Integrative multi-omics and machine learning approach reveals tumor microenvironment-associated prognostic biomarkers in ovarian cancer

Abstract

Background: Ovarian cancer (OC) is a globally prevalent malignancy with significant morbidity and mortality, yet its heterogeneity poses challenges in treatment and prognosis. Recognizing the crucial role of the tumor microenvironment (TME) in OC progression, this study leverages integrative multi-omics and machine learning to uncover TME-associated prognostic biomarkers, paving the way for more personalized therapeutic interventions.

Methods: Employing a rigorous multi-omics approach, this study analyzed single-cell RNA sequencing (scRNA-seq) data from OC and normal tissue samples, including high-grade serous OC (HGSOC) from the Gene Expression Omnibus (GEO: GSE184880) and The Cancer Genome Atlas (TCGA) OC cohort, utilizing the Seurat package to annotate 700 TME-related genes. A prognostic model was developed using the least absolute shrinkage and selection operator (LASSO) regression and independently validated against similarly composed HGSOC datasets. Comprehensive gene expression and immune cell infiltration analyses were conducted, employing advanced algorithms like xCell to delineate the immune landscape of HGSOC.

Results: Our investigation unveiled distinctive immune cell infiltration patterns and gene expression profiles within the TME of HGSOC. Notably, the prevalence of exhausted CD8⁺ T cells in high-risk patient samples emerged as a critical finding, underscoring the dualistic nature of the immune response in OC. The developed prognostic model, incorporating immune cell markers, exhibited robust predictive accuracy for patient outcomes, showing significant correlations with immunotherapy responses and drug sensitivities.

Conclusions: This study presents a groundbreaking exploration of the OC TME, offering vital insights into its molecular intricacies. By systematically deciphering the TME-associated gene signatures, the research illuminates the potential of these biomarkers in refining patient prognosis and guiding treatment strategies. Our findings underscore the necessity for personalized medicine in OC treatment, potentially enhancing patient survival rates and quality of life. This study marks a significant stride in understanding and combatting the complexities of OC.

Keywords: Ovarian cancer (OC); machine learning; prognostic biomarkers; single-cell RNA sequencing (scRNA-seq); tumor microenvironment (TME).

Figure 1

Integrative multi-omics and machine learning approach to identify tumor microenvironment-associated prognostic biomarkers in ovarian cancer. This study employs an integrative analysis combining scRNA-seq data from the GSE184880 dataset, which includes 5 normal and 7 OC samples, with bulk RNA-seq data from the TCGA-OC cohort (n=378). The workflow begins with cell annotation and the identification of 700 TME-related genes. These genes are subsequently utilized to develop a prognostic model using the LASSO method (n=8 genes), which is validated using an independent dataset, GSE140082 (n=380 OC samples). The analysis further includes immune cell infiltration profiling via the xCell algorithm, DEG analysis, and functional annotation within TCGA-OC cohort. Additional analyses encompass T-cell subset characterization, immunotherapy response prediction, drug sensitivity assessment, cell trajectory analysis, and gene function enrichment, providing comprehensive insights into the immune landscape and its implications for prognosis and therapy in ovarian cancer. This figure was created using Figdraw. By figdraw.com. scRNA-seq, single-cell RNA sequencing; OC, ovarian cancer; RNA-seq, RNA sequencing; TCGA, The Cancer Genome Atlas; TME, tumor microenvironment; LASSO, least absolute shrinkage and selection operator; DEG, differential gene expression; GO, Gene Ontology.

Figure 2

Comprehensive cell distribution and gene expression analysis in ovarian cancer TME. (A) Cell distribution map of OC and normal tissue samples, showing the spatial localization of cells post t-SNE analysis. This visualization highlights the heterogeneity within the TME of OC. The numbers 0–8 represent specific cell subtypes, corresponding to: 0, B cells; 1, endothelial cells; 2, epithelial cells; 3, fibroblasts; 4, monocytes; 5, plasma cells; 6, SMC; 7, T cells; and 8, other. (B) Differential expression gene markers for distinct cell types identified using the FindMarkers command, serving as unique biological signatures for each cell type. (C) Identification of 8 distinct cell subclusters annotated based on known cell markers, demonstrating the diversity of cell populations within the TME. (D) Proportional representation of each cell subcluster in normal and ovarian cancer samples, revealing significant variations in T cell and fibroblast populations between the two. The color scheme in this panel corresponds directly to the cell types illustrated in (C), ensuring consistency across the visual data representation. (E) Highlighted gene expression profiles and pathways in each cell subgroup specific to ovarian cancer samples, including 7 key TME markers, providing insights into the molecular mechanisms driving OC progression. t-SNE, t-distributed stochastic neighbor embedding; SMC, smooth muscle cells; OC, ovarian cancer; TME, tumor microenvironment.

