Quality of A domain substrate specificity predictions using HMMs and SVMs.

Substrate specificity predictions were made for various sequence data-sets using the published tools NRPSsp [42], NRPSpredictor 2 [43], and our single and ensemble of HMMs. Column 1 indicates the predictor that was tested and Column 2 the data that was used to test. Columns 3 and 4 provide the percentage of correct and false predictions below the set threshold, respectively, and column 5 the percentage of predictions that scored above threshold. Column 6 gives the fraction of sequences from the complete non-redundant data-set that received an annotation. Column 7 provides the fraction of correctly annotated sequences within the set of sequences that was provided with an annotation.$ To test the coverage and check the validity of the predictions, the four predictors were applied to the non-redundant reference dataset of experimentally validated substrate specific A domain sequences collected by us from the reference databases, literature and from [43] (set K = 571 sequences). To compare the performance, the predictors were applied to those sequences that are shared between data-sets. We found 392 sequences to be shared between the data-set used to train NRPSsp [42] and our non-redundant set (P∩K’), and 405 sequences to be shared between the data-set used to train NRPSpredictor2 [43] and our non-redundant set (R∩K’). In this case, K’ indicates that the sequences related to a substrate for which no model was present in either of the predictors, were left out in the comparison. The ensemble of HMMs was also applied to the dataset provided by [42] (P). To test the sensitivity of the ensemble models with respect to the removal of constituent sequences a Leave One Out cross validation was performed (LOO).