Abstract: Prime Editing (PE), a revolutionary technology in precision genome editing, is celebrated for its powerful "search-and-replace" capabilities. However, its efficiency has long been hampered by the cell's endogenous Mismatch Repair (MMR) system. A recent study published in Cell leverages generative AI to design a miniature protein (MLH1-SB) of just 82 amino acids that precisely "locks down" the MMR pathway, elevating PE efficiency to new heights [1]. This article places this paper in the broader context of PE's technological evolution, dissecting how AI has solved a core bottleneck that has plagued the field for years, and explores the profound implications of this new "AI-designed functional module" paradigm for the future of gene editing and synthetic biology at large.

Introduction: The Dilemma of Precision vs. Efficiency in Prime Editing

Since its debut in 2019, the Prime Editing technology developed by David Liu's team has been hailed as the "Swiss Army knife" of genome editing [2]. Unlike traditional CRISPR-Cas9 systems that rely on DNA double-strand breaks, PE employs a gentle "search-and-replace" mechanism to perform all 12 types of single-base substitutions, as well as small insertions and deletions, vastly expanding the scope and precision of gene editing [2]. Yet, this technology of immense potential has consistently faced a central bottleneck: efficiency.

After a PE system executes a precise DNA sequence modification within a cell, the cell's own "quality inspector"—the Mismatch Repair (MMR) system—often identifies this newly introduced "correct" edit as an "error" and diligently reverts it to the original sequence [4], [5]. This tug-of-war between editing and repair significantly diminishes the final efficiency of PE in many cell types and at various genomic loci, severely constraining its path toward clinical applications.

The Evolutionary Path: A Protracted Battle with the MMR System

To overcome the obstacle posed by the MMR system, scientists have undertaken a series of arduous iterative optimizations. This technological trajectory clearly reflects a deepening understanding of the problem and its solutions.

1. Initial Attempts (PE2/PE3): Early PE systems like PE2 laid the foundational framework but generally suffered from low efficiency [2]. The subsequent PE3 system introduced an additional nick on the non-edited strand to bias the cell's repair machinery toward retaining the edited sequence. While this offered a modest improvement, it did not fundamentally resolve the conflict with the MMR system [6].
2. Direct Inhibition (PE4/PE5): Researchers soon realized that they needed to directly "subdue" the MMR system. This led to the development of PE4/PE5 systems, which co-express a dominant-negative mutant of the MLH1 protein (MLH1dn) [4], [7]. As a key component of the MMR pathway, this truncated MLH1dn (753 amino acids) effectively disrupts the function of the MMR complex, leading to a significant boost in PE efficiency [8]. This marked a crucial strategic shift from "guiding" repair to "inhibiting" it. However, this solution introduced a new and more formidable challenge: the large size of MLH1dn. For in vivo gene therapy, commonly used delivery vectors like adeno-associated viruses (AAVs) have a very limited packaging capacity (around 4.7 kb) [9]. The bulky MLH1dn made it difficult to package the entire PE system into a single AAV vector, severely limiting its in vivo therapeutic potential.
3. Parallel Optimizations (PE6/PE7): While contending with the MMR system, scientists also pursued optimizations on other fronts. The PE6 system successfully "slimmed down" the core reverse transcriptase (RT) component through protein engineering, freeing up some space for AAV delivery [8]. Meanwhile, the PE7 system enhanced the stability of the pegRNA template by fusing it with a La protein [10]. Although these improvements were effective, the MMR inhibition component still relied on the cumbersome MLH1dn. The critical challenge became how to create an MMR inhibitor that was both highly potent and compact.

The AI-Designed "Mini-Inhibitor" MLH1-SB

The study from a team at Seoul National University, published in Cell, directly addresses this core conflict [1]. Their objective was clear: to design a compact MMR-inhibiting protein to replace MLH1dn, one small enough for efficient AAV delivery. Instead of traditional protein engineering, they turned to generative AI.

1. The Innovative Solution: An AI-Driven "Design-Screen" Loop

The team employed an elegant and highly efficient AI-driven workflow:

• Target Locking: First, they used AlphaFold 3 to precisely predict the interaction interface between two key MMR proteins, MLH1 and PMS2. This interface is the "Achilles' heel" for the formation and function of the MMR complex [11].
• AI Generation: Next, using the leading protein design model RFdiffusion, they generated nearly 2,800 candidate "mini-protein" scaffolds targeting this interface [12]. These proteins were designed to be extremely small, with the goal of physically blocking the MLH1-PMS2 interaction.
• AI-Powered Virtual Screening: This was the most creative step in the process. The team again utilized AlphaFold 3 to perform "in silico competition assays." They tasked the AI with predicting the structure of a complex containing MLH1, PMS2, and a candidate mini-protein. If the AI predicted that the mini-protein could successfully "push away" PMS2 and disrupt the original complex, that candidate was deemed highly promising [1]. This virtual showdown narrowed the pool to 43 candidates.
• Experimental Validation: Finally, they experimentally tested the ability of these 43 candidates to enhance PE efficiency in cells, ultimately identifying the top performer: MLH1-SB [1].

