Survival prediction and molecular subtyping of squamous cell lung cancer based on network embedding Authors: Dingjie Guo, Jing Chen, Yixian Wang, Xin Liu Published: 2024 ABSTRACT AbstractSquamous cell lung cancer (SQCLC), which is fatal to humans, is heterogeneous with different genetic and histological features. We used SBMOI, a multi-omics data integration method from previous study, to integrate clinical, gene expression, and somatic mutation data of SQCLC to construct new patient features. Next, random survival forest (RSF) model and SimpleMKL model were constructed to predict the survival of SQCLC patients, and K-means model was constructed to perform molecular subtyping. The results of the RSF model showed that when the dimension of the patient features were 11 × 364 and the hard threshold was 0.2, we obtained the best results, and the AUC value of the 1-year time-dependent ROC curve was 0.706. The SimpleMKL model, constructed using the same patient features, performed exceptionally well, with 1-year, 5-year, and 10-year survival prediction AUC values of 0.944, 0.947 and 0.950, respectively. We used K-means analysis to identify three SQCLC molecular subtypes with significant survival differences. The patient features constructed by SBMOI were used to effectively predict the survival and molecular subtyping of SQCLC patients. In addition, our study further confirmed the effectiveness in multi-omics data integration task and broad applicability of SBMOI.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-024-81199-z.Keywords: Multi-omics data integration, Survival prediction, Molecular subtyping, Squamous cell lung cancerSubject terms: Cancer, Computational biology and bioinformatics Introduction Squamous cell lung cancer (SQCLC) is the second most common histological subtype of lung cancer, accounting for approximately 30% of all lung cancer cases1. The overall prognosis for SQCLC is poor, with a 5-year survival rate of only 12.4% for patients who have undergone surgery2. Most patients with centrally located tumors that are locally aggressive, and often do not have actionable genetic alternations3. Although many studies of SQCLC have been reported in the medical literatures, the characteristics, prognosis and molecular typing remain unclear4. A large number of studies have used gene expression profiles to establish survival prediction models and molecular typing of SQCLC. For example, a previous study successfully predicted survival for SQCLC using COX proportional risk models based on gene expression data and identified 79 genetic signatures that predicted cancer-related death5. Besides, The Cancer Gene Atlas (TCGA) used gene expression profiles to classify SQCLC into primitive (P), secretory (S), basal (B), and classical(C)6. SQCLC is a complex disease characterized by high heterogeneity at the molecular level and this heterogeneity continues to intensify during the course of its development7. Currently, most studies have relied solely on single-omics data for survival prediction and molecular subtype identification of SQCLC, resulting in inadequate outcomes. However, different molecular omics data are not independent; there are interactions and influences among the omics data. Exploring effective data integration methods not only addresses the limitations of single-omics data but also enables a more comprehensive study of SQCLC from multiple perspectives. Therefore, it is essential to design effective multi-omics data integration methods for survival prediction and molecular subtype identification to facilitate data-driven research on SQCLC. Whole-genome sequencing analysis showed that non-small cell lung cancer, including SQCLC, is one of the tumors with the highest frequency of nonsynonymous mutations, with an average of about 150 somatic mutations per patient8. Therefore, understanding the interaction between gene mutations and gene expression profiles may provide more information for survival prediction and molecular subtyping studies of SQCLC. The current development of cancer treatment drugs primarily focuses on identifying individual genetic targets and using these targets to discover corresponding therapeutic drugs9–11. However, by studying various potential genetic defects and identifying key pathways that influence cellular phenotypic behavior, we can develop treatments that target not only the mutated oncogenes or tumor suppressor genes within these pathways but also any component within them12,13. Nevertheless, due to the complex interactions between cellular regulatory pathways and the fact that cancer is caused by the dysregulation of multiple pathway networks, this approach still has significant limitations in cancer treatment14,15. Therefore, network-based analysis has gradually become a popular field in cancer-related research. Extracting simpler, more effective, and meaningful features from complex raw data using advanced techniques is essential for subsequent analysis and machine learning model training. A substantial body of research has focused on identifying