 80:100; 98:100; and 100:0. Eluents were collected and combined into 12 fractions ranging from 5 to 75.5 min. Similar to the proteome analysis, these fractions were acidified with TFA. Both the fractionated peptides and phosphopeptides were dried using a Speedvac and subsequently desalted using stage tip columns. The dried peptides and phosphopeptides were then resuspended in 15 μl of 4% can/0.1% FA for nanoLC-MS/MS analysis, with 4-μl injection volume per run. LC-MS/MS for proteomic and phosphoproteomic analysis The LC-MS/MS analysis for both proteomic and phosphoproteomic analysis was conducted using an EASY-nLC 1200 UHPLC system from Thermo Fisher Scientific. Fractionated peptides were loaded onto an in-house packed capillary nanospray column (15 cm × 75 μm; 1.9 μm particle size; 10 μm tip opening). The mobile phase consisted of two components: 0.1% FA (mobile phase A) and 90% acetonitrile with 0.1% FA (mobile phase B), with a flow rate of 250 nl/min. For proteome analysis, the LC-MS/MS method encompassed 110 min and included a 10-min column equilibration step, a 20-min sample loading step, and the subsequent gradient profile: (min: %B) 0:8; 95:28; 100:100; and 110:100. For phosphoproteome analysis, the same LC system and column configuration were used. For better separation of phosphopeptides, the gradient was extended to 180 min to get better separation, using the following profile: (min: %B) 0:5; 165:27; 170:100; and 180:100. Global proteome analyses and phosphoproteome analyses were conducted on the Orbitrap Fusion Lumos mass spectrometer (Thermo Fisher Scientific) using the MS3-based TMT centric mass spectrometer method (73). Data-dependent acquisition was performed in positive ion mode using Xcalibur v4.1 software. The spray voltage was set at 2.2 kV, and the ion transfer tube temperature was maintained at 275°C. MS1 data were acquired in the Orbitrap analyzer, featuring a resolution of 120,000, an automatic gain control (AGC) target of 4 × 105, and a mass/charge ratio (m/z) range from 400 to 1400. The maximum ion injection time was set to 50 ms. Charge states between 2 and 6 were selected, and a 60-s exclusion duration with isotopes excluded was applied. For MS2 analysis, ion trap detection was used, using a rapid scan rate. Quadrupole isolation was performed with a collision-induced dissociation (CID) collision energy set at 35% for 10 ms. The m/z window for precursor selection during sequencing was set to 400 to 1200, incorporating a low mass exclusion width of 18 and a high mass exclusion width of 5. MS3 quantification scans were carried out using SPS (synchronous precursor selection), where 10 precursor fragment ions from the MS2 spectrum were selected. These selected precursor ions underwent fragmentation by HCD (higher-energy collision dissociation) with a collision energy of 65% and were detected in the Orbitrap analyzer. The MS3 scans were conducted at a resolution of 50,000, with an AGC target of 5 × 104 and a maximum ion injection time of 86 ms. The MS isolation window was set to 1.2 m/z, and the MS2 was set to 2.0 m/z. Genomic and transcriptomic data analysis Somatic mutation detection and filtering After excluding low-quality reads, the exome sequencing data were first aligned to GRCh38/hg38 sequence reference using BWA MEM (version 0.7.10) (http://bio-bwa.sourceforge.net/). The resulting BAM files were further processed including marking duplicates with Picard (version 2.22.8) (https://broadinstitute.github.io/picard/) and base quality recalibration with BaseRecalibrator from GATK (version 4.1.7.0) (https://gatk.broadinstitute.org/hc/en-us). SNVs and indels were detected using Strelka2 (74) (version 2.9.x) from paired LCNEC samples with default parameters, then filtered out short tandem repeat region downloaded from the UCSC table browser (https://genome.ucsc.edu/cgi-bin/hgTables), and lastly annotated using Funcotator from GATK (version 4.1.7.0). Deciphering of mutational signatures in LCNEC The r package “Sigminer” (version 2.2.2) (75) was used to extract mutational signatures from the WES data. The Bayesian variant of non-NMF algorithm was used to decipher mutational signatures in cancer somatic mutations stratified by 96 base substitutions in trinucleotide sequence contexts, and the optimal factorization of k value is selected when the magnitude of the cophenetic correlation coefficient begins to decrease apparently. The gradient boosting machine learning technique yields an individual score for each mutational signature, combining likelihood with cosine similarity and exposure of signatures using the non-negative least squares (NNLS) algorithm. After decomposing into four signatures, signatures were compared against validated single base substitution mutational catalogs retrieved from COSMIC (https://cancer.sanger.ac.uk/signatures/), and cosine similarity was calculated to identify the best match. Analysis of significantly mutated genes The significance of mutated genes was evaluated in tumors by MutSigCV (version 1.4) (76) in GenePattern, and mutated genes were deemed significant above the background mutation rate if FDR for the genes was less than 0.1. Comparison of significant