Protein language models predict ionisation from amino acid sequence

A protein’s sequence can reveal more than how it might fold. Researchers report that protein language models can predict pKa values, which describe how readily particular chemical groups gain or lose a proton. The result extends sequence-based AI towards the electrical chemistry that helps determine binding, catalysis and stability.[1]

The relevant quantity is pKa, an equilibrium measure associated with how readily a chemical group releases a proton. A residue's pKa in an isolated chemical environment is not necessarily its pKa when buried or exposed inside a folded protein. Neighbouring charges, local structure and solvent access influence the answer. Predicting these values accurately is useful because enzymes frequently depend on individual side chains changing charge at the right point in a reaction.[1]

Mingzhe Shen and colleagues developed KaML-ESMs, sequence-based neural models built on protein language models in the ESM family. Instead of treating sequence as a plain identifier for a known protein, these systems use patterns learned across evolutionary sequences to predict chemically meaningful properties. The training data incorporated the authors' GAINES-augmented PKAD-3r dataset. That technical detail matters because a model with an attractive architecture can still fail when it lacks reliable examples of difficult local environments.[1]

Reading electrostatic clues from sequence

Across four benchmarks, the models achieved root-mean-square errors of about 0.5 pKa units for several residue types in native proteins. That is an error on a logarithmic chemical scale, not a percentage accuracy score. Its practical importance depends on the pH range and the role of the residue being studied.[1]

What makes the result striking is the information path. Conventional electrostatic calculations often require a structural model and approximations about the protein's surrounding solvent. A sequence language model starts from amino acid order and an evolutionary representation. The reported success suggests that the history encoded in sequences indirectly contains clues about chemical environments, because evolutionary selection preserves residues that work under particular local physical conditions.[1]

The study tested precisely that boundary: performance weakened on 89 buried engineered mutations lacking strong evolutionary support. These difficult cases matter for protein design, where the aim is often to create behaviour outside a natural pattern. Targeted experimental supervision could therefore be more useful than assuming a larger model will resolve the gap automatically.[1]

The engineered mutations are the harder design problem

The researchers also applied the method across the human proteome to identify possible functional sites and catalytic mechanisms. This is a computational route for prioritising biochemical questions. A suggested site still needs evidence of its function in the relevant protein and cellular setting.[1]

Sequence, structure, dynamics and charge describe connected aspects of a protein. A useful approximation at one layer can accelerate work at another, provided its scope is clear. Here the encouraging native-residue results and the weaker engineered-mutant results belong together: they show both where the representation helps and which experimental data would most improve it.[1]

KaML-ESMs gives researchers another way to explore a protein before committing to more expensive calculations or experiments. The most productive next step is the one the hard mutants identify—learning the chemistry of environments that evolution has supplied less often.[1]

Report an error or suggest a correction