Fig 1.
Overview of PDRWH for prioritizing personalized cancer driver genes.
(A) Model input. i) The somatic mutation profiles from the TCGA; ii) The gene expression data of patients; iii) Gene interaction network. (B) Pro-processing the gene expression profiles and determining the abnormally expressed genes for each sample. (C) Construction of a hypergraph model for each patient. In this model, each hyperedge represents a patient and the vertices incident to each hyperedge represent the mutated genes and abnormally expressed genes in the corresponding patient. (D) Computing the transition probability matrix of the random walks on the weighted hypergraph. (E) The process of generating PDRWH-scores through random walks on the personalized hypergraph.
Fig 2.
Comparison of the PDRWH with the other personalized prediction methods.
The average precision, recall, and F1-score for (A) the BRCA dataset, (B) the KIRC dataset, and (C) the LIHC dataset, are plotted as a function of the number of top-n ranked genes involved in the calculation of the scores. The general driver gene list is used as the reference set.
Fig 3.
Prediction performance of five personalized prediction methods as well as five cohort prediction methods.
(A-C) ROC plots of results on the different cancer types based on the general reference driver set. The solid lines represent the personalized prediction methods (PDRWH, PersonaDrive, Prodigy, DawnRank, and SCS). The dashed lines indicate the cohort-level prediction methods (OncodriveFML, MinNetRank, MutsigCV, Subdyquency and DriverRWH). The numbers in parentheses behind the methods are the AUC values of the corresponding method.
Fig 4.
The performance of PDRWH and other four methods for identifying the known tumor-specific driver genes.
(A) The percentage of patients whose predicted personalized drivers are significantly enriched in the known tumor-specific driver gene list. (B) Comparison of the number of predicted tumor-specific driver genes by various methods and the recall ratio. (C) Overlap among the tumor-specific cancer drivers predicted by different methods for BRCA, KIRC, and LIHC.
Fig 5.
PDRWH identifies both common and rare drivers.
(A-C) Distribution of mutation frequency of top genes predicted by PDRWH. The i-th column in the plot represents the distribution of mutation frequency of the genes which ranked at the i-th in the predicted drivers. Each range of mutation frequency is further classified into whether the genes are known drivers in the reference set. (D-F) Scatter plots about mutation frequency of potential drivers and the occurrence of genes as predicted driver gene. Known tumor-specific driver genes are represented as red dots and others are represented as black dots. Purple lines constructed by known tumor-specific driver genes are the regression lines.
Fig 6.
The survival curves for subtyping BRCA, KIRC, and LIHC using the gene expression data.
(A) In different cancer types, the expression data of known tumor-specific drivers with mutation frequency ≥ 2% were used in subtyping patients. Different subtypes (S1, S2,…) are indicated by different colored lines. (B-D) The similar analysis based on expression data of genes that are known tumor-specific drivers with mutation frequency < 2%, predicted driver genes with mutation frequency ≥ 2%, and predicted driver genes with mutation frequency < 2% respectively.
Fig 7.
Distribution of the number of predicted personalized driver genes in TARGET and DGIdb.
(A) For cancer type BRCA, the first pie chart shows the distribution of the number of predicted personalized driver genes in TARGET. Restricted to predicted personalized drivers predicted by PDRWH, there are 17.88% of patients with not less than three actionable driver genes. The second pie chart shows the distribution of the number of predicted personalized driver genes in DGIdb. There are more than 50% of patients with not less than three druggable personalized drivers. The third pie chart is the distribution of the number of predicted personalized driver genes in the union of the two sets. (B-C) The similar pie charts display for cancer type KIRC and LIHC.
Fig 8.
In vitro assays of a novel driver gene LRP1 predicted by PDRWH.
(A) The expression of LRP1 was detected between GES-1 and GC cells. (B) The expression of LRP1 was detected between GC tissues and adjacent tissues using Immunohistochemical analysis. (C) Overall Survival analysis of LRP1 based on GEPIA. (D) HGC-27 cells transfected with siRNA by real-time PCR and Western Blot. (E) Wound healing assay following knockdown of LRP1 in HGC-27 cells. (F) Apoptosis detection for HGC-27 cells transfected with siRNA. (G) Proliferation detection for HGC-27 cells transfected with siRNA using EdU assay. (H) Cell cycle profile of control and LRP1 knockdown cells. GAPDH protein is used as control. All cell assays were performed in triplicate. The error bars indicate SD of three independent experiments. *P < 0.05, **P < 0.01 using the two-sided Student’s t test.