“ Enhanced Recognition of Transmembrane Protein Domains with Prediction - based Structural Profiles ”

In addition to a simple NN-based classifier developed and assessed in the cross-validation study (see the Systems and Methods section of the main body of the paper), we also developed a multistage protocol for enhanced prediction of transmembrane (TM) helices. For the final predictor we do not consider the MA-based representation, which is shown using cross-validation to yield a lower accuracy compared with the proposed compact prediction-based " structural profiles ". We also excluded hydropathy scales from the representation of an amino acid residue, as it was observed to lead to a higher level of confusion with globular proteins and signal peptides. Moreover, statistical propensities of amino acids (e.g. to lipid – aqueous phase interfaces) that could be used, in principle, to improve the prediction of TM segments, are not taken into account at this stage. Consequently, each residue is initially represented by five numbers: the predicted real valued RSA, confidence of RSA prediction and probabilities of each of the three secondary structures (as predicted by SABLE, http://sable.cchmc.org). SABLE predictions are derived from the multiple alignment, hydropathy scales and other attributes, which are commonly used by other state-of-the-art methods. Therefore, it is expected that other accurate methods for RSA and SS prediction will be useful in that regard as well. This is a subject of a future investigation. Following in the footsteps of other studies (Rost et al., 1995), we use a two-stage prediction system, with the second layer (structure-to-structure) NNs allowing one to " average " and smooth over the initial classification obtained using the first (sequence-to-structure) layer predictor. The architecture of the first and second layer NNs is similar to that used for the cross-validation study. Namely, a simple feed-forward topology with one hidden layer, fully interconnected with the input and output layers, is employed. The choice of the sliding window size, the number of nodes in the hidden layer, training protocols and other characteristics of these NNs are discussed below. In order to reduce the danger of overfitting and achieve regularization as well as improvement in accuracy, a consensus of twenty different networks was used to generate predictions at each stage. These different networks were trained on different subsets of the training set that were also used for the cross-validation study. Multiple NNs were trained on each subset of the data, with different number of nodes in the hidden layer and with different size of the sliding …