Introduction

Recent advancements in oncology have leveraged machine learning techniques to enhance the analysis of transcriptomic data across various cancer types1,2. Integrating transcriptomic data with advanced algorithms has enhanced accuracy of downstream tasks such as mutation prediction, cancer subtyping and survival prediction3,4,5. In particular, models like gradient boosting and neural network architectures, including multilayer perceptrons (MLPs) and graph neural networks (GNNs), have demonstrated their capability in extracting meaningful patterns from complex genomic datasets6,7. These approaches enable more precise predictions and provide new clinical insights, supporting the development of targeted therapies and personalized cancer therapies.

RNA sequencing (RNA-seq) is a powerful technique in cancer research, providing profound insights into the genetic landscape of tumors. However, the high dimensionality and complexity of RNA-seq data present significant analytical challenges, requiring robust feature selection methods to identify biologically relevant genes8. Feature selection is critical for both biological interpretation and predictive model performance, particularly in deep learning applications9. Although various feature selection methods have been developed for RNA-seq data in cancer studies, no generalized method performs optimally across all studies. For instance, Chen et al.10, utilized pathway-based gene sets relevant to cancer from the MsigDB hallmark database. While Elbashir et al.11, focused on differentially expressed genes, and Silva et al.12, selected genes with the highest variance across samples. This diversity underscores the lack of standardization in feature selection, making it challenging to identify consistent patterns across studies. This challenge is particularly notable in pan-cancer studies, where identifying commonalities across cancer types is essential.

Single-cell RNA sequencing (scRNA-seq) has revolutionized our understanding of the tumor microenvironment by revealing the cellular diversity and dynamics often obscured in bulk RNA sequencing data13. It provides detailed views of immune cell heterogeneity and tumor-immune interactions that strongly influence cancer progression and treatment response14. Recent advances in scRNA-seq have illuminated specific cellular mechanisms, enhancing our understanding of the immune landscape across multiple cancers15,16,17. These insights enable high-resolution biomarker discovery and predictive modeling. By capturing gene expression at the cellular level, scRNA-seq allows feature selection enriched with detailed biological context, in contrast to bulk RNA-based approach. Despite these advances in scRNA-seq technology, most current pan-cancer studies still rely on bulk RNA-seq data or predefined pathway gene sets, overlooking the opportunity to leverage single-cell heterogeneity for improved predictive modeling.

In this study, we propose a novel framework that integrated high dimensional weighted gene co-expression network analysis (hdWGCNA) from pan-cancer scRNA-seq data with XGBoost-based feature refinement. To our knowledge, this represents the first application of this combined approach in pan-cancer feature selection. We hypothesize that gene sets derived from scRNA-seq data outperform those from bulk RNA-seq in predictive modeling and downstream cancer tasks. To demonstrate this hypothesis, we analyzed scRNA-seq data from 181 tumor biopsies across 13 cancer types, identifying co-expression modules through hdWGCNA and subsequently refining them using XGBoost for task-specific feature selection. The resulting gene sets were evaluated on TCGA pan-cancer RNA-seq data across multiple downstream tasks. We compared our approach against six established reference gene sets and OncoKB curated oncogenes using two different deep learning models. Our framework demonstrated improved predictive accuracy and also identified biologically meaningful pan-cancer biomarkers, offering a robust strategy for gene selection in cancer genomics.

Results

Overview of study

Our study workflow, as illustrated in Fig.1, we conducted a comprehensive analysis of scRNA-seq data from 181 tumor biopsy samples, encompassing 13 distinct cancer types (Fig.1a). This dataset represents 13 tumor types, with their proportional distribution shown in Fig. 1b. From the scRNA-seq data, we applied hdWGCNA to construct a gene co-expression network (Fig. 1c, d). This analysis allowed us to identify clusters of co-expressed genes with shared biological functions. These co-expressed genes served as gene sets for downstream tasks. We selected genes participating in the co-expression network and excluded those without network connectivity. These selected genes were then applied to feature selection processes of RNA-seq data for downstream modeling tasks, as shown in Figure 1e (Supplementary Table 1).

