Joint steering of protein generation across multiple target properties
Ideally, protein design could optimize several properties at once while ensuring the protein satisfies basic biophysical constraints such as folding, stability, solubility, and expression. However, most guided generation methods optimize only one or two simple objectives. We developed a workflow that uses predictive models of several desired properties to steer sequence generation. Using fluorescent proteins as a test case, we show computationally that guiding generation toward target excitation and emission peaks produces designs closer to those targets than unguided generation with subsequent filtering, particularly when starting from parent proteins the guiding model was never trained on. We also find that a family-specific sequence profile, a probability distribution derived from multiple sequence alignments, provides a more useful generative prior for fluorescent proteins than ESM-2. Out of the ten designs we tested experimentally, one came back as a working fluorescent protein, but didn't fluoresce near our target wavelength. We're sharing this work for researchers using generative protein models who want to incorporate multiple, potentially high-dimensional experimental measurements directly into design. We discuss generalizable lessons on protein search space and predictor-based guidance that we hope will inform similar projects.