Predicting TSS usage. (A) Observed TSS usage in a TSS cluster at the start of the MFSD4 gene. The number of CAGE tags (from all libraries) initiating from each nucleotide is shown as a bar plot. This is for comparison with the predicted initiation propensity in the next panel. (B) Predicted TSS propensity of each nucleotide in the above TSS cluster. The transcription initiation propensity of each nucleotide, calculated as a likelihood ratio predicted from the surrounding DNA sequence using a second-order Markov model (see main text and Methods), is shown as a bar plot. Note the high correlation between predicted rates in this panel and the observed counts in A. (C) Classification of nucleotides within TSS clusters as active or inactive. The receiver operating characteristic (ROC) curve plots sensitivity vs. specificity (see Methods) for classification methods. The area under the curve (AUC) statistic is shown within the plot for the different prediction methods. An AUC of 100% corresponds to ideal performance, while a random classifier (shown as a dotted line) will have an AUC of 50%. We use the prediction scores, as exemplified in B, to classify each nucleotide in a cluster as active or inactive, for the test clusters on chromosome 1. With no additional scaling of these scores, the predictive power is adequate (gray line with boxes). Normalizing the nucleotide scores by the sum of prediction scores within the cluster (black line with triangles) does not improve the prediction. However, after scaling the prediction scores by the overall expression level (number of observed CAGE tags) of the cluster (black line), the AUC reaches an impressive 87%. Thus, knowing the expression output of a given promoter region adds additional predictive power.