Fig. 1
figure 1

Schematic overview of the study workflow. a. Each cancer type is indicated along with the number of samples analyzed: Uveal melanoma (UM, n=9), esophageal adenocarcinoma (EA, n=3), non-small cell lung cancer (NSCLC, n=8), breast invasive carcinoma (BRCA, n=8), hepatocellular carcinoma (HCC, n=18), intrahepatic cholangiocarcinoma (ICC, n=11), renal cell carcinoma (RCC, n=3), pancreatic adenocarcinoma (PAAD, n=24), colorectal cancer (CRC, n=27), ovarian cancer (OC, n=4), squamous cell carcinoma (SCC, n=4), cutaneous melanoma (CM, n=52), basal cell carcinoma (BCC, n=10). b. Proportional distribution of the 13 cancer types within the scRNA-seq dataset, highlighting the diversity and representation of different tumor types used in the study. The dataset includes 317,111 tumor immune cells classified into 25 distinct cell types. c. Schematic UMAP visualization of single-cell expression profiles, which were prepared for hdWGCNA. d. Schematic of the co-expression gene network derived from hdWGCNA. The genes comprising this biological module were used for various downstream analyses through the target gene process. e. (e) Application of selected co-expressed gene modules to various pan-cancer downstream tasks. These tasks include mutation prediction, cancer subtyping, tumor mutation burden (TMB) assessment, microsatellite instability (MSI) classification, cancer subtyping, and grading.

Full size image

We further refined these features using XGBoost to select the most relevant genes for each downstream task. Downstream tasks performed using TCGA pan-cancer RNA-seq data included tumor mutation burden (TMB) assessment, microsatellite instability (MSI) classification, mutation prediction, cancer subtyping, and grading. Our approach integrates single-cell data with advanced machine learning techniques to improve predictive modeling and facilitate the development of targeted therapies.

Combining high-resolution scRNA-seq data with advanced analytical methods refines biomarker selection and enhances model accuracy. These results highlight the importance of advanced feature selection methods combined with powerful analytical tools in advancing cancer research.

High-dimensional co-expression network through scRNA-seq

We generated a scRNA-seq dataset from 181 tumor biopsy samples, including 87,659 genes and 317,111 cells18. This dataset captures distinct immune cell subsets and provides a comprehensive view of the tumor microenvironment. UMAP visualization (Fig. 2a) illustrates the cellular diversity and distribution of 25 immune cell types across 181 samples.

Fig. 2
figure 2

Comprehensive gene co-expression networks reveal key functional modules in the tumor immune microenvironment. a. UMAP embeddings display the distribution of 25 immune cell types across 181 tumor biopsy samples, visualizing all 317,111 cells. Each color represents a distinct immune cell type. b. UMAP representation of the gene co-expression network. A total of 11 modules were identified. Nodes indicate individual genes, and edges signify co-expression relationships between genes and hub genes within modules. Node sizes are proportional to their kME (eigengene-based connectivity) values. Colors denote different co-expression modules. c. Visualization of hub gene networks for each spatial co-expression module. The 25 highest-ranked hub genes based on kME are presented. In the network, nodes represent genes, while edges indicate co-expression links. d. Heatmap summarizing GO pathway enrichment analysis for each module. Each row represents a biological process, and columns correspond to modules, with the color scale indicating Z-score values.

Full size image

For hdWGCNA preprocessing, genes expressed in fewer than 5% of cells were excluded, yielding 6,617 common genes across all cancer types. This filtering step reduced noise and ensured that downstream analyses focused on consistently expressed genes. This processed resulted in 11 modules containing 1,857 genes (Fig. 2b). Each module is represented by the top 25 hub genes with the highest interconnectivity, ranked by kME (Fig.2c), and all module-associated genes are listed in Supplementary Table 2.

To assess the biological significance of the identified modules, we conducted a gene ontology (GO) enrichment analysis. The heatmap in Fig. 2d summarizes the enrichment results, highlighting major biological processes associated with each module. Module 1 is involved in immune and apoptotic processes, which regulate tumor immunity and cell death across cancer types. Module 2 focuses on mitochondrial energy metabolism, including cellular respiration and oxidative phosphorylation, which support the high metabolic demands of proliferating tumor cells. Module 4 is involved in ribosome biogenesis and protein synthesis, including translation and gene expression, which are fundamental to cancer cell growth and proliferation. Module 5 focuses on protein folding and stability, such as the cellular response to unfolded proteins and protein stabilization. Module 6 involves vascular and immune cell migration, such as endothelial cell movement and actin filament formation, which are crucial for tumor angiogenesis and metastasis. Module 7 focuses on cell adhesion and phagocytosis, involving leukocyte adhesion and plasma membrane invagination, highlighting the tumor microenvironment’s role in cancer immunity and inflammation. Module 8 is linked to T cell activation and cytokine responses, including T cell activation pathways and interleukin signaling, emphasizing the importance of adaptive immunity in tumor recognition and elimination. Module 9 covers vesicle transport, such as golgi vesicle transport and endoplasmic reticulum to golgi transport, vital for protein trafficking and secretion in cancer cells. Module 10 involves stress response and growth factor signaling, including responses to fibroblast growth factors and interferon-beta, reflecting adaptive mechanisms for cancer cell survival. Finally, module 11 is associated with transcriptional regulation and RNA processing, including mRNA splicing, which is essential for gene expression regulation and tumor adaptability under stress.

