Deep learning multi-omics integration identifies new molecular subtypes of lung cancer Authors: Bianca Gonda, Derek Wang, Sayed Mehedi Azim, Gabriele Romano, Iman Dehzangi Published: 2026 ABSTRACT AbstractCancer is a heterogeneous disease, with numerous subtypes differing in molecular profiles, risk factors, clinical outcomes, and tumor locations. Lung cancer, the third most diagnosed cancer in the United States, is driven by a complex combination of molecular alterations, which influence tumor behavior and patient outcomes. Large-scale consortia such as The Cancer Genome Atlas (TCGA) have generated comprehensive data sets encompassing genomic, transcriptomic, proteomic, and clinical data that present a unique opportunity to uncover novel subtypes through multi-omics integration. While clustering methods have been used to identify subtypes in lung cancer, many integrations and avenues remain underexplored. Previous studies mostly focused on a single data type for this task, potentially limiting biological insight. To address this issue, we propose a novel analytical pipeline that combines gene expression and somatic mutation data in lung cancer to identify distinct molecular subtypes. We also use an autoencoder deep neural network, as well as model-based clustering for feature extraction and clustering purposes. Integrating gene expression data, representing dynamic tumor processes with somatic mutation data, reflecting invariant genomic characteristics, enables a more comprehensive understanding of the molecular profile. We identify molecularly defined subtypes, enhancing our understanding of lung cancer’s molecular heterogeneity. All source code, figures, latent features, cluster labels, and instructions for data download from the GDC are available at https://github.com/MLBC-lab/LungClust. We believe this serves as a reliable tool that offers valuable biological and clinical insights to advance precision oncology approaches in lung cancer, specifically biomarker discovery, which will aid in patient stratification and treatment response prediction.Supplementary InformationThe online version contains supplementary material available at 10.1186/s13040-026-00563-z.Keywords: Lung cancer, Molecular subtyping, Deep learning, Clinical annotation, Integrative clustering, Gene expression, Somatic mutation, Computational oncology Introduction Lung cancer is the leading cause of cancer-related death worldwide, accounting for more fatalities each year than breast, colon, and prostate cancers combined, with non-small cell lung cancer (NSCLC) accounting for 85% of cases [1]. Despite advances in targeted therapies and immunotherapy, survival outcomes remain relatively low [1]. Cancer is a collection of heterogeneous alterations resulting from a variety of genetic events. Lung cancer specifically is marked by its notable prevalence of heterogeneity, which arises from diverse genetic alterations, transcriptomic programs, and epigenetic modifications, highlighting the need for refined molecular subtyping strategies that can capture these multidimensional differences [2]. Molecular subtyping has transformed the management of several malignancies, such as breast cancer, by identifying biologically and clinically meaningful subgroups that inform treatment selection [3, 4]. In lung cancer, histopathologic classification remains the foundation for clinical classification. However, molecular classification frameworks have also been proposed, including multi-omics and gene expression-based frameworks [5, 6]. Prior gene expression profiling identified distinct, reproducible transcriptional subtypes in lung adenocarcinoma and lung squamous cell carcinoma [7, 8]. Additionally, more recent work has also described gene expression-based subtypes with potential therapeutic relevance [9]. Therefore, because gene expression data reflects the dynamic downstream effects of mutations and gene regulation, whereas mutation data provides a stable record of the tumor’s genomic landscape, integrating these data types offers a more comprehensive view of tumor biology and a promising avenue for discovering subtypes with clinical and therapeutic relevance. Traditional approaches in dimensionality reduction, such as principal component analysis (PCA), have limited capacity to model nonlinear relationships within and between omics layers [10]. Deep learning, particularly through autoencoders (AEs), is a powerful alternative for dimensionality reduction and feature extraction. Autoencoders can learn complex, non-linear representations of input data through an unsupervised neural network framework that compresses high-dimensional information into a low-dimensional latent space [11, 12]. These latent features can then be used for clustering, providing a biological interpretation of each cluster. Current lung cancer subtyping studies have taken a variety of approaches, each with varying reproducibility and reliability [10, 13]. Several recent studies have utilized autoencoder-based frameworks across various cancer types. For instance, it has been shown that leveraging autoencoder frameworks using gene expression data improves performance and robustness