IPC 2.0 (isoelectric point calculator)

What is IPC 2.0?

The isoelectric point (pI) is the pH where a protein or peptide carries zero net charge. It determines how a molecule behaves during isoelectric focusing, ion exchange chromatography, and 2D gel electrophoresis. Predicting pI from sequence requires knowing the pKa of every ionizable group, which varies depending on the pKa scale used.

IPC 2.0 (Isoelectric Point Calculator 2.0) by Kozlowski combines 19 pKa scales with machine learning to predict pI more accurately than any single Henderson-Hasselbalch calculation. The SVR models trained on 2,324 proteins and 119,092 peptides achieve RMSD values of 0.85 and 0.23 respectively. A SepConv2D deep learning model pushes peptide accuracy further (RMSD 0.22). Version 2.0 also introduced per-residue pKa prediction, estimating the dissociation constant of each ionizable residue in context using an MLP-SVR ensemble.

When to use IPC 2.0 vs simpler calculators

For a quick single-sequence estimate, the pI Calculator or Protein Parameters tool runs instantly in the browser using the Bjellqvist scale. IPC 2.0 is the better choice when:

• Accuracy matters (SVR and DL models outperform any single pKa scale)
• Comparing predictions across multiple scales helps assess confidence
• Per-residue pKa values are needed for understanding charge distribution
• Batch processing hundreds of sequences at once

How to use IPC 2.0 online

IPC 2.0 on ProteinIQ predicts the isoelectric point of proteins and peptides from amino acid sequence. Paste sequences in FASTA format (or fetch them from UniProt), choose a prediction method, and get pI values across up to 19 pKa scales plus machine learning predictions, along with per-residue pKa for every ionizable site.

Inputs

Settings

Predictor options

Results

The Results tab shows a spreadsheet with one row per sequence:

The Per-Residue pKa tab shows predicted pKa values for every ionizable site:

Interpreting pI values

Most proteins have pI between 4 and 10. Knowing where a protein falls relative to pH 7 determines its behavior in common buffers:

When multiple pKa scales agree within 0.3 pH units, the prediction is likely reliable. Disagreement of more than 1 pH unit suggests the protein has unusual charge properties (e.g., many histidines) where scale choice matters. In those cases, the SVR prediction is more trustworthy because it was trained on experimental data and can capture non-additive effects.

Interpreting per-residue pKa

Per-residue pKa values show how sequence context shifts each ionizable group's dissociation constant away from its "standard" textbook value. For example, an aspartate surrounded by other negative residues will have a higher pKa (harder to deprotonate), while one near positive charges will have a lower pKa.

These values are useful for identifying residues with unusual protonation behavior, which matters for enzyme active sites, pH-dependent conformational changes, and designing mutations that shift pI.

How does IPC 2.0 work?

Henderson-Hasselbalch method

The classical approach calculates net protein charge as a function of pH using:

$\text{pH} = \text{p}K_a + \log\frac{[\text{A}^-]}{[\text{HA}]}$

Starting at pH 6.51, a bisection algorithm adjusts pH up or down until the sum of positive charges (Lys, Arg, His, N-terminus) equals negative charges (Asp, Glu, Cys, Tyr, C-terminus) within a precision of 0.01 pH units.

Each pKa scale assigns different dissociation constants to these groups, producing different pI estimates. The Bjellqvist scale additionally uses position-specific pKa values for N-terminal and C-terminal residues (e.g., N-terminal Ala has pKa 7.59 instead of the generic 7.5), which matters for proteins starting with A, M, S, P, T, V, or E.

ProMoST uses a separate model with three pKa values per residue (N-terminal, middle, C-terminal positions), giving it slightly different behavior from the other 18 scales.

pKa scales

IPC 2.0 includes 19 pKa scales:

The Avg_pI column averages all scales except Patrickios, whose simplified model (ignoring cysteine, histidine, and tyrosine) makes it an outlier.

Support vector regression

The SVR models take a 19-dimensional feature vector (pI predictions from all 18 H-H scales + ProMoST) and feed it into a support vector machine with RBF kernel. The protein SVR was trained on 2,324 proteins from SWISS-2DPAGE and PIP-DB. The peptide SVR was trained on 119,092 peptides from high-resolution isoelectric focusing (HiRIEF) experiments.

Because the SVR sees all 19 scale predictions simultaneously, it learns which scales are most informative for different sequence compositions. It consistently outperforms any individual scale.

SepConv2D deep learning

The deep learning model uses a separable convolutional architecture with four input channels:

• One-hot encoding: 22 x 60 matrix of amino acid identity
• AAindex features: 15 physicochemical indices per position plus hydrophobicity
• Composition vector: Amino acid counts repeated across sequence length
• IPC predictions: pI from all scales + SVR, giving the model calibrated priors

Sequences longer than 60 residues undergo truncation that preferentially removes non-charged residues (Ala, Gly, Leu, etc.) while preserving ionizable residues that determine pI. This truncation uses random selection among non-charged positions, so predictions for long sequences may vary slightly between runs.

Per-residue pKa (MLP-SVR ensemble)

The per-residue pKa predictor estimates the dissociation constant of each ionizable residue (D, E, H, K, Y) plus the N- and C-terminal groups. It uses a stacking ensemble of 9 multilayer perceptrons (MLPs) feeding into a final SVR:

• 6 sequence-only MLPs operating on 5-mers through 15-mers around the target residue
• 3 AAindex-augmented MLPs operating on 3-mers through 7-mers

Each MLP encodes the local sequence context around an ionizable residue using one-hot encoding (and optionally AAindex physicochemical features). The 9 MLP predictions are stacked as features for the final SVR, which outputs a single pKa value. Charged N- and C-terminal residues are flanked with alanine padding before prediction to ensure they are treated as terminal groups.