Gene Expression Clustering of Lung Adenocarcinoma Tumors as a Prognosis Tool to Infer Progression-Free Survival

Abstract

Introduction

Lung cancer is the leading cause of cancer-related death. Studies have reported distinct tumor type based on gene expression signatures and attempted to use this information to predict survival, with limited replicability due to tumor heterogeneity, technical artefacts, and cohort size.

Methods

RNA expression profiling was performed on 515 early stage resected lung adenocarcinomas from a single institution. We then used this data set to explore the potential of gene expression to stratify patients into distinct classes predictive of survival using unsupervised clustering and Cox proportional hazards model.

Results

Several clinicopathologic variables, including smoking history, tumor grade, and stage, were predictive of survival. Gene expression clustering analysis supports two distinct groups of tumors as the optimal solution. Despite weak clustering of tumors based on gene expression, clusters were predictive of survival, independent of clinicopathologic variables. We also tested gene expression signatures from the literature, which were again predictive of survival, but not more than pathologic staging. Nevertheless, when restricted to early stage tumors, gene expression clustering had better predictive power than pathologic staging.

Conclusion

We revealed that gene expression is naturally highly variable in tumors and that the optimal number of clusters is closer to two, instead of three as often reported. We argue that although pathologic stage remains an important predictive variable, gene expression-based biomarkers can serve as a useful prognostic tool for early stage lung adenocarcinomas.

Introduction

Lung cancer is the leading cause of cancer-related death. Adenocarcinomas are the most common type of lung tumors accounting for approximately 50% of all cases.^1,2 Although most patients are diagnosed at advanced stages which are usually inoperable, early stage NSCLC is potentially curable by complete surgical resection. Given its prevalence and the limited survival of patients, the scientific and medical communities have continued to seek clinical, pathologic, and molecular factors that can stratify patients into distinct predictive classes and/or that can identify specific patients who may benefit from adjuvant therapy to reduce the risk of relapse.

Different clinical and molecular factors have been evaluated to develop and validate a practical and reliable post-surgery assay that would improve risk stratification compared with a conventional tumor staging approach, and this includes the use of pathologic parameters, such as histology, immunohistochemistry, and genetic markers. Although the TNM staging provides a consistent framework for describing cancer and guiding treatment based on tumor size, regional lymph node involvement, and the presence or absence of distant metastasis,³ patients may respond very differently to therapies due to molecular differences.

Numerous studies using gene expression quantification (including microarrays, reverse-transcriptase polymerase chain reaction, and RNA-sequencing (RNAseq)–based approaches) to stratify tumors into distinct classes have been conducted. For example, studies have reported the presence of three distinct types of tumors based on their expression patterns: namely bronchioid, squamoid, and magnoid.^4,5 However, results are not always consistent and do not always explicitly account for variations due to technical factors, which can influence gene expression. Therefore, it is important to explicitly test the number of clusters based on gene expression values and other clinicopathologic variables.

Beyond the goal of identifying distinct gene expression-based clusters, several studies have identified tumor-based RNA expression signatures associated with patient outcome in NSCLC.6, 7, 8, 9, 10, 11, 12 In one of the most promising studies, researchers from the study of Biswas et al.¹¹ have revealed that although transcriptomic intratumor heterogeneity can confound existing expression-based biomarkers, using signature genes expressed homogeneously within individual tumors provides a good strategy to overcome tumor sampling bias and predict survival independently of clinicopathologic risk factors. However, to date, no gene expression signatures have gained sufficient support to be integrated as standard of care, especially because biomarker signatures typically exhibit relatively small reproducibility in independent validation cohorts.^11,13 As such, there is little evidence that any of the reported gene expression signatures are ready for clinical application.¹⁴

Here, using a large, well-characterized cohort of 515 early stage resected lung adenocarcinomas from a single institution, who underwent bulk RNAseq in a single processing laboratory and sequencing facility, we aim to explore the potential of gene expression to stratify patients into distinct classes predictive of survival. We hypothesized that (1) although heterogenous, gene expression levels derived from bulk RNAseq data can be used to stratify tumors into clusters predictive of survival; (2) gene expression-based clustering offers an independent and complementary signal to currently used clinicopathologic variables; and, finally, (3) gene expression-based clustering remains useful for risk stratification of early stage cancers, which are harder to classify into prognostically relevant substages using current pathologic staging approaches.

