Datasets used for isoelectric point optimization

Note:

• all datasets presented in the table are available as hyperlinks, final datasets were divided randomly into 75% training and 25% testing subsets (hyperlinks denoted [75] and [25] respectively)
• NA - not available refers to the situation where given dataset was not created as merged version was used
• links in second column, Initial no. entries, points to files in native format as presented by Authors (different formats e.g., pdfs, Excel etc.), remaining files are in fasta format with informative headers
• both PIP-DB and SWISS-2DPAGE datasets contain many queries without sequence information (such cases were excluded from the final dataset)
• for SWISS-2DPAGE only first isoforms from Uniprot were used (splicing variants were not analyzed)
• duplicates and extreme outliers (MSE > 3 between average predicted pI and experimental pI) were removed
• if multiple values of isoelectric point were reported than in further studies average was used
• to remove redundancy CD-HIT program was used (99% sequence identity cut off was used as even single point mutations of charged residues can lead to big change of isoelectric point)