Transcriptomic Gene Network Profiling and Weak Signal Detection for Prediction of Ovarian Cancer Occurrence, Survival, and Severity by Integrating Bulk and Single-cell RNAseq Data

Background Ovarian cancer (OC) is a significant gynecological malignancy characterized by its high mortality rate, poor long-term survival rate, and late-stage diagnosis. OC is the 5th leading cause of cancer death among woman and counts 2.1% of all cancer death. OC survival rates are much lower than other cancers that affect woman. Its 5-year survival rate is less than 50%. Only ∼17% of OC patients are diagnosed within the early stage. The majority are diagnosed at an advanced stage, making early detection and effective treatment critical challenges. Currently, the identified OC predictive genes are still very sparse, resulting in pool prognostic performance. There exists unmet needs to identify novel prognostic gene biomarkers for OC occurrence, survival, and clinical stages to promote the likelihood of survival and to perform optimal treatments or therapeutic strategies at the earliest stage possible.

Methods Previous RNAseq analysis on OC focused on detecting differentially expressed (DE) genes only. Many genes, although having weak marginal differential effects, may still exude strong predictive effects on disease outcomes though regulating other DE genes. In this work, we employed a new machine learning method, netLDA, to detect such predictive coregulating genes with weak marginal DE effects for predicting OC occurrence, 5-year survival, and clinical stage. The netLDA detects predictive gene networks (PGN) containing strong DE genes as hub genes and detects coregulating weak genes within the PGNs. The network structures of the detected PGNs along with the strong and weak genes therein are then used in outcome prediction on test datasets.

Results We identified different sets of signature genes for OC occurrence, survival, and clinical stage. Previously identified prognostic genes, such as EPCAM, UBE2C, CHD1L, TP53,CD24, WFDC2, and FANCI, were confirmed. We also identified novel predictive coregulating weak genes including GIGYF2, GNPAT, RAD54L, and ELL. Many of the detected predictive gene networks and coregulating weak genes therein overlapped with OC-related biological pathways such as KEGG tight junction, ribosome, and cell cycle pathways. The detection and incorporation of the gene networks and weak genes significantly improved the prediction performance. Cellular mapping of selected feature genes using single-cell RNAseq data further revealed the heterogeneous expression distributions of the signature genes on different cell types.

Conclusions We established a transcriptomic gene network profile for OC prediction. The novel genes detected provide new targets for early diagnostics and new drug development for OC.