Contribution of different features in Methylphet Sample Clauses

Contribution of different features in Methylphet. It is important to understand the relative predictive power of methylation levels and other genomic features used in Methylphet. We present the Gini importance of predictors in the RF model for CTCF and SOX2 using bar plots in Figure 2.6(A, C). Gini importance is the measurement of classification efficiency for each feature, which is defined as the level of decrease in the class impurity. Based on these analyses, we found that 5mC scores play the most important role in predicting CTCF binding, whereas the CH score is the most important predictor for SOX2. Motif is the second most important for CTCF, but its importance is very low for SOX2. This is because the SOX2 motif has much lower specificity compared to the CTCF motif. For both TFs, CH and 5hmC methylations play rather important roles. We compared the predictive performances with or without 5hmC and CH methylations. Figure 2.6(B, D) shows the ROC curves from such comparisons. These results demonstrate that including both 5hmC and CH methylation scores in the model improves the prediction power. Figure 2.6(E) shows the distributions of Gini importance of each predictor across all TFs we tested. We can see clearly that the 5mC score contributes the most on average among all features, next being the CH and motif scores, followed by sequence conservation and distance to the closest TSS.