Brain tumour genetic network signatures of survival

Abstract

Tumour heterogeneity is increasingly recognized as a major obstacle to therapeutic success across neuro-oncology. Gliomas are characterized by distinct combinations of genetic and epigenetic alterations, resulting in complex interactions across multiple molecular pathways. Predicting disease evolution and prescribing individually optimal treatment requires statistical models complex enough to capture the intricate (epi)genetic structure underpinning oncogenesis. Here, we formalize this task as the inference of distinct patterns of connectivity within hierarchical latent representations of genetic networks. Evaluating multi-institutional clinical, genetic and outcome data from 4023 glioma patients over 14 years, across 12 countries, we employ Bayesian generative stochastic block modelling to reveal a hierarchical network structure of tumour genetics spanning molecularly confirmed glioblastoma, IDH-wildtype; oligodendroglioma, IDH-mutant and 1p/19q codeleted; and astrocytoma, IDH-mutant. Our findings illuminate the complex dependence between features across the genetic landscape of brain tumours and show that generative network models reveal distinct signatures of survival with better prognostic fidelity than current gold standard diagnostic categories.

Keywords: brain tumours; graph modelling; machine learning; representation learning; survival modelling; tumour genetics.

Figure 1

Data: distributions by geography, tumour (epi)genetics, diagnoses and demographics. (A) Geographical distribution of all neuro-oncology patient data in the UK referred to our Division of Neuropathology between 2006–2020 for molecular diagnostics, in logarithmic axis per the colour bar. (B) Number of mutant samples across the n = 4023 glioma patient cohort. (C) Distribution of WHO CNS5 diagnoses in cohort. (D) Age kernel density estimators for male and female, subdivided to the diagnoses with corresponding colours, as in C.

Figure 2

Method: graph modelling of brain tumours. (A) Rich genetic feature sets are extracted from patient histopathology report data (y-axis) and fractionated into individual genetic lesions (x-axis). The approach yields patient feature ‘barcodes’ corresponding to the complete molecular data available for a given patient's tumour. Note only a subset of features is labelled owing to visualization constraints. (B) Heat map of the conditional probability of one genetic feature given the presence of another, [P(A|B)], derived across the feature space, yielding an asymmetric adjacency matrix to be modelled as a directed Bayesian graph. In A and B, only a subset of features are labelled on the axes for visualization purposes. (C) Histogram of edges in the Bayesian genetic network, with the number of edges present (y-axis, logged), and the corresponding probability assigned to the weighted edge (x-axis). The conditional weighted graph (blue bars) exhibits greater variation in interrelatedness compared with the intersection between features (orange bars). (D) Schematic illustrating the application of graph modelling for two purposes: (i) graph feature genetic mapping, where tumour genetics are modelled as nodes and their relations as probabilistically weighted, directed edges; and (ii) graph patient genetic mapping, where individual patients are nodes and edges are weighted by individual genetic features.

Figure 3

Graph feature genetic mapping identifies characteristic genetic interrelations. (A) Radial graph of layered, nested, degree-corrected and exponentially weighted stochastic block model revealing the community structure of tumour genetics and their influence upon overall network topology. Communities are colour-coded by the first level community of the hierarchical community structure. Edges are sized according to their conditional probability. Nodes are sized according to their weighted-eigenvector centrality. Hierarchical levels are annotated from level 0 (L₀) to level 4 (L₄). (B) Fits in accordance with link probability by conditional probability, measured frequency and the comparative null illustrate the description length of the layered model lower than the null, evidencing it a more suitable structure. Shown below is the visualization of the first level hierarchy (L₁) with node colour as per that of A. Edge size and colour is proportional to the incidence of edges linking mutations both between and within a given community. Node size is proportional to the degree of the corresponding community. (C) There is a significant difference in weighted eigenvector centrality of tumour genetic factors (P < 0.0001), page rank (P < 0.0001) and hub centrality of tumour genetic factors when organized by stochastic block model community (P < 0.0001). In C, block colour as per that of A, points are labelled by their corresponding abbreviation: A = ATRX; B = BRAF, E = EGFR; H = Histone; I = IDH; M = MGMT; O = 1p/19q; T = TERT. Supplementary Figs 3–7 also accompany this plot with additional results.

