Systems-level network modeling deciphers the master regulators of phenotypic plasticity and heterogeneity in melanoma

Abstract

Phenotypic (i.e. non-genetic) heterogeneity in melanoma drives dedifferentiation, recalcitrance to targeted therapy and immunotherapy, and consequent tumor relapse and metastasis. Various markers or regulators associated with distinct phenotypes in melanoma have been identified, but, how does a network of interactions among these regulators give rise to multiple "attractor" states and phenotypic switching remains elusive. Here, we inferred a network of transcription factors (TFs) that act as master regulators for gene signatures of diverse cell-states in melanoma. Dynamical simulations of this network predicted how this network can settle into different "attractors" (TF expression patterns), suggesting that TF network dynamics drives the emergence of phenotypic heterogeneity. These simulations can recapitulate major phenotypes observed in melanoma and explain de-differentiation trajectory observed upon BRAF inhibition. Our systems-level modeling framework offers a platform to understand trajectories of phenotypic transitions in the landscape of a regulatory TF network and identify novel therapeutic strategies targeting melanoma plasticity.

Keywords: Cancer systems biology; Cell biology; Network modeling; Transcriptomics.

Graphical abstract

Figure 1

Existence of two distinct groups of genes associated with phenotypic heterogeneity in melanoma (A) Experimentally reported phenotypes of melanoma. (B) Spearman's correlation of regulators of phenotypic heterogeneity for GSE7127 (n = 63), GSE80829 (n = 54), GSE137391 (n = 24) (left to right, top panel) and CCLE (n = 59 – melanoma cell lines), GSE10916 (n = 50) and GSE4843 (n = 45) (left to right, bottom panel). Crosses indicate p > 0.05. Colorbar denotes the values of correlation coefficient. ‘n’ stands for number of samples in each dataset.

Figure 2

Two distinct classes of cells exist across multiple datasets (A–C) K-means clustering for K = 2 yields two distinct clusters along principal component 1 (PC1). GSEA for Hoek proliferative geneset (B) and Hoek invasive geneset (C) confirms that the two clusters correspond to the respective phenotypes for i. GSE7127 (n = 63) ii. GSE80829 (n = 53) iii. GSE137391 (n = 24) iv. CCLE (n = 59 – melanoma cell lines) v. GSE10916 (n = 50) vi. GSE4843 (n = 45). ‘n’ stands for number of samples in each dataset.

Figure 3

Identification of master regulators driving phenotypic heterogeneity in melanoma (A) Pipeline followed to identify nodes for underlying TF regulatory network. (B) Modules identified from WGCNA for GSE4843: i. Proliferative module ii. Invasive module. Module eigengene values are shown for each sample (top) and the corresponding expression level heatmap of genes in each module (bottom) for proliferative (orange) and invasive (cyan) samples are given for GSE4843. (C) Spearman's correlation matrix among pairs of master regulators in: i. GSE4843 and ii. GSE137391. Crosses indicate p > 0.05. Colorbar denotes correlation coefficient. ‘n’ stands for number of samples in each dataset.

Figure 4

Dynamic simulations recapitulate phenotypic heterogeneity in melanoma (A) Interaction network identified for master regulators. (B) i. PCA plot and ii. Heatmap for simulated data forming two distinct clusters iii. Percentage contribution of each gene to PC1. The red line indicates the value for uniform contribution (= 1/17) of all genes considered (n = 17). (C) Bimodality test using Bimodality coefficient (BC) (Green boxes indicate BC > 0.555 and red boxes indicate BC < 0.555) and Hartigan's dip test (Green boxes indicate p < 0.05 and red boxes indicate p > 0.05). (D) Histograms for individual gene expression distribution in the simulated dataset.

Figure 5

A subset of master regulators explain the existence of two phenotypes (A) Differential expression of subset of master regulators (n = 11) in proliferative and invasive samples in GSE4843 and GSE137391. (B) Projection of proliferative and invasive clusters on the principal components for expression levels of these 11 master regulators. (C) Spearman's Correlation coefficient matrix for the 11 master regulators in i. RACIPE simulated dataset ii. GSE4843. Crosses indicate p > 0.05. Colorbar denotes correlation coefficient. ‘n’ stands for number of samples in each dataset.

Figure 6

Master regulator network explains the existence of proliferative and invasive subpopulations and dedifferentiation trajectory followed during BRAFi (A) i. PCA plot and ii. Heatmap for four clusters identified in simulated data. (B) Z-scores for discriminant genes can distinguish phenotypes within the proliferative sub-clusters (Melanocytic – M and Transitory –T) and invasive sub-clusters (Undifferentiated –U and NCSC –N) in GSE4843 and GSE134432. (Significance is represented based p value for Student's t-test, ∗ for p < 0.1, ∗∗ for p < 0.01, ∗∗∗ for p < 0.001 and ∗∗∗∗ for p < 0.0001). (C) Trajectory of cells in GSE134432 i. Phenotype of cells lying in the trajectory ii. Pseudo-time trajectory.

Figure 7

MITF knockdown explains phenotype switching caused by BRAFi and MAPKi (A) Barplot representing percentage of solutions representing each phenotype (B) Density of data points form a multimodal distribution along PC1 in Control (Top) and MITF-KD (bottom) networks [left]. MITF-KD increases the coefficient of variance of the two Gaussian distributions corresponding two the proliferative and invasive phenotypes [right]. (n = 3, error bars represent SD, Significance is represented based p value for Student's t-test, ∗ for p < 0.05, ∗∗ for p < 0.01, ∗∗∗ for p < 0.001 and ∗∗∗∗ for p < 0.0001). (C) Cellular plasticity gives rise to phenotypic heterogeneity in melanoma.