Comparative performance analysis of gene sets for downstream tasks

After identifying gene modules through hdWGCNA, we applied XGBoost-based feature selection to refine these modules for specific downstream tasks (Supplementary Table 3). We evaluated feature importance and retained genes with scores higher than 0.001 for each task. These tasks included TMB assessment, MSI classification, mutation prediction for TP53, EGFR, and KRAS genes, cancer subtyping, and grading. For validation, we compared our refined gene sets against six reference gene sets and the OncoKB curated oncogene set (Supplementary Table 4 and 5)9,19,20,21,22,23,24. We performed stratified 5-fold cross-validation to ensure robust performance evaluation. Results from each fold were averaged and bootstrapped to calculate 95% confidence intervals (CI) for the mean area under the receiver operating characteristic curve (AUROC). Model performance was evaluated using two different models: MLP and GNN (see more details in the method).

Overall, the XGBoost-refined hdWGCNA gene sets outperformed the reference gene sets across downstream tasks, except for the BRCA-TP53 mutation prediction task (Fig. 3, Supplementary Table 6). To access statistical significance, we performed pairwise z-tests with Bonferroni correction, showing that our method achieved statistically significant gains (p < 0.005) in 88% of all comparisons across downstream tasks (Supplementary Table 7)

Fig. 3
figure 3

Comparative performance analysis of gene sets across downstream tasks. Bar plots showing the performance of nine different gene selection methods evaluated across downstream tasks using the TCGA pan-cancer RNA-seq dataset. Each bar represents the mean AUROC across 5-fold cross-validation, with error bars indicating 95% confidence intervals obtained from bootstrapping. Dashed horizontal lines mark the average AUROC across all methods within each task category for visual reference. a. Tumor mutation burden (TMB) assessment performances across four cancer types: CRC, LUAD, LUSC, and SKCM. b. Microsatellite instability (MSI) classification for CRC and STAD. c. Mutation prediction (MUT) performance for six cancer gene combinations: BRCA-TP53, LUAD-EGFR, LUAD-KRAS, LUAD-TP53, PAAD-KRAS, and STAD-TP53. All models were trained using a multi-layer perceptron (MLP) architecture with identical hyperparameters to ensure fair comparison. Higher AUROC values indicate stronger discriminative performance of the corresponding gene set.

Full size image

The results are as follows:

Tumor mutation burden (TMB): As shown in Fig.3a, our refined gene sets yielded consistently high performance. Specifically, for lung adenocarcinoma (LUAD), the average AUROC was 0.791 (95% CI: 0.782–0.800.782.800) with the MLP model and 0.764 (95% CI: 0.754–0.773.754.773) with the GNN model. In the case of lung squamous cell carcinoma (LUSC), the MLP model achieved an average AUROC of 0.647 (95% CI: 0.636–0.659.636.659), while the GNN model achieved 0.604 (95% CI: 0.592–0.616.592.616). For skin cutaneous melanoma (SKCM), the average AUROC was 0.772 (95% CI: 0.762–0.782.762.782) with the MLP model and 0.747 (95% CI: 0.736–0.758.736.758) with the GNN model. Lastly, for colorectal cancer (CRC), the average AUROC was 0.826 (95% CI: 0.818–0.835.818.835) with the MLP model and 0.808 (95% CI: 0.799–0.817.799.817) with the GNN model.

Microsatellite instability (MSI): For stomach adenocarcinoma (STAD), our refined gene sets achieved the highest AUROC with the MLP model, 0.990 (95% CI: 0.988–0.991.988.991). However, with the GNN model, the OncoKB gene set slightly outperformed ours, achieving an AUROC of 0.982 (95% CI: 0.979–0.985.979.985). For CRC, the average AUROC was 0.931 (95% CI: 0.925–0.937.925.937) with the MLP model and 0.936 (95% CI: 0.930–0.942.930.942) with the GNN model, both outperforming the reference gene sets.