compared to traditional clustering methods in breast cancer [14]. In addition, recent models have incorporated more complex architectures, such as convolutional autoencoders, to integrate multiple omics layers for better prognostic prediction [15]. Similarly, multi-branch and attention-based frameworks have been proposed to fuse diverse omics types, such as mRNA, CNV, and proteomics, to enhance classification performance and biomarker identification [16]. Finally, recent studies have also demonstrated the utility of autoencoder-based multimodal integration in non-small cell lung cancer specifically [17]. However, prior work in lung cancer has focused on predictive modeling, while our work focuses on identifying and characterizing molecular subtypes via integrated representation learning using somatic mutations and gene expression data specifically. In this study, we apply an integrative deep learning-based autoencoder approach to a publicly accessible lung cancer cohort from The Cancer Genome Atlas (TCGA) to jointly analyze gene expression and somatic mutation profiles. By combining these complementary data sources, we aim to uncover molecular subtypes that reflect the biological diversity of lung tumors. We hypothesize that this integrative clustering framework will identify distinct subtypes characterized by unique transcriptional and mutational signatures, potentially including variation in immune-related pathways, smoking status, and mutation counts, all features crucial to the development progression of lung cancer [18]. Beyond its immediate biological insights, this work also demonstrates the feasibility and utility of deep learning integration for multi-omics cancer subtype discovery and biomarker discovery, providing a foundation for improved cancer stratification across different cancer types. All source code, figures, and instructions for data download from the GDC are available at https://github.com/MLBC-lab/LungClust. We believe this serves as a reliable tool that offers valuable biological and clinical insights to advance precision oncology approaches in lung cancer, specifically biomarker discovery, which will aid in patient stratification and treatment response prediction. Methods Overview The general architecture of our proposed workflow is illustrated in Fig. 1. Open-access, publicly available data were downloaded from the Genomic Data Commons (GDC) using the GDC tool on December 20, 2024 [19]. Samples that had a primary site and origin site of lung/bronchus were chosen. Masked somatic mutation data and quantified gene expression data specifically were selected for download. After data cleaning, we optimized two autoencoders for each data type and concatenated equal numbers of latent features for a downstream model-based clustering to identify subtypes. We then performed biological and clinical analysis to identify biomarkers of each subtype and interpret clinical correlations. The general pipeline of our proposed model. We first collect the gene expression and somatic mutation from the TCGA database. We then use an autoencoder for feature extraction and reduction. After that, we apply a model-based clustering approach. Finally, we conduct a comprehensive analysis to investigate the attributes of our identified clusters Data acquisition and preprocessing We accessed RNA-Seq and somatic mutation data for all tumor samples with lung as the primary organ. It primarily consists of lung adenocarcinoma and lung squamous cell carcinoma cohorts from the Genomic Data Commons with a small number of cases from different histology (Table S1). Case identifiers were harmonized using the GDC manifest and metadata files to ensure one-to-one correspondence between RNA-Seq and mutation files. Gene expression matrices were downloaded in .tsv format and filtered to include only protein-coding genes normalized to TPM, and log₂-transformed + 1 for all analyses. Somatic mutation data were downloaded as .maf formats and processed into a binary matrix, where each entry indicated the presence (1) or absence (0) of at least one non-synonymous mutation per gene per sample. The following variant classifications were excluded: Silent’, ‘Intron’, ‘IGR’, ‘3\‘UTR’, ‘5\‘UTR’, ‘3\‘Flank’, ‘5\‘Flank’, ‘RNA’, ‘lincRNA’. Samples lacking paired data across modalities were excluded, creating a unified cohort of 299 tumor samples, and 19,938 total protein-coding genes were retained. The final cohort features are in Fig. 2A, and cluster-sample labels can be found in Supplementary File B. A shows the distribution of clinical and demographic features in the data set. The majority of the patients are male, white, current or reformed smokers, and older than 60 years old. 61.8% of samples are categorized as adenomas or adenocarcinomas, and the majority of the samples are stage I or II. B shows the clustering hierarchy tree, with 9 total clusters, with 1 final cluster at the third level and the final at the fourth. C is a UMAP projection of the latent features of each cluster Deep autoencoder integration RNA-seq and binary mutation data were integrated using two parallel autoencoders trained independently for each omics type and subsequently