A prognostic signature for lung adenocarcinoma in patients who have never smoked

Abstract

Understanding tumor cell dynamics can improve prognosis and treatment but remains limited for lung adenocarcinoma in people who have never smoked (NS-LUAD). With RNA-seq data from 684 NS-LUAD and validation in an independent dataset, we identified three subtypes with distinct phenotypic traits and cell compositions. Additional genomic and histological data further characterized the subtypes. ‘Steady’, marked by low proliferation, high alveolar cell fraction, moderate-to-well differentiation, and fewer driver genes’ alterations, is linked to prolonged survival and low immune evasion. ‘Proliferative’ shows high proliferation markers, TP53 mutations, and gene fusions. ‘Chaotic’, with high epithelial-to-mesenchymal transition markers, has the worst prognosis even within stage I tumors. Lacking known molecular or histological characteristics, this aggressive subtype is solely identified by transcriptomic data. A 60-gene signature recapitulates the classification and predicts survival even within subgroups based on tumor stage or known genomic features, emphasizing its potential for improving early-stage NS-LUAD prognostication in clinical settings.

Introduction

Approximately 10–25% of lung cancers occur in people who have never smoked(1,2), constituting one of the leading cause of cancer mortality worldwide(3). The majority of these are lung adenocarcinomas (LUAD)(4). Compared to smoking-related lung cancers, LUAD in people who have never smoked (NS-LUAD) is most common in women and the Asian population, and appears to differ in morphological presentations, molecular features and clinical outcomes(5,6). Many NS-LUAD cases have no identifiable risk factors and remain dramatically understudied.

In a previous analysis(5), we characterized the genomic alterations and evolutionary history of lung cancers in people who have never smoked using bulk whole-genome sequencing (WGS) data. We further expanded the analysis incorporating RNA-seq data, since accumulating evidence indicates that genomic changes alone may not be sufficient to drive tumor initiation and progression(7). In fact, somatic driver mutations have been identified in healthy tissues(8,9) and the presence of rare altered cell states has been identified in tumors with no clear genetic or somatic alterations(10,11). Acquisition of stem cell-like features by reactivating developmental programs contributes to metastasis, therapeutic resistance, and poor outcomes in lung cancer(12–15). Moreover, the tumor microenvironment (TME) is now widely appreciated as a pivotal player during tumor initiation, development, metastasis, and therapeutic resistance. A comprehensive investigation of the cell state plasticity and the complex cellular composition of the NS-LUAD and its TME through transcriptome profiling is critical for understanding tumor evolutionary patterns, improving the prediction of clinical outcomes, and informing treatment decision making.

Previous LUAD gene expression studies were predominantly conducted in patients’ samples with smoking history(16–21) and full transcriptomic sequencing (RNA-seq) was only examined in a few dozens of NS-LUAD per cohort, mostly of Asian ancestry (Supplementary Table S1). Here, we analyzed RNA-seq from 684 treatment-naïve NS-LUAD, including 386 (56.4%) stage I tumors, of both European and East Asian ancestry. With this unprecedented sample size, we identified three gene expression-based subtypes that encapsulate a combination of features, such as cell type composition, histological lineages, morphological characteristics, and genomic drivers. Importantly, we found substantial differences in survival and predicted treatment responses across these subtypes. Further, we developed a 60-gene signature that recapitulates the expression subtype classification and its prognostic capability, which could be easily used in clinical settings, even in the absence of driver gene or detailed morphological data. We validated our findings in an independent cohort of 110 NS-LUAD samples from Chen et al.(22).

Results

The Sherlock cohort

We assembled RNA-seq data from 684 tumor samples, including 610 from the Sherlock-Lung study and 74 from the TCGA-LUAD study, hereafter referred to as the Sherlock cohort (Supplementary Table S2). Paired lung adjacent normal samples were also profiled by RNA-seq from 491 subjects from the Sherlock-Lung study and 7 subjects from the TCGA-LUAD study. The patients’ clinicopathological characteristics are summarized in Supplementary Table S3. The cohort has a median age at diagnosis of 64 years and includes 129 (18.9%) male and 555 (81.1%) female patients, similarly between the Sherlock-Lung and TCGA-LUAD studies (Supplementary Fig. S1A–B). Forty subjects (5.8%) had stage IV disease at the time of diagnosis. In non-mucinous NS-LUAD, histologic grade based on the International Association for the Study of Lung Cancer (IASLC) grading system³² was evaluated in 301 cases with both RNA-seq data and H&E morphological images by a team of lung pathologists, resulting in 34 (11.3%) well, 69 (22.9%) moderate, and 198 (65.8%) poorly differentiated tumors (Supplementary Table S3).

Gene expression pathways do not differ by ancestry/ethnicity

Using WGS-derived genetic ancestry, among the 610 subjects from Sherlock-Lung, 209 are of European ancestry, not Hispanic or Latino, from the US, Canada and Europe (referred to as EUR); 320 are of East Asian ancestry from East Asia, US and Canada (referred to as EAS); 15 are of native American/mixed ancestry from Europe and Canada (referred to as AMR or Mixed); and one is of African ancestry from the US (referred to as AFR). For the subjects with no WGS data, we used self-reported race/ethnicity and identified an additional 27 EUR, 20 EAS, 17 AMR or Mix and one patient of unknown ancestry. Among the 74 subjects from TCGA-LUAD, genetic ancestry was derived from genotyping data, but the geographical location is unknown. This cohort includes 59 EUR, two EAS, three AFR, and ten patients of unknown ancestry. In total, we examined 295 EUR, 342 EAS, 4 AFR, 32 AMR/Mix, and 11 with unknown ancestry or race/ethnicity (Supplementary Fig. S1C). Only 80 genes were differentially expressed between EUR and EAS tumors (the two groups with the largest sample size) and there was no significant difference in gene expression pathways between them (Supplementary Fig. S2). Therefore, we combined all ancestry/ethnic groups in further analyses.

New transcriptomics-based classification identifies three major subtypes in NS-LUAD

We defined gene expression-based subtypes in NS-LUAD using non-negative matrix factorization (NMF), an approach that does not assume lack of correlation or independence and provides results with high interpretability(23). We assessed the stability of the decomposition results and found that a factorization rank of three and four (corresponding to three and four clusters) provides similar robust clustering (Fig. 1A–B, Supplementary Fig. S3, Methods). As we aim to identify the minimum number of classes that could explain the inter-patient heterogeneity, we clustered the cohort into three groups.

