Developing predictive mechanistic models in biology is challenging. Elektrum uses neural architecture search, kinetic models and transfer learning to discover CRISPR–Cas9 cleavage kinetics, achieving high performance and biophysical interpretability.