Property guidance for protein sequence generative models with ProteinGuide

No principled framework exists for conditioning sequence generative models for protein engineering on auxiliary information, such as experimental data, without additional training of a generative model. Here we present ProteinGuide, a method for such 'on-the-fly' conditioning. ProteinGuide is amenable to a broad class of protein generative models including masked language models such as ESM3, any-order autoregressive models such as ProteinMPNN and diffusion and flow-matching models on discrete state-spaces such as MultiFlow. ProteinGuide stems from a unifying statistical framework for these model classes. As proof of principle, pretrained generative models are used to design proteins with user-specified properties, such as higher stability or activity. Proteins are additionally designed to optimize for two desired properties that are in tension with each other. Lastly, we apply ProteinGuide jointly with wet-lab data generation to increase the editing activity of an adenine base editor in vivo, resulting in a base editor with higher editing efficiency than was previously achieved using seven rounds of directed evolution.