The effects of sparsification. (A) The figure plots accuracy as a function of the percentage of dlinks eliminated. Dlinks are the weight parameters that are assigned to the edges in the graphical model representation of the DBN. In an auto-regressive model, the majority of the model parameters become dlink coefficients. It is possible to remove significant proportion of the dlinks while maintaining the overall predictive accuracy such that 70% of the dlink parameters can be removed for the DBN model that uses PSI-BLAST PSSMs only, 80% of the dlink parameters can be removed for the DBN model that uses HHMAKE PSSMs only and the 95% of the dlink parameters can be removed for the DSPRED method (DBN combined with SVM) that uses PSI-BLAST PSSMs, HHMAKE PSSMs and posterior distributions of secondary structure labels. (B) The figure plots the percentage of dlinks that are retained as a function of the sparsity of the model. The three series correspond to all dlinks, only the current dlinks and only the auto-regressive dlinks. In the auto-regressive and the current series, the percentages are computed with respect to the total number of dlinks within each of these series separately. For both panels, results are computed via seven-fold cross-validation on SD576. The hyperparameters of the DBN are L
= 5, L
= 3, α = 0.035, ω = 0.4.