Figure 3

Comparison of T cell characteristics between OCT and normal T cells. This figure presents box plots comparing various functional and signaling pathways between OCT and normal T cells. Key pathways include T cell exhaustion, activation/effect, TCR signaling, cytokine/chemokine receptor expression, adhesion molecule expression, IFN response, Treg signature expression, costimulatory molecule expression, OXPHOS, glycolysis, and apoptotic processes. Statistical significance between the groups is assessed using Wilcoxon test P values, with most comparisons showing highly significant differences (P<2.2e−16), highlighting the distinct functional profiles of T cells within the tumor microenvironment compared to those in normal tissue. OCT, ovarian cancer T cells; TCR, T cell receptor; IFN, interferon; Treg, regulatory T cells; OXPHOS, oxidative phosphorylation.

Figure 4

Establishment and validation of the ovarian cancer prognostic model using TME-related genes. (A) Cross-validation for tuning parameter selection in the LASSO model shows the minimum criteria and the 1 standard error of the minimum criteria (1−SE criteria). (B) LASSO coefficient profiles of the TME-related genes plotted against the log(lambda) sequence, illustrating the shrinkage of coefficients as the penalty increases. (C) Box plots representing the distribution of the gene expression levels for key biomarkers in low-risk and high-risk groups, indicating significant differences between the two, where *, P<0.05 and ****, P<0.0001. (D) Risk score distribution, patient survival status, and heatmap of gene expression profiles. The top panel displays risk scores with a threshold defining low and high-risk groups, the middle panel shows the survival status of patients, and the bottom panel is a heatmap of key genes differentially expressed between the two groups. (E) Kaplan-Meier curves for the low-risk and high-risk groups with a significant difference in survival (P<0.001) and time-dependent ROC curves demonstrating the model’s predictive accuracy. (F) Validation of the prognostic model with a Kaplan-Meier survival curve for an external dataset showing a significant difference between the two risk groups (P=0.03) and its respective ROC curves. (G) Further validation with another external dataset illustrating significant survival differences (P<0.001) and the predictive accuracy of the model with ROC curves. AUC, area under the curve; TME, tumor microenvironment; SE, standard error; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic.

Figure 5

Differential gene expression and enrichment pathway analyses between high-risk and low-risk groups in the TCGA-OC cohort. (A) PCA of the TCGA-OC cohort showing a clear distinction between high-risk (red) and low-risk (blue) groups along the PC1 and PC2 axes, which explain 8.9% and 5% of the variance respectively. (B) Volcano plot illustrating the differentially expressed genes between high-risk and low-risk groups, with significant genes highlighted; red dots represent upregulated genes and blue dots represent downregulated genes in the high-risk group. (C) Heatmap of the top differentially expressed genes with hierarchical clustering, showing patterns of gene expression across the two risk groups. (D) Bar chart of KEGG pathway enrichment analysis, where the y-axis represents the pathways and the x-axis represents the count of genes involved, indicating pathways such as Toll-like receptor signaling and NF-kappa B signaling are prominently enriched. (E) GSEA showing enriched pathways with negative or positive NES and their statistical significance, demonstrating a significant association with immune-related pathways. TCGA, The Cancer Genome Atlas; OC, ovarian cancer; PCA, principal component analysis; NF-kappa B, nuclear factor kappa-light-chain-enhancer of activated B cells; KEGG, Kyoto Encyclopedia of Genes and Genomes; TRP, transient receptor potential; GSEA, gene set enrichment analysis; NES, normalized enrichment score.