Method

Patient Cohort Description

Clinically and pathologically annotated biological samples were obtained from 515 patients in the LORD cohort.^15,16 Patients who underwent curative-intent lung surgery for a lung adenocarcinoma at the Institut universitaire de cardiologie et de pneumologie de Québec—Université Laval (IUCPQ-UL) from 2008 to 2020 were selected. After surgery, lung samples were immediately evaluated and staged by a pathologist. Samples were divided into smaller fragments (approximately 0.5 cm³), placed in 2-mL cryovials, and snap frozen in liquid nitrogen. They were then stored at −80°C at our local biobank until further processed for RNA extraction. The time from surgical removal to storage was less than 30 minutes. All study procedures were approved by the institutional review board of IUCPQ-UL (ethics number 21871). Patients provided written informed consent to participate in the local biobank. Patient data were anonymized and analyzed in accordance with ethical standards.

For all patients, clinical characteristics including age, sex, and smoking status were recorded (Supplementary Table 1). Only patients with available nontumoral and tumor lung tissues were processed, and, in this study, we focused on tumor samples specifically. Nearly all patients were treatment naive (507 patients), whereas eight patients had received neoadjuvant chemotherapy in the months before surgery. Pathology review included pathologic stage (updated to ninth edition), predominant histologic pattern, grade, and tumor size. Finally, we also calculated the tumor largest dimension in either height, length, or width (mm) as a predictive variable. The post-surgery primary outcome analyzed was progression-free survival (PFS), which was calculated as the interval starting at the date of surgery and ending at the date of relapse or death of any cause, whichever occurred first. In the absence of either, the date of the last follow-up was used as the ending point.

Next-Generation Sequencing and Gene Expression Quantification

RNA was extracted from approximately 30 mg of snap-frozen tissue at our institution following the RNeasy Plus Universal Tissue Mini Kit (Qiagen, Hilden, Germany) including DNAseI treatment as per manufacturer’s protocol. Sequencing libraries were prepared using the Illumina Stranded mRNA Prep Ligation Kit (San Diego, CA). Sequencing was performed at Genome Quebec (Montreal, Quebec, Canada) on an Illumina NovaSeq 6000, aiming for a minimum of 75 million paired-end (2 × 100 bp) reads per library. The tumor libraries of 515 patients were sequenced in five batches.

We used a pseudo-alignment approach KALLISTO¹⁷ to match all reads (k-mers unique match) against all transcripts of the genome (GENCODE version 44). Following which, we associated transcripts to genes for gene-level summarization of expression (62,700 genes) and estimated counts and TPM (transcripts per million mapped reads) using the R packages ANNOTATIONDBI¹⁸ and TXIMPORT.¹⁹

Patient Stratification

Unsupervised molecular structuring was conducted using the normalized (variance stabilizing transformation) expression values of the top 25% most variable genes (15,135 genes). Note that we also tested different gene sets (top 1%, 5%, 10%, and 25% of the most variable or most expressed genes). However, because these yielded similar results, we settled on a representative fraction of the most variable genes, as is often done.^8,11 We used a k-means unsupervised clustering approach (R package stats, version4.3.3),²⁰ a widely used partitioning iterative algorithm that minimizes the distance from centroids for a predefined number of clusters k (here k = 2–10 were tested), while maximizing the distance between. We calculated a Silhouette score and the total within sum of squares to assess the optimal number of clusters k as typically done (factoextra, version1.0.7).²¹ To quantify the robustness of this result, we developed a resampling strategy by subsampling 100 genes randomly (with replacement) from the top 25% most variable genes, which we then repeated 1000 times. This enabled us to assess whether the specific genes chosen influenced the final number of clusters identified. We calculated the Adjusted Rand Index (ARI, aricode, version 1.0.3)²² between different clustering results, including clinicopathologic variables to quantify the degree of similarity between clustering methods (ARI of 1 indicates perfect agreement between the clustering results and 0 indicates random agreement). Finally, we conducted Principal Component Analyses using factominer (version2.10)²³ and umap (version0.2.10)²⁴ to analyze visually the clustering of tumors.