mutation frequency in LCNEC with previous studies Combining previously reported WES or targeted sequencing cohorts (10, 12, 19, 20), we found differences in the distribution of the three significantly mutated genes (RB1, KEAP1, and STK11) across the cohorts. TMB and cancer-associated driver mutations TMB was calculated via the R package “maftools” (version 2.16.0). Cancer-associated driver mutations were acquired from DriverDBv3 (21). Estimation of APOBEC enrichment score Deaminase effects of APOBEC were preferentially found in TCW motif, especially resulting in C-to-G or C-to-A substitution. On the basis of these characteristics, APOBEC enrichment score was calculated using the method reported by Roberts et al. (77). Briefly, for each sample, the APOBEC enrichment score was calculated as the ratio of the proportion of C>T and C>G substitutions occurring within TCW motifs to the background frequency of TCW motifs in the 20-bp regions surrounding the mutated bases. Reverse-strand effects were also taken into account in the score estimation. Copy number variation calling Somatic copy number variation (CNV) with discrete copy number segments was determined using the copy number calling pipeline of CNVkit (version 0.9.6) (78). These were further examined using GISTIC (version 2.0.23) (79) to identify gained or lost focal-level and arm-level events as well as CNV at the gene level. Q value less than 0.25 was considered significant. To exclude more false positives, a log2 ratio greater than 0.4 or less than −0.4 was deemed as the threshold for gain or loss, respectively. Default value was used for other parameters. RNA quantification After the removal of adaptors and low-quality reads, RNA-seq cleaned sequence data were aligned to human reference sequence hg38 using STAR (version 2.7.6a) (80). The resulting BAM files were analyzed for gene expression using Salmon (version 1.3.0) (81) against a transcriptomic reference GENCODE (version 22, https://gencodegenes.org). Then, the resulting gene-level counts for protein-coding genes were normalized and transformed into transcripts per million values, followed by applying log2 transformation. Data rows of redundant gene symbols were aggregated by calculating the average expression values. Batch effects were adjusted using the “combat” function in the R package “sva” (version 3.48.0). Proteomic and phosphoproteomic data analysis MS database searching MS raw files generated by LC-MS/MS were searched against the UniProt human proteome database (version 2019-03-07, 20,404 sequences) using MaxQuant (version 1.6.5.0) software enabled with Andromeda search engine. TMT 11-plex–based MS3 reporter ion quantification was chosen with reporter mass tolerance set at 0.003 Da. The precursor intensity fraction (PIF) filter value was set at 0.5 to reduce the interference of precursor cofragmentation. Carbamidomethyl (C) was considered as a fixed modification. In proteomic data analysis, variable modifications were oxidized methionine (M), acetylation (protein N-term), and asparagine and glutamine (NQ) deamidation. In phosphoproteomic data analysis, oxidized methionine (M), acetylation (protein N-term), asparagine and glutamine (NQ) deamidation, and phospho (S/T/Y) were set as variable modifications. The cutoff of FDR by using a target-decoy strategy was 1% for peptides, proteins, and phosphorylation sites. A maximum number of five modifications per peptide were allowed for each peptide. Enzyme specificity was set as trypsin. The maximum missing cleavage site was set as 2. The tolerances of the first search and main search for peptides were set at 20 and 4.5 parts per million (ppm), respectively. The minimal peptide length was set at 7. For each set of TMT labeling data, the purities of TMT labeling channels were corrected according to the kit lot number. Batch effect and data quality analysis for proteomic data The batch effect due to TMT multiplexes was assessed by performing unsupervised PCA on the proteomic and phosphoproteomic data. The leading PCs of the global proteomic and phosphoproteomic data clearly separated the tumor from normal samples, and no obvious batch effect was observed among the 21 TMT batches. The quantification repeatability of longitudinal benchmark samples shows the robust and accurate proteome/phosphoproteome platform (to take the ratios of 127N to 126 as an example, Pearson’s correlation coefficients, 0.77 to 0.82). Data normalization We corrected sample loading differences by normalizing data using a median centering method across total proteins or phosphorylation sites. The log2 TMT ratio for these proteins or phosphorylation sites should be zero in normalized samples. Normalized proteins/phosphorylation sites with SwissProt ID were converted to Human Genome Nomenclature Committee’s HUGO symbols provided by HGNC (https://genenames.org). A total of 8142 proteins and 8932 phosphorylation sites quantified in at least half of the samples were imputed for the missing values using the “RegImpute” algorithm and included in subsequent analyses. The imputation method was implemented using DreamAI tool, an ensemble algorithm developed during the NCI-CPTAC Dream