PeptideRanker is a server for the prediction of bioactive peptides based on a novel N-to-1 neural network.
Users may submit a list of peptides to PeptideRanker which will be returned to the user ranked by the probability that the peptide will be bioactive. It is important to note that this is not a prediction of the degree of bioactivity.
This server is designed to evaluate user designed peptides. If you want to study a protein, you must decide in advance what peptides within that protein you wish to study, as in the example above. Simply pasting in a protein sequence, with for example, line breaks every fifty residues will only explore the predicted activity of that set of peptides of fifty residues long each, which is unlikely to represent your intended goal. Submit a list of peptides and PeptideRanker will predict the probability that each of these peptides will be bioactive. The list will then be returned ranked by the predicted probability of bioactivity for each peptide. When the user is interpreting the results is important to note that the server predicts how likely the peptide is to be bioactive, and not how bioactive the peptide is likely to be.
Running the example series of peptides
We have chosen as an example a series of user-defined peptides derived from the Bovine milk protein beta-casein (UniProt accession number P02666). These were derived from a series of overlapping 7 mer peptides which span the location of a known bioactive 7-mer, beta-casomorphin 7 (YPFPGPI), which may have a potential role in human disease such as ischaemic heart disease, diabetes mellitus, sudden infant death syndrome, autism and schizophrenia. This peptide was not included in our training or test sets. To run the provided example click on "example" and then click the "submit" button. The results show the beta-casomorphin 7 peptide strongly predicted to be bioacitve (0.92). Only one of the overlapping peptides scored slightly higher, and it overlaps for 6 of its 7 residues. Many of the overlapping peptides are predicted not to be bioactive, since they are returned with values of under 0.5. Note that the web server automatically uses the short peptide predictor for peptides of less than 20 amino acids, and the long peptide predictor for peptides of 20 or more residues. To run this example click on "example" and then click the "submit" button.
PeptideRanker was trained at a threshold of 0.5 i.e. any peptide predicted over a 0.5 threshold is labelled as bioactive. However, the user may decide to chose a higher threshold to reduce the number of false positives. From our testing (see paper) we would expect that choosing a threshold of 0.8 will reduce the false positive rate from 11% and 16% at 0.5 threshold to 2% and 6% at a 0.8 threshold for long and short peptides respectively. However, increasing the threshold to 0.8 from 0.5 also reduces the true positive rate. The user needs to choose a threshold carefully based on their needs. Is it more important the reduce the number of false positives or capture all the true positives?
Another factor that will reduce the number of true positives predicted by PeptideRanker is if the peptide has cysteine content lower than that of most secreted proteins/peptides i.e. less than 4%. When we tested PeptideRanker we observed that the false negatives in our independent test set (bioactive peptides incorrectly predicted as non-bioactive) had an average cysteine content of just 1.7%.
PeptideRanker was trained and tested on amino acid sequences without any modifications. Peptide modifications such as the acetylation of the N-terminus, the amidation of the C-terminus, the cyclisation of the peptide or the inclusion of one or more D-amino acids may increase the bioactivity and the specificity of the peptide. The inclusion of D-amino acids may also increase protease resistance. We would suggest that after using PeptideRanker to discover bioactive peptides in silico that various peptide modifications would be experimented with in vitro to increase the bioactivity.