2. Key Results and Performance Validation

This AI-forged MLH1-SB, at a mere 82 amino acids, is less than one-ninth the size of MLH1dn, perfectly solving the AAV delivery problem. Its performance is equally impressive:

• Massive Efficiency Boost: In HeLa cells, fusing MLH1-SB with the PE7 system resulted in editing efficiencies 18.8 times higher than the classic PEmax system and 2.5 times higher than PE7 alone [1].
• Successful In Vivo Validation: In mouse liver experiments, the new system demonstrated a 2.3 to 3.4-fold increase in efficiency with manageable toxicity [1].
• High Modularity: MLH1-SB acts as a "plug-and-play" enhancement module, compatible with various PE architectures (PE2, PEmax, PE6, PE7) and stackable with other enhancement strategies, like the La protein, for maximum effect [1].

The entire AI design and screening pipeline was completed in just a few days without relying on local supercomputers, showcasing the immense potential of AI to accelerate biological discovery.

A New Paradigm of AI-Enabled Functional Module Design

The significance of this research extends far beyond providing a powerful "add-on" for Prime Editing. It unveils a new research paradigm: using generative AI to design functional biological modules that can upgrade existing complex biological systems in a modular, plug-in fashion.

Previously, optimizing a biological tool often involved a repetitive, trial-and-error process of modifying its core components. Now, we can focus on designing entirely new, compact, and precise "functional plug-ins" to address specific bottlenecks. This approach has broad applicability and could be used in the future to:

• Design inhibitors for other DNA repair pathways to further enhance the efficiency of different types of gene editing tools.
• Create novel protein "glues" or "insulators" to regulate metabolic pathways or signal transduction.
• Develop "protective shields" that stabilize specific RNAs or proteins, improving the efficacy of gene therapies or protein drugs.

To fully realize this paradigm, however, we must accelerate the entire "Design-Build-Test-Learn" (DBTL) cycle. AI has already dramatically sped up the "Design" phase. To match the incredible speed of AI-driven design, the "Build" and "Test" phases of biology must also accelerate. High-throughput vector screening platforms, such as Ailurus Bio's Ailurus vec®, could offer a viable path forward, enabling the parallel testing of thousands of AI-designed variants to rapidly close the DBTL loop.

Of course, the long-term safety of MLH1-SB, particularly the potential risks associated with sustained MMR inhibition in vivo, requires more extensive and thorough evaluation. Nevertheless, this work has opened a window to a future where AI and synthetic biology are deeply intertwined—a future where designing and creating macromolecules with novel functions becomes as efficient and precise as writing code.

References

1. Lee, S., et al. (2025). AI-generated MLH1 small binder improves prime editing efficiency. Cell. https://www.cell.com/cell/fulltext/S0092-8674(24)01246-8
2. Anzalone, A.V., et al. (2019). Search-and-replace genome editing without double-strand breaks or donor DNA. Nature. https://www.nature.com/articles/s41586-019-1711-4
3. Gaudelli, N.M., et al. (2017). Programmable base editing of A•T to G•C in genomic DNA without DNA cleavage. Nature. https://www.nature.com/articles/nature24644
4. Chen, P.J., et al. (2021). Enhanced prime editing systems by manipulating cellular determinants of editing outcomes. Cell. https://www.cell.com/cell/fulltext/S0092-8674(21)01192-5
5. Front Line Genomics. (2022). Improving prime editing by studying DNA repair. https://frontlinegenomics.com/improving-prime-editing-by-studying-dna-repair/
6. Wang, Y., et al. (2023). Current advancement in the application of prime editing. Cell & Bioscience. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9978821/
7. Ferreira da Silva, J., et al. (2022). Prime editing efficiency and fidelity are enhanced in the absence of mismatch repair. Nature Communications. https://www.nature.com/articles/s41467-022-28442-1
8. Lee, J., et al. (2025). Emerging trends in prime editing for precision genome editing. Experimental & Molecular Medicine. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC12322275/
9. Broad Institute. (2023). Evolved prime editors are smaller and more efficient for therapeutic applications. https://www.broadinstitute.org/news/evolved-prime-editors-are-smaller-and-more-efficient-therapeutic-applications
10. Gao, R., et al. (2024). Improving prime editing with an endogenous small RNA-binding protein. Nature. https://www.nature.com/articles/s41586-024-07259-6
11. DeepMind. (2024). AlphaFold 3 predicts the structure and interactions of all of life’s molecules. https://deepmind.google/discover/blog/alphaproteo-generates-novel-proteins-for-biology-and-health-research/
12. Watson, J.L., et al. (2023). De novo design of protein structure and function with RFdiffusion. Nature. https://www.nature.com/articles/s41586-023-06415-8