concatenated. For the expression autoencoder, log-transformed TPM values were min–max scaled across samples. The encoder consisted of three dense layers (800, 400, 200 units; ReLU activation) and a symmetric decoder with a sigmoid output. The model was compiled using Adam optimizer (learning rate = 1 × 10⁻⁴) and mean-squared error for 50 epochs. For the mutation autoencoder, binary matrices (0 = wild-type, 1 = mutated) were used directly. Dropout layers (rate = 0.4) were applied to reduce overfitting, and binary cross-entropy was implemented to balance the large proportion of non-mutated entries. The encoder and decoder each contained three dense layers (800, 400, 200 units; ReLU activation). Training used Adam optimizer (learning rate = 5 × 10⁻⁵, decay = 1 × 10⁻⁵) for 50 epochs. Latent features (200 dimensions per -omics type) were extracted from each encoder and concatenated to create a 400-dimensional combined feature matrix for downstream clustering. Latent sizes of 50, 100, 200, 400, and 600 were tested for each -omics layer separately. We chose 200 dimensions as the bottleneck layer for both encoders, as it achieved the most balanced clustering performance across both modalities and was the most consistent among dimension sizes (Supplementary File C). Therefore, it was used for all subsequent analyses. Hierarchical model-based clustering We employed unsupervised clustering on the 400-dimensional combined latent feature matrix using model-based clustering. Specifically, we used a recursive Gaussian mixture model (GMM) based clustering [20]. As a model-based clustering, it is a probabilistic, parametric method. In this method, the model maps the best Gaussian distribution to the latent features to determine the optimal cluster number [21]. At each level, the optimal number of mixture components is chosen by minimizing the Bayesian Information Criterion (BIC) across potential numbers of clusters. This is applied hierarchically/iteratively until the clusters fall below a size threshold. In this study, this led to two clusters of 107 and 192 at the first level, and these clusters split again at the second level, and three out of those four clusters split again at the third level, leading to 9 total clusters of sizes 36, 31, 40, 33, 21, 39, 31, 25, and 43 samples (Fig. 2). Differential expression and enrichment analyses Gene set enrichment analysis was performed using GSEApy, a Python implementation of the original Broad Institute algorithm [22]. Differentially expressed genes for each cluster were ranked by the signed log10-adjusted p-value obtained from pairwise comparisons against all other clusters. Ranked gene lists were generated separately for each cluster and analyzed using the Hallmark gene sets from the Molecular Signatures Database. Ranked gene lists were generated separately for each cluster and analyzed using the Hallmark gene sets from the Molecular Signatures Database [23]. For heatmap visualization, expression values were log2 +1 transformed and genes were ranked within each cluster based on the difference between mean expression in that cluster and the mean expression across all remaining clusters. The top three genes from each cluster were selected, combined, and deduplicated to generate the final gene set for heatmap visualization. Z-scores were computed per gene across cluster means to highlight relative overexpression patterns, and horizontal boundaries were drawn to delineate cluster groups. To investigate the immune environment, relative expression values of immune checkpoint genes (PD-L1, CTLA4, PD-1, PD-L2) were also evaluated for each cluster. Nonsynonymous mutated gene count and mutational signature analysis For each sample, a nonsynonymous mutated gene count is calculated as the total count of mutated genes per case, where a mutated gene is a gene with at least one nonsynonymous mutation. This displays the relative mutational burden across the nine molecular clusters to illustrate the different average per cluster. To characterize driver gene alterations, a per-gene mutation frequency profile was derived by computing the proportion of samples mutated for each gene within each cluster. The top recurrently mutated genes, including TP53, TTN, KRAS, EGFR, MUC16, CSMD3, USH2A, LRP1B, and ZFHX4, were visualized across clusters for comparison. An Oncoplot-style summary was generated directly from the binary mutation matrix. Each cell in the summary visual matrix was colored according to binary mutational status. To further characterize mutational processes, we used the SigProfiler package to determine single-base substitution (SBS) patterns. SigProfiler uses non-negative matrix factorization (NMF) to estimate these signatures [24]. These were then summarized for each cluster, and contributions of established COSMIC mutational signatures (v3.3) were estimated as well [25] . Clinical association and visualization Corresponding clinical data and smoking exposure data were downloaded from the GDC for all lung samples in our cohort. Associations between cluster membership and clinical variables (tumor stage, histology, smoking status, gender) for samples with associated data available were