key features related to anatomical structures in medical images for downstream tasks such as medical image segmentation, enhancement, and registration16–21. However, the high dimensionality, sparsity, and intricate topology of biological network data present significant challenges for its analysis and modeling. Network embedding methods have broad applicability and demonstrate considerable competitiveness across various graph analysis tasks. Therefore, in prior research, our team proposed a multi-omics data integration method called struc2vec-based multi-omics data integration. This method uses the network embedding technique struc2vec to integrate gene co-expression networks, somatic mutation data, and clinical information to construct mutation gene vectors, thereby better representing patient characteristics. Then, the patient characteristics constructed by SBMOI were utilized to build a random survival forest (RSF) model, which successfully predicted survival of lung adenocarcinoma patients22. Researches indicated that higher embedding dimensions correlate with better model performance. But, as the embedding dimension increases beyond a certain threshold, the performance of downstream analysis models tends to plateau or saturate23. However, previous research only explored 128-dimensional network embedding features and lacked further investigation into the data integration capabilities of the SBMOI algorithm and its applicability to different downstream analysis tasks. Therefore, this study utilized the SBMOI algorithm from previous research to construct patient features of SQCLC at different dimensions. These patient characteristics will be used for modeling SQCLC survival prediction and molecular subtyping. Figure 1 showed the flow chart of this study. The flow chart of this study. Methods Datasets and data preprocessing The data of our study was carried from the LUSC project of The Cancer Genome Atlas (TCGA) (https://portal.gdc.cancer.gov/repository). After excluding patients with missing survival data, survival times less than 30 days, and those with fewer than 10 mutated genes, a total of 337 SQCLC patients with complete and valid gene expression data, somatic mutation data, and clinical data were included. Additionally, 51 normal samples with complete and valid gene expression data were included. For the gene expression data, genes with an expression value of zero in all patients, genes with the same expression value in more than 80% of patients, and genes with more than 10% missing values were excluded. The remaining missing gene expression values were filled in using the median. For the somatic mutation data, genes that did not have any mutations in any patient were excluded. For the clinical data, clinical features with more than 10% missing values were excluded. Missing values in continuous clinical feature (Age) was filled in using the median, while missing values in categorical clinical features (Gender, Stage, T, N, History of malignancy, Year at diagnosis, Treatment method) were filled in using the mode. Construct features for SQCLC patients using the SBMOI algorithm Our study employed the SBMOI data integration method from prior research22. The specific calculation processes of SBMOI algorithm are as follows. Scale-free gene co-expression networks (FCNs) Initially, using the gene expression matrix of SQCLC patients, we determined the absolute value of the Spearman correlation coefficient between gene j and gene k to assess their similarity. The resulting similarity matrix, S, is as follows: In network analysis, a network is typically represented using an adjacency matrix. The power adjacency function was applied to convert the similarity matrix S into an adjacency matrix , In our network, increasing the soft threshold β improved the differentiation between genes with high and low connectivity. The criteria used to determine whether networks follow a scale-free distribution are as follows: the logarithm of the mean connectivity of a node is negatively correlated with the logarithm of the node’s occurrence probability, and the linear regression model fitting index is greater than 0.9. The optimal soft threshold found to be 3 (Supplement Fig. 1). Finally, we applied a hard threshold to obtain the list of directly linked neighbors of a node in the FCN, which served as input data for the struc2vec algorithm. The matrix was transformed into : Vectorize the FCN using the struc2vec algorithm Struc2vec, as a novel and flexible graph embedding technique, can be used to simplify, visualize and facilitate the analysis of FCN. Its goal is to find a low dimensional representation of the gene nodes in a FCN in some metric space so that the given similarity function (or distance) can be preserved as much as possible. Struc2vec algorithm was adopted to complete the vectorization of the FCN of SQCLC. The steps are as follows: (1) Structural similarity measurement. Let denote the structural similarity between node i and j when considering their k-hop neighborhoods (all nodes at