pKALM

Cysteine 46 in the bacterial peroxiredoxin AhpC titrates at pH 5.94, more than three units below the 9.0 an isolated cysteine would show, because residues scattered along the sequence pack against it once the protein folds. Structure-reading predictors miss that almost entirely — PROPKA returns 9.0, PKAI 9.64, PKAI+ 9.08. pKALM returns 6.67 from the amino-acid sequence alone, and the attention maps of its frozen language-model backbone show that position attending to His78, Trp81, Arg119, Ala120 and Asp143: residues far apart in sequence, adjacent in space.

That case is the argument the model makes. Protein pKa has been treated as a structure-based problem, because titration behaviour is set by a residue's local electrostatic environment. pKALM shows that a language model pretrained on large sequence databases has already absorbed enough of that environment for a small supervised head to recover the shift. Shijie Xu and Akira Onoda at Hokkaido University posted the method to bioRxiv in September 2024 and published it in the Journal of Chemical Theory and Computation in March 2025. It covers six titratable side chains — Asp, Glu, His, Lys, Cys and Tyr — plus the N- and C-termini, a wider set than most structure-based tools attempt.

Dropping the structure requirement also lifts the throughput ceiling: the language model runs once per protein rather than once per residue, and pKALM annotated the entire human proteome — 20,598 sequences, 3,554,768 protonatable residues — in 11 minutes 56 seconds. The Shen lab reached the same sequence-only construction independently seven months later in KaML-ESM2, an unrelated artifact with its own training pipeline and benchmarks.

#Key Features

• Sequence-only inference: No PDB entry, no predicted structure, no conformational sampling — proteins that never crystallized are as tractable as ones that did.
• Predicts the shift, not the value: The network outputs a perturbation added to the tabulated standard pKa for that residue type, so the easy baseline is built in and the model spends its capacity on the deviation.
• Isoelectric point as an auxiliary signal: Two pI predictors — one for peptides, one for whole proteins — are trained separately, frozen, and concatenated into the pKa input. Removing them raises test error, with the protein-level pI mattering most.
• Coverage of the rare titration sites: Cys, Tyr and both termini are included, where several established tools return nothing. Cysteine RMSE is roughly 2.1 pH units against more than 4 for PKAI+ and more than 5 for PKAI and PROPKA.
• Proteome-scale throughput: 4,965 pKa predictions per second — 37 times faster than PKAI+, 130 times faster than PROPKA, 5,773 times faster than the Poisson-Boltzmann solver PypKa.

#Technical Details

A protein sequence is encoded in parallel by the frozen language model and the two frozen pI models; the concatenated features enter a bidirectional LSTM, a residue-type embedding is added to its outputs, and a fully connected layer emits the pKa shift. Nine backbones were compared — ESM-2 at 8M through 15B parameters, ESM-1b at 650M, and two ProtTrans models — with the 35M-parameter ESM-2 selected on test RMSE over the abundant residues. Larger backbones did better on shifts under 1 pH unit, smaller ones on shifts above 1.5. Sequences over 960 tokens are segmented with overlap.

Training used a manually corrected PKAD-2 (1,450 pKa values across 165 wild-type proteins and 262 across 47 mutants), clustered with CD-HIT at 30% sequence identity to give 376 training and 152 test measurements; the pI models used 119,092 peptide and 2,324 protein values. Test RMSE is 0.6929 for Asp, 0.6641 for Glu, 1.0468 for His, 0.7701 for Lys, 0.6479 for Tyr, 0.7419 for the N-terminus and 0.8038 for the C-terminus. Buried residues, the hard case, come out lowest among the compared methods; exposed residues are second to PKAI+.

#Applications

Protonation states left at their pH 7 defaults are a standing source of error in molecular dynamics setup, structure-based drug design and enzyme mechanism work; pKALM makes assigning them cheap enough to do proteome-wide. On chymotrypsin it recovers His57 of the catalytic triad at 7.65 against a measured 7.5, and on ribonuclease T1 its five titrated residues land within 0.25 units of experiment. A batch FASTA script and a hosted web server let a wet-lab group screen mutations for charge effects with no local GPU.

#Impact

pKALM adds a quantitative biophysical property to what protein language model embeddings turn out to carry, alongside secondary structure, contacts and disorder — and it does so on 376 training measurements, which is an argument about the backbone rather than the dataset. The limits are correspondingly narrow: for pKa shifts above 1.5 units the model drops to second or third best, cysteine error stays near 2 pH units, and the released version accepts single-chain input only, with a multi-chain version the authors describe as underway. Code is CC BY-NC-SA 4.0 and the checkpoints sit behind a gated academic-use agreement, so commercial use needs the authors' permission.