assessed using chi-squared tests. We employed the ‘lifelines’ [26] package in Python to conduct Kaplan-Meier curve analysis and calculate significance. Results We performed integrative analysis of TCGA lung cohorts using paired gene-expression and binary somatic-mutation data. General cohort statistics are found in Fig. 2A. In order to reduce the feature dimension of a dataset of the total ~ 20,000 genes, we employed two parallel autoencoders. The recursive clustering framework we utilized has been applied in other multi-omics contexts to refine complex tumor landscapes [20]. As a result, nine molecular clusters (C1–C9) were identified (min cluster size = 20) (Fig. 2B). UMAP visualization (Fig. 2C) shows distinct cluster separation, indicating that the autoencoder framework preserved relevant information from both data types. Distinct mutational landscapes Each cluster exhibited a meaningful, distinct mutational profile as shown in Fig. 3 [21]. Mutations such as oncogenes EGFR and KRAS, as well as tumor-suppressor TP53, are commonly mutated genes in lung cancer that drive tumorigenesis and progression [11]. Clusters C1 and C2 were enriched with TP53 mutations (Fig. 3A and C), and also contained frequent cancer gene mutations, potentially passenger mutations such as TTN [22] and others like MUC16, and LRP1B. As validated cohorts have shown, TP53 is mutated in 46.1% of lung adenocarcinomas and 78.4% of lung squamous cell carcinomas. It indicates that mutations in TP53 are a crucial factor in lung cancer [23]. However, its distribution throughout our clusters indicates that clusters were driven by more than driver mutation status alone as mutations in notable driver genes were noted across clusters (Fig. 3B). For instance, samples in C3 exhibited both a KRAS-driven signature with STK11 co-mutation and occasional EGFR alterations (Fig. 3B), potentially indicating a metabolic or oxidative-stress-linked phenotype, as both genes lead to metabolic changes [24]. KRAS, KEAP1, STK11, and their different co-occurrences have been found to have compounding effects on the tumor microenvironment, either sensitizing tumors to combined or single immunotherapy [25]. A is an oncoplot-style map of the top mutated genes and their prevalence per cluster, where the presence of a mutation is indicated by shading. C shows the relative contribution (contribution of each gene in the top mutated genes shown) of the top mutated genes per cluster of the overall most frequently mutated genes per cluster. B shows the average nonsynonymous mutated gene count per cluster, in mutations per Megabase In addition to recapitulating known biology, mutational analysis reflected novel or evolving phenotypes or mechanisms in lung cancer. C4 combined TP53 and KRAS mutations with moderate RYR2 mutation frequency (Fig. 3B), representing a potentially novel or mixed phenotype driven by compounding mutations. For instance, prior analysis of TCGA data has illustrated that RYR2 mutations may be predictive of survival rates, and in our analysis, C3, which was one of only two clusters without RYR2 in its top ten most mutated genes [26]. C5 retained the TP53-dominant pattern with recurrent RYR2 and LRP1B alterations. Across clusters, RYR2 mutations appeared repeatedly, particularly in C2, C4, C5, C8, and C9, suggesting a potential context-dependent biomarker of calcium-signaling dysregulation, as RYR2 encodes a calcium ion channel [26]. Additionally, while Wnt-signaling is not fully characterized in lung cancer, there is growing evidence that APC, a key negative regulator of this pathway, may play a role in the tumor progression of EGFR mutant lung cancers [27]. C6 was uniquely characterized by EGFR mutations, with samples in this cluster also having TP53 or tumor suppressor gene APC, potentially indicative of a canonical receptor tyrosine kinase–activated subgroup, which could be responsive to tyrosine-kinase inhibitors [28]. C7 exhibited overlapping KRAS, EGFR, and TP53 mutations together with KEAP1, indicating a possible hybrid state of both proliferative and oxidative-stress-resistant samples [29]. C8 and C9 largely followed the TP53-driven pattern observed in earlier clusters, though both also displayed recurrent RYR2 and KRAS events. Furthermore, of the top 10 genes mutated in each cluster, 18 genes were mutated in only 1 cluster (TLR4, ADGRB3, ZNF804A, ZNF536, RBM10, OBSCN, PKHD1, FLG, FAM135B, KMT2D, CCDC168, APC, NALCN, KEAP1, RYR1, PPP4R3C, NAV3, UNC80). Nonsynonymous mutated gene count Nonsynonymous mutated gene count estimates varied across the nine molecular clusters (Fig. 3C). Clusters C2 and C9 exhibited the highest overall mutation counts. This is consistent with their high frequencies of driver mutations such as TP53, and passenger mutations, for instance TTN, and others. In contrast, clusters such as C3 and C6 showed comparatively lower mutated gene values, reflecting more genomically stable profiles dominated by KRAS or EGFR mutations, respectively. Notably, our cluster, C9, has a high nonsynonymous mutated gene count as well as the highest PD-L1 expression and immune and inflammation pathway enrichment [30]. These differences in mutational