distance ≤ k and all edges among them), where denotes the set of notes at distance k from the node j, denotes the ordered degree sequence of a set , measures the distance between the ordered degree sequence and , . Next, the Dynamic Time Warping (DTW) is performed to measure the distance between and , because DTW is a good way to compare the similarity between two sequences with different lengths, If , the distance between and will be zero. (2) Multilayer weighted graph construction. Each layer ( is the diameter of the original network.) of the multilayer weighted graph is formed by a weighted undirected complete graph, the edge weight between nodes i and j in a layer is: Then, directed edges connect the layers as follows, each node i of layer k is connected to its corresponding node i in the layer and layer . The edge weight between layers is defined as: where denotes the number of edges connecting to i that have weight larger than the average edge weight in layer k. (3) Generating and training sequence of nodes. Struc2vec model uses biased random walk that moves around the multilayer weighted graph to generate sequence of nodes. Before each step, the random walk determines whether to stay in the current layer based on probability q (if the random walk stays in the current layer). If the random walk stays in the current layer, the probability of walking from node i to node j in layer k is defined as: where is the normalization factor for node i in layer k, which is defined as: The random walk tends to choose nodes with structurally more similar to the current nodes. Since the probability of random walk staying in the current layer is q, the probability of changing layers is . If the random walk changes layers, the probability of choosing or layers are defined as: Then, a node is randomly selected in layer 0 and a random walk is started to generate a sequence of nodes. Finally, language learning model Skip-Gram which is effective at learning meaningful representations for a variety of data, is performed to train the node sequence. In this study, each node in the scale-free gene co-expression network is embedded as a low dimensional vector (the key parameters are showed in Supplement Table 1). Construct SQCLC patient features To better represent SQCLC patients, this study integrated low-dimensional gene vectors, gene mutation data, and patient clinical information to construct mutation gene vectors that capture patient features. We employed Student’s t-test to identify clinical characteristics of SQCLC patients with significantly different gene mutation frequencies, which were then included in subsequent analyses. Patients with similar clinical characteristics were grouped together. Student’s t-test indicated significant differences in somatic mutation frequencies among SQCLC patients with varying stages, T stages, N stages, history of malignancy, and year of diagnosis (Supplement Table 2). Based on these clinical variables, we classified SQCLC patients into 16 categories. Additionally, our study identified significant differences in gene mutation frequencies among SQCLC patients across various categories (Supplement Fig. 2). Some genes were mutated in most patients, while others were mutated only in patients with specific clinical information. Therefore, we defined a weight to mitigate the effects of varying gene mutation frequencies: where n is the total number of clinical categories of SQCLC patients, is the number of mutations to gene u within clinical type v, is the gene with the highest number of mutations in clinical type v, and is the number of mutations in gene u across all clinical types. The weight and low dimensional gene vector were utilized to construct patient features. Genes of patients within the same clinical type were sorted based on the weight and divided into blocks containing the same number of genes. The number of blocks was based on the minimum number of gene mutations among the patients (In this study, the minimum number of gene mutations in the SQCLC patients was 11). Ultimately, we combined the features of the mutated genes within each block to create the SQCLC patient features. Survival prediction of SQCLC The random survival forest (RSF) model, used in survival analysis to predict survival time or event probability, is a machine learning model that integrates multiple decision trees within the RF model framework to handle survival data. It offers robust survival prediction capabilities and reliability24. We randomly assigned SQCLC patients to a training set and a validation set in a 7:3 ratio. Then, based on the constructed SQCLC patient features, we used RSF model to predict 1-year, 5-year, 10-year survival in patients. The accuracy of the RSF model was evaluated by time- dependent receiver operator characteristic curve (ROC) analysis. This study also planed to directly construct the RSF model using clinical, gene expression, and somatic mutation data of patients with SQCLC. However, given the high-dimensional character of gene expression and somatic mutation