Mutation prediction (MUT): In mutation prediction tasks, the OncoKB gene set achieved the highest performance for BRCA-TP53, with an average AUROC of 0.951 (95% CI: 0.949–0.954.949.954) using the MLP model, and 0.935 (95% CI: 0.932–0.938.932.938) with the GNN model. For LUAD-TP53, the average AUROC was 0.869 (95% CI: 0.862–0.876.862.876) with the MLP model and 0.872 (95% CI: 0.866–0.880.866.880) with the GNN model. Similarly, for LUAD-EGFR, the average AUROC was 0.868 (95% CI: 0.858–0.878.858.878) with the MLP model and 0.865 (95% CI: 0.855–0.875.855.875) with the GNN model. For LUAD-KRAS, the MLP model achieved an average AUROC of 0.845 (95% CI: 0.835–0.854.835.854), outperforming the GNN model, which had an AUROC of 0.792 (95% CI: 0.782–0.801.782.801). In the case of pancreatic adenocarcinoma (PAAD)-KRAS, the average AUROC was 0.904 (95% CI: 0.894–0.915.894.915) with the MLP model and 0.892 (95% CI: 0.879–0.904.879.904) with the GNN model. For STAD-TP53, both the MLP and GNN models exhibited similar performance, achieving average AUROC of approximately 0.790 (MLP: 95% CI: 0.779–0.800.779.800; GNN: 95% CI: 0.780–0.799.780.799).

Grading (GRAD): For prostate adenocarcinoma (PRAD) grading, our refined gene sets showed the highest performance among all gene sets with the MLP mode, achieving an average AUROC of 0.779 (95% CI: 0.771–0.787.771.787). The GNN model exhibited slightly lower performance, with an AUROC of 0.695 (95% CI: 0.686–0.704.686.704).

Despite these consistent results, the XGBoost-refined hdWGCNA gene sets did not outperform reference sets in all tasks. In MSI classification for STAD, the GNN model using OncoKB performed slightly better, with an AUROC only 0.003 higher than ours. Similarly, for BRCA-TP53 mutation prediction, the OncoKB gene set performed best, but the differences were minimal. The MLP model had an AUROC difference of only 0.006 and the GNN model had an AUROC difference of only 0.001. For cancer subtyping, all evaluated gene sets already exhibited high AUROCs, resulting in minimal performance gaps. Thus, detailed results are provided separately in Supplementary Table 8.

In summary, the XGBoost-refined hdWGCNA gene sets achieved robust and statistically significant performance across most downstream tasks, highlighting their effectiveness in optimizing gene selection for pan-cancer analyses.

Biological significance of XGBoost-refined hdWGCNA gene sets

To evaluate the biological relevance of the XGBoost-refined hdWGCNA gene sets, we analyzed feature importance of selected genes within each downstream task. For each task, the top 10 genes with the highest importance values were identified (Fig. 4a-d). Many of these genes are known to be associated with cancer biology, supporting the biological validity of our findings25,26,27,28,29,30,31,32,33,34,35,36,37.

Fig. 4
figure 4

Feature importance analysis and overlap of key genes across downstream tasks. a-d. Feature importance scores for the top 10 genes in each downstream task. Bar plots display XGBoost-derived feature importance scores (x-axis) for genes (y-axis). Panels represent: a. Mutation prediction (MUT) tasks across six gene-cancer combinations, b. Microsatellite instability (MSI) classification in CRC and STAD. c. Tumor mutation burden (TMB) assessment in four cancer types, d. Cancer grading (GRAD) in PRAD. e. Upset plot illustrating gene overlap across all 13 downstream tasks. The left vertical bar plot shows the total number of features selected for each task. The matrix indicates which tasks share common genes (connected dots), showing the size of each intersection. f. Kaplan-Meier survival curves for 14 genes showing significant pan-cancer prognostic value (p < 0.005). Each panel shows overall stratified by gene expression level: high expression (red line, top quartile) vs low expression (blu line, bottom quartile). Each plot displays the log-rank p-value, hazard ratio (HR) with 95% confidence interval, and HR p-value. All survival analyses were performed across 33 TCGA cancer types using the GEPIA platform.