Fig. 1. Transcriptomic clusters in the Sherlock cohort.

(A) Consensus matrices for NMF rank 2 to 10. Cophenetic coefficient from consensus matrix (left) and silhouettes from basis, coefficient and consensus matrices (right) were computed from 80 runs for each rank value. (B) Clustering of subtype-specific genes in the three expression subtypes are defined by NMF. The top panel shows cluster assignment, driver mutations and fusions, data set, RNA-based tumor purity, genetic ancestry, sex and S-LUAD subtypes. The right panel indicates the subtype specific genes.

The top genes contributing to the transcriptomics clustering include genes involved in lung development, cell-cell adhesion, epithelial-to-mesenchymal transition (EMT) pathways and immune responses (Supplementary Table S4). We performed pathway enrichment analysis in each subtype using the Ingenuity Pathway Analysis (IPA) (RRID:SCR_008653) and Gene Set Enrichment Analysis (GSEA) (RRID:SCR_003199) (24) (Supplementary Fig. S4–5, Supplementary Table S5, Methods). The first cluster, named here ‘steady’, has a ‘normal-like’ expression pattern compared to other NS-LUAD tumors, with features similar to non-neoplastic tissues, including low expression levels of genes involved in proliferation, cell cycle and cell movement (Supplementary Fig. S4A–B, S5A–B), and low proliferation marker MKI67 (Supplementary Fig. S5C). This subtype was also associated with significant upregulation of alveolar epithelial cell genes (Supplementary Fig. S5D). The second cluster, named ‘proliferative’, shows upregulation of proliferation signatures (e.g., genes regulated by MYC) and pathways involved in cell cycle and stemness features (Supplementary Fig. S4C–D, 5E). The third cluster, ‘chaotic’, shows high expression of genes associated with cell movement, EMT and cell adhesion (Supplementary Fig. S4E–F, S5F–G). The canonical mesenchymal markers CDH2 and TWIST1 were also highly expressed in chaotic (Supplementary Fig. S5H–I). To exclude the possibility that the gene expression differed purely as a consequence of different tumor purities, we repeated the analysis of the association of canonical proliferation and mesenchymal markers with the three clusters within tumors with estimated purity > 0.8 (Methods). The analysis showed consistent results and supported differentially regulated pathways in the NS-LUAD clusters (Supplementary Fig. S6A–C).

Previous studies had defined three gene expression-based subtypes by unsupervised clustering of smoking-related LUAD (S-LUAD), including terminal respiratory unit (TRU), proximal-proliferative (PP), and proximal-inflammatory (PI)(25–27), with biological and clinical significance(17,28). We compared the three NS-LUAD expression subtypes to the previously defined S-LUAD expression subtypes (Supplementary Fig. S7A). We found that 100%, 67% and 46% of steady, proliferative and chaotic subtypes, respectively, were classified as TRU (Supplementary Fig. S7B), while in the TCGA S-LUAD(17), only 35% were TRU. Only 43 samples in this study were classified as PP and were mostly included in proliferative subtype. These results demonstrate that NMF clustering provides a more informative subtyping, and that NS-LUAD whole transcriptome substantially differs from S-LUAD.

Clinical and morphological features of NS-LUAD transcriptomics-based subtypes

Steady tumors were enriched with stage I (74%, Fig. 2A, stage I proportions in steady vs. others: p=2.75×10⁻⁸), while 2/3 of chaotic had advanced stage (proportion of advanced stages vs stage I tumors in chaotic vs. other subtypes, p=1.44×10⁻⁸). Moreover, steady had significantly lower grade than the other tumors, with 53%, 27% and 23% of tumors being well/moderately differentiated in steady, proliferative and chaotic subtypes, respectively (Fig. 2B, steady vs others, p=5.52×10⁻⁶).

Fig. 2. Comparison of clinical and histopathological features across NS-LUAD expression subtypes.

A-C, Comparison of (A) tumor stages, (B) histologic grades, and (C) signet ring cell features across the NS-LUAD expression subtypes. P-values for comparison of morphological features from chi-squared test are shown. D-E, Violin plots of (D) LUAD histology scores and (E) mix-histological lineage scores across the NS-LUAD expression subtype. Mean values are indicated by the yellow lines. P-values from two-sided Mann–Whitney U-test are shown.

We assessed the presence of six histologic/cytologic features in the three subtypes, including lymphoid aggregate, tumor spread through air spaces, and clear cell, morular, mucinous, and signet ring cell (SRC) features (Methods). In line with the previous findings that the SRC feature in lung cancer is associated with high histologic grade and worse survival(29–31), we observed that the chaotic tumors had higher fractions of SRC⁺ tumors (Fig. 2C); they also tended to lack clear cell features (Supplementary Fig. S8). No other features showed consistent association with subtypes, partially due to the limited sample size.

Cell lineage infidelity and cellular plasticity increases from steady to chaotic subtypes

Disruption of lineage integrity has been identified as an important driver of lung cancer initiation and progression(13,15). We analyzed the compositions of cell lineages typical of three histological types - LUAD, lung squamous cell carcinoma (LUSC) and neuroendocrine tumor (NET) - in each tumor using lineage-specific genes defined by single-cell RNA-seq (scRNA-seq) (Supplementary Table S6)(32). We assessed the presence of mix-lineage based on scores from the three histological types, with high scores representing high lineage-mixing features. While all the tumors had higher scores for LUAD compared to LUSC or NET, as expected, steady had significantly higher LUAD scores compared to the other subtypes (Fig. 2D, p=6.84×10⁻²²), indicating a high fidelity to the LUAD lineage. In contrast, chaotic had higher mixed-lineage scores, even though histologically they were LUAD (Fig. 2E, p=2.31×10⁻¹⁸). Interestingly, while the LUAD lineage component slightly decreased with stage in general, among the three subtypes the decreasing trend was observed only in chaotic (Supplementary Fig. S9A–B).

Overall, steady was associated with lower tumor stage and histologic grade, and exhibited higher histological lineage fidelity, while chaotic was associated with advanced tumor stage, signet ring cells, and mixed lineage.