data, they are not directly suitable for constructing the RSF model. To enhance computational efficiency, we utilized univariate COX proportional hazards regression analysis to identify clinically relevant gene expression and somatic mutation features associated with survival. These selected features will then be incorporated into the RSF model to predict survival outcomes in patients with SQCLC. The SimpleMKL model is an advanced data integration method. Its core idea is to linearly combine multiple kernel functions and learn the optimal weights for these combinations through an optimization process, thereby enhancing the performance and generalization capability of kernel methods25. To further demonstrate the effectiveness of the SBMOI algorithm in integrating data, we used the patient features constructed by the SBMOI algorithm, survival-related clinical features, gene expression features, somatic mutation features, and combinations of survival-related clinical, gene expression, and somatic mutation features to build SimpleMKL models for survival prediction. Prior to the SimpleMKL model analysis, SQCLC patients were divided into training and validation sets in a 7:3 ratio. However, since the SimpleMKL model cannot directly use survival data, this study converted survival data of SQCLC patients into binary classification data. Specifically: For 1-year survival prediction, patients with survival time greater than 1 year were labeled as 1, otherwise labeled as 0. For 1-year survival prediction, patients with survival time greater than 1 year were labeled as 1, otherwise labeled as 0. For 5-year survival prediction, patients with survival time greater than 5 years were labeled as 1, otherwise labeled as 0. For 5-year survival prediction, patients with survival time greater than 5 years were labeled as 1, otherwise labeled as 0. For 10-year survival prediction, patients with survival time greater than 10 years were labeled as 1, otherwise labeled as 0. For 10-year survival prediction, patients with survival time greater than 10 years were labeled as 1, otherwise labeled as 0. We constructed one kernel for the SimpleMKL model based on patient features derived from the SBMOI algorithm, survival-related clinical features, survival-related gene expression features, and survival-related somatic mutation features. Three different kernels were built for the SimpleMKL model based on combinations of survival-related clinical, gene expression, and somatic mutation features. For the Gaussian kernel, parameter values ranged from 0.25 to 20 (specifically: 0.25, 0.5, 1, 2, 5, 7, 10, 12, 15, 17, 20), and for the polynomial kernel, degrees ranged from 1 to 3. All SimpleMKL models were subjected to 10-fold cross-validation, with the average area under the ROC curve (AUC) of the 10-fold cross-validation used as the primary evaluation metric for prediction performance, and the average accuracy and recall of the 10-fold cross-validation used as supplementary evaluation metrics. Subtype discovery of SQCLC For established cancer molecular subtypes, it is essential to describe their characteristics in molecular, clinical, and histopathological terms26. However, the study of molecular subtyping in cancer is still in the exploratory stage, so there is no standardized gold standard for the identification of molecular subtyping results for specific cancers. Currently, survival analysis is widely used in the evaluation of disease molecular subtyping results. By comparing the survival times between different molecular subtypes, established subtypes are determined to have not only statistical significance but also biological significance27. In this study, K-means analysis was used to cluster SQCLC patient features into different subtypes. Kaplan-Meier survival curves for the SQCLC subtypes were calculated, and comparisons were made using the Log-rank test. The number of clusters showing the most significant survival differences was selected as the optimal subtyping result. In order to further prove the validity of the clustering results of SQCLC patients, we selected the top 50 genes with frequently mutated and assessed their mutational patterns across all the subtypes of SQCLC. We screened the differentially expressed genes (DEGs) of each subtype comparing to health people and performed set difference operations on these gene sets to extract the subtype-specific DEGs. Subsequently, we analyzed the somatic mutation patterns of these subtype-specific DEGs and explored their biological functions through Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) functional enrichment analyses. Results We analyzed 337 SQCLC patients, 74.2% of whom were male (Table 1). Additionally, the overall survival rate of SQCLC patients was 76.9%. The cumulative hazard figure (Fig. 2) showed that the cumulative risk of death for patients with lung squamous cell carcinoma stabilizes at around 160 months. Therefore, clinical monitoring for SQCLC patients should not be relaxed over a follow-up period of at least 10 years. Table 1Clinical