The AlphaPred server predicts the alpha turn residues in the given protein sequence. The method is based on the neural network training on PSI-BLAST generated position specific matrices and PSIPRED predicted secondary structure. Two neural networks with a single hidden layer having 10 units have been used where the first sequence-to-structure network is trained on PSI-BLAST obtained position specific matrices. The filtering has been done using second structure-to-structure network trained on output of first net and PSIPRED predicted secondary structure. The training has been carried out using error backpropagation with a sum of square error function(SSE). The network is trained and tested on a set of 193 non-homologous protein chains with 5-fold cross-validation. It predicts alpha turns in proteins with prediction accuracy of 78.0% and MCC value of 0.16.
The input is a single letter-code amino acid sequence either in fasta or plain text. The residues in the query sequence predicted as alpha turns are marked as a and non-turn residues are marked as '.'. The PSIPRED predicted secondary structure (H: Helix; E: Strand and C: Coil) is displayed along with the predicted turn/non-turn results.
If you are using this server then please site Kaur H & Raghava GP. (2004). Prediction of alpha-turns in proteins using PSI-BLAST profiles and secondary structure information. Proteins. 55: 83-90