Full size image

For TP53 mutation prediction in BRCA, CIRBP, a regulator of DNA repair and cell proliferation has been implicated in triple-negative breast cancer25. Similarly, BTF3 has been reported to be up-regulated in individuals with TP53 mutations across various cancer types26. In LUAD, KPNA2, a nuclear export protein essential for tumor formation, has been recognized for its significant role in human tumors27. Additionally, mutations in CDKN2A have been associated with TP53 mutations, underscoring its relevance in this context28. For TP53 mutation prediction in STAD, DDB2, closely linked to the p53 pathway, acts as a key regulator of p21 Waf1/Cip1 following DNA damage or p53-mediated induction29. In the LUAD-EGFR mutation prediction task, GGA2 was identified as a critical gene due to its interaction with EGFR, which increases EGFR protein levels and modulates its degradation30. For KRAS mutation prediction in LUAD, RHOB, known to be down-regulated in non-small-cell lung cancer (NSCLC), is directly related to lung cancer progression31. In PAAD, S100A11, a well-established pancreatic tissue marker, was identified as crucial due to its association with tumor stage, drug resistance, and shorter overall survival32. YWHAZ, also identified in PAAD, is recognized as a promising therapeutic target due to its role in promoting cell proliferation, migration, and invasion37. For MSI classification tasks, RPL22L1 was significant in MSI-STAD due to its up-regulation in high-MSI groups across various cancer types33. In MSI-CRC, PMEPA1 was notable for its association with tumor immunity in pan-cancer studies34. In TMB tasks, PSMB9 was identified as a strong predictor of immune response in melanoma patients, significantly influencing their response to checkpoint therapy35. In TMB-CRC, CCL5 was identified as a prognostic biomarker associated with the efficacy of immunotherapeutic interventions in various cancers36.

To identify the overlap of selected genes across downstream tasks, we visualized them using upset plots (Figure 3e). These plots highlight genes that consistently ranked high in feature importance across multiple tasks. In particular, DPM1 was significant in all 13 tasks, highlighting its critical role in pan-cancer studies. DPM1 has previously been identified as a prognostic indicator in hepatocellular carcinoma (HCC), further supporting its importance in cancer biology38. Moreover, BAD ranked high in all tasks except TMB-CRC, indicating its broad involvement across multiple cancer pathways. FKBP4, which appeared in 10 tasks, has been associated with prognostic and immunological roles in several cancers39. DUSP4, associated with treatment resistance in breast cancer was identified40. Similarly, LASP1 which is involved in metastasis in pancreatic cancer, was present in 9 tasks41. These recurrent genes highlight shared oncogenic mechanisms across cancer types and represent promising targets for future pan-cancer research.

To further validate the clinical relevance, we performed pan-cancer survival analysis using GEPIA with TCGA clinical data. We focused on genes with high feature importance across multiple tasks (Fig. 4e), stratifying patients into high and low expression groups using quartile cutoffs. Remarkably, 14 genes showed significant associations with overall survival (p < 0.005) at the pan-cancer level (Fig. 4f). These results provide independent clinical validation of our computation findings, showing that the hdWGCNA-XGBoost approach identifies features that are both statistically robust and biologically meaningful markers with direct clinical implications.

Beyond statistical associations, thirteen of the fourteen survival-associated genes identified in our analysis have prior evidence supporting their clinical and biological relevance. High expression of DPM1 has been associated with poor prognosis in hepatocellular carcinoma, reflecting its involvement in aberrant protein glycosylation42. LASP1 has been shown to drive metastasis and unfavorable clinical outcomes in several malignancies, including pancreatic adenocarcinoma43. HDGF has been widely recognized as a prognostic marker and therapeutic candidate across solid tumors44. In contrast, reduced expression of RHOB and CYFIP2 correlates with aggressive phenotypes and poorer survival, consistent with their tumor-suppressive functions45,46. DDIT4 mediates therapy resistance through mTORC1 inhibition and metabolic adaptation47. Upregulation of PFDN2, TPD52L2, and IFRD1 has been reported in association with tumor aggressiveness, recurrence, or resistance phenotypes48,49,50. Expression of ATP1B3 correlates with survival outcomes in hepatocellular and gastric cancers51, while C1orf174 shows diagnostic and prognostic potential in colorectal cancer52. RPL22L1 promotes invasion and metastasis, contributing to adverse clinical outcomes in multiple cancer types. Although SLC25A5 has been implicated in tumor metabolism and treatment response in preclinical studies53, there is currently no direct evidence that genetic variation in this gene modifies the clinical benefit of antibody-based therapies. Accordingly, SLC25A5 is conservatively presented as a potential biomarker warranting further validation.

