Signal with unpredictable trends

Here, in order to remove unpredictable trends in the hybridization signal we will use sliding windows to fit localy the HMM (as in (6)). Thus, independent run of the HMM will be successively applied in window of 1kb (i.e. $\sim$250 probes) all along the genome. The model parameters and posteriors of all windows containing a fixed probe will be then averaged and used for a global computation of both state probabilities and most-likely states (among well-positioned nucleosome, fuzzy nucleosome and linker). As we used only probes with unique and perfect matches, we also allowed the HMM to deal with missing data. State probabilities and most-likely states of ``missing probes'' will be computed in the same way than for observed probes, taking advantage of neighboring observed information.

Inferring nucleosome positionning with NucleoMiner requires two steps. The first one consists in fitting the HMM to the data (NMhmmfit) and the second one applies the Viterbi algorithm (NMhmmvit) to the averaged posterior probabilities as computed and outputed by NMhmmfit.