Fig. 4.
A summary of results of learning NCFs from synthetic datasets. Synthetic sets were sampled over a 500 000-bp-long sub-sequence of yeast chromosome 4, using the model with each of the NCFs: Exp*, S-Exp*, S-ES*, Off Step*, On Step* and No Coop* (shown in Fig. 2). To each sampled set different levels of noise (different SDs for Gaussian perturbations of sampled nucleosome locations, denoted Stdnoise) were introduced. On each resulting synthetic set, parameters of five types of NCFs were learned (Exp, S-Exp, S-ES, Step and No Coop), together with the model's temperature and nucleosome concentration parameters, in a 5-fold cross validation manner. The results are organized in a table-like fashion, with rows per synthetic data type and columns per noise level introduced into the synthetic set. Each cell shows results attained for each of the learned NCFs, along with results attained for the original NCF (with original temperature and nucleosome concentration) used for sampling the synthetic data. Results per learned NCF are color coded according to a color legend appearing in the left part of the respective row. For each learned NCF shown are: in the bar plot, the cross validation mean (bar) and SD (blue error bar) of the test R2 statistic (quantifying the fraction of the variance in the test data that is explained by the model with the learned NCF), as well as the cross validation mean and SD of the train R2 statistic (light green error bar). In the graphs plot, shown are the cross validation mean and SD (per linker length) of the linker lengths distribution (over linker lengths 0–50) sampled using the model with the learned NCF.