The proportion of epithelial cells and cancer-associated fibroblasts differ across subtypes

Previous studies have identified substantial tissue heterogeneity in lung cancer(15,33–40). Here we estimated the abundances of major cell types by cell deconvolution using bulk RNA-seq data and expression signature genes from a LUAD scRNA-seq study(40) (Methods). Proliferative and chaotic had higher proportions of epithelial cells (Fig. 3A–B, mean=38.33%, p=2.03×10⁻¹² in proliferative vs. others), and fibroblasts (Fig. 3C, mean=21.55%, p=6.88×10⁻⁴⁰ in chaotic vs. others), respectively. An inverse association between fibroblasts and morphological lepidic components was present as expected (Supplementary Fig. S10). The association of chaotic with fibroblasts remained statistically significant after accounting for tumor lepidic features (p=2.57×10⁻¹⁹, Methods)

Fig. 3. Cell type deconvolution in the Sherlock cohort.

A, Summary of the average proportions of cell types in NS-LUAD expression subtypes. B-C, Violin plots of the proportions of (B) epithelial cells and (C) fibroblasts across NS-LUAD expression subtypes. D, Summary of the average proportions of epithelial cell types relative to all epithelial cells. E-F, Violin plots of the proportions of (E) AT2 and (F) basal cells relative to all epithelial cells across NS-LUAD expression subtypes. G-I, Comparison of the relative proportions of (G) COL10A1+, (H) COL4A1+, and (I) non-malignant lung fibroblasts across NS-LUAD expression subtypes. The fibroblast estimates are relative differences in abundances and cannot be compared across cell types. Mean values are indicated by the yellow lines in the violin plots. P-values from two-sided Mann-Whitney U-test are shown.

Among epithelial cells, steady had substantially higher proportions of alveolar AT2 cells (Fig. 3D–E, mean=34.94%, p=5×10⁻²⁹ in steady vs. others), consistent with the observed upregulation of the alveolar cell signature in steady (Supplementary Fig. S5D) and the high score for LUAD lineage (Fig. 2D). Correspondingly, steady was depleted of basal cells (Fig. 3F, mean=3.40%, p=1.09×10⁻⁵³ in steady vs. others). Among the alveolar cell types, the high levels of SFTPB, SFTPC and other markers(40) shows preponderance of AT2 features. For simplicity, we define these cells as AT2 throughout the manuscript; however, some AT1 cells undergoing reprogramming towards AT2 cells may be included in this category.

To dissect the fibroblast cell types in chaotic, we decomposed the data based on a scRNA-seq study of stromal cells in lung cancer(35) using a computational algorithm(41) that allows estimation of relative cell abundances with high accuracy (Fig. 3G–I, Methods). Compared to the other subtypes, we found that chaotic had significantly higher relative abundances of two classes of cancer-associated fibroblasts (CAFs), including the fibroblasts expressing COL10A1 (COL10A1⁺) and fibroblasts expressing COL4A1 (COL4A1⁺), and was depleted of non-malignant fibroblasts. Notably, given the different stage distribution across subtypes, we verified that the association of cell compositions with subtypes remained significant within each stage (Supplementary Fig. S11A–D). The expression of canonical cell type markers confirmed the cell composition across the three subtypes (Supplementary Fig. S12A–J).

Driver genes and fusions strongly differ across transcriptomics subtypes

To verify whether the gene expression-based subtypes harbored mutational driver events and to explore possible therapeutic opportunities in NS-LUAD subtypes, we integrated RNA-seq data with WGS data from 512 tumors, including 499 tumor samples from the Sherlock-Lung study and 13 from the TCGA-LUAD (Fig. 1A–B). Among them, 385 (75.2%) samples harbored non-synonymous mutations or fusions in 7 common oncogenic driver genes in NSCLC (Supplementary Fig. S13). EGFR was the most frequently altered gene (55%), followed by ALK (7%) and KRAS (6%). TP53 mutations were significantly enriched in proliferative (Supplementary Fig. S14A–C and Supplementary Fig. S15A–D, p=4.99×10⁻⁶, FDR=9.73×10⁻⁴), while were highly depleted in steady (p=4.63×10⁻¹⁴, FDR=9.03×10⁻¹²). Whole genome doubling (WGD), often found in association with TP53 mutations across cancers(5), was not enriched in proliferative. ALK fusions, often associated with a tumor aggressive behavior(42), were negatively associated with steady (p=1.73×10⁻⁴, FDR=0.0168). Co-occurrence of ALK fusions and TP53 mutations, known to be associated with poor prognosis and treatment resistance(43,44), was observed in seven proliferative tumors. Although not significantly after multiple testing correction, EGFR mutations were more frequent in steady (p=3.65×10⁻³, FDR=0.315), consistent with our previous observation of a slow tumor growth in EGFR-positive tumors(5). No driver mutations or fusions (e.g., MET exon 14 skipping) were significantly enriched in chaotic, although TP53 mutations and WGD were frequently found.

We further investigated gene fusions beyond driver events, most of which were validated with WGS data (Supplementary Table S7). A total of 11,947 fusion events were identified using RNA-seq in 638 out of 684 tumors, including 64.6% intrachromosomal and 35.4% interchromosomal fusions (Supplementary Fig. S16A, Supplementary Table S7). Substantially more fusion events were detected per sample in proliferative (mean=22.1) compared to steady (mean=13.9) and chaotic (mean=10.7) tumors (p=1.11 × 10⁻¹³, Fig. 4A). In tumors with available WGS data, 54.3% of fusions were supported by structural variants detected by WGS, similarly across the three subtypes (Supplementary Fig. S16B). Among the fusions, 3,422 (28.6%) were detected in pairs of protein-coding genes and 953 (8.0%) were in-frame (Fig. 4B, Supplementary Fig. S16C). We detected 39 recurrent protein-coding fusion events (defined as fusions between pairs of protein coding genes detected in more than one tumor, Fig. 4C). The most frequently occurring in-frame fusions were EML4-ALK (33 tumors) and KIF5B-RET (8 tumors). The PARG-BMS1 fusion identified in five tumors is a novel recurrent fusion event in lung cancer. It involves PARG, whose protein product is important for DNA damage repair(45), and has been previously reported in breast cancer(46). Consistent with previous studies(17,22), genes that were most frequently involved in in-frame fusions include ALK (35 tumors), EML4 (33 tumors), RET (12 tumors) and ROS1 (9 tumors) (Fig. 4D). Tumors in the proliferative subtype harbored significantly more protein-coding and in-frame fusions per sample than the other subtypes (Supplementary Fig. S16D–E), and among them, a higher frequency of in-frame ALK fusions (8.2%) compared to steady (2.3%) and chaotic (1.6%) (Fig. 4E), showing consistency between RNA-seq and WGS-based results.