Figure 4

Tumour heterogeneity. (A) Sankey plot illustrating the variety of genetic features, under both coarse and finely granular descriptors, aligned to patient diagnosis. Only the most frequent links in our dataset are shown for readability. (B) Principal component analysis (PCA) of all tumour genetic data, which clusters individuals into patient groups reproducible of the diagnostic labels, colour-coded as per the key. (C) Principal component analysis of all tumour genetic data with patient survival projected onto the plot illustrates a qualitatively poor representation of clusters of individuals with systematically better or worse survival. (D) Minimum spanning tree of patients with edges weighted by the similarity between individual genetic tests appears to create a more richly structured representation of a tumour-genetic landscape, colour-coded as in the key in A. (E) Minimum spanning tree of patients with edges weighted by the similarity between individual genetic tests with survival projected onto the plot illustrates a clearly superior segregation of individuals with better or worse prognosis, colour-coded by survival as per C.

Figure 5

Graph patient genetic mapping enables richer, more informative phenotyping. Radial graph of nested, degree-corrected and multivariate binomially weighted stochastic block model revealing the community structure of patients and the genetics of their brain tumour. Hierarchical levels are annotated from level 0 (L₀) to level 7 (L₇). For visualization purposes, communities are colour-coded by the second level blocks (L₂) of the hierarchical community structure. Around the radial graph are the breakdown of median survival and box and whisker plots for the coefficients and 95% credible intervals of genetic loadings, where the coloured border of the plots depicts the corresponding community. All box plots where the error-bar does not cross the vertical zero-line are significant, with features left of the vertical zero-line favouring the wildtype, and right of the zero-line favouring mutation. Supplementary Figs 3 and 8–10 accompany this plot with additional results.

Figure 6

Graph patient genetic mapping offers higher fidelity prognosis. (A) Radial graph of a nested, degree-corrected and binomially weighted stochastic block model revealing the community structure of patients based on tumour genetics (as also shown in Fig. 5 and Supplementary Figs 8 and 9). Communities are colour-coded by the hazard ratio of the survival model of the second level blocks (L₂). Note only the minimum spanning tree of the graph is shown, owing to visualization constraints. (B) Pie chart of brain tumour diagnoses colour-coded by the hazard ratio of the survival model with the diagnostic label. In A and B, darker colours convey a poorer prognosis (hazard ratio > 1), and conversely lighter colours a more favourable one (hazard ratio < 1). (C) Box and whisker plot illustrating the hazard ratios with 95% confidence interval (CI) of the second level blocks of the stochastic block model community structure. (D) Box and whisker plot illustrating the hazard ratios with 95% confidence interval of the tumour diagnoses, illustrating only a crude discrimination of glioblastoma, IDH-wildtype from the remainder. Also shown is a box and whisker plot of hazard ratios with 95% confidence interval of the raw tumour genetics. (E) Survival plot of the second level blocks of the stochastic block model community structure illustrates a rich variation in survival, colour-coded by the blocks in both A and C and in Fig. 5. (F) Survival plot of the tumour diagnoses shows coarser forecasting of patient prognosis, colour-coded by the diagnoses of B. In C and D, all points where the whiskers do not cross the vertical line at 1 are statistically significant. Supplementary Figs 9–11 also accompanies this plot with additional results.

Figure 7

Network signatures forecast survival better than WHO CNS5 diagnosis or raw genetic and epigenetic features. (A) Model predictive performance evaluated by cross-validated concordance index of Cox's proportional hazard model shows network signatures outperform both models of diagnosis and the original genetic information in forecasting survival. (B–D) Model predictive performance evaluated by pseudo-R² and widely applicable information criterion (WAIC) of Bayesian logistic regression survival models for 12-month (B), 24-month (C) and 36-month (D) survival shows network signatures outperform both models of diagnosis and the original genetic information in forecasting survival, and with more favourable fits by WAIC (lower is better). Supplementary Fig. 11 accompanies this plot with additional results.