Nonparametric Biologic Design via Discrete Drifting

Computational optimization of therapeutics requires moving candidate populations through combinatorial spaces in a manner that retains properties of the inital population, avoids toxic regions of design space and achieves optimization goals. Recent methods train discrete diffusion/flow models to form a prior over this movement. However, these priors are task-specific and cannot transport between arbitrary regions of design-space. In this work, we introduce NimblDrift, a discrete feature-space drifting method that uses the structure from a frozen representation encoder to guide therapeutic optimization. At each step, we compute a drift field in the continuous feature-space, then project the update onto a neighborhood in the discrete design-space. Testing on peptide and small-molecule design tasks, we find that representation geometry can guide property-relevant updates and that different encoders contain different biases. We conclude with discussion on the promises and limitations of the method.