Fig. 4. Gene fusion events in the Sherlock cohort.

(A) Comparison of the total numbers of gene fusions per sample across NS-LUAD expression subtypes. Mean values are indicated by the yellow lines. P-value from two-sided Mann-Whitney U-test is shown. (B) Summary of types of all detected gene fusions. (C) Summary of frequent gene fusions between pairs of protein-coding genes. The bar plot indicates the number of tumors containing gene fusion events. Colors indicate fusion types. Recurrent in-frame fusions of protein-coding genes are indicated by the star signs(*). (D) Summary of genes most frequently involved in fusions between protein-coding genes. The bar plot indicates the number of tumors in which gene fusions involving the listed genes are detected. Colors indicate fusion types. (E) Summary of genes most frequently involved in fusions between protein-coding genes within NS-LUAD expression subtypes. The bar plots indicate the percentage of tumors in which gene fusions involving the listed genes are detected. Colors indicate fusion types.

Dramatic survival differences across subtypes even within stage I

To evaluate the prognostic power of the NS-LUAD expression subtypes beyond established prognostic factors, we tested mortality risk using multivariable Cox models adjusting for age at tumor diagnosis, sex and tumor stage in all Sherlock subjects and within subgroups defined by molecular and clinical classifiers (Fig. 5A–D). We analyzed the 10-year overall survival rates for all subtypes and further analyzed the 5-year overall survival rates for chaotic due to its association with advanced tumor stage. We found that steady had better survival (HR=0.43, 95% CI=0.3–0.63, p=1.3×10⁻⁵), while chaotic had worst survival (HR_10-year =1.4, 95% CI=0.99–2, p=0.056; HR_5-year =1.5, 95% CI=1.0–2.3, p=0.031). Given the high prevalence of early and late tumor stages within steady and chaotic, respectively, we restricted the analyses to stage I tumors only (n=374) and confirmed the subtype associations with survival (Steady: HR=0.39, 95% CI=0.22–0.68, p 8.8×10⁻⁴; chaotic: HR_10-year =2.5, 95% CI=1.3–4.7, p=0.0043; HR_5-year =3.7, 95% CI=1.8–7.3, p=2.2×10⁻⁴). Even within stage IA tumors (n=249), the chaotic subtype remained significantly associated with poor survival (Supplementary Fig. S17A), indicating that the transcriptomic profiling captures subgroups of stage I tumors that are more aggressive and may benefit from ad hoc treatments. To account for the potential impact of adjuvant treatment, we further evaluated prognostic associations in patients with stage I tumors harboring no EGFR mutations or ALK fusions, or with no clinical records indicating treatment with tyrosine kinase inhibitors (TKIs). In this ‘non-TKI’ subset, the steady subtype remained significantly associated with longer overall survival, whereas the chaotic subtype showed a borderline association with poorer outcomes (Supplementary Fig. S17B).

Fig. 5. Association between NS-LUAD expression subtypes and overall survival in the Sherlock cohort.

Kaplan-Meier (KM) survival curves for overall survival stratified by NS-LUAD expression subtypes in (A) all patients, (B) patients with stage I tumors, (C) patients with tumors classified as TRU subtype²⁶, and (D) patients with TP53-wt tumors. P-values and hazard ratios (HR) were calculated using Cox proportional hazards models adjusted for age, sex, and tumor stage. For analyses restricted to stage I tumors, stage was not included as a covariate. The number of samples in each subgroup is indicated on the KM curves. Unless otherwise specified, p-values are based on 10-year overall survival; for comparisons involving the chaotic subtype, 5-year overall survival rates were used owing to its association with advanced tumor stage.

To test whether the NS-LUAD subtypes add significant prognostic information to the previously defined S-LUAD subtypes, we restricted the analyses to the 303 tumors that were classified as TRU based on the S-LUAD classification. Even within the TRU tumor group, we could clearly see consistent prognostic differences across transcriptomic-based subtypes although the statistical significance was borderline for chaotic likely due to the small sample size (n =56) (Steady: HR=0.53, 95% CI=0.35–0.8, p=0.0025; chaotic: HR_10-year =1.5, 95% CI=0.85–2.5, p=0.17; HR_5-year =1.8, 95% CI=0.99–3.4, p=0.054). Moreover, clinical studies have suggested that non-small cell lung cancers with TP53 alterations have worse prognosis(47,48). As the TP53 mutations were enriched in proliferative and depleted in steady, we analyzed 169 tumors with no TP53 mutations to verify whether the distinction in survival between proliferative and steady was solely driven by the TP53 mutations. We found that tumors in proliferative were still associated with shorter survival than steady (HR=2.1, 95% CI=1.2–3.6, p=0.011).

Furthermore, analysis of the relapse-free survival of the three subtypes showed consistent results in all tumors and within subgroups of stage I, TRU and TP53-wt patients (Supplementary Fig. S18A–D).

Predicted response to immune checkpoint blockade in the steady subtype

Compared to LUAD from patients with smoking history, NS-LUAD are thought to be less sensitive to immune checkpoint blockade (ICB) due to differences in tumor mutational burden (TMB) and immune microenvironment, including PD-L1 expression(49) – two established markers predictive of ICB responses(50–55). We tested whether transcriptional classes could identify NS-LUAD patients that could potentially benefit from ICB therapy. We predicted the responses to ICB using the Tumor Immune Dysfunction and Exclusion (TIDE) algorithm(56) in 498 tumors with paired tumor and normal RNA-seq data. A lower TIDE score predicts a better ICB response and has been shown to be more accurate than PD-L1 levels or mutation load(56). Surprisingly, steady had lower TIDE scores than the other subtypes (Supplementary Fig. S19A, p=6.21×10⁻¹⁴). Compared to the paired normal samples, most NS-LUAD tumors (77.1%) had low levels of cytotoxic T lymphocytes (CTL) across all three NS-LUAD subtypes (Supplementary Fig. S19B). Among them, steady had low TMB compared to other subtypes (Supplementary Fig. S19C), consistent with its lower proliferative activity. However, steady had lower abundances of suppressive cells prohibiting T cell infiltration, including myeloid-derived suppressor cells (MDSCs, p=4.77×10⁻¹⁹) and cancer-associated fibroblasts (CAFs, p=1.44×10⁻¹⁰), which are known mechanisms for tumor immune evasion through T cell exclusion(57,58) (Supplementary Fig. S19D–E). These findings are in line with the TIDE score results but warrant further validation.