All data visualizations presented here were made using, in combination, the R packages dplyr (version 1.1.4),²⁵ ggplot2 (version3.5.0),²⁶ COMPLEXUPSET (version1.3.3),²⁶ and patchwork (version1.3.0).²⁷

Reference Signatures

Through a literature search, we settled on ten gene expression–based signatures that have been reported to dichotomize patient into a low-risk (high survival) and a high-risk (low survival) group or into a three (low, medium, and high) survival group.5, 6, 7, 8, 9, 10, 11^,28 We used the expression values of these signature genes (4–31 genes, with little overlap between signature gene sets; Supplementary Table 2) to cluster patients into two or three groups and test prognostic value (PFS) as compared with our own set of the top 25% most variable genes identified from our current data set.

Comparisons With The Cancer Genome Atlas Data Set

In addition, we also downloaded a data set from The Cancer Genome Atlas (TCGA) of 497 lung adenocarcinoma samples with transcriptome profiling. This gene-expression data set was generated from RNAseq next-generation sequencing (Illumina) and available as normalized counts (transcripts per million mapped reads) (https://portal.gdc.cancer.gov/). This control reference data set was used to repeat the same clustering analyses and test the replicability of the results in an independent cohort.

Cox-Based Models and Hazard Ratios

PFS probabilities were calculated using the Kaplan–Meier method. Differences in survival curves were assessed using the log-rank test. Hazard ratios (HRs) and 95% confidence intervals (CIs) were generated using a Cox proportional hazards model (R packagesurvival version3.5.8).²⁹ Each factor was assessed through separate univariate Cox regressions models with time censored at 10 years.

Specifically, clinical and pathologic variables were categorized as such: age divided into four equal quantiles (<59, 59–66, 66–71, and ≥71 y old); smoking status categorized as non, former, and current smoker; smoking quantity (number of pack-year) divided into four equal quantiles (<23, 23–38, 38–49, and >49 pk/y); tumor largest dimension in length, height, or width divided into four equal quantiles (<22, 22–27, 27–38, and >38 mm); tumor stage (1A, 1B, 2A, 2B, and stages 3 and 4 combined together due to low sample size for stage 4); and tumor grade (grade 1, grade 2, grade 3). Predominant histologic pattern is known to separate adenocarcinoma into the three following prognostic groups: low grade (lepidic predominant); intermediate grade (acinar or papillary predominant tumor); and high grade (solid, micropapillary, or complex glandular patterns [fused gland or cribriform]).³⁰ As such, these were analyzed together and labeled as low (lepidic), intermediate (acinar|Pap), and high (solid|Mpap|CGP). In addition, 25 samples with missing histologic pattern were categorized as unknown, given that these were predominantly mucinous adenocarcinomas, with respectively, four, two, and three tumors demonstrating more than or equal to 5% of fetal, colloid, and enteric components.

Differential Expression Analyses

We quantified the total number of differentially expressed genes (DEGs) between the high- and low-risk groups using a negative binomial distribution (DESeq2 R Package, version 1.40.2).³¹ Finally, we performed enrichment analyses (Gene Ontology Biological Process) using the R package topGO (version 2.52.0)³² to look at concordance in functional terms among DEGs.

Data Availability Statement

All analytical codes used to produce the results of this study are available at https://github.com/Yohan-Bosse-Lab/clustering_gexpr. Gene expression data have been deposited to the European Genome-Phenome Archive under the accession code EGAD50000002405.

Results

Clinicopathologic Variables Are Predictive of Survival

Variables known to affect survival³³ were analyzed in this cohort of 515 adenocarcinomas. Specifically, sex, age, smoking status, and smoking quantity were analyzed in addition to the tumor-specific pathologic variables (tumor largest dimension, stage, predominant histologic patterns, and finally grade). HRs with 95% CI were calculated for these clinical (Fig. 1A) and pathologic (Fig. 1B) variables and plotted on a natural log scale, with sample sizes and statistical significance values. Consistent with previous results for patients with early stage NSCLC, tumor stage was a strong prognostic factor of PFS in the cohort (Fig. 1B), as were pack-year, predominant pattern, and tumor grade. The maximum and most significant HR (1.67) for clinical variables was for heavy smokers (pack year > 49; Fig. 1A) and the stage 3 to 4 group (HR = 7.4; Fig. 1B) for pathologic variables.

Figure 1.

Prognostic value for known risk variables (A: clinical variables sex, age, smoking status, and pack-year; B: pathologic variables stage, predominant feature, grade, and tumor largest dimension). Multivariate Cox analysis was performed and hazard ratios with 95% confidence intervals were found for each predictor and plotted on a natural log scale. Number (n) of samples in each group on the left (numbers do not necessarily sum up to the total number of samples in the cohort, because there were missing values for some variables). The p values for each group compared with the reference group on the right.