load paralleled observed transcriptional distinctions, with high-mutation clusters enriched for immune-response pathways and low-mutation clusters displaying more immune-silent expression patterns [31]. Mutational signature analysis revealed pronounced heterogeneity in underlying mutagenic processes. Smoking-associated single-base substitution signature 4 (SBS4) was most prevalent in clusters corresponding to squamous histology, whereas age-related signatures such as SBS1 were broadly distributed across the cohort (Demonstrated in Fig. 4 and Figure S3) [25]. Together, these findings underscore that both exogenous exposures and intrinsic mutational processes may contribute to the distinct genomic and transcriptional architectures defining each cluster. A. We performed mutational signature analysis using the SigProfiler tool. Using this tool and somatic mutation data, we identified a total of 6 signatures, with 3 out of 6 signatures aligning incredibly similarly (85%, 97%, 99%, Figure S3) to those identified by Alexandrov et al. [1]. B shows the prevalence of each signature per cluster, with SBS96A correlating to smoking related signature, SBS4 Transcriptional and immune signatures Differential expressions and pathway (Hallmark) enrichment analyses underscored clear functional divergence across clusters (Fig. 5) [31]. Clusters C1 and C5, to a lesser degree, showed strong cell-cycle programs (G2M checkpoint/E2F targets), with C1 additionally appearing immune-cold and marked by down-regulation of several histone H2/H3 genes, consistent with chromatin/state shifts rather than immune activation (Fig. 5B) [32]. Epigenetic and methylation events are of increasing importance in tumors without known actionable targets, so this subtype may be of interest for future therapeutic targets at the epigenetic level [33]. In contrast, C3 and C7, which were the KRAS-context groups that also contained KEAP1 mutations, showed enrichment in pathways consistent with RAS/MAPK/PI3K–AKT signaling (KRAS, G2M, E2F targets) [5, 6] (Fig. 5A). A. Differential expression of the most representive genes per cluster performed as explained earlier. Note histone methylation gene downregulation in. B. Heatmap of GSEA pathways per cluster. Enrichment analysis shows differential enrichment of interferon alpha response, interferon gamma response IL-2/Stat5 signaling, IL-6/Jak/Stat3 signaling in C4, C5, C8, and C9, as well as proliferative enrichment (Myc Targets, G2M Checkpoint) in C1 Immune-related pathways were also differentially enriched between clusters. The immune-enriched clusters were C4, C8, and C9, and are characterized by elevated interferon signaling, inflammatory response, and immune checkpoint gene expression upregulation (Fig. 6A-E). LAG-3 expression was also notably higher in C1 than the other immune markers in that cluster. The understanding role of LAG3 in lung cancer is currently evolving, with some studies exhibiting LAG3 as a potential predictor of checkpoint inhibitor resistance, others stating that it is not expressed on tumor cells in NSCLC [34, 35]. Consistent with the mutational profiles, the EGFR-dominant cluster C6 was low-mutation and was negatively enriched in immune pathways, indicating a likely immune-cold phenotype, whereas C2 and C9 showed higher mutation rates. Among these, C9 also aligned with the immune-active transcriptional state [28]. Overall, the transcriptional programs (proliferative vs. immune-active vs. immune-cold) tracked with the dominant genotypes (TP53-, KRAS-, or EGFR-context) and the observed mutation patterns (C2/C9 high; C3/C6 low) [33]. Heatmap displaying expression levels across clusters of genes related to checkpoint inhibition and immunotherapy response, PD-L1 (PDCD1), PD-L2 (CD274), PD-1, CTLA-4, and LAG3 illustrated. Higher expression of all PD-1, PD-L1, PD-L2, and CTLA-4 are found in Clusters 4 and 9, which correlates with GSEA pathway prediction of associated immune and inflammatory response pathways Clinical and survival associations Survival analyses did not identify meaningful differences in overall survival across the nine clusters (Figure S1, Figure S2). However, several clinical and etiological features varied across clusters (Fig. 7). Smoking status differed notably in C1, showing the highest concentration of current and former smokers. Additional smoking-influenced clusters (e.g., C2, C9) highlight the heterogeneity of smoking-associated tumors. Treatment was concentrated almost entirely in C1, consistent with a progressive, immune-cold, and proliferative profile [36]. Four out of the nine clusters contained mixed histology. Because the integrated dataset included samples from multiple source cohorts with differing clinical characteristics, some cluster structure may reflect cohort-related variation in addition to underlying biological differences. Associations between cluster membership and clinical variables were evaluated using chi-squared tests, with corresponding p-values reported in Table S2. As shown in this table, histology, gender, smoking status, and stage all differed significantly across clusters. Collectively, these findings suggest that the nine clusters reflect