Discussion

This study highlights the potential of integrating scRNA-seq data with advanced feature selection to improve predictive modeling in cancer genomics. By combining hdWGCNA with XGBoost-based refinement, we generated scRNA-seq-derived gene sets that outperformed conventional bulk RNA-seq gene sets across multiple pan-cancer tasks. The high-resolution single-cell data captured cellular heterogeneity and tumor microenvironment interactions, thereby enhancing biological interpretability. Additionally, XGBoost refinement enabled the selection of the most significant genes tailored to specific tasks, improving model accuracy and robustness. Many of the identified genes are well-established in cancer research, confirming the validity of our framework. In addition, several novel genes emerged as a potential pan-cancer biomarkers, offering new avenues for therapeutic exploration. Overall, our findings underscore the value of integrating scRNA-seq data with machine learning based feature selection, providing a scalable and biologically grounded strategy for discovering key genetic drivers in cancer.

Traditional feature selection approaches for RNA-seq data, such as those using MsigDB or OncoKB, typically rely on predefined pathways or cancer gene-based sets24,54. These databases have shown high performance due to their curated biological foundation, effectively capturing relevant pathways and processes. However, methods that depend solely on bulk RNA-seq or static gene lists often fail to reflect the complexity of the tumor microenvironment. For example, Mohammed et al. used LASSO regression to screen 173 genes with 10-fold cross-validation11. Similarly, Chen, Joe W., et al. selected the top 500 differentially expressed genes by p-value21, and Silva., et al chose genes with the highest variance across samples12. While these methods are statistically robust, those approaches overlook cellular heterogeneity and gene interactions within tumors, potentially missing genes active in rare but critical cell populations. Furthermore, the resulting feature sets are typically tailored to single-cell analyses, limiting their generalizability for identifying shared biomarkers across multiple cancer types.

Our framework introduces several key advances over existing approaches. First, by leveraging scRNA-seq data with hdWGCNA, we capture cell-type specific co-expression patterns that remain hidden in bulk analyses. Second, the integration with XGBoost enables task-specific gene prioritization, moving beyond static pathway- or variance-based selection. Unlike previous studies that relied on statistical filters, differential expression, or variance-based selection, our method combines the biological insights from hdWGCNA with the predictive power of XGBoost. To our knowledge, this represents the first application of this integrated approach in pan-cancer features selection. Furthermore, our refined gene sets consistently outperformed reference gene sets across most tasks, demonstrating their robustness and broad applicability. This approach improves predictive accuracy while enhancing biological interpretability and generalizability across cancer types. The success of our gene sets in these tasks underscores their potential utility in multi-modal research, where combining diverse omics data could further enhance predictive modeling.

Our study highlights several genes, including DPM1, BAD, and FKBP4, as promising pan-cancer biomarkers. FKBP4 has been previously reported in a pan-cancer context, supporting its relevance across multiple cancer types39. Similarly, DPM1 has been identified as a prognostic marker in hepatocellular carcinoma and breast cancer38, and also showed consistent importance across all 13 tasks in our analysis, suggesting its potential as a pan-cancer biomarker. Furthermore, our pan-cancer survival analysis revealed that 14 high-importance genes such as DPM1, LASP1, and RHOB, demonstrated significant associations with patient survival. This independent clinical validation confirms that our computational approach identifies not only statistically significant features but also biomarkers with direct prognostic relevance. Notably, many of the survival-associated genes identified in our analysis correspond to known prognostic or therapeutic biomarkers in clinical oncology. Their consistent recurrence across multiple cancer types underscores the translational value of our framework, bridging computational discovery with clinically actionable insights that may guide patient stratification and targeted therapy development.

Like other studies, our study has certain limitations. Frist, although the scRNA-seq dataset included 13 cancer types, it did not cover all possible tumor types, which may limit the generalizability of our findings. However, our framework maintained strong performance even in unseen cancers, such as the grading task in PRAD, highlighting its broad applicability. Second, while we propose that our gene sets could enhance multi-modal model performance, this remains to be experimentally validated. Recent pan-cancer studies have increasingly adopted multi-modal approaches, integrating transcriptomic data with other multi-omics or whole-slide pathology images55,56. In particular, foundation models trained on whole-slide images (WSI) have shown that incorporating RNA features through self-supervised fine-tuning can improve downstream task performance. Furthermore, as our downstream tasks align with precision medicine goals, recent advances in contrastive learning show promise in biomedical applications. Contrastive learning has been successfully applied in both drug-target affinity prediction, and broader computational drug repositioning under label sparsity highlight the potential of representation learning in biomedical domains57,58,59. While these methods have primarily focused on drug-molecular interactions, incorporating transcriptomics features from targeted gene sets could enable these frameworks to account for patient-specific molecular contexts, potentially improving their predictive accuracy in clinical settings.