Since ICB response rate is approximately 20% in unselected non-small cell lung cancer patients(59) and is unknown in NS-LUAD, we focused on the bottom 20% of the TIDE scores. Steady was enriched (Supplementary Fig. S20A–C, p=1.09×10⁻⁹) and chaotic was depleted (p=3.58×10⁻⁴) in this lowest TIDE quintile. We evaluated the association between predicted ICB response based on the TIDE scores and known biomarkers, including TMB and the expression of PD-L1(CD274) and PD-1(PDCD1). As expected, the TIDE scores were positively correlated with TMB and inversely correlated with PD-L1/PD-1 expression. However, TMB and PD-L1/PD-1 expression alone accounted for only a small proportion of the TIDE score variance (R²_PD-L1=0.081, proportion of the variance explained by TMB, p_TMB=1.46×10⁻⁵; R²_PD-L1=0.081, p_PD-L1=1.06×10⁻¹⁰; R²_PD-1=0.037, p_PD-1=1.46×10⁻⁵; R²_PD-L1+PD-1=0.093, p_PD-L1+PD-1=3.87×10⁻¹²; Supplementary Fig. S20A–D, Methods). Incorporating transcriptomic subtype information improved the prediction of ICB response beyond that that achieved by TMB or PD-L1/PD-1 expression alone or in combination (R²_{subtype+TMB+PD-L1+PD-1}=0.266, p=3.00×10⁻¹⁷ for prediction improvement with subtype inclusion, Supplementary Fig. S20D).

A 60-gene signature of NS-LUAD subtypes with prognostic significance

To identify the minimum subset of genes capable of distinguishing the three subtypes for potential clinical application, we applied a centroid-based prediction method in the Sherlock cohort, which estimated that 20 genes for each subtype provide robust results (Supplementary Fig. S21A–B, Methods). To assess the gene set accuracy for predicting subtypes, we adopted a cross-validation approach within the Sherlock cohort and estimated the error rate as 13.2% and 15.1% in the training and testing sets, respectively (Methods). To evaluate the robustness of gene selection, we applied an orthogonal method, the random forest–based recursive feature elimination, which demonstrated strong concordance with the gene signature (Supplementary Fig. S22A–C, Methods). Among the 60-gene signature, genes associated with the steady, proliferative and chaotic subtypes were enriched for lung tissue-specific genes, MYC signaling and cell cycle regulators, as well as genes involved in EMT and myogenesis, respectively. (Supplementary Table S8, Supplementary Fig. S23, Methods).

The 60-gene signature correctly predicted expression subtypes of 595 (87.0%) samples consistently with the NMF classification results based on the whole transcriptome (Supplementary Table S9). The tumors predicted as steady were significantly associated with prolonged overall survival after adjusting for age, sex and tumor stage, and remained significant when restricting the analysis to patients with stage I, TRU or TP53-wt tumors (Supplementary Fig. S24A–D). The association between predicted chaotic tumors and worst survival was also confirmed in all tumors and within subgroups of stage I and TP53-wt patients.

We validated the performance of the 60-gene signature in an independent data set from the Genomic Institute of Singapore (referred to as the GIS cohort)(22) (Supplementary Table S10). Although a much smaller sample size (n=110 samples), the GIS cohort was the largest publicly accessible RNA-seq dataset from lung cancer in people who have never smoked and was limited to the Asian population. In the GIS cohort, the 60-gene signature classified 35, 64 and 11 tumors as steady, proliferative, and chaotic, respectively. Excluding chaotic due to limited statistical power, particularly the small number of deceased patients (n = 4), the 60-gene signature confirmed mortality risk prediction after adjusting for age, sex and tumor stage (c-index=0.682), with steady significantly associated with prolonged survival overall (Fig. 6A, HR=0.32, 95% CI=0.12–0.87, p=0.026) and within TP53-wt tumors (Fig. 6B).

Fig. 6. Association between NS-LUAD expression subtypes and overall survival in the Genome Institute of Singapore (GIS) cohort²².

A-B, Kaplan-Meier (KM) survival curves for overall survival stratified by NS-LUAD expression subtypes in (A) all patients and (B) patients with TP53-wt tumors. P-values and hazard ratios (HR) are calculated using Cox proportional hazards models adjusted for age, sex, and tumor stage. The number of samples in each subgroup is indicated on the KM curves. (C) C-indexes of Cox models predicting the overall survival using different sets of predictors in the training (Sherlock) and testing (GIS) data sets. P-values testing for the effect of adding the expression subtypes to models of single modality were calculated using likelihood ratio test.

To develop an overall prognostic risk score, we trained Cox regression models in the Sherlock cohort using (A) tumor stage (‘Stage model’), (B) driver mutations in TP53, EGFR or KRAS and EML4-ALK fusions (‘Mutation model’), (C) NS-LUAD expression subtypes quantified as the distance to the centroids of the three expression subtypes (‘Expression Subtype model’), (D) histologic grade (‘Histologic grade model’), (E) tumor stage plus mutations/fusions, (F) expression subtype plus stage, (G) expression subtype plus mutations/fusions, (H) expression subtype plus histologic grade, (I) expression subtype plus both stage and mutations/fusions. Each model was tested in the GIS set (Fig. 6C). Since the GIS cohort lacks morphological data, the predictive power of histologic grade compared to other models was assessed based on the performance in the training set. We found that among models of single modality (model A-D), the stage and expression subtype models had similar performance and were superior to the mutation and histologic grade models. Combining more features improved the predictive performance compared to single modality models. Notably, adding transcriptomics-based subtypes significantly improved the predictive power of the models based on stage, mutation, or histologic grade, respectively (Fig. 6C). Across all models, the one combining expression subtype plus stage provided the most accurate prediction of survival in the testing set (c-index=0.68).