RNAseq-Based Gene Expression

On average, samples included in this study had 86.2 M paired-end sequencing high-quality reads (Supplementary Fig. 1A) (mean Phred score = 35.8; Supplementary Fig. 1B). These were aligned to the genome (83.3% reads aligned per sample; Supplementary Fig. 1A and C), and gene expression values were quantified for 62,700 genes, with a mean of 15.9 TPM per gene (median = 0.12) and 9794 genes with a TPM above 10 (Supplementary Fig. 1D). These metrics indicate that we have generated an RNAseq data set of the highest standards to quantify gene expression for tens of thousands of genes in hundreds of samples.

Gene Expression–Based Clustering

Next, we tested for k = 2 to 10 using 100 random genes out of the top 25% most variable genes (Fig. 2). Although the optimal number of clusters was two, the relatively low Silhouette score (Silhouette width = 0.086 for k = 2) implies that the signal of clustering based on gene expression profile was relatively low. In addition, based on the elbow plot in Figure 2B, the support for k = 2 was not clearly distinctive from other k values (no sharp drop in the WSS which usually indicate strong support for a specific k value), although it clearly diminished with increasing values of k. Despite this low signal of clustering, our resampling approach (1000 resamples) consistently identified two clusters as the optimal solution based on the Silhouette score and total within sum of square (red lines in Fig. 2A and B).

Figure 2.

Clustering metrics based on the clustering of all 515 tumors. (A, B) Average Silhouette score and TWSS for k = 2 to 10. Each black dot connected by a red line represents the average Silhouette width or TWSS for each k for each resample (1000 resamples). Boxplot represents the median and interquartile range for all resamples. (C) Adjusted Rand index comparing clustering between k values and several frequently used variables to predict survival. (D–F) Principal component analyses with each sample colored according to its respective cluster identified through K-means analyses (the first two dimensions are found for k = 2, 3, or 4). TWSS, total within sum of square.

In Figure 2D to F, principal component analyses depict the stratification of the patients based on two, three, or four clusters (k = 2, 3, 4). In Figure 2C, the correlation (adjusted Rand index) between the different gene expression clusters (k = 2, 3, or 4) was relatively moderate (more than approximately 0.2), implying that clustering results were correlated. However, it was near zero when comparing k = 2, 3, or 4 with other clinicopathologic variables such as the smoking group, sex, stage, predominant pattern, and grade. This consequently implies that gene expression clustering was largely independent from the clinicopathologic variables. In addition, we also confirmed that samples did not cluster based on a sequencing batch effect (Silhouette score ≈ 0, one-sample t test p value = 1, ARI score with other clustering approaches ≈ 0; Fig. 2C). Finally, the few patients (n = 8) who received neoadjuvant chemotherapy before surgery were not significantly over-represented in any gene expression–based clusters (chi-square test, k = 2, p value = 0.33; k = 3, p value = 0.34; k = 4, p value = 0.07).

The Number of Clusters Compared With Previous Results in TCGA

We repeated clustering analyses using a gene expression set of 497 lung adenocarcinomas from the TCGA cohort and identified a similarly weak signal of clustering with an optimum at two clusters (average silhouette width = 0.0965 for k = 2; Fig. 3A and B). Based on the elbow plot in Figure 3B, the support for k = 2 was not clearly distinctive to other k values, although it also diminished with increasing values of k. In addition, when compared with our own data set of adenocarcinoma, the TCGA cohort did not consistently identify two clusters as the most optimal solution based on the Silhouette score. In fact, based on 1000 resamples of 100 genes, k = 2 was the optimal solution 57% of the time; then k = 3, 9% of the time; k = 4, 31%; and k > 4, 3% of the time. Finally, the correlations (adjusted Rand index) between the different gene expression clustering results (k = 2, 3, 4) were modest (>0.3) but small when compared with the pathologic stage (≤0.02; Fig. 3C). In Figure 3D to F, principal component analyses depict the stratification of patients into two, three, or four clusters.

Figure 3.

