AI and the World's Hard Problems · Life sciences · Nature (machine learning) · 2026-10-24
Predicting protein electrostatics with protein language models
Ionization states are crucial to protein function, but predicting their equilibrium constants (pKa values) remains difficult. The researchers built KaML-ESMs, neural networks based on ESM protein language models that predict pKa values from sequence alone, reaching errors of about 0.5 units for common residues in native proteins. Applied to the entire human proteome, the predictions helped identify functional sites and infer catalytic mechanisms.