Property guidance for protein sequence generative models with ProteinGuide
On-the-fly conditioning of pretrained protein generative models guides protein generation toward specific properties.
A principled framework for conditioning sequence generative models for protein engineering on auxiliary information, such as experimental data, does not currently exist. To address this gap, researchers have developed ProteinGuide, a method for 'on-the-fly' conditioning that is applicable to a broad class of protein generative models, including masked language models like ESM3, any-order autoregressive models such as ProteinMPNN, and diffusion and flow-matching models on discrete state-spaces like MultiFlow. ProteinGuide is based on a unifying statistical framework for these model classes. By leveraging this framework, pretrained generative models can be used to design proteins with user-specified properties, such as increased stability or activity. Furthermore, proteins can be designed to optimize for two desired properties that are in tension with each other. The effectiveness of ProteinGuide has been demonstrated through its application in designing proteins with specific properties, as well as its use in increasing the editing activity of an adenine base editor in vivo. In this latter application, ProteinGuide was used in conjunction with wet-lab data generation to achieve a base editor with higher editing efficiency than was previously achieved using seven rounds of directed evolution. The use of ProteinGuide in protein engineering applications is limited to research-use only and should not be used for any clinical or therapeutic purposes without further investigation and regulatory approval.