characteristics of SQCLC patients.Clinical characteristicsN(%)Alive (%)Dead (%)Total33725978Gender Male250 (74.2)189 (73.0)61 (78.2) Female87 (25.8)70 (27.0)17 (21.8)Age (y) ≤ 65132 (39.2)108 (41.7)24 (30.8) > 65205 (60.8)151 (58.3)54 (69.2)Stage 1–2282 (83.7)220 (84.9)62 (79.5) 3–455 (16.3)39 (15.1)16 (20.5)T T1–T2280 (83.1)217 (83.8)41 (80.8) T3–T457 (16.9)25 (16.2)15 (19.2)N N0–N1312 (92.6)238 (91.9)74 (94.9) N2–N325 (7.4)21 (8.1)4 (5.1)History of malignancy No299 (88.7)233 (90.0)66 (84.6) Yes38 (11.3)26 (10.0)12 (15.4)Year at diagnosis Before 200680 (23.7)39 (15.1)41 (52.6) 2007 or later257 (76.3)220 (84.9)37 (47.4)Treatment method Radiation148 (43.9)113 (43.6)35 (44.9) Chemotherapy189 (56.1)146 (56.4)43 (55.1) Clinical characteristics of SQCLC patients. The cumulative risk curve for SQCLC patients. When using the RSF model for analysis, due to the imbalance in the study sample, we adjusted the sample weights based on actual weights using the “case.wt” parameter in the rfsrc() function in R language. Based on the SQCLC patient features with different dimensions (11 × 32, 11 × 64, 11 × 128, 11 × 256, 11 × 512), we made 1-year, 5-year, 10-year survival prediction for SQCLC patients by RSF model. The results of RSF model showed that both dimensions of patient features and the value of the threshold of FCN affect the accuracy of survival prediction. The Fig. 3 showed the value of AUC of 1-year, 5-year, and 10-year time-dependent ROC curve. It was observed that when the threshold of FCN was 0.4 and the patient feature dimension was 11 × 32, the RSF model performs well in predicting 1-year and 10-year survival (1-year AUC = 0.791, 10-year AUC = 0.815). However, the corresponding prediction error rates are 0.794 and 0.729, respectively (Supplement Table 4). Under these conditions, the 5-year survival prediction AUC value of the RSF model is only 0.499. Therefore, considering the survival prediction performance and prediction error rates of the RSF model at various time points, we selected the analysis results with the threshold of FCN was 0.2 and the dimensions of SQCLC patient features was 11 × 64 as the optimal predictive performance for the SBMOI combined RSF model. At this point, the AUC of 1-year, 5-year, and 10-year time-dependent ROC curve in the validation dataset were 0.706, 0.616 and 0.684, respectively (Fig. 4). Other parameters of the RSF model were shown in Supplement Table 5. The AUC of time-dependent ROC curve of RSF based on patient features with SQCLC of different dimensions. When the dimensions of SQCLC patient features was 11 × 64 and the threshold of FCN was 0.2, the AUC of 1-year, 5-year, and 10-year time-dependent ROC curve in the train dataset (A) and validation dataset (B). We identified 1 significant survival-related clinical features, 1092 significant survival-related gene expression features, and 999 significant survival-related somatic mutation features. The analysis results of RSF model using survival-related clinical, gene expression, and somatic mutation features of SQCLC patients indicated the following: the best performance of 1-year survival was the RSF model based on somatic mutation features, and the AUC of validation dataset was 0.978. The best performance of 5-year survival was the RSF model based on combined survival-related clinical, gene expression, and somatic mutation features, and the AUC of validation dataset was 0.998. The best performance of 10-year survival was the RSF model based on somatic mutation features, and the AUC of validation dataset was 0.909 (Fig. 5). The key parameters of the RSF model were shown in Supplement Table 6. Compared to the performance of the SBMOI combined RSF model, although better survival prediction performance was achieved, it is noteworthy that there is a certain degree of correlation and complementarity between patient survival-related clinical, gene expression, and somatic mutation data. Relying on a single data feature or directly integrating multiple data features for modeling may obscure the complex relationships between these features, thereby compromising the model’s robustness and generalization ability. The AUC of time-dependent ROC curve of the RSF based on clinical, gene expression, and somatic mutation features (clin clinical features, exp gene expression features, muta somatic mutation features). Based on the analysis above, when the dimensions of SQCLC patient features was 11 × 64 and the threshold of FCN was 0.2, the RSF model had the best survival prediction performance. Therefore, when analyzing with SimpleMKL, other hard thresholds of the FCN network within the SBMOI framework and patient features of different dimensions constructed by the SBMOI algorithm are no longer considered. Due to the transformation of patient survival data into binary classification data, there exists a class imbalance issue during SimpleMKL analysis, which is particularly pronounced in the case of 5-year and 10-year survival predictions. Therefore, this study addressed the issue of data imbalance by performing random