Proteomics Imputation Challenge (version 0.1.0) (82). Integrated analysis mRNA-protein correlation analysis For each gene-protein pair across all 124 LCNEC tumor and NAT samples, Spearman correlation coefficient was calculated and adjusted using the Benjamini-Hochberg (BH) method. A total of 5393 gene-mRNA matched genes were calculated, with a median Spearman correlation of 0.48. A cutoff P value of 0.05 was used to determine the significance of the correlation pairs. Defining CAGs CAGs were compiled using the Census website https://cancer.sanger.ac.uk/census. CNA-driven cis and trans effects Correlations between CNAs and mRNA, proteome, and phosphoproteome were determined using Pearson correlation. In addition, P values (corrected for multiple testing using the BH method) for assessing the statistical significance of the correlation values were also calculated. CNA trans-effects for a given gene were determined by identifying genes with statistically significant (FDR < 0.05) positive or negative correlations. Tumor-NAT differential expression analysis Differential expression analysis was performed for tumors and paired NATs using the Wilcoxon rank-sum test. P values were adjusted using the BH method. Each feature was required to be nonmissing values (n = 5683). Proteins differentially expressed between tumors and NATs (Wilcoxon rank-sum test, FDR < 0.05, |LFC| > 1) were further used for KEGG enrichment analysis. LCNEC-associated protein identification First, we filtered proteins with at least twofold up-regulation in more than 90% tumor-NAT pairs (Wilcoxon rank-sum test, FDR < 0.01, |LFC| > 1, and concordant differential in ≥90% of all 103 tumor-NAT pairs). Second, for the patients with an ESTIMATE score for both tumor and NAT samples, a linear mixed model was used to correct for nonepithelial content. The “lmerTest” (version 3.1-3) package in R was used. The tumor type and z-scored ESTIMATE score were fixed effects and the patient was a random effect. FDR < 0.01 and β value >1.5 for expression were used to filter LCNEC-associated proteins. Then, proteins of a specific type or function including cancer-related genes, enzymes, plasma proteins, and drug targets (Food and Drug Administration–approved drug targets and potential drug targets) were annotated by the Human Protein Atlas (https://proteinatlas.org) (83). Kinase substrate enrichment analysis KSEA was conducted by using the “KSEAapp” R package (version 0.99.0) (84). Substrates of every kinase were collected from the PhosphoSitePlus database and NetworKIN (85–87). A kinase score was given for each kinase based exclusively on the collective phosphorylation status of its substrates and transformed into z-score. For the kinase enrichment analysis, the threshold for significantly enriched kinases was P value <0.05. Prognostic biomarker analysis for LCNEC To identify potential protein prognostic biomarkers, four filter criteria were used: (i) The proteins should be quantified in all samples; (ii) The correlation coefficient between mRNA and protein expression should be >0.7; (iii) the candidate proteins should be differentially expressed between tumor and NATs with adjusted P value <0.01 (Wilcoxon rank-sum test, BH adjusted) and |log2 fold change| > 1 at the protein level; and (iv) we stratified the patients with LCNEC into two groups based on the median expression of each protein. Kaplan-Meier curve with Log-rank test was used to visualize the survival difference and the Cox proportional hazard model was used to evaluate the hazard ratio (HR) for each protein. The candidate proteins should significantly correlate with the OS [Log-rank P value < 0.05, and HR (high/low) > 2 for up-regulated or <0.5 for down-regulated proteins]. Gene set enrichment analysis For GSEA, results for all proteins were ranked by log2-fold change and evaluated with the “GSEA” algorithm (88). “Hallmark” and “KEGG” gene sets were acquired from MSigDb (https://www.gsea-msigdb.org/gsea/msigdb). The NRF2 gene set was acquired from a previous study (14). Stem-like signature was acquired from the Wong et al. study (89) and EZH2 targets were available in the Kondo et al. study (90). We filtered GSEA results based on the criterion of P value < 0.05 and visualized the candidate pathways based on the normalized enrichment score. Immune subtype identification and downstream analysis The abundances of 64 different cell types in LCNEC tumors and NAT samples were computed via xCell (version 1.1.0) (38) (https://xcell.ucsf.edu/) using the RNA-seq data. Consensus clustering was performed using the R package “ConsensusClusterPlus” (version 1.64.0) (91). Specifically, 80% of the original samples were randomly subsampled without replacement and partitioned into three major clusters using the k-means algorithm, which was repeated 200 times. Furthermore, the ESTIMATE algorithm (92) was used to derive the overall immune score and stromal score, and tumor purity for each sample based on RNA-seq data. Immune repertoire and TIDE score estimation We used a computational algorithm named TRUST4 to evaluate the immune repertoire and to extract TCR and BCR CDR3 sequences. Then we used the TIDE algorithm (39) to