Clustering metrics based on the clustering of 497 TCGA lung adenocarcinoma tumors. (A, B) Average Silhouette score and TWSS for k = 2 to 10. Each black dot connected by a red line represents the average Silhouette width or TWSS for each k for each resample (1000 resamples). Boxplot represents the median and interquartile range for all resamples. (C) Adjusted Rand index comparing clustering between k values and pathologic staging. (D–F) Principal component analyses with each sample colored according to its respective cluster identified through K-means analyses (the first two dimensions are found for k = 2, 3, or 4). TCGA, The Cancer Genome Atlas.

Gene Expression Clustering

Patient stratification based on unsupervised gene expression clustering found a distinct group of tumors with low survival for k = 2, 3, or 4 (PFS curves in Fig. 4A–C, respectively). Next, we identified a subgroup of 103 tumors (Fig. 4D) consistently present in the high-risk (i.e., low PFS) cluster (cluster 1 for k = 2, cluster 3 for k = 3, and cluster 3 for k = 4, which correspond to the red survival curves in Fig. 4A–C). We then proceeded to analyze clinicopathologic variables in these 103 high-risk tumors compared with all others (n = 412). We found that they had a distinct pattern of clinical variables (Fig. 4E–J), with later stage tumors, more smokers, more male, more high-grade architecture patterns, higher grade, and larger tumors in the poor prognostic cluster. At the same time, these 103 high-risk tumors could not be identified by a specific clinicopathologic variable alone (Fig. 4E–J), highlighting that gene expression clustering is largely complementary (orthogonal) to the clinicopathologic variables. The 103 high-risk tumors as defined in Figure 4 also appeared to have more samples with a high tumor content, which may be associated with the resection procedure, although this was not statistically significant (chi-square test, p = 0.071; Supplementary Fig. 2).

Figure 4.

Patient survival according to gene expression clusters. (A–C) Survival curves for gene expression clusters (k = 2, k = 3, k = 4), with red curve always depicting the cluster with the highest risk (i.e., lowest PFS). Upset plot (D) identified a high risk group of 103 tumors, invariably found in the high-risk clusters (red curves in A–C). (E–J) Clinicopathologic characteristics for the high-risk group (103 tumors) versus all other (412 tumors) samples.

Differential gene expression analysis and an analysis of enrichment of Gene Ontology terms (biological process) of high versus low group were performed to further characterize the gene expression and biological signal. This analysis revealed a large number of differentially expressed genes (2756 genes at absolute log2 fold change > 1 and adjusted p value < 0.01; Supplementary Fig. 3A) in addition to Gene Ontology biological process mainly associated with the immune response (Supplementary Fig. 3B).

Comparison With Previously Reported Gene Expression Signatures

HRs with 95% CI were calculated based on the gene expression clustering results (two clusters; Fig. 5). Here, unsupervised clustering approach was done using either the top 25% most variable genes or using a list of marker genes from the literature (Method section) previously used to identify high-/low-survival groups. In nearly all cases, the low-survival group had a significantly elevated risk (maximum HR of 2.06 using the gene set identified by Shukla et al.²⁸; Fig. 5). Interestingly, these HRs never surpassed the prognostic values identified with pathologic variables, such as staging (Fig. 1B). We repeated this analysis but using unsupervised clustering to define three clusters and found similar results with HR ranging from 1 to a maximum of 2.30 for cluster 3 derived from the gene set by Shukla et al.²⁸ (Supplementary Fig. 4).

Figure 5.

Gene expression–based predictors. Prognostic value for tumors divided into two groups according to gene expression (top 25% most variable genes from this study) and to different gene expression signatures from the literature. Multivariate Cox analysis was performed and hazard ratios with a 95% confidence interval were found for each predictor and plotted on a natural log scale. Number (n) of samples in each group on the left. On the right, p values for each group compared with the reference group (arbitrarily chosen as the group with the lowest hazard ratio).

Detection of Early Stage Cancer

Finally, the ability to split patients into prognostically informative clusters predictive of PFS was tested for the early stage (stage 1A versus 1B) tumors solely (Supplementary Fig. 5A–C). Then, this was compared with a gene expression–based clustering on the same early stage tumors (Supplementary Fig. 5D–G). The goal here was twofold: test if either or both clustering approaches (based on staging or based on gene expression) were predictive of survival and then test if gene expression–based clustering could divide the tumors in more groups than staging and still be predictive of survival. Biomarker-based stratification of these small tumors is important given that it might identify specific patients more likely to benefit from adjuvant treatment after surgical resection. Dividing this cohort by stage 1A versus 1B (n = 215 tumors for 1A; n = 117 tumors for 1B) was prognostically informative (p value = 0.023) with a HR of 1.9 for stage 1B tumors. In addition, gene expression–based clustering (p value = 0.019) also offers a prognostically informative signal, with finer grained resolution into a greater number of clusters (e.g., four clusters, Supplementary Fig. 5D–G), with a HR of 2.8 for the lowest survival cluster (cluster 4, Supplementary Fig. 5F). These gene expression–based clusters did not have a statistically significant enrichment of stage 1A or 1B tumors (chi-square test, p value = 0.08; Supplementary Fig. 5G).

