A Hybrid Large Language Model and Diffusion Framework for Rational Promoter Design

Abstract High-performance promoters are essential tools for precisely regulating gene expression, yet their rational design within the vast combinatorial sequence space remains a major challenge. Here, we present a hybrid framework that integrates a large language model (LLM) with a diffusion model to enable data-driven and interpretable promoter design. The fine-tuned LLM predicts promoter strength with high accuracy and, through systematic sequence mutations, identifies biologically essential core regions. A diffusion model is then conditioned on these core regions to reconstruct non-core regions and generate complete promoter sequences. We experimentally validated this approach in Escherichia coli by high-throughput barcoded promoter activity sequencing: 7.78% of the designed promoters outperformed the benchmark promoter BBa_J23119, with the best variants reaching approximately 2.5-fold the expression level of BBa_J23119. By explicitly coupling interpretability with generative design, this strategy provides a practical framework for promoter design in E. coli and may be adaptable to other regulatory sequence engineering tasks.