estimate a TIDE score and the predicted response to ICB. IFN-α and IFN-γ activity calculation To identify lung cancer–specific IFN-α and IFN-γ signatures, GSE5542 (93) was acquired from the GEO database (https://ncbi.nlm.nih.gov/geo/). Differentially expressed analysis between IFN-α or IFN-γ treated samples and untreated samples was conducted. Genes with FDR < 0.05 and LFC > 1 were used to generate a gene matrix transposed (GMT) file for IFN-α/IFN-γ activity score calculation using single-sample GSEA (ssGSEA; https://github.com/broadinstitute/ssGSEA2.0) (94). Regulon analysis We reconstructed transcriptional regulatory networks (regulons) using the R package RTN (95) for 95 patients with RNA-seq data. We investigated 1639 TFs acquired from http://humantfs.ccbr.utoronto.ca/download.php (96). Potential associations between a regulator and all possible target genes were inferred from the expression matrix by Mutual Information and Spearman’s correlation, and permutation analysis was used to remove associations with FDR > 0.01. Unstable associations were eliminated by bootstrap analysis (200 resamplings, consensus bootstrap >95%), and the weakest associations in triangles of two regulators and common target genes were removed by data processing inequality (DPI) filtering (tolerance = 0). For calculating regulon activity profiles across a cohort, we required a regulon to have at least 15 positive and 15 negative targets. Regulon activity scores were calculated by two-tailed GSEA. The direction of TF-target gene regulation (positive or negative) was inferred using Pearson correlation. TF-target edge P values were calculated by permuting the Mutual Information matrix 1000 times, retaining edges with FDR < 0.01. The network was bootstrapped 200 times and TF-target edges were found in 95% of the bootstrap samples retained. Last, indirect TF-target edges were removed using the DPI filter with a tolerance of 0. Unsupervised proteomic clustering using NMF We conducted non-NMF–based proteomic clustering in LCNEC. Briefly, given a factorization rank k (where k is the number of clusters), NMF decomposes a p × n data matrix V into two matrices W and H such that multiplication of W and H approximates V. Matrix H is a k × n matrix whose entries represent weights for each sample (1 to N) to contribute to each cluster (1 to k), whereas matrix W is a p × k matrix representing weights for each feature (1 to p) to contribute to each cluster (1 to k). Matrix H was used to assign samples to clusters by choosing the k with maximum score in each column of H. For each sample, we calculated a cluster membership score as the maximal fractional score of the corresponding column in matrix H. To remove uninformative features from the dataset before NMF clustering, we removed features with the lowest SD (bottom fifth percentile) across all tumor samples. Each column in the data matrix was further scaled and standardized such that all features from different data types were represented as z-scores. The resulting data matrix of z-scores was converted to a nonnegative input matrix required by NMF as follows: 1. Create one data matrix with all negative numbers zeroed. 2. Create another data matrix with all positive numbers zeroed and the signs of all negative numbers removed. 3. Concatenate both matrices resulting in a data matrix twice as large as the original, but containing only positive values and zeros and hence appropriate for NMF. The resulting matrix was then subjected to NMF analysis leveraging the “NMF” R package (version 0.26) (97) and using the factorization method described in the Brunet et al. study (98). To determine the optimal factorization rank k (number of clusters), we calculated cophenetic correlation coefficient measuring how well the intrinsic structure of the data was recapitulated after clustering across 50 iterations. The optimal k was defined as the maximum of cophenetic score for cluster numbers between k = 2 and 7. Having determined the optimal factorization rank k, to achieve robust factorization of the proteomic data matrix V, we repeated the NMF analysis using 200 iterations with random initializations of W and H and performed the partitioning of samples into clusters as described above. CMAP analysis Differentially expressed genes among the LCNEC proteomic subtypes were identified using the Wilcoxon rank-sum test. The identified differentially expressed genes and proteins were then filtered for gene symbols measured in the L1000 assay (978 landmark genes). These NMF-specific signatures were used as input to calculate connectivity scores against the LINCS L1000 perturbation-response signatures using CLUE (https://clue.io). Inferred HDACi therapy response scores The HDACi response signature was acquired from a previous study (50), and the response score was calculated via ssGSEA (https://github.com/broadinstitute/ssGSEA2.0). NE score inference LCNEC NE score was calculated based on a gene set of 25 neuroendocrine and 25 non-neuroendocrine genes established in the Zhang et al. study (99). For each sample, the NE score was calculated with the following formula: NE score = (correl NE − correl non-NE)/2, where