Kako se Europa može natjecati u umjetnoj inteligenciji

Scientists have developed a groundbreaking AI-designed gene-editing enzyme that significantly enhances the CRISPR toolkit. Published in Science on 16 July 2026, the study introduces a new class of synthetic enzymes called SynTnpBs, which outperform their natural counterparts in both efficiency and versatility. These enzymes, derived from the TnpB family, ancestors of widely used CRISPR-Cas12 systems, are engineered using an advanced AI model known as the ESM Inverse Folding (ESM-IF1) model. This breakthrough marks a major step forward in the quest to create entirely novel gene-editing tools with unique functionalities. The research team, led by Jennifer Doudna, a biochemist at the University of California, Berkeley, and a co-recipient of the 2020 Nobel Prize in Chemistry, sought to overcome a key limitation of existing CRISPR technologies. While traditional CRISPR systems rely on naturally occurring enzymes like Cas9 and Cas12, these enzymes are constrained by evolutionary history. Their structures and functions are optimized for survival in microbial environments, limiting their adaptability for applications in human cells or plants. By leveraging AI, the researchers aimed to generate enzymes with tailored properties that could perform tasks previously unattainable with natural proteins. The process began with the known three-dimensional structure of the TnpB enzyme. Using the ESM-IF1 model, the team inputted this structural data into the AI system, which then proposed numerous variations of the enzyme's amino acid sequence. The goal was to identify modifications that would preserve the enzyme's essential functions while introducing new capabilities. The AI model analyzed vast datasets of evolutionary relationships and protein interactions to determine which mutations would retain functionality without compromising performance. The resulting designs were then tested in bacterial cultures, followed by trials in human cell lines and Arabidopsis thaliana plant cells. Among the 1,980 candidate designs evaluated in bacterial screens, 466 demonstrated measurable activity. Notably, approximately 8% of these variants exhibited superior performance compared to the original TnpB enzyme. In human cells, two of the most effective variants achieved editing efficiencies of up to 50%, surpassing the 28% efficiency of the native enzyme. Some variants matched the original enzyme's effectiveness, while others showed remarkable improvements at specific DNA targets, achieving nearly four times greater editing accuracy. These findings underscore the potential of AI-driven protein design to revolutionize gene-editing capabilities. To ensure the new enzymes were sufficiently distinct from their natural predecessors, the researchers assessed the degree of divergence from the original TnpB. Unlike previous AI approaches that produced proteins with over 99% similarity to natural homologs, the ESM-IF1 model enabled the creation of enzymes with only 72% to 83% similarity to their closest natural counterparts. This reduced resemblance suggests that the AI-generated enzymes possess fundamentally different molecular interactions, potentially opening new avenues for therapeutic and agricultural applications. Despite these successes, the study acknowledges limitations. The current approach focuses on a single family of nucleases, meaning the methodology may not apply universally to all CRISPR systems. Nevertheless, the results highlight the promise of combining AI with structural biology to expand the designable protein space. As the field continues to evolve, further refinements could lead to even more efficient, smaller, and versatile gene-editing tools. The long-term vision, according to the researchers, is to develop fully synthetic enzymes capable of performing functions beyond what nature has already provided.