ProteinGuide enables on-the-fly conditioning of protein design models with experimental data
ProteinGuide provides an inference-time conditioning framework that lets a broad class of protein sequence generators (including masked language models like ESM3, autoregressive models like ProteinMPNN, and discrete-space diffusion/flow models) be guided by auxiliary experimental data without retraining. Built on a unifying statistical approach, it enables designs targeting properties such as stability and activity, supports multi-property Pareto optimization, and, when combined with wet-lab data, can enhance in vivo base-editor activity beyond multi-round directed evolution. The work provides code and data (Zenodo, GitHub) and a ProteinGen package to facilitate design workflows.