oversampling on the minority class. The results indicated that the SimpleMKL model, constructed using patient features derived from SBMOI with a hard threshold of 0.2 and an embedding dimension of 64 in the FCN network, performed optimally for 1-year and 5-year survival prediction tasks. Under 10-fold cross-validation, the mean AUC values were 0.944 and 0.947, respectively, although the mean accuracy and recall were relatively lower. The SimpleMKL model, built using survival-related gene expression, somatic mutations, and combined clinical features, performed best in the 10-year survival prediction task, with a mean AUC of 0.956 under 10-fold cross-validation (Table 2). Although the SimpleMKL model, incorporating survival-related gene expression, somatic mutations, and combined clinical features, performed best in predicting 10-year survival for SQCLC patients, it involved three distinct kernels, and the number of survival-related gene expression, somatic mutation, and clinical combined features was considerably higher compared to the patient features constructed by SBMOI (2092 vs. 704 [11 × 64]). In contrast, the SimpleMKL model with these combined features had a higher computational burden, and the increase in AUC for 10-year survival prediction was relatively modest (0.956 vs. 0.950). In conclusion, while the SBMOI algorithm combined with the SimpleMKL model had a somewhat weaker capacity to distinguish patient categories, it demonstrated excellent overall performance in both short-term and long-term survival prediction tasks for SQCLC patients. This further highlighted the effectiveness of SBMOI in multi-omics data integration tasks. Table 2The prognostic performance of the SimpleMKL model based on different data features of SQCLC (mean ± standard deviation).Data typeAUCAccuracyRecall1-year Clinical0.544 ± 0.0360.535 ± 0.2340.594 ± 0.087 Gene expression0.943 ± 0.0680.721 ± 0.1930.700 ± 0.189 Somatic mutation0.912 ± 0.0660.784 ± 0.1730.798 ± 0.180 Clinical and gene expression and somatic mutation0.934 ± 0.1130.772 ± 0.2150.768 ± 0.215 SBMOI0.944 ± 0.0600.740 ± 0.2180.728 ± 0.2085-year Clinical0.728 ± 0.0310.675 ± 0.0680.777 ± 0.128 Gene expression0.936 ± 0.0480.722 ± 0.2100.705 ± 0.201 Somatic mutation0.907 ± 0.0690.800 ± 0.1860.789 ± 0.178 Clinical and gene expression and somatic mutation0.857 ± 0.1450.850 ± 0.1570.839 ± 0.161 SBMOI0.947 ± 0.0500.785 ± 0.1160.720 ± 0.10210-year Clinical0.899 ± 0.0200.835 ± 0.1450.805 ± 0.128 Gene expression0.947 ± 0.0700.838 ± 0.2050.820 ± 0.207 Somatic mutation0.946 ± 0.0350.784 ± 0.1840.774 ± 0.175 Clinical and gene expression and somatic mutation0.956 ± 0.0490.955 ± 0.0480.955 ± 0.048 SBMOI0.950 ± 0.0660.829 ± 0.2080.810 ± 0.207 The prognostic performance of the SimpleMKL model based on different data features of SQCLC (mean ± standard deviation). The K-means clustering analysis of the SQCLC patient feature models based on SBMOI with different dimensions (11 × 32, 11 × 64, 11 × 128, 11 × 256, 11 × 512) shows that patient features from each dimension can cluster SQCLC patients into 2 or 3 molecular subtypes with significant survival differences. By comparing the p-values from the Log-rank test, the clustering results were optimal when the FCN network threshold was set to 0.3 and the embedding dimension was 128 (11 × 128 patient features). In this case, the K-value for the K-means clustering analysis was 3, and the p-value from the Log-rank test was 0.020 (Table 3). The number patients of subtypes 1 was 205; The number patients of subtypes 2 was 128; The number patients of subtypes 3 was 4. Subtypes 1 had a poor survival (Fig. 6A), and we found that the top 50 most frequently mutated genes exhibited significantly different mutation patterns across different subtypes (Fig. 6B). Compared to healthy controls, subtype 1 had 1,000 significantly differentially expressed genes (DEGs); subtype 2 had 1,022 significantly DEGs; and subtype 3 had 1,253 significantly DEGs. The set difference of differentially expressed genes for each molecular subtype was calculated, and the results showed that subtype 1, 2, and 3 contained 64, 86, and 317 uniquely significant DEGs, respectively (Supplement Fig. 3). Their expression was shown in Supplement Fig. 3. The mutation patterns of SQCLC-specific DEGs varied significantly across the subtypes, with subtype 1, which had the worst survival prognosis, showing fewer somatic mutations in its unique DEGs (Supplement Fig. 4). GO enrichment analysis indicated that the DEGs in subtype 1 were significantly enriched in the biological process (BP) of negative regulation of extrinsic apoptotic signaling pathway via death domain receptors, the cellular component (CC) of ficolin-1-rich granule, and the molecular function (MF) of glucose binding signaling pathways (Supplement Fig. 5). KEGG enrichment analysis revealed that the DEGs in subtype 1 were significantly enriched in the Pentose phosphate pathway, Bacterial invasion of epithelial cells, and Tight junction pathways, with key genes such as PFKL, DERA, and TKT involved (Supplement