AMS 3.0: prediction of post-translational modifications

Abstract

Background: We present here the recent update of AMS algorithm for identification of post-translational modification (PTM) sites in proteins based only on sequence information, using artificial neural network (ANN) method. The query protein sequence is dissected into overlapping short sequence segments. Ten different physicochemical features describe each amino acid; therefore nine residues long segment is represented as a point in a 90 dimensional space. The database of sequence segments with confirmed by experiments post-translational modification sites are used for training a set of ANNs.

Results: The efficiency of the classification for each type of modification and the prediction power of the method is estimated here using recall (sensitivity), precision values, the area under receiver operating characteristic (ROC) curves and leave-one-out tests (LOOCV). The significant differences in the performance for differently optimized neural networks are observed, yet the AMS 3.0 tool integrates those heterogeneous classification schemes into the single consensus scheme, and it is able to boost the precision and recall values independent of a PTM type in comparison with the currently available state-of-the art methods.

Conclusions: The standalone version of AMS 3.0 presents an efficient way to identify post-translational modifications for whole proteomes. The training datasets, precompiled binaries for AMS 3.0 tool and the source code are available at http://code.google.com/p/automotifserver under the Apache 2.0 license scheme.

Figure 1

MLP algorithm. A block diagram of an MLP shown as a feed forward layered neural network.

Figure 2

AUC for four kinase families. Scope of AUC values for the kinase families PKA, PKC, CDK and CK2, computed on sample train and test dataests using AMS3.

Figure 3

AUC values for predictors. Comparison of scope of AUC best values for the kinase families PKA, PKC, CDK and CK2, using AMS3, GPS, KinasePhos, NetPhosK, PPSP, PredPhospho, Scansite and Meta Predictor.

Figure 4

ROC values for four kinase families. Comparison of ROC values for the kinase families PKA, PKC, CDK and CK2, using GPS, KinasePhos, NetPhosK, PPSP, PredPhospho, Scansite and Meta Predictor with the corresponding ROC curves for training and test datasets using AMS3.