Fig 1.
Overview of personalized network control model (PNC) for identifying personalized driver genes.
(a) The motivation of PNC. By integrating sample specific network theory and structure based network control theory, a small number of personalized driver genes that are altered in response to oncogene activations is detected for triggering the state transition of an individual patient from the healthy state to the disease state at expression level. The main consideration of PNC is to design methods for i) constructing personalized state transition networks to capture the phenotypic transitions between healthy and disease states / attractors, and ii) developing network control method on personalized state transition networks where we identify personalized driver genes for driving individual system in cancer from normal attractor to disease attractor through oncogene activation. (b) The main workflow of PNC. Our PNC consists of two main parts. One is a paired single sample network construction method (Paired-SSN), which we used to construct personalized state transition networks. In these networks, edges denote the significant difference of gene interactions between normal sample and tumor sample of an individual patient. For the second part of PNC, we designed a novel nonlinear structural control method (namely, NCUA) for identifying personalized driver genes. This was done to ensure that the state in the personalized state transition network would asymptotically be changed from normal state to disease state through oncogene activations.
Fig 2.
Overview of Paired-SSN for constructing personalized state transition networks.
For a given cancer patient, TCGA.EJ.7781 in Prostate Adenocarcinoma (PRAD) cancer data, we chose the expression data of all normal samples in PRAD as the reference data and respectively constructed the co-expression network of tumor sample (white color) and normal sample (green color) with the reference data by using SSN method. Then the personalized state transition network was constructed where the nodes represent the genes and edges denote the significant difference of gene interactions between normal sample and tumor sample of an individual patient in the disease development. Here we showed the individual specific sub-networks related with driver gene TP53 which contain its first-order neighboring genes as an example in this Fig.
Fig 3.
The significant enrichment F-scores of PNC and other methods for identifying cancer driver genes.
Fig 4.
The performance evaluation of different single sample network construction methods and network control methods.
(a) The significant enrichment F-scores of NCUA on personalized state transition networks constructed by Paired-SSN (the first step of our PNC), SSN and LIONESS. (b)The enrichment F-scores of NCUA (the second step of our PNC) and other structure based network control methods (MMS and MDS and DFVS) on the paired SSN networks (the first step of our PNC) in the list of CCG and NCG genes. (c) The enrichment F-scores of PNC with gene interaction network used in this work (Network 1) and other references (Network 2, Network 3) and the gene interaction network from STRING data set (scores>900) (Network 4) and the gene interaction network with top 10000 high scores from STRING data set (Network 5) and our filleted reliable gene interaction network (Network 6) and our filleted unreliable gene interactions network (Network 7).
Fig 5.
The F-measures on different cancer data sets by using PNC on gene interaction network used in this work (Network 1) and our filleted reliable gene interaction network (Network 6) and our filleted unreliable gene interactions network (Network 7).
Fig 6.
Statistic analysis of the personalized driver genes.
(a) The fractions of high-frequency personalized driver genes (yellow, fh), medium-frequency personalized driver genes (red, fm), and low-frequency personalized driver genes (blue, fl) on 13 cancer datasets. (b) The fraction of personalized driver genes in CCG and NCG with low mutation frequency and high mutation frequency respectively on 13 kinds of cancer data.
Fig 7.
The p-value of personalized driver genes enriched in CCG and NCG genes on 13 kinds of cancer data.
The red line denotes the significant threshold value 0.05.
Table 1.
The statistic information of individual specific sub-networks of TP53 in cancer.
Fig 8.
Heat map of the patient frequency (different colors) of these 20 intersected pathways in 13 cancer data sets which was previous reported in other literatures.
Fig 9.
Controllability evaluation of individual patient system by using PNC.
(a) Box plot with the distribution of the personalized controllability in 13 cancer datasets. (b) Heat map of the personalized controllability (different colors) which is as a function of the average degree and power law degree exponent.