Fig. 6). Table 3The P value of the log-rank test for the survival curve between subtypes of squamous cell lung cancer.Number of subtypesThe Hard threshold m0.20.30.40.50.60.70.811 × 32-dimensional patient features 20.063 0.037 0.037 0.0680.0860.0580.051 3 0.031 0.022 0.030 0.043 0.024 0.058 0.025 40.0850.0650.0790.0600.1300.1420.150 50.0800.0600.0950.0920.1210.1380.247 60.1230.1730.0760.2330.1300.2240.124 70.2860.1150.2290.2870.2550.3690.115 80.3580.3570.3720.3020.3010.3390.239 90.3550.4450.3800.2910.2820.3360.567 100.4260.4400.4250.2730.3900.3070.32211 × 64-dimensional patient features 2 0.035 0.037 0.0790.0500.0960.0580.053 3 0.039 0.068 0.036 0.025 0.021 0.032 0.038 40.0620.0780.1160.0760.1020.2090.147 50.0910.0800.1160.0870.1140.1710.162 6 0.048 0.1420.1150.0680.0890.2640.081 70.2620.1670.3750.1580.2210.1810.207 80.3230.2090.3150.2640.2130.2750.232 90.4230.3310.4280.3090.4550.2570.286 100.4930.4100.4870.2050.5140.2020.44611 × 128-dimensional patient features 2 0.027 0.037 0.040 0.0580.0810.0580.064 3 0.033 0.020 0.039 0.039 0.022 0.035 0.024 40.0880.0940.0810.0770.0820.1750.122 50.0830.1270.1200.088 0.038 0.1980.138 6 0.035 0.0770.056 0.044 0.0840.1200.128 70.1530.1060.1760.2050.2210.2470.083 80.3260.3190.3660.1830.2300.3120.140 90.4250.4310.5330.1910.3560.4720.274 100.3960.2000.2320.4060.5370.6920.39211 × 256-dimensional patient features 2 0.037 0.037 0.027 0.048 0.0790.058 0.049 30.115 0.029 0.024 0.086 0.057 0.046 0.027 40.1010.0640.0770.0910.1200.2000.159 50.1090.1270.0990.0950.1300.1710.138 60.053 0.045 0.105 0.038 0.1030.1040.101 70.0970.4020.1180.2350.1990.1850.155 80.2970.3810.4190.2120.5000.2110.251 90.0870.5330.2230.3710.5900.2470.148 100.2380.2690.5250.3080.6970.2680.19511 × 512-dimensional patient features 2 0.031 0.031 0.031 0.0790.079 0.031 0.026 30.064 0.028 0.027 0.040 0.022 0.035 0.043 40.0960.0730.0730.0730.0690.1360.171 50.1020.0930.1120.0830.0800.1320.124 6 0.042 0.039 0.044 0.043 0.0700.0800.075 70.1970.1760.1580.1160.1050.1110.149 80.4530.3420.3960.2410.2930.1270.117 90.5880.3160.3350.3540.3890.2650.081 100.2020.2030.4770.5810.1660.4040.083Font bold indicates statistically significant at alpha = 0.05. The P value of the log-rank test for the survival curve between subtypes of squamous cell lung cancer. Font bold indicates statistically significant at alpha = 0.05. Survival curves of different molecular subtypes (A) and mutation patterns of the top 50 most frequently mutated genes (B). Discussion In this study, we integrated gene expression, somatic mutation, and clinical data of SQCLC patients to construct patient feature vectors based on network embedding technology. Based on patient feature vectors, we conducted survival predictions using the RSF model and the SimpleMKL model, respectively. The results of RSF model showed that we get optimal results when the patient feature vector was 11 × 64 and the hard threshold of FCN was 0.2, and the AUC value of 1-year time-dependent ROC curve was 0.706, which suggested that when the dimension of patient feature vector was 11 × 64, our method has good stratification ability. Moreover, when the patient feature vector was 11 × 64 and the hard threshold was 0.2, the SBMOI algorithm combined with the SimpleMKL model exhibited exceptionally strong performance in both short-term and long-term survival prediction tasks for patients with SQCLC. Our research further demonstrated that the SBMOI algorithm can effectively integrate gene expression, somatic mutations, and clinical data of SQCLC patients. The features of SQCLC patients constructed using the SBMOI algorithm can accurately represent the SQCLC patient population. Using SBMOI to construct multidimensional features of SQCLC patients, K-means clustering model exhibited a certain degree of stratification ability for features of SQCLC patients across all dimensions. The clustering result with an FCN network hard threshold of 0.3 and an embedding dimension of 128 in SBMOI was selected as the optimal subtyping. This approach divided SQCLC patients into three molecular subtypes with significantly different survival outcomes, with subtype 1 showing poorer survival. Therefore, the SBMOI combined with the K-means model can effectively identify clinically meaningful molecular subtypes of SQCLC. Previous studies on graph embedding have found that machine learning models downstream of graph embedding algorithms, such as Random Forest, Naive Bayes, Logistic Regression, Multilayer Perceptron, LightGBM, and XGBoost, exhibit similar performance in classification tasks, suggesting that the choice of classifier does not significantly impact the results28. In this study, although the models constructed using patient features based on the SBMOI algorithm achieved excellent performance in SQCLC survival prediction tasks, the downstream prediction model SimpleMKL outperformed the RSF model. The SimpleMKL model primarily improves model performance by optimizing combinations of multiple kernel functions, providing flexibility and adaptability that allow it to achieve better results when handling