AI设计人类增强子！超越天然增强子，短至50bp也能实现细胞特异性

Core Viewpoint - The research conducted by Washington University and Altius Biomedical Science Institute successfully designed synthetic enhancers that are more efficient and simpler than natural enhancers, achieving unprecedented cell-type specificity in human cells through iterative deep learning technology [2][6]. Group 1: Research Challenges - Traditional enhancer discovery faces three major challenges: the vast number of candidate enhancers in the human genome, the lack of precision in existing enhancers that often activate multiple cell types, and the complexity of regulatory rules involving various transcription factor combinations and spatial arrangements [6]. Group 2: Research Methodology - The research team developed an iterative deep learning design system, which underwent two cycles of "design-experiment-optimize," starting from 29,891 natural enhancer MPRA activity data to train the model, resulting in the design of 1,037 synthetic enhancers [6]. - The model was refined using real measurement data of synthetic enhancers, reducing the training data volume by 30 times compared to previous generations, and introducing L2 regularization to prevent over-reliance on a single transcription factor [6]. - The second generation achieved a breakthrough with the design of 688 new enhancers, significantly increasing median expression levels in specific cell types, such as a 46.2-fold increase in HepG2 cells and a 6.7-fold increase in K562 cells [6][7]. Group 3: Research Highlights - The specificity of the deep learning-designed enhancers surpassed that of natural controls, and the sequence grammar used for synthetic enhancers was more compact than that of natural enhancers [8]. - Iterative retraining of synthetic enhancers led to designs with superior specificity, and the activity of synthetic enhancers was correlated with single-cell transcription factor expression [8]. Group 4: Applications - The research opens three major application directions: targeted gene therapy for liver cancer, customized tissue-specific enhancers for rare genetic diseases, and the construction of cell-type-specific biosensors in synthetic biology [10]. - This study marks a fundamental shift in the design paradigm of gene regulatory elements, moving from traditional methods to an AI-driven approach that significantly increases success rates [10].