Leveraging pre-trained AI models for robust promoter sequence design in synthetic biology

Although artificial intelligence (AI) has begun to be applied in synthetic biology, it is limited by its reliance on large amounts of high-quality data, which presents a significant challenge in synthetic biology. Pre-trained models have profoundly influenced natural language processing by enabling systems to understand and generate human language with remarkable accuracy and efficiency by capturing complex linguistic patterns and contextual nuances. This study applies the concept of pre-trained models to promoter sequence analysis through an innovative pre-training and fine-tuning paradigm. Our analysis reveals that pre-trained DNA models, particularly DNABERT, consistently outperform non-pre-trained models in predicting promoter expression levels across various dataset sizes. Building on DNABERT's strengths, we developed the AI model Pymaker, which specializes in predicting yeast promoter expression levels. Additionally, we introduced a novel base mutation model to simulate promoter mutations, enabling the generation of new promoter sequences. By integrating Pymaker with this mutation model, we effectively screened for high-expression, mutation-resistant promoters. Experimental validation in Saccharomyces cerevisiae showed that these selected promoters significantly enhanced LTB protein expression. Notably, Pymaker's predictions demonstrated superior accuracy, achieving a three-fold increase in protein expression compared to traditional promoters. Our findings highlight the potential of Pymaker not only to identify robust promoters but also to significantly reduce reliance on conventional, labor-intensive experimental methods, heralding a new era in synthetic biology and genetic engineering with practical applications in biopharmaceuticals.