Activity-feature-guided discrete diffusion with direct preference optimization for strong promoter design in actinomycetes

Abstract

Strong promoters enhance transcriptional output and are important for natural product discovery and industrial strain engineering in actinomycetes. However, host-compatible strong promoters are scarce, and AI-assisted design of high-activity promoters remains challenging. Here, we develop DPOCondD3PM, an activity-feature-guided conditional discrete diffusion framework for strong promoter design in actinomycetes. We construct a promoter dataset from three actinomycete species and derive relative activity labels from RNA-seq data. These labels are aligned onto a unified relative activity scale. We then learn multi-scale activity-related features from natural promoters and use them to condition discrete diffusion. We further apply direct preference optimization (DPO) using within-condition differences in predicted activity among generated candidates. Across the actinomycete dataset and an Escherichia coli benchmark, DPOCondD3PM generates sequences with higher predicted activity than those from representative methods. In silico analyses suggest that the generated sequences exhibit promoter-like compositional characteristics, predicted structural profiles and core-promoter organization while remaining diverse and distinct from natural promoters. In vivo fluorescence assays showed measurable activity for all tested generated promoters, with several exceeding actinomycete strong constitutive promoters kasO*p and ermE*p. This work provides an activity-directed framework for strong promoter design in actinomycetes and extends generative promoter design beyond model bacteria.