KaML-ESM2

The pKa of a titratable side chain determines whether it is charged at physiological pH, and therefore whether it can act as a catalytic acid, a base, a nucleophile, or a salt-bridge partner. Predicting those values has been a structure-based problem for decades: Poisson-Boltzmann solvers, empirical schemes such as PROPKA3, constant-pH molecular dynamics, and machine learning models trained on structures. All of them need a structure, and all are sensitive to which structure you give them.

KaML-ESM2, from Jana Shen's lab at the University of Maryland School of Pharmacy, drops that requirement. It is a small multilayer perceptron task head trained on the frozen per-residue embeddings of ESM-2, predicting the pKa shift of Asp, Glu, His, Cys, Tyr, and Lys residues directly from sequence. The underlying claim is a biological one: because protein language models learn from sequences shaped by hundreds of millions of years of selection, their representations encode not only structure and function but, indirectly, electrostatic character.

A sibling head, KaML-ESMC, was trained the same way on ESM Cambrian embeddings; this variant is the more accurate of the two in distribution and the one that runs without an external embedding service. Both succeed the lab's structure-based KaML family. First posted as a preprint in April 2025 and revised since, the work ships as an end-to-end platform: a command-line tool, a hosted web server, and a precomputed pKa atlas covering the human proteome.

#Key Features

• Sequence-only prediction: No experimental or predicted structure is needed, making proteome-scale application feasible for targets that lack structures or whose predicted models contain large low-confidence regions.
• Frozen backbone, thin head: Parameter-efficient fine-tuning of the backbone with qLoRA gave a negligible net change in error, so the released model trains only the task head on frozen embeddings.
• GAINES data augmentation: An augmentation strategy named auGment dAta wIth lateNt spacE Sampling retrieves residues from a protein database whose embeddings closely match a labelled query residue and transfers the experimental pKa label, addressing the scarcity of measured values.
• Separate acid and base channels: Independent heads handle acidic (Asp, Glu, Cys, Tyr) and basic (His, Lys) residues, each deployed as an ensemble of 500 models to give per-prediction standard errors.
• Fully local inference: Embeddings come from openly downloadable ESM-2 650M weights, so the whole pipeline runs offline.

#Technical Details

The task head is a four-layer MLP with three hidden layers, taking per-token embeddings from ESM-2 650M and regressing the pKa shift relative to model-peptide reference values. Layer choice matters: a head on first-layer representations gave an RMSE of 1.22 versus 0.75 from the second-to-last layer. Training used the PKAD-3r experimental database expanded by GAINES, with 50 stratified random holdout splits grouped by unique residue.

Across those holdouts the GAINES-trained model reached an overall RMSE of 0.48 ± 0.02 with a Pearson correlation of 0.98. Partitioning the data with CD-HIT at 70% sequence identity, which removes homology leakage, gave 0.43 ± 0.02, and the external EXP67S benchmark gave 0.52. Against ten alternatives the margin is large: the strongest competitor, the structure-based CNN DeepKa, sits just under 1.0 RMSE, and PROPKA3 is comparable to a null model that simply assumes solution pKa values. Critical error rates — residues assigned the wrong protonation state at pH 7 — stay below 1.2% on stratified holdouts and 0.6% on CD-HIT splits, about four times lower than DeepKa.

The clear failure mode is engineered buried substitutions in staphylococcal nuclease, which have no evolutionary precedent: RMSE rises to 2.76, notably worse than the 1.98 its ESM Cambrian sibling manages on the same set. Where a design campaign is producing residues evolution never sampled, that gap is the reason to prefer the sibling.

#Applications

Ionization states feed directly into enzyme mechanism assignment, molecular dynamics setup, protein engineering of pH-dependent stability, and structure-based drug design where a mispredicted protonation state changes a binding pose. The human proteome scan produced 1,877,704 predicted values, from which 130 cysteines predicted to be deprotonated at physiological pH were mapped to Pfam families; of 30 such cysteines in ubiquitin hydrolases, 29 are annotated catalytic nucleophiles, illustrating use as a functional-site annotation screen. Proteins longer than 1,022 residues fall outside the ESM-2 context limit and were excluded.

#Impact

This is the strongest current evidence that residue-level electrostatics can be read out of a protein language model without a structure, and the proteome-wide scan demonstrates a scale no structure-based predictor can match. The practical caveats are concrete: the code is released under CC BY-NC 4.0, so commercial use requires contacting the authors, and the work remains a preprint. The command-line interface downloads pretrained weights from Zenodo and can optionally add structure-based KaML-CBT2 predictions, folding an input sequence with ESM-3 when no structure is supplied.