Distribution and Reliability of SVM Scores

<div><p>(A) The gray shading indicates prediction accuracy as a function of SVM score (left <i>y</i>-axis). Predictions are most accurate (>96%) when SVM scores are 0.5 or higher (for coding) or lower than −1 (for non-coding). Also shown are the score distributions (right <i>y</i>-axis) of three datasets: the training/testing set, random DNAs, and FANTOM3 transcripts. The solid vertical line in the middle indicates the SVM decision boundary (score = 0).</p> <p>(B) Most SVM predictions for the training/testing set are in the high specificity range. FANTOM3 transcripts have a slightly larger fraction of low accuracy predictions.</p>