SPROUT: Steered Plant Promoter Editing via Rollout-Guided Utility Tilting of Edit Flows

Engineering plant cis-regulatory elements (CREs) has high-impact applications in domains including agriculture, bioremediation, and bioengineering. Existing CRE engineering frameworks typically use either naive generative models, including random sequence generation or genetic algorithms, and/or reductive descriptions of expression context. The common formulation for cell-type specific expression -- expression enrichment in one cell type over another -- ignores all other cell types and axes of expression. For plant promoters, functional utility depends upon controlling a richer expression context across tissues, developmental stages, and environmental conditions. We thus propose SPROUT, an alternative formulation for CRE engineering based on Steered plant Promoter editing via ROllout-guided Utility Tilting of Edit Flows, pairing a recently introduced discrete flow matching framework for variable-length sequence editing with multi-objective optimization across oracles predicting expression context and strength. We evaluate SPROUT against the common formulations and show that our Edit Flow-based formulation produces more realistic sequences than random generation, while optimization with oracles trained on only two conditions does not fully constrain other expression axes.