correl NE or non-NE was the Pearson correlation between expression of the 50 genes in our LCNEC sample and the mRNA expression of these genes in the NE or non-NE cell line group. The resulting NE score ranged from −1 to +1, which indicated non-NE phenotype to NE phenotype. Identification of differentially regulated events in PI3K-NEU mutant tumors To identify C3 subtype mutations, Fisher exact test was conducted and mutations with OR > 5 and P < 0.05 were identified as the C3 subtype-specific mutations. After performing enrichment analysis and literature checking for these C3 subtype mutations, we found that “STK11”, “KRAS”, and “RELN” were enriched in PI3K pathway and 14 mutations were related to neurodegeneration diseases. Then, we combined these two lists into a PI3K-NEU mutation list and assigned LCNEC tumors into “PI3K-NEU mutated” (n = 48) or “PI3K-NEU wild” (n = 59). The PI3K-NEU signature contains a total of 54 genes up-regulated in PI3K-NEU mutant samples both at the protein and mRNA level (P value <0.05 and LFC > 1.5). These 54 genes were used to generate a GMT file for PI3K-NEU activity score calculation using ssGSEA (https://github.com/broadinstitute/ssGSEA2.0) (94). Differential expression analysis for IGF2BP3 protein expression The proteomic data filtered with no missing values were used as input data for differential expression analysis. The expression of IGF2BP3 in different tumor samples is sorted. We define the top 30% of samples as high expression, and the bottom 30% as low expression. Wilcoxon rank-sum test was used to examine whether proteins were differentially expressed between the high-IGF2BP3 group and the low-IGF2BP3 group. TF plot A list of human TFs with DBD annotation was retrieved from http://humantfs.ccbr.utoronto.ca/download.php (96). Kinase plot A list of human kinases was retrieved from the human kinome project at https://science.org/doi/10.1126/science.1075762 (100). MitoCarta 3.0 analysis Version 3.0 of the human MitoCarta was retrieved from https://broadinstitute.org/mitocarta/mitocarta30-inventory-mammalian-mitochondrial-proteins-and-pathways. Complexes I to IV lists from MitoCarta were further analyzed (101). Identification of high-potential candidate samples for targeted therapies For each of the three targeted therapies, i.e., RTK inhibitors, HDAC inhibitors, and complex Ii, we proposed samples to be high-potential candidates for response to these three therapeutic strategies if they have both a high level of pathway activity targetable by the drug and a high level of candidates reflecting the pathway activity. For RTK inhibitors, the pathway and candidates are AXL, PDGFRA, and PDGFRB. For HDAC inhibitors, the pathway and candidates are HDAC sensitivity score and HDAC1 to 6 protein expression, respectively. For complex Ii, the pathway and molecular signature is the PI3K-NEU score in the protein level. We required that the level of both pathways and molecular signatures from high-potential samples be significantly higher than the rest of tumor samples (i.e., low-potential candidates). For each feature in the molecular signatures and pathways mentioned above, we performed univariate k-means clustering to cluster samples into low or high groups (k = 2 for HDACi, k = 4 for RTKi, and k = 3 for complex Ii) using the “Ckmeans.1d.dp” R package (102). Samples were assigned to the “high potential” group if they belonged to the “high” group based on the pathway clustering and simultaneously belonged to the “high” group based on the clustering of at least one of the molecule signatures. Genomic subtyping According to the conclusion in the Rekhtman et al. study (10), LCNECs characterized by TP53 and RB1 comutation were assigned to SCLC-like LCNEC; LCNECs characterized by the lack of co-altered TP53 and RB1 and nearly universal occurrence of NSCLC-type mutations (STK11, KRAS, and KEAP1) were assigned into NSCLC-like LCNEC. The remaining LCNECs were assigned to other categories. RNA subtyping According to the results of the George et al. study (12), 2271 subtype-specific signatures were derived for LCNEC type I, type II, and type SCLC using data from the George et al. LCNEC cohort. Then, we performed NMF clustering in the TU-LCNEC cohort using transcriptome data of these subtype-specific signatures, and 95 tumor samples were assigned to type I, type II, or type SCLC, separately. Defining phosphoproteomic subtype in the pathway level The phosphorylation sites from the same phosphoproteins were collapsed by calculating the median ratio, and then samples with less than 50% missing values in phosphoprotein were performed, resulting in a dataset of 3136 phosphoproteins. In addition, a total of 653 differentially expressed phosphoproteins were identified by Wilcoxon rank-sum test with statistical significance [FDR < 0.05 and LFC > log2(1.5)]. The phosphoprotein dataset of these differential proteins across tumor samples was subjected to ssGSEA to obtain enrichment scores over the MSigDB c2 pathway database with at least five overlapping genes (GSVA R package, version 1.36.3) (94). Next, the pathway-mapped phosphoprotein data were clustered into three robust groups, using k-means consensus