Integrated molecular and clinical characterization of pulmonary large cell neuroendocrine carcinoma

Abstract

Pulmonary large cell neuroendocrine carcinoma (LCNEC) is a rare, aggressive lung tumor marked by significant molecular heterogeneity. In a study of 590 patients across two independent cohorts, we observe comparable overall survival across treatment regimens (chemotherapy, chemoimmunotherapy, immunotherapy) without unexpected adverse events. Genomic analysis identifies distinct non-small cell lung cancer-like (NSCLC-like, KEAP1, KRAS, STK11 mutations) and SCLC-like (RB1, TP53 mutations) LCNEC subtypes, with 80% aligning with SCLC transcriptional profiles. Serial sampling reveals stable mutational but shifting transcriptomic landscapes over time. Here we show, elevated FGL-1 (a LAG-3 ligand) and SPINK1 expression in NSCLC-like LCNECs, and higher levels of DLL3 in SCLC-like LCNECs. Immunofluorescence confirms FGL-1 expression in NSCLC-like LCNECs, and H&E slide analyses indicates fewer tumor-infiltrating lymphocytes in LCNECs versus other lung cancers. These findings highlight LCNEC's distinct immunogenomic profile, supporting future investigations into LAG-3, SPINK1, and DLL3-targeted therapies.

Fig. 1. Clinical outcomes of first-line treatment options in Cohorts 1 and 2.

Kaplan–Meier analysis of A overall survival (OS) in Cohort 1, B OS in Cohort 2, and C real-world progression-free survival (rwPFS) in Cohort 1, comparing patients with pulmonary large cell neuroendocrine carcinoma treated with chemotherapy (n = 119 for Cohort 1, n = 47 for Cohort 2), chemoimmunotherapy (n = 81 for Cohort 1, n = 99 for Cohort 2), or immunotherapy (n = 14 for Cohort 1, n = 12 for Cohort 2). Survival distributions were compared using a two-sided log-rank test. D Tornado plot depicting treatment-related adverse events for patients treated with any first-line systemic therapy in Cohort 1 (n = 216). Any grade (right) and ≥grade 3 (left). HR hazard ratio, ref reference. Statistical significance is defined as p < 0.05.

Fig. 2. Genomic blueprint of LCNECs in Cohorts 1 and 2.

A CoMut plot for 85 patients with LCNEC. For each tumor, from top to bottom, the molecular subtype, sex, age at first-line systemic treatment, first-line systemic treatment, and prevalent molecular alterations. B CoMut plot for 373 patients with LCNEC. For each tumor, from top to bottom, the tumor mutational burden (mutations/Mb), LCNEC molecular subtype, sex, age, and prevalent molecular alterations. C Heatmap depicting the genomic driver and transcriptional profile evolution of two temporally different biopsies from four LCNECs in Cohort 1 and five LCNEC patients in Cohort 2^*.IHC-PD-L1 (22c3) positivity ≥1^†. TMB-High >19 mutations (muts)/megabase (Mb). D Scatter plot showing the prevalence of genomic alterations and FDA-approved ICI biomarkers prevalence across NSCLC-like (n = 89) and SCLC-like (n = 136) LCNEC in Cohort 2. A two-sided Chi-Square test was employed with statistical significance defined as p < 0.05. E Bar-and-whisker plot comparing FDA-approved ICI biomarkers prevalence across NSCLC-like (n = 89), SCLC-like (n = 136), and unclassified (n = 148) LCNECs in Cohort 2. A two-sided chi-squared test was employed with statistical significance defined as p < 0.05. ****<0.0001 (p value for unclassified vs NSCLC-like comparison: 000013; p value for unclassified vs SCLC-like comparison: 0.00002). 1L First-line, TF Transcription factor, mut mutation, dMMR Mismatch repair deficient, MSI-H microsatellite instability high, TMB Tumor mutational burden, PD-L1 Programmed death-ligand 1.

Fig. 3. Transcriptomic modeling resolves unclassified LCNECs into NSCLC-like and SCLC-like molecular subtypes.

A Receiver operating characteristic (ROC) curve demonstrating the performance of a support vector machine (SVM) classifier trained to distinguish NSCLC-like from SCLC-like LCNECs based on 2168 transcriptomic features (AUC = 0.98). B Confusion matrix showing classification accuracy within the validation cohort. C Unsupervised UMAP projection of transcriptomic profiles reveals three distinct molecular clusters. D Overlay of classifier-derived labels onto the UMAP demonstrates concordance between predicted subtypes and transcriptomic clustering, enabling reclassification of previously unclassified LCNECs into biologically coherent groups. NSCLC non-small cell lung cancer, Unc Unclassified.

Fig. 4. Comparative analysis of transcriptional subtypes, genomic alterations, and FDA-approved ICI biomarkers in SCLC and LCNEC molecular subtypes.

