pKALM: Accurate and Rapid Protein pK_a Prediction from Sequence

pKALM^[1] predicts protein pK_a values directly from amino-acid sequence using a protein language model and transfer learning — no structure required. It reaches state-of-the-art accuracy (0.8658 RMSE) at a throughput of ~4,965 pK_a/s, covering six ionizable side chains (Asp, Glu, His, Lys, Cys, Tyr) plus the N- and C-termini.

GitHub | Paper | License

Run a Prediction

Paste one or more sequences in FASTA format or upload a FASTA file, pick a model, and submit. Each run opens a dedicated, bookmarkable result page.

Model

Sequence (FASTA) Up to 2,000 residues per job. For longer sequences, please contact us.

Upload FASTA file Accepts .fasta, .fa, .txt, or .seq.

Predictions currently run on a CPU server. pKALM is sequence-based and fast, so most jobs finish in seconds to a minute. The result page shows live progress, can be bookmarked or shared, and is retained for one week.

How pKALM Works

Architecture of pKALM: a protein language model, protein and peptide pI models, and a residue embedding feed a BiLSTM whose output is added to standard pKa values; a separate pI model predicts isoelectric point. — Figure 1: Overview of pKALM. **(a)** The pK_a predictor concatenates features from a frozen **protein language model** (ESM2)^[2], **protein** and **peptide pI models**, and a residue embedding, feeds them to a **BiLSTM**, and adds the learned shift to the standard pK_a. **(b)** The shared pI model architecture (residue embedding → BiLSTM → average → linear).

Performance

pKALM is benchmarked against widely used pK_a predictors — the empirical method PROPKA^[3], the Poisson–Boltzmann solver PypKa^[4], and the deep-learning methods pKAI / pKAI+^[5] and DeepKa^[6].

Bar charts comparing pKALM RMSE against PKAI+, DeepKa, PKAI, PypKa and PropKa for abundant residues (Asp, Glu, His, Lys) and less abundant residues (Cys, Tyr, N- and C-terminus). — Figure 2: pK_a RMSE vs. existing methods. **(a)** Abundant residues — pKALM gives the lowest RMSE on Asp, His and Lys. **(b)** Less abundant residues — pKALM uniquely predicts Cys and leads on Tyr and the termini, where several structure-based methods are unavailable (marked "X").

Case Studies

Beyond aggregate metrics, pKALM recovers chemically meaningful sites directly from sequence — including catalytic residues whose pK_a is strongly perturbed by their environment.

Three protein structures with predicted ionizable residues highlighted: ribonuclease H, the chymotrypsin catalytic triad, and the bacterial peroxiredoxin AhpC with residues having high attention to the catalytic Cys46. — Figure 3: Exemplary predictions. **(a)** The five titrated residues of ribonuclease H. **(b)** The catalytic triad of chymotrypsin — the shifted His57 pK_a (predicted 7.65 vs. experimental 7.5) is captured from sequence alone. **(c)** The bacterial peroxiredoxin AhpC: the catalytic Cys46 (red) and residues with high attention to it (orange) are distant in sequence but close in 3D, suggesting the PLM encodes long-range structural context.

Supported Ionizable Groups

Side chains

AspGlu HisLys CysTyr

Termini

N-terminusC-terminus

Each job also returns predicted isoelectric points (pI) for the peptide and protein models, appended to the results.

References

^ Xu, S.; Onoda, A. Accurate and Rapid Prediction of Protein pK_a: Protein Language Models Reveal the Sequence–pK_a Relationship. J. Chem. Theory Comput. 2025, 21 (7), 3752–3764. DOI: 10.1021/acs.jctc.4c01288
^ Lin, Z.; Akin, H.; Rao, R.; Hie, B.; Zhu, Z.; Lu, W.; Smetanin, N.; Verkuil, R.; Kabeli, O.; Shmueli, Y.; et al. Evolutionary-Scale Prediction of Atomic-Level Protein Structure with a Language Model. Science 2023, 379 (6637), 1123–1130. DOI: 10.1126/science.ade2574
^ Olsson, M. H. M.; Søndergaard, C. R.; Rostkowski, M.; Jensen, J. H. PROPKA3: Consistent Treatment of Internal and Surface Residues in Empirical pKa Predictions. J. Chem. Theory Comput. 2011, 7 (2), 525–537. DOI: 10.1021/ct100578z
^ Reis, P. B. P. S.; Vila-Viçosa, D.; Rocchia, W.; Machuqueiro, M. PypKa: A Flexible Python Module for Poisson–Boltzmann-Based pKa Calculations. J. Chem. Inf. Model. 2020, 60 (10), 4442–4448. DOI: 10.1021/acs.jcim.0c00718
^ Reis, P. B. P. S.; Bertolini, M.; Montanari, F.; Rocchia, W.; Machuqueiro, M.; Clevert, D.-A. A Fast and Interpretable Deep Learning Approach for Accurate Electrostatics-Driven pKa Predictions in Proteins. J. Chem. Theory Comput. 2022, 18 (8), 5068–5078. DOI: 10.1021/acs.jctc.2c00308
^ Cai, Z.; Luo, F.; Wang, Y.; Li, E.; Huang, Y. Protein pKa Prediction with Machine Learning. ACS Omega 2021, 6 (50), 34823–34831. DOI: 10.1021/acsomega.1c05440

Please contact shijie.xu@ees.hokudai.ac.jp for any questions.

Changelogs

2026-05-28Rebuilt the web interface with a live result page.
2025-11-09Updated citation.
2024-10-22First release.

pKALM: Accurate and Rapid Protein pKa Prediction from Sequence