Discussion

Understanding tumor cell plasticity and composition and its TME can offer important insights into evolutionary trajectories, prognosis and treatment options. In this study, whole transcriptomic analysis of 684 lung adenocarcinomas from subjects who have never smoked identified three subtypes that encompass tumors’ genomic, clinical, and morphological features. The steady subtype is characterized by low proliferation markers, high proportion of alveolar epithelial cells, low histologic grade and depletion of TP53 mutations and ALK fusions. The proliferative subtype, on the contrary, features upregulation of proliferation and cell cycle pathways and enrichment of TP53 mutations, and ALK and other protein-coding gene fusions. The chaotic subtype is characterized by elevated levels of mesenchymal cell state, disrupted cell-cell adhesion, increased proportions of cancer associated fibroblasts and histologic features of mixed-lineage, but no significant enrichment of driver mutations or fusions, suggesting that the transcriptomic data capture features beyond the known drivers. Notably, this NMF clustering provides a more informative subtyping than the original TCGA transcriptional classification based on 500 genes and highlights large differences between NS-LUAD and from LUAD in subjects with smoking history. These subtypes have superb prognostic capability even in patients with stage 1A tumors, identifying tumors with otherwise unrecognizable aggressive behavior that may require ad hoc treatment, even when the standard of care is surgery, and adjuvant chemotherapy or targeted therapy is not routinely recommended (60–62). We further confirmed this prognostic potential in subjects with no tyrosinase kinase drivers or no evidence of treatment with TK-inhibitors. However, the development and increasing use of targeted therapies may substantially influence NS-LUAD clinical management in the future. The clinical significance of the transcriptomics-based subtype therefore requires evaluation in large cohorts of patients receiving such therapies. Previous studies, mostly on subjects with smoking history, identified gene expression signatures for LUAD prognosis (mostly based on a handful of genes and AI-approaches) and developed models to stratify LUAD patients into different prognostic groups(63–67). Here, the NS-LUAD transcriptomics-based subtypes show robust prognostic performance also within subgroups defined by tumor stage, driver gene mutations, and S-LUAD transcriptional classification, with steady associated with prolonged survival and chaotic with the poorest survival. Remarkably, including the information on the three subtypes into prognostic predictive models strongly improved survival prediction compared to the models using solely stage, grade differentiation or driver gene mutations. The combination of stage information with the expression subtypes provided the best prognostic predictive power. Furthermore, the subtype information significantly improves the prediction of ICB response beyond that based on PD-L1/PD-1 expression, identifying a subgroup in steady with strong predicted response to ICB therapy. This indicates that the transcriptomics-based subtyping could be a tool for the ICB response prediction in NS-LUAD but needs validation in studies with both transcriptomic and ICB response data.

Finally, we demonstrated that the expression of just 60 genes can provide information on the underlining NS-LUAD architecture, encompassing histological, genomic and cell composition features, and to predict survival even within subgroups based on tumor stage or known molecular features. All major findings were validated in an independent cohort exclusively of East Asian ancestry. Validation in larger date sets representative of more heterogeneous genetic ancestries is warranted in future studies. Importantly, like the whole transcriptomics-based subtype classification, the 60-gene signature can discriminate aggressive from more stable tumors even within stage I, providing an easily adoptable tool for treatment decision making in hospital settings. In fact, multiple signature genes characterizing the proliferative and chaotic subtypes are drug-development targets; for example, XPO1 is elevated in proliferative tumors and is being targeted by SINE compounds in ongoing trials for solid tumors, including lung cancer. In chaotic tumors, upregulation of collagen genes (COL1A1, COL3A1) and metalloproteinases (MMP2, MMP14) suggests potential benefits from therapies targeting MMPs or cancer-associated fibroblasts. These findings highlight distinct therapeutic vulnerabilities across subtypes, with potential implications for tailored treatment strategies and future translational research.

Methods

Ethics declarations

The NCI exclusively received de-identified samples and data from collaborating centers, had no direct interaction with study subjects, and did not use or generate any identifiable private information, therefore the Sherlock was classified as “Not Human Subject Research (NHSR)” according to the Federal Common Rule (45 CFR 46; eCFR.gov).

Some tissue specimens were obtained from the IUCPQ Tissue Bank, site of the Quebec Respiratory Health Network Biobank or the FQRS (www.tissuebank.ca) in compliance with Institutional Review Board-approved management modalities. All participants provided written, informed consent.

Some samples and data from patients included in this study were provided by the INCLIVA Biobank (PT17/0015/0049), integrated in the Spanish National Biobanks Network and in the Valencian Biobanking Network, and they were processed following standard operating procedures with the appropriate approval of the Ethics and Scientific Committees.

Sample collection, RNA sequencing, and public data preprocessing

We collected 620 fresh-frozen tumor samples from 610 treatment-naïve lung cancer patients. Among these patients, matched fresh-frozen normal lung tissue samples were also obtained from 491 subjects. All tumors were reviewed by seven lung pathologists using FFPE tissue blocks to determine the histological subtypes and morphological features and were confirmed to be lung adenocarcinomas. RNA-seq was performed using the Illumina NovaSeq6000 platform (RRID:SCR_016387) and Roche KAPA RNA HyperPrep with RiboErase protocol, generating 2×151bp paired-end reads.

The FASTQ files were aligned to the human reference genome GRCh38/hg38 using STAR v2.7.3 (RRID:SCR_004463)(68), and were quantified using HTSeq v2.0.4(RRID:SCR_005514)(69) and GENCODE v35 (RRID:SCR_014966) (70). Quality control was performed at three levels: (1) at FASTQ file level, the raw data was analyzed by FastQC (RRID:SCR_014583 ). We assessed five main QC metrics (base-wise quality, k-mer overrepresentation, guanine-cytosine content, content of N bases and sequence quality) defined by the PCAWG study(18) and excluded samples if three or more metrics failed. (2) At alignment level, we assessed the BAM files using STAR and PicardTools(RRID:SCR_006525) and excluded samples if more than 50% PF base pairs (base pairs that passed Illumina quality filtering) were unmapped or the total reads were fewer than 1 million. (3) At gene quantification level, we excluded samples with fewer than 5 million total counts. For each subject with multiple tumor samples, the sample with the largest number of PF reads (reads that passed Illumina quality filtering) was included in the analysis. The resulting expression data is based on 610 tumor samples and 491 matched normal samples.