Beyond these advances, several recent studies in omics-driven drug discovery, such as OmicsTweezer and GraphCL-DTA60,61, have demonstrated that effective feature selection and representation learning from omics data can substantially enhance predictive modeling in drug-target interaction, drug repositioning, and multi-omics integration. These frameworks share conceptual similarities with our hdWGCNA-XGBoost approach, as they all rely on extracting biologically meaningful molecular features that drive downstream predictive performance. Integrating scRNA-seq-derived gene sets such as those proposed in our study could further complement these models by introducing patient-specific transcriptomic signals, thereby enriching molecular context in drug response prediction and other translational applications.

In conclusion, our study provides a robust framework to optimize feature selection in cancer genomics by integrating high-resolution scRNA-seq data with advanced analytical techniques. This approach improves predictive performance across diverse pan-cancer tasks and identifies gene sets with strong biological and clinical relevance. By revealing biomarkers with pan-cancer significance, our framework supports the development of more accurate and reliable diagnostics and prognostic models. Moreover, our results offer valuable insights that can be applied to multi-modal frameworks, thereby advancing personalized cancer treatment strategies and contributing to the broader field of cancer genomics.

Methods

Datasets

To characterize the gene expression landscape across a wide range of cancers, we obtained scRNA-seq data for 181 tumor biopsy samples from 13 different cancer types, which were collected by Nieto., et al.18. We obtained RNA-seq data for 7,178 tumor samples across 16 cancer types from The Cancer Genome Atlas (TCGA), accessed through the NIH Genomic Data Commons Portal62. All datasets used in this study were downloaded using the GenomicDataCommons package in R.

To facilitate various downstream tasks, cancer types were grouped as follows. For the MSI classification, CRC, specifically colon adenocarcinoma (COAD) and rectal adenocarcinoma (READ), were collectively analyzed as a combined TCGA project. Similarly, for subtype classification, lung cancers were grouped as non-small cell lung cancer (NSCLC), combining both LUAD and LUSC. Kidney cancers were classified as renal cell carcinoma (RCC), comprising kidney renal clear cell carcinoma (KIRC), kidney renal papillary cell carcinoma (KIRP), and kidney chromophobe (KICH). Melanomas (MEL), which include skin cutaneous melanoma (SKCM) and uveal melanoma (UVM), were classified separately due to their distinct anatomical origins, despite both being melanomas. Finally, gynecologic cancers (GYN) were comprised into ovarian serous cystadenocarcinoma (OV), uterine corpus endometrial carcinoma (UCEC), cervical squamous cell carcinoma, and endocervical adenocarcinoma (CESC).

RNA-seq data pre-processing

RNA-seq data were pre-processed to ensure consistency across samples and to facilitate accurate comparative analysis. Read counts were first normalized by adjusting for gene length to accounts for differences in transcript length. The normalized counts were then scaled to counts per million to standardize for sequencing depth, providing a standardized comparison of gene expression levels. Then, the read counts were log-transformed to stabilize variance63.

Data preprocess for pan-cancer downstream tasks

Based on this processed data, TMB scores were calculated for each sample by further normalizing read counts against gene lengths and adjusting for sequencing depth. First, the RNA-seq data were mapped to barcodes for each sample. Then, read counts were divided by gene lengths and were scaled by a factor of \({10}^{6}\) bases to derive TMB scores for each sample. Samples with a TMB score higher than 10 were categorized as TMB-high, while those with a TMB score below 10 were classified as TMB-low64.

To classify MSI status, we accessed missense mutations in key DNA mismatch repair (MMR) genes, including MLH1, MLH3, MSH2, MSH3, MSH4, MSH5, MSH6, PMS1, and PMS265. Samples were classified as microsatellite unstable (MSI-H) or microsatellite stable (MSS).

For prostate cancer grading, we used gleason scores to categorize samples into three groups: low, moderate, and high. Clinical data were downloaded from cBioPortal, and samples were grouped based on their gleason scores66. Samples with a score of 6 were assigned to the low-grade group, a score of 7 in the moderate group, and scores of 8, 9, and 10 in the high group.