To further explore the ability of gene expression to split early stage tumors into prognostically informative groups, we quantified survival and HR for substage 1A specifically (stage 1A1, the tumor in the lung is 1 cm or smaller; stage 1A2, larger than 1 cm but not more than 2 cm; and stage 1A3, larger than 2 cm but not more than 3 cm; Fig. 6A–C) compared with gene expression–based clustering (Fig. 6D–G). Although dividing this cohort per substage (n = 2, 51, and 44 for 1A1/1A2/1A3, respectively) was not prognostically informative (p value = 0.89), interestingly gene expression–based clustering provided a significantly finer grained resolution (p value = 0.0096, four clusters). Clusters C1 and C3 were predictors of the PFS risk with a HR of 3.8 and 5.9 (p value of 0.015 and 0.0022, respectively; Fig. 6F). A chi-square test (p value = 0.02) revealed a non-independent relationship between stages 1A1 to 1A3 and clusters 1 to 4 (i.e., cluster 1 was slightly enriched for stage 1A1 and 1A3 tumors; Fig. 6G).

Figure 6.

The ability of substaging to split early stage (stage 1A1, 1A2, and 1A3) tumors into prognostically informative groups (A: Kaplan–Meier plots, B: risk table, and C: hazard ratios), compared with the same tumors clustered by expression (D: Kaplan–Meier plots, E: risk table, F: hazard ratio, G: barplots revealing number of stage 1A1/1A2/1A3 samples in each gene expression–based cluster).

Discussion

We observed that gene expression clustering was consistently predictive of PFS using both gene expression signatures from the literature and our own set of the most variable genes identified in this cohort. In addition, gene expression clusters were mainly independent of clinicopathologic factors known to influence survival.³³ Gene expression typically exhibits large natural variation, and this quantitative variation may play a major role in both a species evolutionary trajectory and the pathogenesis of numerous diseases.³⁴ Biological variation in gene expression is especially large in tumors due to several factors such as tumor purity³⁵ and genomic intratumor heterogeneity¹¹ and explains why gene expression clustering observed here is relatively weak (large within versus among group variance).

The process of partitioning samples into an optimal number of clusters can be performed according to different methodological approaches, but the K-means approach, where each n observation is assigned to the cluster k with the nearest mean (therefore minimizing within-cluster variance), followed by a measure of clustering quality (such as the Silhouette score or the total within sum of square) remains one of the most widely used methods to partition observations into an optimal number of clusters.³⁶ Previously, several studies have stated that, according to gene expression, three classes of tumors appear to be a common observation in lung adenocarcinomas.^4,5,37,38 One of the seminal studies which observed this is the TCGA study from 2014. When we reanalyzed the publicly available TCGA gene expression data set used in that study, but without considering prior knowledge of the optimal number of clusters, we observed a weak signal for two clusters, although the overall signal was only slightly more optimal than for k = 3 or 4. In their study, the authors assigned samples to three predefined clusters labeled terminal respiratory unit (formerly bronchioid), proximal-inflammatory (formerly squamoid), and proximal-proliferative (formerly magnoid), which they described as a hallmark of lung adenocarcinomas. In fact, the cluster definition arises from the work of Wilkerson et al.,⁵ which identified three distinct groups of tumors based on the unsupervised clustering of 442 samples reported by the Director’s Challenge Consortium for the Molecular Classification of Lung Adenocarcinoma.¹³ Interestingly, these 442 adenocarcinomas were collected from six different hospitals, and samples (RNA extraction, microarray hybridization) were prepared at four different sites. Although the authors undoubtedly took all precautions to ensure sample standardization as discussed in their study, it is difficult to rule out any unwanted or unforeseen effects related to the hospital or sample preparation, which may have conflated the actual population clustering effect. Unfortunately, this was not explicitly tested. In contrary, our current cohort was collected from a single site and all RNA were prepared by a single team and sequenced at the same facility, thus avoiding a large source of potentially confounding technical biases. At the same time, it could be argued that this design introduces other biases, so both types of studies (large multicenter collaborations and single-facility data sets) are likely needed.