A Heatmap illustrating hierarchical clustering of SCLC (n = 1643, Caris Life Sciences) and LCNECs (n = 361, Cohort 2) for established SCLC transcriptional subtypes (ASCL1, NEUROD1, POU2F3, and YAP1). B Bar plot showing the distribution of SCLC transcriptional subtypes across LCNECs (n = 361, Cohort 2). The non-parametric two-sided Wilcoxon rank sum test was used with statistical significance defined as p < 0.05. C Comparison between the prevalence of genomic alterations and FDA-approved ICI biomarkers between SCLC (n = 1643, Caris Life Sciences) and SCLC-like LCNEC (n = 136, Cohort 2). A two-sided Chi-Square test was employed with statistical significance defined as p < 0.05. D Bar plot illustrating the prevalence of NSCLC-like genomic drivers and FDA-approved ICI biomarkers between SCLC (n = 1643, Caris Life Sciences) and SCLC-like LCNEC (n = 136, Cohort 2). The non-parametric two-sided Wilcoxon rank sum test was used with statistical significance defined as p < 0.05. The p value for STK11 mutations was 0.003, while p values for both KEAP1 mutations and tumor mutational burden (TMB) were <0.0001. E Comparison of DLL3-transformed gene expression across NSCLC-like LCNEC (n = 84), SCLC-like LCNEC (n = 134), unclassified LCNEC (n = 143), and SCLC (n = 1643). The non-parametric two-sided Wilcoxon rank sum test was used with statistical significance defined as p < 0.05. Dot plots with median values are shown. *<0.05; **<0.01; ****<0.0001. The p value for unclassified versus SCLC-like LCNEC was <0.0001; for unclassified versus NSCLC-like LCNEC, 0.04; and for NSCLC-like versus SCLC-like LCNEC, 0.04. dMMR Mismatch repair deficient, MSI-H microsatellite instability high, TMB Tumor mutational burden, NSCLC non-small cell lung cancer, SCLC small cell lung cancer, LCNEC large cell neuroendocrine carcinoma.

Fig. 5. FGL-1 and SPINK1 are potential vulnerabilities in NSCLC-like LCNECs.

A Volcano plot showing differentially expressed genes between NSCLC-like (n = 89) and SCLC-like (n = 136) LCNECs in Cohort 2. Y-axis displays the −log10 p value derived from a two-sided Kolmogorov–Smirnov test. Genes with a False discovery rate of 5% and an absolute value of the log10 fold change of 0.5. B Heatmap of the top differentially expressed genes identified in (A), applicable to LCNEC and SCLC molecular subtypes. C Comparison of FGL-1 and SPINK1 log-transformed gene expression across LCNEC subtypes: NSCLC-like (n = 19), SCLC-like (n = 16), and unclassified (n = 31) LCNECs, using previously published data from George et al.. The non-parametric two-sided Wilcoxon rank sum test was used with statistical significance defined as p < 0.05. For FGL-1, the p value for NSCLC-like versus SCLC-like LCNEC was 8.4 × 10⁻⁶; for NSCLC-like versus unclassified LCNEC, 0.0003; and for unclassified versus SCLC-like LCNEC, 0.01. For SPINK1, the p value for NSCLC-like versus SCLC-like LCNEC was 1 × 10⁻⁶; for NSCLC-like versus unclassified LCNEC, 8.3 × 10⁻⁵; and for unclassified versus SCLC-like LCNEC, 0.003. D Comparison of relative FGL-1 protein expression across 54 cell lines from various cancer types, using data from the DepMap dataset. E Box-and-whisker plots comparing median FGL-1 expression across 20 cancer types from Caris Life Sciences (n = 125,632 tumor samples). Dashed lines from top to bottom represent median FGL-1 expression in NSCLC-like, all, and SCLC-like LCNECs, respectively. For the box-and-whisker plots, the center line indicates the median, the bounds of the box represent the 25th and 75th percentiles (interquartile range), and the whiskers extend to the minimum and maximum values. Each point represents an individual patient tumor (biological replicate). F GSEA plots showing pathways enriched in FGL-1 high versus FGL-1 low NSCLC-like LCNECs. G Representative immunofluorescence staining of FGL-1 (green) and DAPI (white) in 2 NSCLC-like LCNECs, 1 SCLC-like LCNEC, 3 NSCLC, and 4 SCLC (H). 20× magnification is shown. The experiment was repeated using independent biological replicates (no technical replicates). Dot plot comparing tumor-infiltrating lymphocyte (TIL) counts among patients with lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), small cell lung cancer (SCLC), and large cell neuroendocrine carcinoma (LCNEC). Median values are shown per group. The non-parametric two-sided Wilcoxon rank sum test was used with statistical significance defined as p < 0.05. TIL tumor-infiltrating lymphocytes, LUSC lung squamous cell carcinoma, LUAD lung adenocarcinoma, NSCLC non-small cell lung cancer, SCLC small cell lung cancer, LCNEC large cell neuroendocrine carcinoma. The p value for SCLC versus NSCLC-like LCNEC was 0.005; for LUAD versus NSCLC-like LCNEC, 0.009; and for LUSC versus NSCLC-like LCNEC, 0.006.

Fig. 6. Suggested model for a therapy approach based on expression and subtypes, to be used for testing ideas in future clinical trials.

The dashed line corresponds to a potential therapeutic target. NSCLC non-small cell lung cancer, SCLC small cell lung cancer.