Figure 3

How PARma works. PARma is an iterative algorithm, which repeatedly executes three steps: based on a current model of the PAR-CLIP characteristics (left; see also Figure 6), scores are computed for each position in each cluster, which express the likelihood that the cluster is explained by the activity of the k-mer at this position (top right; see also Figure 7). These scores are fed into kmerExplain as prior probabilities, which then estimates k-mer activity probabilities using an EM algorithm (bottom). These k-mer activities in conjunction with data from the PAR-CLIP experiment (T to C conversions and RNase cleavage sites) are used to estimate the parameters of the PAR-CLIP model. We start this procedure by running kmerExplain on uniform scores and end it as soon as the model converges. EM, expectation maximization.