clustering (“ConsensusClusterPlus” R package, version 1.52.0) (91) [number of repetitions = 1000 bootstraps; pItem = 0.8 (resampling 80% of any sample); pFeature = 1 (resampling 100% of any genes); a range of clusters between k = 2 and 8]. Survival analysis OS was measured from the date of surgery to the date of last follow-up or death. We used Kaplan-Meier analysis to explore survival differences associated with TMB status, SCNA status, immune subtypes, and proteomic subtypes. For continuous variables, including mRNA, protein, and phosphosites expression, “survmier” R package (version 0.4.9) was used to determine the optimal cutpoint for selected samples. For protein biomarker selection and validation, samples were stratified based on the median protein abundance. Log-rank tests and Cox proportional hazards regression for statistical significance were implemented via the “survmier” (version 0.4.9) and “survival” (version 3.5-5) R packages. Functional experiments FBP1 experiment in the H727 cell line We performed proteome profiling for 15 cell lines including 2 SCLC cell lines, 1 carcinoid cell line, and 12 NSCLC cell lines. NE score for each cell line was calculated as defined in a previous study (99). The NE score of a carcinoid cell line (H727) was as high as those of SCLC cell lines (table S3), and thus, H727 was chosen to validate the function role of FBP1. Construction of H727-FBP1 stable cell lines Full length human FBP1 genes were amplified from cDNA and cloned into gateway cloning vector plenti6/V5/DEST (Thermo Fisher Scientific no. V49610). Empty pLenti6 vector was used as control for following lentivirus packaging and stable cell line construction. VSV-G and delta-8.9 vectors were cotransfected for packaging lentivirus in the 293FT cell line by using Lipofectamine 3000 reagent. Lentivirus-containing medium was collected after 48 hours and was used to infect H727 supplemented with polybrene (8 μg/ml). After infecting the H727 cell lines with pLenti6-FBP1-V5 and pLenti6-V5 (empty vector) lentiviruses, stable cell lines were generated using blasticidin selection at a concentration of 5 μg/ml. Immunoblotting assay For SDS–polyacrylamide gel electrophoresis (SDS-PAGE), cells were lysed with 1% Triton X-100 buffer (Beyotime, ST1723) on ice for 30 min and then centrifuged at 12,000 rpm for 10 min to remove cellular debris. The supernatants were boiled in 5× reducing sample buffer (Epizyme, LT101) for 10 min. Proteins on polyvinylidene fluoride (PVDF) membranes were immunoblotted with the antibodies targeting FBP1 (Absin, abs159391) and GAPDH (Abclonal, AC002). Immunoblots were visualized by a chemiluminescence imaging system (Tanon). Cell proliferation assay H727-pLenti6-EV/FBP1 cells were maintained in RPMI 1640 medium supplemented with 10% FBS and blasticidin (5 μg/ml). Cells in logarithmic growth phase were seeded into a 96-well plate at densities of 1 × 103 and 2 × 103 cells per well, with six replicates per condition. After overnight attachment, 10 μl of CCK8 was added to each well. Absorbance at 450 nm was measured after 2.5 hours, recorded as day 0. Subsequently, 10 μl of CCK8 was added to each well daily at the same time, and absorbance at 450 nm was measured after 2.5 hours, with data recorded from day 1 to day 5. The cell growth curve was plotted with time on the x axis and absorbance on the y axis. In vivo tumorigenesis assay H727-pLenti6-FBP1 or H727-pLenti6-EV stable cells (1 × 106 cells per mouse) were inoculated into the right flank of female nude mice aged from 6 to 8 weeks (GemPharmatech Co., Ltd., Nanjing, China). Tumor size was measured with calipers and was calculated using the formula: Volume (V) = 0.5 × length × width2. Xenograft mouse model and experiments for IL-33 Purification of IL-33 protein and PEG modification To produce rIL33 protein, mouse IL-33 (amino acids S109-I266) was cloned into the pMAL-C5x-6xHis vector, with a GGGGS spacer and tobacco etch virus (TEV) protease cleavage site. The resulting pMAL-6xHis-MBP-TEV-IL33 vector was transformed into Escherichia coli BL21 (DE3) strain. Cultures in LB media were grown at 37°C until the OD600 (optical density at 600 nm) reached 0.8 to 1.0. Subsequently, the temperature was lowered to 16°C, and induction was initiated by adding 0.5 mM isopropyl-β-d-thiogalactopyranoside, which allowed the expression of the 6xHis-MBP-IL33 protein for 20 hours. Following this, the E. coli culture was harvested for purification. Ni-NTA resin was used in the initial purification step. The eluent from the Ni-NTA resin was then loaded onto amylose resin. After washing with a wash buffer (20 mM tris-Cl and 100 mM NaCl, pH 7.5) for five column volumes, proteins were eluted using the same buffer containing 10 mM maltose. To remove the 6xHis-MBP fragment, TEV protease was used to cleave the 6xHis-MBP-IL33 fusion protein. Subsequently, Ni-NTA resin was used to deplete the 6xHis-MBP fragment. The final purified IL-33 protein was assessed for purity using SDS-PAGE. Before the animal experiments, we conducted the endotoxin testing to ensure that the final endotoxin concentration was less than 0.01 EU per 1 μg of the protein, using the Limulus amebocyte