Figure 1: Integrative modelling of local DNA methylation levels.

(a) Schematic summarizing the integrative approach utilized for modelling local DNA methylations. DNA methylation at a given 10 kb region was predicted by incorporating relevant gene expression, somatic mutation, copy number alteration (GISTIC⁵⁶ values) and clinical information into integrative models (see Supplementary Data 1 for the complete list of variables included for each cancer type). (b) An example of an Elastic Net model performance in lung cancer. The x axis shows true values of DNA methylation in each sample, and the y axis shows the value predicted by the integrative modelling in the same sample when it was in the test subset. (c) Summary of overall model performance. For each cancer, the MSEs of test set predictions by Elastic Net and Random Forest were averaged across all models of local DNA methylation. (d) Comparison of original gene expression variables with randomly selected variance-matched genes. The y axis shows the average rank of each gene expression category based on average variable usage score across all Elastic Net models (left) and average variable importance score across all Random Forest models (right) of local DNA methylation in brain cancer (boxes extend from 25th to 75th percentiles, centre lines represent the median and whiskers show the minimum and maximum value in each group). P values associated with the Mann–Whitney test between the ranks across all models are shown (a higher rank corresponds to higher contribution; see Methods).