histologic composition and also capture distinguishable molecular, clinical, and immune states. However, clinical and biological interpretation should be considered in the context of cohort composition. Clinical data associated with each of the different clusters. The first panels show the histological stage and the histopathologic classification of the cluster. The second panel shows gender composition. The final panel illustrates the proportion of smokers present in each subgroup; this is further broken into subcategories based on current, former, never, or unknown status Discussion In this section, we comprehensively discuss the results that we presented in the previous section. Translational significance of clusters The integrative deep learning pipeline identified nine subtypes of non–small cell lung cancer (NSCLC) with distinct immune, transcriptional, and clinical features. By combining autoencoder-based dimensionality reduction with recursive Gaussian mixture model (GMM) clustering, this framework identified subgroups interpretable across multiple layers. Since we use a hierarchical clustering model, some overlap between lower-level clusters was expected among subgroups derived from the same higher-order cluster. In this study, cluster identities were interpreted based on the combined pattern of genomic, pathway, immune, mutational, and clinical/pathological composition rather than any single feature alone. Although histological composition differed significantly between clusters, several clusters remain histologically mixed, indicating that subtypes are not explained by histology alone. As a result of this clustering, several clear biological themes emerged. For instance, clusters C1 and C5 were dominated by TP53 alterations and enrichment of cell-cycle and DNA-damage related pathways, suggestive of high genomic instability, while demonstrating limited immune engagement [37]. However, C5 also exhibited notable downregulation of immune-related gene sets, indicating an immunologically cold subtype. Similarly, C3 and C7 were defined by KRAS mutations with frequent STK11 or KEAP1 co-mutations, consistent with activation of RAS/MAPK and PI3K–AKT–mTOR signaling and transcriptional signatures. These pathways are linked to oxidative stress adaptation and metabolic distress [38]. In contrast, the EGFR-enriched cluster (C6) was downregulated for immunogenic pathways aligning with the well-documented immune-excluded phenotype of EGFR-driven NSCLC [39]. Conversely, clusters C4, C8, and C9 showed notable immune activation, including upregulation of interferon signaling and checkpoint molecules such as PD-L1 and PD-1. Collectively, these findings underscore a functional continuum between immune-cold, proliferative clusters (C1/C5) and immune-hot, interferon-rich subtypes (C8/C9). Integration of mutation and burden data further supported these biological distinctions. Mutated gene count varied markedly across clusters, with C2 and C9 showing the highest gene-based Mutated gene count, driven by widespread TP53 and TTN alterations, whereas C3 and C6 exhibited the lowest burdens. Smoking-associated mutational signature SBS4 was most prevalent in squamous histology–enriched clusters, and smoking-related signatures were highly enriched in clusters C2 and C9, which had a high proportion of current smokers. Age-related signature SBS1 was broadly distributed. These findings suggest that variation in mutational burden and smoking exposure contributes to the different immune phenotypes observed across the subtypes. Beyond canonical drivers, several recurrent alterations also showed cluster-specific enrichment. RYR2 emerged as a mutated gene across multiple clusters, with especially high representation in C2, C4, C5, C8, and C9. Given its role in calcium homeostasis and stress signaling, RYR2 may serve as a marker of tumor stress adaptation or immune–metabolic crosstalk within the NSCLC microenvironment [40]. Other cluster-specific alterations—including TLR4 (C1), RBM10 (C3), KEAP1 (C7), and RYR1 (C8) highlight potential subtype-specific events that warrant further mechanistic investigation. Overall, the coherence between mutational, transcriptional, and immune features suggests that integrative deep learning can capture biologically pertinent structure from complex multi-omics data. The identified clusters recapitulate known NSCLC biology such as histology, immune, and proliferation-driven subtypes. It also further expands them, linking genomic alterations and mutational processes to immune phenotypes with significant resolution [5, 7]. Clinically, these results suggest that molecular subtype-based stratification may help refine NSCLC classification in future studies. Highly mutated, interferon-rich clusters (C2, C9) may represent immunotherapy-responsive phenotypes, whereas KRAS- and EGFR-driven subtypes (C3, C6, C7) could benefit from combination approaches that integrate pathway inhibition with immune modulation [28, 31, 41]. Finally, the discovery of underappreciated mutations such as RYR2 underscores the capacity of this framework to reveal new biology from large-scale genomic data, highlighting the potential utility of