lysate (LAL) method. Recombinant human IL-33 proteins containing various cysteine mutations were modified using Methoxy PEG5000 (PEG5K) or PEG20000 (PEG20K) Maleimide (JenKem, Beijing, China). PEG5K and PEG20K therefore refer to the 5- and 20-kDa PEG reagents, respectively, used for protein modification. Usually, the longer PEG will have a longer half-time in the circulation system. The conjugation reaction was performed in phosphate buffer at pH 7.2 for 2 hours with a reaction ratio of PEG to protein set at 2:1. The PEGylated IL-33 was further purified with diethylaminoethyl (DEAE) ion exchange chromatography. Syngeneic tumor models and treatments For subcutaneous tumors, murine LLC cells (1.0 × 106 cells per mouse) or CT26 cells (4 × 105 cells per mouse) were inoculated into the right flank of female C57BL/6J mice aged from 6 to 8 weeks (GemPharmatech Co., Ltd., Nanjing, China). Starting from day 6 following tumor inoculation, mice were administered daily intraperitoneal treatments of recombinant mouse IL-33 at a dosage of 0.25 mg/kg, which was dissolved in sterile saline. This treatment regimen was sustained for 10 consecutive days. Tumor measurements were performed every 2 days using calipers, and tumor volume was calculated using the formula: Volume (V) = 0.5 × length × width2. Immunohistochemistry of xenograft tumor tissues Tissues were fixed in 4% paraformaldehyde, embedded in paraffin, cut into sections, and placed on adhesion microscope slides. Sections were subjected to IHC staining according to standard procedures. The IHC experiments were performed by using the CD45 antibody (Servicebio, GB11066), RORγt antibody (Biorbyt, orb385620), TTF-1 antibody (ZSGB-BIO, ZM0270), SYN antibody (ZSGB-BIO, ZM0246), P40 antibody (ZSGB-BIO, ZA0483), CGA antibody (Abcam, ab11232), FBP1 antibody (Absin, abs159391), PDHA1 antibody (Abcam, ab168379), and PDHA1 (phospho S293) antibody (Abcam, ab92696), respectively. The above primary antibodies were incubated at 4°C overnight followed by using the BOND Polymer Refine Detection Kit (Leica, DS9800) according to the manufacturer’s instructions. Whole slide scanning was performed using panoramic MIDI under a 20× objective lens. Luciferase reporter assay A total of 2 × 106 HEK293 cells were seeded in 100-mm culture dishes (Corning) and cultivated for 48 hours. The cells were then transiently transfected with the pGL4.32[luc2P/NF-κB-RE/Hygro] plasmid using polyethyleneimine for a duration of 24 hours. Transfected cells were inoculated into 96-well plates (1 × 105 cells per well) and cultured for 24 hours. Cells were incubated with various concentrations of proteins for 6 hours in triplicate. After incubation, the cells were collected and lysed for luciferase assays. Graphic representation of data and statistical analysis was performed using Prism8. Stability study in nonhuman primates PEGylated-IL33 (GSSC-PEG) was prepared at gram scale and was used for the study of stability and toxicology in M. fascicularis (cynomolgus monkey). Two animals were given one dosage with 0.4 mg/kg to explore half-life and response of PEGylated-IL33. Animal serum was collected and analyzed with DIA methods to study the half-life and response in cynomolgus monkey. DIA analysis of nonhuman primate serum Serum samples (20 μl) from M. fascicularis were used for DIA analysis. Samples were reduced with 5 mM DTT (Thermo Fisher Scientific, 20291) for an hour at 37°C and alkylated with 10 mM iodoacetamide (Sigma-Aldrich, A3221) for 45 min in the dark at room temperature. Protein was precipitated with methanol/chloroform, then washed with 1 ml of methanol, and resolved with 8 M urea. Digestion was performed with LysC (Wako, 129-02541) for 2 hours and with trypsin (Promega, V511X) overnight, both at a 1:50 enzyme-to-protein ratio and at room temperature. Digested samples were acidified with FA (Sigma-Aldrich, F0507) to achieve a final volumetric concentration of 1% (final pH of 2 to 3) and centrifuged at 1500g for 15 min to clear precipitated urea from peptide lysates. Samples were desalted on C18 SepPak columns (Waters, WAT054925) and lyophilized using a SpeedVac apparatus. After reconstitution with 0.1% FA, 25-μg proteome samples were separated using an EASY-nLC 1200 system with the same gradient and analytical column (180 mm by 100 μm) in-house packed with C18-AQ 1.9 μm C18 resin (Dr. Maisch GmbH, Germany) with a gradient of 5 to 28% solvent B (0.1% FA/80% acetonitrile) in 45 min, at a flow rate of 250 nl/min. MS data acquisition in DIA mode was performed on Orbitrap Fusion Lumos using 30 variable windows covering a mass range of 350 to 1500 m/z. The resolution was set to 60,000 for MS1 and 15,000 for MS2. The AGC was 3 × 106 in MS1 and 5 × 105 in MS2, with a maximum injection time of 100 ms in MS1 and auto in MS2. The normalized collision energy (NCE) was set to 28%. In the analysis of data, search parameters of DIA-NN (version 1.8.1) were set as follows: precursor FDR 1%; mass accuracy at MS1 and MS2 set to 5 and 15 ppm for LC-MS data; scan window set to 0; isotopologues and match-between-runs (MBR) turned on; protein inference at the gene level; heuristic protein inference enabled; quantification