Several clinical variables were associated with survival (sex, age, smoking). Surprisingly, smoking status was not predictive of survival, as was observed by the TRACERx Consortium.^11,39 This can be explained by the highly heterogeneous nature of this response variable in addition to the fact that many former smokers typically stop smoking during the early diagnosis of the disease, a few weeks/months before surgery. However, smoking quantity (pack-year), which probably reflects more accurately the actual tobacco consumption, was associated with survival, as were pathologic variables typically found to be predictive of survival such as tumor stage, predominant histologic pattern, and grade.

The differential gene expression analysis revealed a large number of differentially expressed genes. This is likely due to several factors. First, we have a large gene expression data set (515 samples) and therefore a lot of statistical power to detect even small changes (i.e., we detect significant differences in expression as small as 5% between high versus low groups). Second, gene expression was used to define the groups (i.e., k-means clustering maximizes differences among groups while minimizing differences within groups), and, therefore, it is expected to find differentially expressed genes. Third, as found in Figure 2A, the formation of two clusters does not depend on the identity of the genes themselves, as repeated analyses using random subsets of genes consistently yield two clusters as the best solution. In fact, previous work has revealed that a high proportion of gene expression signatures based on a random set of genes significantly predict survival probability when assessed in independent validation data sets.^40,41 The corollary of this is that gene expression difference among tumors is not the consequences of a handful of genes, but an overall reprogramming of the cell that has consequences on the whole transcriptome dynamics.

Gene expression clustering was predictive of survival, independently of clinicopathologic risk factors, but never more than pathologic staging itself. This was consistent for all previously published signatures we tested. Biswas et al.¹¹ found similar gene expression clustering results looking at overall predictive power of their ORACLE signature. In addition, they also alluded to the idea that the benefit of gene expression clustering might be more for early stage cancers and that prospective clinical trials should test whether offering adjuvant therapy to early stage, but higher risk patients, could improve postsurgical outcomes. Here, we revealed that based on the gene expression signal, we can distinguish a subgroup of completely resected stage 1 NSCLC with poorer postoperative outcomes. Although relapse risk for stage 1, and especially stage 1A is lower, there is in fact, approximately one in five stage 1 patients with NSCLC who will have disease recurrence and prognosis after this event is poor.⁴² Thus, there is a need to identify stage 1 patients at higher risk of relapse, and gene expression–based clustering remains a promising tool to do this.

In conclusion, we have generated a unique cohort with well-characterized clinical, pathologic, and gene expression values. We revealed that gene expression is naturally highly variable in tumors and that the optimal number of clusters might be two, instead of three as often reported. Moreover, we revealed here that clinicopathologic variables were predictive of PFS, but so were the gene expression clusters, and these signals were mainly independent from one another. Finally, gene expression clustering has the ability to provide a more granular view of the tumors for example by predicting the survival of early stage patients. Further work should likely focus on refining predictions for these specific cases.

CRediT Authorship Contribution Statement

Sébastien Renaut: Conceptualization, Formal analysis, Visualization, Writing - original draft.

Victoria Saavedra Armero: Investigation, Writing - review & editing.

Dominique K. Boudreau: Investigation, Writing - review & editing.

Nathalie Gaudreault: Investigation, Writing - review & editing.

Hanie Abolfathi: Investigation, Writing - review & editing.

Michael Maranda-Robitaille: Investigation, Writing - review & editing.

Fabien C. Lamaze: Conceptualization, Writing - review & editing.

Sabrina Biardel: Investigation, Writing - review & editing.

Andréanne Gagné: Investigation, Writing - review & editing.

Patrice Desmeules: Writing - review & editing, Funding acquisition.

Philippe Joubert: Supervision, Writing - review & editing, Funding acquisition.

Yohan Bossé: Supervision, Writing - review & editing, Funding acquisition.

Supplementary Data

Note: To access the supplementary material accompanying this article, visit the online version of the JTO Clinical and Research Reports at www.jtocrr.org and at https://doi.org/10.1016/j.jtocrr.2026.101008.

Disclosure

The authors declare no conflict of interest.