For the TCGA-LUAD data set, raw RNA-seq FASTQ files were downloaded from the GDC legacy archive (RRID:SCR_014514). 74 tumor samples and 7 matched normal samples from NS-LUAD subjects passed QC and were included in the analysis. For the GIS data set, raw RNA-seq FASTQ files from 110 tumor samples and 59 matched normal samples from 110 NS-LUAD subjects were downloaded from the European Genome-Phenome Archive (EGA, RRID:SCR_004944, http://www.ebi.ac.uk/ega/) under the accession code EGAD00001004421. All data sets were processed using the same pipeline for alignment, quantification and QC used for the Sherlock cohort.

The expression read counts from the three sets were processed by the ComBat-Seq (RRID:SCR_010974) (71) for batch adjustment followed by TMM normalization using DESeq2(72).

Expression subtype detection

TMM normalized data with median gene levels greater than 1 was sorted by variance. The top 5000 genes with large variance excluding the mitochondrial genes were used for expression subtype detection. Subtypes were identified using the R package NMF (RID:SCR_023124) (73) and the Brunet’s algorithm(74). To estimate the factorization rank (equals to the number of clusters in this analysis) and to assess the stability of the results, we performed 80 runs of NMF analysis for the rank value 2 to 10. The rank of 3 corresponding to three clusters was selected based on the cophenetic coefficient(74). We identified genes specific to the subtypes by computing the basis-specificity. The score for gene $i$ was defined as:

$$S_i=1+\frac{1}{{log}_2\left(k\right)}\sum _{q=1}^{k}p\left(i,q\right){log}_2\left(p\left(i,q\right)\right),$$

Where $k$ is the factorization rank, and $p(i,q)$ is the probability that gene $i$ contributes to cluster $q$, i.e. $p\left(i,q\right)=W(i,q)/{\sum }_{q=1}^{k}W(i,q)$.

The TCGA-LUAD expression subtype was identified as previously described(25). Briefly, the TMM normalized data was the median centered. The 506-gene signature for LUAD was applied by calculating the Pearson’s correlation of each sample to the centroids of the three TCGA-LUAD expression subtypes, respectively. The samples were assigned to the subtype with the highest correlation coefficient.

Differential expression and gene set enrichment analyses

Differential expression analysis was performed using DESeq2 (RRID:SCR_015687) to compare between one subtype versus all the other NS-LUAD tumors and repeated for each of the three subtypes. Genes with FDR < 0.05 and 50% change in expression level constituted the differentially expressed gene sets and were analyzed by IPA (RRID:SCR_008653) to identify the enriched pathways. GSEA (RRID:SCR_003199) (24) was applied to test the gene sets associated with each subtype using the MSigDB (RRID:SCR_016863) hallmark gene sets and an additional 155 curated gene sets associated with lung development(15). To balance sensitivity and specificity in detecting relevant biological pathways, we applied a two-tiered approach. An FDR threshold of 0.1 was used in the initial GSEA to increase sensitivity for capturing potentially meaningful enrichment signals. This was followed by downstream analysis of canonical pathway markers, for which a more stringent threshold of 0.05 was applied to ensure robustness and interpretability of the findings. To exclude the possibility that the differential gene expression was the consequence of different tumor purity, we inferred the tumor purity from RNA-seq data using the R package ESTIMATE (RRID:SCR_026090) (75). The canonical marker genes were analyzed across three subtypes in samples with high tumor purity, defined as the purity estimate greater than 0.8.

Inference of cell type compositions and histological lineages

For major cell types and alveolar epithelial cells, the gene signatures for deconvolution were obtained from a previous scRNA-seq analysis of lung cancer(40). The R package CAMTHC (RRID: SCR_027561) was used to infer cell type abundances because of its high accuracy and robustness in a benchwork study of RNA-seq deconvolution methods(76). The deconvolution results of CAMTHC allows comparison across cell types.

For fibroblast cells, the gene signatures for deconvolution were obtained from a scRNA-seq study of stromal cells in lung cancer(35). The R package Bisque (RRID:SCR_005564) (41) was used to infer cell type abundances for its high accuracy. Of notes, this method estimates relative differences in abundances across samples that cannot be compared across cell types.

Cell lineage scores of LUAD, LUSC and NET subtypes were inferred from the RNA-seq data as described previously(32). Cancer histological type-specific genes were identified from scRNA-seq data(32). The score of each predicted histological type component in each LUAD was defined based on the total expression of genes specific to that histological type divided by the total expression of histological type-specific genes for all three subtypes. The mix-lineage score is defined as:

$$S_{mix}=(1-\max \left(S_{LUAD},S_{LUSC},S_{NET}\right))/\max \left(S_{LUAD},S_{LUSC},S_{NET}\right))$$

Analysis of genomic driver events

504 RNA-seq tumor samples had WGS data available from the same subject. SNVs, SCNAs and global genomic features, including WGD and kataegis, were obtained from previous Sherlock-Lung publications(5,77,78). Fisher’s exact test was used to assess the association between genomic features and expression subtypes.

Analysis of pathology image data

The diagnosis and morphological features of grade differentiation were reviewed by seven pathologists using H&E from the Sherlock-Lung FFPE tumor samples. Each tumor was reviewed by two pathologists independently. If there was not a consensus between the two pathologists, a third pathologist provided the final reading. In addition, six cytologic and immunophenotypic features were analyzed, including lymphoid aggregates, tumor spread through air spaces, clear cell feature, morular features, mucinous features and signet ring features. To ensure both sensitivity and specificity of detection, the presence of features was defined in two ways: (1) ‘OR’ logic: either of the two pathologists reported presence, and (2) ‘AND’ logic: both pathologists reported presence. The proportions of histologic components estimated by two pathologists were considered as consistent if the discrepancies were less than 20% and the mean values of the two estimates were used for the association analysis.