strategy set to Robust LC (high precision); neural network classifier double-pass mode; cross-run normalization off. The main search settings were the same with additional settings for in silico library generation as follows: Trypsin/P with maximum 1 missed cleavage; protein N-terminal M excision on; Carbamidomethyl on C as fixed modification; no variable modification; peptide length from 7 to 30; precursor charge 1 to 4; precursor m/z from 300 to 1300; fragment m/z from 300 to 1800. scRNA-seq analysis of xenograft tumor tissues Cell isolation Mouse LLC tumors were collected and mechanically dissociated into single-cell suspensions. Fresh samples were stored in the GEXSCOPE Tissue Preservation Solution (Singleron Biotechnologies, Nanjing, China) at 2° to 8°C immediately after being collected by needle biopsy or bronchoscopy. Before dissociation, tissue samples were washed with Hanks’ balanced salt solution three times, minced into small pieces, and digested in 2 ml of GEXSCOPE Tissue Dissociation Solution (Singleron Biotechnologies) following the manufacturer’s instructions. Briefly, the specimens were digested at 37°C for 15 min with continuous agitation. Tumor-infiltrating leukocytes were positively selected by magnetically activated cell sorting using mouse CD45 MicroBeads (Miltenyi Biotec). Single-cell RNA library preparation and sequencing A 40-μm sterile strainer (Corning) was used to separate single cells after digestion. The cells were then centrifuged at 300g at 4°C for 5 min and cell pellets were resuspended in 1 ml of PBS (HyClone). Cell suspensions were counted with a TC20 automated cell counter (Bio-Rad) to determine cell concentration and viability. The concentration of single-cell suspension was adjusted to 1 × 105 cells/ml in PBS. Single-cell suspension was then loaded onto a microfluidic chip (GEXSCOPE Single Cell RNA-seq Kit, Singleron Biotechnologies) and scRNA-seq libraries were constructed according to the manufacturer’s instructions (Singleron Biotechnologies). The resulting scRNA-seq libraries were sequenced on an Illumina HiSeq X10 instrument with 150-bp paired-end reads. scRNA-seq data processing The matrix of read counts per sample was further analyzed by the “Seurat” R package (version 4.2.0) (103) in R software (version 4.1.2). For each cell, we used four quality control measures. Cells falling into any of the following four categories were excluded: (i) nFeature_RNA < 500 or >6000, (ii) > 15% unique molecular identifiers (UMIs) of mitochondria genes, (iii) nCount_RNA < 500 or >50,000, and (iv) housekeeping score (defined as the sum of the UMIs of three canonical housekeeping genes: ACTB, GAPDH and MALAT1) < 1. To exclude data from droplets containing more than one cell, doublet detection and removal were performed using “DoubletFinder” (version 2.0.3) (104). An expected doublet rate parameter of 0.025 was used, and doublet score thresholds were chosen manually to divide putative singlet and neotypic doublet modes in the score distribution. Predicted doublets were then removed from gene-by-barcode matrices. Subsequently, we performed data normalization, identification of highly variable genes, and PCA with Seurat’s classic workflow. Then, harmony algorithm was used to correct the potential batches among sample resources. Uniform Manifold Approximation and Projection (UMAP) was facilitated for dimension reduction with the parameter “reduction” set as “harmony”. The Seurat “FindNeighbors” and “FindClusters” functions were applied to detect communities and find cell clusters, and then the “FindAllMarkers” function with default settings was used to find markers for each identified cluster. To annotate the identity and lineage of each cell type, the differentially expressed genes from each cluster were analyzed to a previously reported set of cell-type markers. The cell-cell communications and signaling pathways were analyzed using the “CellChat” R package (version 1.1.3). Reanalyzed public data The Sequencing data and Affymetrix 6.0 SNP array data of LCNEC were downloaded from the European Genome-phenome Archive under the accession number EGAS00001000708 and reanalyzed. Statistical analysis Statistical and computational analysis methods for single-omics and multi-omics analyses were mainly described in the respective method sections and figure legends. Statistical analyses were performed using R. The t test or Wilcoxon rank-sum test was used to compare the difference between two groups, and the BH procedure was applied for multiple testing correction. The analysis of variance (ANOVA; one-way or two-way) test was used to analyze the difference in means of multiple groups, and comparisons between all conditions were further tested by post hoc Tukey test. Fisher’s exact test or chi-square test was used to analyze categorical data. Pearson or Spearman’s correlation test was used for correlation analysis. Kaplan-Meier plots (Log-rank test) were used to describe OS. Variables associated with OS were identified using univariate Cox proportional hazards regression models. The experiment results are shown as mean ± SD or SEM. All statistical tests were two-sided unless otherwise indicated.