Fig 1.
The computational method of DryNetMC (differential regulatory network-based modeling and characterization) developed to prioritize key genes responsible for drug resistance.
(I) The TCGs were selected as core genes from time-course RNA-seq data of sensitive and resistant cells. (II) The dynamic GRNs for sensitive cells and resistant cells were reconstructed via an integrated approach that incorporates prior information, data interpolation, dynamic systems modeling and regularized regression methods. (III) Subsequently, a differential network was then extracted and its functional enrichment was performed. (IV) Moreover, the features of network topology, local entropy and adaptation dynamics were analyzed to measure the importance of each node in the differential network for prioritizing key genes responsible for drug resistance. (V) In addition, the above node importance measurement was incorporated into a differential regulatory network-based biomarker (DryNB) model for predicting drug response of clinical patients. (VI) Furthermore, experimental data and statistical significance test were used to validate the effectiveness of the key genes prioritized by DryNetMC.
Fig 2.
Distinct temporal gene expression profiles and patterns of dbcAMP-sensitive and dbcAMP-resistant glioma cells.
(A-B) Principal component analysis of RNA-seq transcriptomic data from the sensitive and resistant cells. The resistant cells showed an adaptive response, whereas the sensitive cells did not show this type of response. (C) Heat maps showing the distinct expression profiles of temporal changing genes (TCGs) in the sensitive (left panel) and resistant cells (right panel). (D-E) Clustering of the dynamic trends of the TCGs in the sensitive and resistant cells. The TCG profiles were divided into nine clusters. Most TCGs in the sensitive cells showed monotonic (increasing or decreasing) patterns (D), whereas most TCGs in the resistant cells exhibited adaptive dynamics (E). (F) Scatter plot showing scores for monotonic response (SA) and adaptive response (SM) of TCGs in the sensitive cells and resistant cells, respectively. The TCGs in the resistant cells tended to have a higher adaptive response score but a lower monotonic response score compared to the TCGs in the resistant cells. The assessment of statistical significance was given in S2C and S2D Fig.
Fig 3.
The reconstructed sensitive and resistant networks as well as the differential network.
(A) Sensitive GRN. (B) Resistant GRN. (C) The differential network. The genes are represented as circled nodes, and the activating and repressing interactions are represented as red arrows and blue lines, respectively. (D) Top 10 significantly enriched pathways involving genes in the differential network. Cell cycle and chromosome segregation were significantly enriched.
Fig 4.
Characterization of the difference between sensitive and resistant GRNs.
(A) Degree distributions of the sensitive and resistant networks. One-tailed Wilcoxon rank sum test p-value = 4.939×10−5. (B) Two- and three-node feedback loops involved in the GRNs. PF: positive feedback; NF: negative feedback; PPF: positive-positive feedback; PNF: positive-negative feedback; NNF: negative-negative feedback. (C) The percentages of various feedback loops in the resistant network were substantially larger than those in the sensitive network. (D-E) Violin plots comparing the differences of local network entropies (D) and dynamic changes in gene expression (E) between the sensitive and resistant networks, with one-tailed Wilcoxon rank sum test p-values of 7.094×10−13 and 1.029×10−15, respectively.
Fig 5.
The top-ranked genes in the differential network were predictive of the targeted therapeutic response and prognosis of glioma patients.
(A-B) The areas under the ROC curves (AUCs) were computed for both the training datasets (A) and the test datasets (B) of glioma patients. (C) Differential expression profiles of the seven identified genes in the normal and tumor tissues of glioma patients. The p-values were assessed using the one-tailed Wilcoxon rank sum test. (D) The seven identified genes were associated with the survival probability of glioma patients (N = 610). A log rank test was used to assess the significance between two curves, with a p-value less than 0.0001. (E) Statistical significance of the association of the five top-ranked genes identified by the DryNetMC with prognosis of glioma patients, in comparison with a conventional differential expression analysis method, DEseq2. Wald test p-value was used to assess the prognostic significance of each gene in the table.
Fig 6.
Temporal pattern similarity of the prioritized 5 genes was predictive of drug sensitivity/resistance.
RNA-seq data from the U87MG cell line (U) and its response to dbcAMP were used for testing. (A) Heat map showing the expression profiles of the genes in the differential network of the sensitive, resistant and tested cells (i.e., U87MG cell line). (B) Time-course expression patterns of the 5 genes (i.e., KIF2C, CCNA2, NDC80, KIF11 and KIF23) in the three cell lines. (C) The similarity of temporal pattern of the 5 genes between the tested cell line and the sensitive or resistant cell line. Dynamic time warping (DTW) distance was used to measure the temporal expression similarity of the 5 genes between different cellular states. Boxplots showed significant differences in the distributions of pair-wise distances, with p-value equal to 0.03125 (one-tailed Wilcoxon signed rank test), indicating that the tested cell line was more similar to the sensitive cells. (D) Experimental data of the dose-response of three cell lines to dbcAMP treatment validated that the response of the tested cell line was much closer to that of the sensitive cells.
Fig 7.
Comparison of the effectiveness of the top 5 genes prioritized by the DryNetMC with that by other methods, including DEseq2 and GSNCA.
RNA-seq data of U87MG cells was used to test its distance to sensitive DBTRG-05MG cells or resistant LN-18 cells, based on the expression pattern similarity of the selected genes evaluated using pair-wised DTW distance. (A-B) The similarity between the tested cells and the sensitive or resistant cells was evaluated based on the temporal expression patterns of (A) the top 5 genes prioritized by DEseq2 (i.e. OCIAD2, FLG, SPP1, SOX3 and NEFL) and (B) the top 5 genes prioritized by GSNCA (i.e. C5orf58, COL4A3, MUC16, DKFZp686O1327, GAL). Boxplots show the difference in distance distributions, with the statistical significance evaluated by two-sided Wilcoxon signed rank test. (C) The significance test for the top 5 genes prioritized by the DryNetMC using a bootstrapping approach (see details in S1 Text). The resulting p-value was 0.025, indicating non-randomness of the DryNetMC-prioritized 5 genes.