14 May 2025

AI Designs Synthetic DNA to Control Genes in Healthy Cells for First Time

A groundbreaking study published recently in Cell reveals the first successful use of generative AI to design synthetic DNA sequences that can control gene expression in healthy mammalian cells. Scientists at the Centre for Genomic Regulation (CRG) developed an AI model capable of producing novel DNA regulatory sequences that do not exist in nature.
The AI can be instructed to create synthetic DNA fragments with specific functions, such as activating a gene only in stem cells that will become red blood cells but not platelets. The model predicts the sequence of DNA bases needed for specific gene expression patterns in various cell types. Researchers then synthesize the roughly 250-letter DNA segments and introduce them into cells via viruses.
To demonstrate its capabilities, the researchers instructed the AI to design DNA that activates a gene encoding a fluorescent protein in some mouse blood cells, without disrupting normal gene activity. The DNA fragments were inserted at random locations in the genome and performed exactly as the model predicted.
According to Dr. Robert Frömel, lead author of the study, this technology is comparable to writing software for cells, offering new ways to precisely direct cellular development and behaviour. He emphasized the potential for medical applications, such as enhancing gene therapy by precisely targeting gene activity in specific tissues while minimizing side effects.
Until now, gene expression was controlled using naturally occurring enhancers, which limited design possibilities. AI-generated enhancers allow the creation of ultra-selective genetic switches that evolution has not produced, enabling therapies tailored to individual cell needs.
Developing the model required vast datasets. According to Dr. Lars Velten, the study senior author, understanding the grammar of the DNA language was essential to generating meaningful sequences. The team conducted thousands of experiments in healthy cells, instead of the more commonly used cancer cell lines, to build a biologically accurate model.
Over five years, the team synthesized more than 64,000 synthetic enhancers, testing them in various blood cell stages. Each enhancer included combinations of binding sites for 38 transcription factors—proteins that help regulate genes. This effort produced the largest library of synthetic enhancers in blood cells to date.
The experiments revealed that most enhancers fine-tune gene activity like a volume dial, while some act as binary switches depending on their combinations—an effect termed negative synergy. For example, two factors that normally activate a gene separately could shut it down when combined.
These insights shaped the machine learning model, enabling it to design novel enhancers that produce precise gene activation or repression patterns, even if the sequences have never existed naturally.
Although the study was a proof-of-concept, it sets the stage for broader applications. With humans and mice having about 1,600 transcription factors, the researchers believe this is just the beginning. The project was led by Veltens group at CRG and supported by the European Union's ERC Starting Grant and Spains National Agency for Research.