Figure 7

By exploring, filtering, and manually eliminating or saving various dependencies by brushing, a significantly simplified picture of the dependency network emerges. Successive refinement from the raw dependency data shown in Figure 6, to a computationally tractable dependency structure for a CRF that enables accurate identification of other members of the sequence family. A) By applying the reduction of the displayed data to only the unexpected residuals, Figure 6 becomes much more sparse. B) Applying threshold filters to the magnitude of the residuals, further reduces the visual complexity of the model and simultaneously decreases the likelihood of overfitting the data with the CRF model, and brings the dependency network closer to being computationally tractable. C) Finally applying statistical filters, and manual editing of the dependency structure, results in a CRF dependency model that captures the important family sequence signatures. It is also relatively easy to browse and understand in the interactive interface, despite casting the parallel axes in a volume rather than a plane.