Integrative analysis of cell compositions, expression subtypes and histologic components

In samples with consistent estimates of lepidic components, we evaluated the effect of the lepidic components on the association between cell compositions and expression subtypes. We fitted the generalized linear regression models (GLM) predicting the proportions of fibroblasts using the proportions of lepidic components and expression subtypes. P-values for individual subtype were assessed.

Gene fusion analysis

Fusions were detected from RNA-Seq data using SFyNCS v0.15(79) and STAR-Fusion (RRID:SCR_025853) (80). For SFyNCS, the “duplication_like_and_inversion_like_distance” parameter was set to 300,000 so that duplication-like and inversion-like fusions with distance less than 300kb were filtered out. Default values were used for all other parameters. Fusions from 73 TCGA samples included in this study were obtained from the previous study(79). Additionally, fusions in three other TCGA samples (TCGA-44–2665, TCGA-50–5066, and TCGA-62-A46U) were detected using SFyNCS with the same parameters stated above. Fusion events detected by either software were included in the analysis and were annotated by annoFuse(81).

Survival analysis

We investigated the associations of the three NS-LUAD expression subtypes with 10-year overall survival and additionally 5-year overall survival for the chaotic subtype using multivariate Cox proportional-hazards models adjusting for age, sex, and tumor stage (the stage was not included in the analysis of stage I subgroups). When evaluating the overall effect of the three subtypes, p-values were based on a likelihood ratio statistic, comparing models with and without the three subtypes, adjusting for the same covariates. When testing for the effect of a binary variable, p-values were based on a Wald statistic from Cox regression model. Survival analysis based on the 60-gene signature in the Sherlock data and the GIS data followed the same approach.

Prediction of ICB response

The Sherlock data set was normalized by TPM after batch correction. RNA-seq data from 498 tumors with paired normal samples was normalized based on the paired normal samples. TIDE analysis was performed in the normalized data using the TIDE web application (RRID:SCR_026350) (56) (http://tide.dfci.harvard.edu/). The CTL level was estimated as the average expression level of CD8A, CD8B, GZMA, GZMB and PRF1. When evaluating the effect of the three subtypes on predicting the TIDE score, we fitted the GLM models to predict TIDE scores as a continuous variable, while minimizing the impact of cut-off selection for ICB responders. We used TMB and immune gene expression (i.e. PD-L1(CD274), PD-1(PDCD1) or both) with or without the three subtypes. R-squared values for GLM and p-values based on a likelihood ratio statistic were assessed.

Identification of the 60-gene signature

To identify a minimum gene set for expression subtype prediction, we applied a centroid-based method, Classification of Nearest Centroid (ClaNC)(82). ClaNC ranks genes by standard t-statistics and selects N genes for each cluster (aka, a total of 3*N genes for the three subtypes) to build the classifier. To identify the optimal N, we assessed 1 to 30 genes for each cluster by cross-validation. While the mean cross-validation error plateaued across a range of gene numbers, we selected N = 20 for each subtype (i.e., a total of 60 genes for the three subtypes) because it was the lowest number of genes with the smallest variance of the cross-validation error. (Supplementary Fig. S20A). To evaluate the performance of this approach, we used cross-validation in the Sherlock data by randomly selecting 70% of samples as the training set to build the 60-gene signature. The predicted subtype was tested for error rate in the training set and the remaining 30% samples (testing set). The error rates for training and testing sets were reported based on 100 repeats of cross validation (Supplementary Fig. S20B). The final model was built using all Sherlock data and was applied to the GIS set for external validation. Samples were classified to a subtype based on the nearest centroid.

To validate the robustness of the gene selection, we derived the subtype signatures using random forest-based recursive feature elimination (RF-RFE) for each subtype. Genes associated with each subtype were first identified by univariate analysis (p < 0.001), and recursive feature elimination was then applied to refine the gene sets using the caret R package(83). To assess consistency between methods, we examined whether signature genes identified by the nearest centroid approach were captured by RF-RFE or, accounting for potential gene collinearity, were highly correlated (correlation coefficient R > 0.8) with genes in the RF-RFE sets (Supplementary Fig S21A–B).

To explore the functional and therapeutical relevance of the 60-gene signature, we used Enrichr(84) to identify pathways and cell types enriched for gene signatures associated with each subtype. Significantly enriched pathways from the MSigDB Hallmark database were identified for the proliferative and chaotic subtypes, respectively. The steady subtype had gene expression patterns similar to those of non-neoplastic tissues, thus showed no significant enrichment in cancer hallmark pathways. Instead, it exhibited robust enrichment of gene signatures related to lung tissue development and cell composition. Therefore, we assessed enrichment for the steady subtype using the Human Gene Atlas database.

The risk models were trained with Cox regression in the Sherlock cohort using nine sets of variables, including (A) tumor stage, (B) driver mutations in TP53, EGFR and KRAS, and EML4-ALK fusions, (C) NS-LUAD expression subtypes, (D) histologic grade, (E) tumor stage plus mutations/fusions, (F) expression subtype plus stage, (G) expression subtype plus mutations/fusions, (H) expression subtype plus histologic grade, and (I) expression subtype plus both stage and mutations/fusions. For the metrics of expression subtypes, we used the distance to centroids. Specifically, for each subtype, we calculated the centroids of the 60 genes across all samples in this subtype in the training set. Then for each sample, we calculated the Euclidean distance to the centroids of the three subtypes, respectively. The coefficients of the Cox models were then applied to the GIS data to predict the risk scores. The C-indexes were calculated in the training and testing data sets, separately.

Significance.

The transcriptome of 684 lung adenocarcinomas in people who have never smoked (NS-LUAD) identifies three subtypes with different cellular dynamics, and genomic and morphologic features. A 60-gene signature accurately stratifies subjects for mortality risk, even in stage I, offering a potential clinically applicable tool for treatment decision making in NS-LUAD patients.

Data Availability

The RNA-seq data have been deposited in SRA through dbGaP (RRID:SCR_002709) under the accession number phs003955.v1.p1. The corresponding data and major bioinformatic codes for NMF clustering, molecular subtype prediction, and survival analysis are available through Code Ocean at https://codeocean.com/capsule/6608518/tree/v1 and on GitHub (RRID:SCR_002630) at https://github.com/vivianzhao919/Sherlock-Lung.