For the mutation prediction tasks, we used TCGA somatic mutation data to identify mutations. We determined the mutation status of target genes in each sample. This mutation information was then cross-referenced with RNA-seq data to ensure consistency in the samples. Only patients with both mutation data and corresponding RNA-seq data were included in the final dataset for further analysis.

Feature selection through high dimensional weighted gene co-expression network analysis

To identify co-expressed modules in the pan-cancer scRNA-seq data, we performed weighted gene co-expression network analysis (WGCNA)67. First, metacells were constructed for each immune cell subset to optimize the data sparsity and reduce technical noise. We applied principal component analysis (PCA) for dimensionality reduction, followed by the k-nearest neighbors (k=25) aggregation to generate metacells within each cell type-patient group. The metacell gene expression matrix was then normalized and used as input for WGCNA. The soft-thresholding power (β=9) was selected using the TestSoftPower function to achieve approximate scale-free topology, as recommended in the hdWGCNA pipeline. Co-expression modules were then identified using the dynamic tree-cut method with a minimum module size of 30. Harmonized module eigengenes (hMEs) were calculated across patients to summarize module-level activity. For network visualization, we generated a UMAP embedding based on the topological overlap matrix, which illustrating the modular structure and relationships among genes.

Pathway enrichment analysis

Pathway enrichment analysis was performed using the GO Biological Process database68. For each hdWGCNA module, we selected the top 100 genes based on kME scores. Enrichment significance was accessed using log-transformed P-values. The top 5 enriched terms were selected for further analysis. Analyses were performed using the Enrichr package in R69.

Model development

We developed two separate model architectures optimized for RNA-seq data: a multi-layer perceptron (MLP) model and a graph neural network (GNN) model.

MLP

The MLP model comprises of an embedding module and a classifier module. The embedding module reduces input dimensional input features into a lower-dimensional space through a single linear layer. This dimensionality reduction is followed by batch normalization and Leaky ReLU activation, which stabilize training and enhance pattern recognition70. Dropout is applied to prevent overfitting. The classifier module processes these features through three sequential linear layers, with SELU activation and dropout after each layer71. This design preserves feature variability and enhances classification robustness. Additionally, the number of neurons in each layer is adjusted in each layer according to the specific needs of each downstream task.

GNN

The GNN model was adapted from previous study72. The gene embedding module contains an input embedding layer that projects input features to a hidden dimension using a linear layer, followed by dropout to prevent overfitting. This is followed by multiple graph transformer layers that perform message passing and update node features, incorporating batch normalization and layer normalization for training stability. Each graph transformer layer uses multi-head attention to capture diverse relational patterns. Graph layer outputs are reshaped and processed by a fully connected linear layer, followed by batch normalization and dropout. This embedding is then flattened to form the input to the classifier module. The classifier module consists of three sequential linear layers with SELU activation functions and dropout. Batch normalization is applied after each linear layer to maintain feature stability.

Model training

Both models were trained using binary cross-entropy loss for binary classification and cross-entropy loss functions for multi-class classification. Models were optimized using the Adam optimizer with a mini-batch size of 32 and a learning rate of 0.0173. Early stopping with a patience of 30 epochs was implemented. Model performance was evaluated using stratified cross-validation74. For each fold, we conducted 100 bootstrap iterations on the test set to obtain 95% confidence interval. This approach ensures robust performance estimates across different patient samples. All experiments were conducted using PyTorch (version 2.0.1) on an NVIDIA A100 GPU (40 Gb) equipped with CUDA version 11.7.

Feature importance through XGBoost

We used XGBoost to access feature importance from genes75. For each downstream task, we performed 100 bootstrap iterations to assess the importance of gene features. In each iteration, we resampled the dataset, trained the XGBoost classifier, and evaluated its performance using metrics such as AUROC and AUPRC. This bootstrap approach identified genes with consistently high feature importance, highlighting their importance in biological processes relevant to our study.

Pan-cancer survival analysis

To validate the clinical relevance of our identified biomarkers, we performed pan-cancer survival analysis using the GEPIA2 platform (http://gepia.cancer-pku.cn/). For each gene, we evaluated overall survival across all 33 available TCGA cancer types. Patients were stratified into high and low expression groups using quartile cutoffs (top 25% vs bottom 25%). Kaplan-Meier survival curves were generated, and statistical significance was accessed using log-rank test. Hazard ratios (HR) with 95% confidence intervals were calculated using Cox proportional hazards regression. Genes with p(HR) < 0.005 were considered to have significant prognostic value.