complex data or nonlinear relationships29. In contrast, while the RSF model is robust and generalizable based on ensemble learning with multiple decision trees, it may not be as flexible in adjusting feature representations when dealing with complex data24. It is important to note that the RSF model is a commonly used machine learning method in the field of survival analysis30. However, when using the SimpleMKL model for survival prediction, the issue of converting survival data into binary classification data leads to a certain degree of data information loss, which can result in inaccurate model predictions. Additionally, the computational time cost of the SimpleMKL model should not be overlooked. Furthermore, this study found that the performance of the SimpleMKL model and RSF model varied across different data features. For example, the RSF model based on somatic mutation features performed better compared to other data types, whereas somatic mutation features exhibited relatively poorer performance in the SimpleMKL model. This suggested that the RSF model may be more robust when handling high-dimensional sparse features. In this study, subtype 1 exhibited poorer survival, and its specific differentially expressed genes had fewer somatic mutations. GO and KEGG enrichment analysis of the subtype-specific DEGs revealed that subtype 1 of SQCLC may have unique biological characteristics in regulating cell survival, immune defense, and cell adhesion. The specific differentially expressed genes were significantly enriched in the pentose phosphate pathway, bacterial invasion of epithelial cells, and tight junction pathways. Numerous studies have confirmed that the pentose phosphate pathway plays a crucial role in cancer cell metabolism and survival by generating key molecules such as pentose phosphates and ribonucleic acids, and by regulating enzymes of the oxidative and non-oxidative branches, supporting rapid cancer cell proliferation and helping them withstand stress conditions31,32. Furthermore, a study by Wang et al. found that genes related to the bacterial invasion of epithelial cells pathway underwent significant mutations in advanced non-small cell lung cancer, which may partly explain the poor survival outcomes in subtype 1 of SQCLC33. Apart from this, we found that TP53, TTN, CSMD3, MUC16, RYR2 had high mutation frequencies in SQCLC, which is consistent with previous studies34–36. The mutational patterns of the top 50 genes with frequently mutated were significantly different between SQCLC. subtypes. For example, TP53 maintained a high mutation frequency in SQCLC subtype 1. This further demonstrated the differences between subtypes in our study and the effectiveness of our method in identifying SQCLC subtypes. We constructed the SQCLC patient feature vectors and based on the patient feature vectors, we successfully performed survival prediction and molecular classification of SQCLC. However, it still had some limitations. For instance, there are inherent limitations of the SBMOI algorithm: there is a certain degree of information loss when setting a hard threshold to obtain the directly adjacent neighbor list of FCN nodes. Additionally, SBMOI algorithm only integrated gene expression, gene mutation, and clinical information for SQCLC patients, without considering other omics data such as DNA methylation and protein expression, which could potentially improve the accuracy of survival prediction and molecular classification. Conclusions In this study, the data integration algorithm SBMOI from previous studies was used to successfully construct patient features of SQCLC, and this patient features was used for effective survival prediction and molecular typing of SQCLC patients. Our study further confirmed the effectiveness of SBMOI in data integration tasks. The SBMOI algorithm could be used not only for supervised machine learning model analysis but also for unsupervised machine learning model analysis, and different downstream analysis models have different performance, which requires careful selection. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements We are thankful for the contribution of The Cancer Genome Atlas (https://portal.gdc.cancer.gov/analysis_page?app=Downloads). Author contributions D.G.: Writing—original draft, Methodology, Formal analysis, revising . J.C.: Writing—review and editing, Supervision, Methodology, Conceptualization, revising. Y.W.: Methodology, Formal analysis. X.L.: Supervision, Project administration, Conceptualization. Funding sources This work was supported by Fundamental Research Funds for the Central Universities (No. 2412023QD033). Data availability The datasets analyzed during the current study are available in The Cancer Genome Atlas (https://portal.gdc.cancer.gov/analysis_page? app=Downloads). Declarations 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 Data Availability Statement The datasets analyzed during the current study are available in The Cancer Genome Atlas (https://portal.gdc.cancer.gov/analysis_page? app=Downloads).