Figure 6

Immune infiltration and core risk gene expression in ovarian cancer. (A) t-SNE plots illustrating the expression patterns of core risk genes, including LGALS2, EDNRB, FCGBP, MATK, FKBP2, CD24, PAX5, and C5AR1, across various cell subpopulations. Notably, LGALS2 and C5AR1 are predominantly expressed in monocytes, while MATK shows significant expression in T cells. (B) Box plots comparing immune cell infiltration levels within the OC microenvironment between low-risk and high-risk groups. Significant differences in immune cell proportions are denoted as follows: *, P<0.05; **, P<0.01; ****, P<0.0001; and ns, no significant difference. (C) Heatmap showing the correlation between core risk gene expression and different immune cell types. The heatmap reveals a strong positive correlation between C5AR1 and both macrophages and monocytes, and a strong negative correlation between FCGBP and Th2 cells. The significance levels are represented as follows: *, P<0.05; **, P<0.01; and ***, P<0.001. t-SNE, t-distributed stochastic neighbor embedding; OC, ovarian cancer; FCGBP, Fc fragment of IgG binding protein; Th2, T helper type 2 cells.

Figure 7

Immune therapy responsiveness and drug sensitivity prediction in high-risk and low-risk ovarian cancer groups. (A) TIDE analysis reveals significant differences in scores between high-risk and low-risk groups, with high-risk groups exhibiting elevated TIDE scores, suggesting a reduced response to immune therapy. (B) Correlation plot displaying a significant positive correlation between TIDE scores and signature risk scores (R=0.33, P=3.6e−08), reinforcing the link between higher risk scores and diminished immune response. (C) Box plots illustrating the distribution of IC₅₀ values for various FDA-approved chemotherapeutic and immune drugs, demonstrating differential sensitivities between high-risk and low-risk groups. Specifically, high-risk groups show increased sensitivity to cisplatin, etoposide, and metformin, while displaying decreased sensitivity to cyclopamine, erlotinib, nilotinib, sunitinib, and tipifarnib. Statistical significance is denoted as follows: *, P<0.05; **, P<0.01; and ***, P<0.001. (D) Spearman correlation analysis between drug IC₅₀ values and risk scores for selected chemotherapeutic agents (rapamycin, sorafenib, MG-132) and immunomodulatory agents (ACY-1215, crizotinib, cediranib, VE821, GSK2830371), with positive correlations indicating potential therapeutic efficacy in high-risk ovarian cancer patients. TIDE, tumor immune dysfunction and exclusion; IC₅₀, half-maximal inhibitory concentration; FDA, Food and Drug Administration.

Figure 8

Analysis of T cell subgroups in ovarian cancer. (A) t-SNE plots identifying five T cell subgroups within the ovarian cancer T cell population: naive T, Treg, exhausted CD8, NKT, and cytotoxic CD8 cells. (B) Stacked bar charts displaying the proportions of each T cell subgroup in normal and ovarian cancer samples, with a higher prevalence of exhausted CD8 cells in cancer samples. The colors used in this panel correspond to those in (A) to maintain consistency in the representation of T cell subgroups. (C) Volcano plot highlighting differentially expressed marker genes across T cell subgroups. (D) Scatter plots showing differential gene expression between normal and ovarian cancer T cells, with key upregulated and downregulated genes identified. (E,F) Enrichment analysis bar graphs and line plots showing the pathways in which the differentially expressed genes are predominantly involved, highlighting their significance in various biological processes. t-SNE, t-distributed stochastic neighbor embedding; Treg, regulatory T cells; NKT, natural killer T cells; OC, ovarian cancer; FC, fold change; NES, normalized enrichment score.

Figure 9

Pseudotime analysis and gene module expression in T cell subgroups. (A) UMAP plot showing different T cell subtypes including naive T, Treg, exhausted CD8, cytotoxic CD8, and NKT cells, with each cell type represented by different colors as indicated in the legend. (B) UMAP plot colored by pseudotime values, illustrating the progression of cells along a developmental trajectory. (C) UMAP plot displaying the log-transformed P value [log₁₀(P value + 0.1)] for the expression of the MATK gene, with darker colors indicating higher expression levels across the T cell subtypes. The numbers 1, 2, 3, and 4 in (A-C) indicate key branch points that correspond to significant changes in the developmental stages of these T cells. (D) Heatmap representing the clustering of gene expression profiles across T cell subtypes, divided into four clusters as indicated by the color scale on the right, which represents the z-score of gene expression. (E) GO enrichment analysis for the different clusters identified in (D), with dot plots showing significantly enriched GO terms, where the size of the dots represents the count of genes and the color indicates the P value. The analysis highlights the biological processes related to T cell differentiation, immune response, and cellular activation. Treg, regulatory T cells; NKT, natural killer T cells; MATK, megakaryocyte-associated tyrosine kinase; UMAP, Uniform Manifold Approximation and Projection; GO, Gene Ontology.