deep learning–based multi-omics integration for translational investigation. Limitations and future directions There are several limitations that can be associated with this study. First, this analysis relied exclusively on publicly available data. Also, the gene expression and mutation data were limited to bulk tumor profiles. Therefore, they cannot capture tumor-immune dynamics or cell-type variation. Additionally, the moderate sample size within certain clusters may reduce power to detect low-frequency alterations or subtle pathway differences. Although we tried several different clustering models and finally chose hierarchical model-based clustering mode, we did not perform a systematic comparison of the proposed framework against established clustering approaches such as consensus, NMF, or iCluster. In the current study, we instead focused on evaluating the biological interpretability of the identified subtypes across multiple molecular layers, which provides a basis for assessing their relevance. In the future, a more comprehensive comparative analysis of different clustering models can be conducted. Finally, although the autoencoder-clustering framework effectively integrated multi-omics features, its performance could further improve with the inclusion of additional omics layers and external validation in independent NSCLC cohorts. Our future directions include expanding the data set to enhance the power of our study and the robustness of clusters. Despite these limitations, the present analysis demonstrates that deep learning–based integration of transcriptomic and genomic data can reveal biologically coherent NSCLC subtypes. The resulting clusters offer a refined molecular classification that connects driver alterations, mutational processes, and immune context, providing a potential foundation for future biomarker-driven therapeutic stratification. Conclusion Lung cancer is the third most diagnosed cancer in the United States. It is driven by a complex combination of molecular alterations, which influence tumor behavior and patient outcomes. Thanks to large-scale consortia such as TCGA, there is a unique opportunity to uncover novel subtypes through multi-omics integration for this cancer. While clustering methods have been used to identify subtypes in lung cancer, many integrations and avenues remain underexplored. By leveraging publicly available data, we developed an analytical pipeline using deep learning and model-based clustering that integrated two-omics types that provide complementary information. By integrating both somatic mutation and gene expression data, we were able to capture a meaningful representation of both via an autoencoder that represented temporal changes as well as static mutation events. Moreover, by combining this with iteratively applied model-based clustering, we define subtypes with distinct molecular characteristics. Gene expression changes, common driver, and functional mutations. Common driver and functional mutations were identified and analyzed to reveal differences in immunogenicity, histology, smoking status, gender, pathologies, mutation count, and biological pathways. These features are key features that provide insight into tumor progression, treatment response, and survival rates [18]. Furthermore, while this work provides insights into the landscape of lung cancer, it also illustrates the potential of deep learning in multi-omics analysis. This pipeline provides solid reasoning for the expansion of future integrative studies, which may lead to advances in cancer subtyping and biomarker discovery across more cancer types, using easily accessible, publicly available data. Electronic Supplementary Material Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Acknowledgements We would like to thank Gennaro Caledo for all his help and support in having a better conceptual understanding of genomics analysis. Author contributions BG and DW preprocessed the dataset. BG cleaned and organized the data and conducted the experiment. DW, BG, and SMA developed and designed the model, ran and tested it on the data, and contributed to writing the paper. GR and ID defined and supervised the project, analyzed the results, and contributed to writing and revising the paper. All authors wrote and reviewed the manuscript prior to submission. Funding This study is funded by NSF-NRT grant number 2152059. Data availability All source code, figures, and instructions for data download from the GDC are available at [https://github.com/MLBC-lab/LungClust]. For further information, you can contact the corresponding author of this paper (Iman Dehzangi: [i.dehzangi@rutgers.edu]). Declarations Ethics approval and consent to participate Not applicable. Consent for publication All authors have read and approved the final manuscript and consent to its publication. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Data Availability Statement All source code, figures, and instructions for data download from the GDC are available at [https://github.com/MLBC-lab/LungClust]. For further information, you can contact the corresponding author of this paper (Iman Dehzangi: [i.dehzangi@rutgers.edu]).