ExportPred: Architecture and performance. (a) The architecture of the ExportPred GHMM. The GHMM progresses from left to right, beginning with an amino-terminal methionine and terminating at a stop codon. Length probability densities are shown for non-geometric states. Tail states and the KLD spacer state are modelled by geometric distributions. (b) ROC curves for the ExportPred model comparing the training against the five described negative sets. (c) False discovery rate as a function of score threshold, calculated using the training set and the P.f negative set, and assuming 10% of the P. falciparum proteome is exported. (d) Comparison of predictions made by ExportPred using the default threshold of 4.3 with those published in [6,7]. The -rifin set is exclusive of any sequence annotated as rifin or stevor, whereas the +rifin set includes these sequences.