Comparision of IPC to other pKa sets (on 25% of datasets not used for training)

* average from all pKa sets without Patrickios (highly simplified pKa set) and IPC sets

& Bjellqvist is implemented as e.g. ExPASy Compute pI/Mw

% - note that pH scale is logarithmic with base 10, thus percent difference corresponds to pow(10, x) where x is equal to delta of RMSD of two error estimates represented in pH units e.g. % difference between Toseland and IPC_protein is pow(10, (0.934-0.874)), for better explanation why this should be calculated like that see ... here

Protein dataset (IPC_protein was trained on 1,743 proteins with 10-fold cross-validation - data in next table, below, tested on 581 proteins not used for training - data in the table above), peptide dataset (IPC trained on 12,662 peptides with 10-fold cross-validation - data in next table below, tested on 4,220 peptides not used for training - data in the table above). Outliers correspond to the number of predictions for which difference between experimental pI and predicted pI was bigger than threshold of mean standard error (MSE) of 3 for protein dataset and MSE of 0.25 for peptide dataset.

Statistics for different pKa sets on 75% training datasets (10-fold cross-validatated)

Changes in method order in comparison to previous table are bolded.

Statistics for different pKa sets for SWISS-2DPAGE and PIP-DB separately

Both SWISS-2DPAGE and PIP-DB were cleaned from outliers (MSE >3 between experimental pI and average predicted pI) and clustered by CD-HIT with 99% sequence identity threshold as described in Datasets (982 and 1307 proteins respectively), but they were not divided into training and testing datasets, thus results for IPC sets can by slightly overestimated, yet this is not a relevant issue as seen from comparison of previous tables. Changes in method order in comparison to Table 1 are bolded.

Experimental versus theoretical isoelectric points for proteins datasets (SWISS-2DPAGE and PIP-DB) calculated using EMBOSS pKa set.

Outliers defined as MSE>3, marked red. Note, plots correspond to datasets as presented by authors before cleaning and removing duplicates (duplicates are defined as records which has the same sequence and yet referred as separate records in database although in both databases authors many times report multiple pI from different experiments in one record, nevertheless for current analysis average pI was used). Solid line represent linear regression fitted to data after removing outliers.

Experimental versus theoretical isoelectric points for the main protein dataset (merge of SWISS-2DPAGE and PIP-DB, including training and test sets) calculated using different pKa sets

R² – Pearson correlation before removing outliers. R²corr – Pearson correlation after removing outliers. Additionally, linear regression models fitted to predictions with outliers (magenta line) and without outliers (blue line) are shown. Outliers (marked magenta) are defined as pI predictions with MSE>3 in comparison to experimental pI, other predictions are represented as heat maps according to density of points. Number of outliers for both training and testing set together.

Interesting facts about protein isoelectric point

It is well known that distribution of isoelectric point for proteins in almost all organisms is very similar and non uniform. Below you can find some plots which should give you general impression (for raw data for generating those plots see "Datasets" section).

Swiss_prot (90% identity, 318k proteins)                                      Escherichia coli (4128 proteins)                                      Homo sapiens (99739 proteins)

As you can see, normally proteomes have bimodal distribution of pI with very low fractions of proteins with pI close to 7.4 This is due the fact that the proteins are mostly insoluble, less reactive and unstable in pH close to pI, and pH inside of the the cell is close to 7.4. Thus, this property of proteomes could be the result of evolutionary selection to avoid this. Of course there are some exceptions. Many of halophilic Archaeons do not even try to force high concentration of salt of their environment, instead they change their physiological pH inside of the cell to more similar to environment (in this way they use less energy to keep the homeostasis). This has dramatic consequence for the amino acid composition and isoelectric point of theirs proteins. For more informations see: doi:10.1186/1471-2164-8-163

Natrialba magadii (3559 proteins)

Implications

The bimodal distribution of pI in the proteomes is a feature that you can see in almost all organisms and you should expect to see it. Let now analyze well known yeast proteome (6720 proteins) with widely used EMBOSS scale (used by e.g Saccharomyces Genome Database (SGD)) and IPC_protein scale.

EMBOSS                                                                           IPC_protein

---

Contact: Lukasz P. Kozlowski