Clustered regularly interspaced short palindromic repeats (CRISPR; a genetic editing tool adapted from the bacterial immune system) and the discovery of associated Cas systems have initiated a comprehensive transformation in the field of genome engineering. First-generation CRISPR-Cas nucleases (enzymes that cut DNA) have been used to inactivate genes by generating DNA double-strand breaks (DSBs; the simultaneous breakage of both strands of DNA) at target sites; however, these breaks have been shown to lead to undesirable consequences such as large DNA deletions, chromosome losses, and p53 activation. Base editing (the chemical conversion of individual nucleotides without cutting both strands of DNA), prime editing (PE; a method that enables targeted insertions, deletions, and base substitutions in DNA through a reverse transcriptase enzyme), and large DNA integration systems, developed to overcome these limitations, have extended the scope of genome editing methods beyond applications based on DNA cleavage. With the diversification of these tools, experimental design processes such as guide RNA (gRNA; an RNA molecule that directs the Cas enzyme to the target DNA region) selection, evaluation of off-target effects (undesired changes occurring in DNA sequences outside the targeted region), and prediction of DNA repair outcomes have become more complex; this has increased the need for advanced computational approaches. This review, published in Molecular Celland authored by Soochul Shin, Heesoo Uhm, Yong-Woo Kim, Hyeon Woo Im, and Sangsu Bae, . The study discusses developments ranging from early machine learning (ML; methods that make predictions by learning patterns from data) and deep learning (DL; learning methods based on multilayer artificial neural networks) approaches, which initially focused on predicting guide RNA activity, off-target effects, and editing outcomes, to the direct design of new molecular components. The limited protein sequence space accessible through random mutagenesis (the introduction of random mutations into genetic sequences), rational design, and large-scale experimental screening approaches used in traditional protein engineering has encouraged the development of approaches such as protein language models (PLMs; models that learn patterns in protein sequences), generative artificial intelligence (generative AI; models that generate new molecular candidates using learned patterns), and foundation models (models trained on large-scale datasets). In this context, the evolution of artificial intelligence from an auxiliary prediction tool into an approach capable of optimizing protein variants, designing new molecular architectures, and generating de novo (designed from scratch) genome editing proteins has been evaluated in light of the existing literature.
The computational approaches discussed in the review were examined under different categories, including predictive models, evolutionary-scale protein language models, structure-based diffusion models, and genomic foundation models. In the first stage, the applications of machine learning and deep learning models trained on experimental datasets in predicting guide RNA activity, off-target effects, and DNA repair outcomes were evaluated. In this context, predictive models for Cas9, Cas12a, and the RNA-targeting Cas13 nucleases, were also discussed. Algorithms predicting editing outcomes and undesired changes in neighboring bases within the editing window (bystander editing) in cytosine and adenine base editors (CBEs and ABEs), as well as models evaluating the design of pegRNA components such as the primer binding site (PBS) and reverse transcriptase template (RTT) in prime editing systems, were examined.
In the second stage, the EVOLVEpro approach, which combines numerical representations (embeddings) obtained from the ESM2 protein language model with random forest regression and is iteratively updated with new experimental data, and the MULTI-evolve framework, which models functional interactions between mutations (epistasis), were examined.
In the third stage, the review examined structure based design workflows combining RFdiffusion for protein-backbone generation, ProteinMPNN for designing amino acid sequences compatible with the generated backbones, and structure-prediction approaches using AlphaFold 2 or AlphaFold 3. In the MLH1-SB example, structural prediction was followed by an in silico competition analysis to assess whether the designed binder could compete with PMS2 for binding to MLH1.
In the final stage, approaches such as ProGen2, used for protein sequence generation, and the genomic foundation models Evo, Evo 2, and EDEN were discussed. ProGen2 was fine-tuned on extensive CRISPR-Cas sequence datasets to generate diverse Cas9-like protein sequences, from which candidate genome editors were prioritized for further evaluation. Evo and Evo 2 are foundation models for modeling and generating genomic sequences. The article states that Evo 2 was trained on approximately 9 trillion DNA base pairs and has a context window of 1 million tokens. EDEN, on the other hand, was discussed as an approach for designing new recombinases using target DNA sequences. One of the distinguishing features of these approaches is that they are not limited to identifying sequences similar to existing proteins but also enable the design and generation of new molecular candidates. The review describes how the role of artificial intelligence in genome editing has expanded from predicting editing outcomes to supporting multistage molecular design and engineering. Predictive models such as the Rule Set series, DeepCRISPR, inDelphi, BE-Hive, and PRIDICT 2.0 have been developed to address different problems, including guide RNA activity, off-target effects, DNA repair outcomes, base editing outcomes, and prime editing efficiency. Some models also take into account additional variables such as chromatin features and the cellular repair context. It has been stated that protein language model-guided evolution approaches can contribute to the prioritization of protein variants to be experimentally tested and the optimization of the functional properties of genome editing proteins. Structure-based generative design has been used to develop a compact MLH1-binding protein (MLH1-SB) that interferes with the mismatch repair (MMR) machinery and improves prime-editing efficiency. In a separate application, DdCBE-TOD uses a rigid orienting domain in place of a flexible linker to constrain the relative positioning of its editing components and reduce undesired bystander editing. Furthermore, sequence-based generative models and genomic foundation models have been reported to enable the design of de novo nucleases such as OpenCRISPR-1, which differ from natural Cas9 proteins at the sequence level, hyperactive transposons (mega-active PiggyBac), and specific serine recombinases (EDEN-LSR) that enable the integration of large gene fragments. Nevertheless, it has been emphasized that current AI-assisted genome engineering workflows largely rely on human guidance, predefined design objectives, manually curated datasets, computational filtering, and extensive experimental validation. The findings indicate that the design processes of genome editing systems could be improved in the future through the integration of sequence, structure, function, and cellular context information and improvements in model reliability. The authors also emphasize the importance of ethical considerations and appropriate regulatory oversight for the safe and responsible development of AI-assisted genome editing technologies.
Translated by: Ziya Burak Ekinci
Editor: Hazel Birgül
Reference: Shin, S., Uhm, H., Kim, Y.-W., Im, H. W., & Bae, S. (2026). Beyond prediction: The evolving role of artificial intelligence in genome editing. Molecular Cell, 86(20), 4118–4130. https://doi.org/10.1016/j.molcel.2026.08.025
-Bioinfocodes Scientific News Service-
News articles prepared by our team members, reviewing and compiling scientific research
published in journals with and impact factor greater than 20 (click here for the list)