Fig. 4

Decision tree of predictive models for the K562 cell line, constructed from the same data as the standard model evaluated earlier (H3K4me3, H3K27me3, H3K9me3, H2A.Z and DNA methylation). This tree uses promoter-localised DNA methylation (MMFS) and the H2A.Z histone variant to classify genes into three latent regulatory classes: MMFS⁺ (high MMFS score), H2A.Z⁺ (low MMFS and high H2A.Z) and H2A.Z⁻ (low MMFS and low H2A.Z). Sub-categorising MMFS⁺ by H2A.Z would be biologically meaningless as the histone variant is mutually exclusive with DNA methylation in vivo [25]. Thresholds were learned directly from the data using our unsupervised methodology. Specifically, the blue, red and black lines illustrate Δadj. R ² (relative to a standard model constructed from the same data) as a function of threshold values for positive (MMFS⁺), negative (MMFS⁻) and cumulative models, respectively, with the optimal value for both forks indicated by a black dashed line. Error bars capture the standard error of the mean (μ ≈ 0) for models constructed from 100 randomly-sampled gene sets of equal size, illustrating the performance variation expected by chance (i.e. fewer genes equals larger variation in model performance, as expected)