AI that learns the mechanics of prime editing picks better guides—and fixes a mutation in mouse brains

Prime editing can rewrite DNA with great precision, but finding the right guide RNA often means testing hundreds of candidates. A new machine-learning tool, OptiPrime, cuts that search down dramatically—and, in a mouse model of a neurological disorder, helped correct a disease-causing mutation in the brain with editing efficiencies above 40% in bulk tissue.

Scientists report in Nature Biotechnology. The work, led by David R. Liu at the Broad Institute of Harvard and MIT, with first author Alvin Hsu, is based on a simple idea: a model built around the actual biological steps of prime editing should generalize better than a black-box predictor.

Rather than throwing all sequence features into one opaque network, OptiPrime divides the editing process into mechanistic steps—target binding, flap synthesis, DNA mismatch repair—and learns a “pseudorate” for each step. The team trained it on nearly 300,000 prime editing measurements, including new high-throughput screens of about 75,000 editing outcomes, and then tested it on datasets the model had never seen. OptiPrime outperformed two existing prediction tools, PRIDICT2.0 and DeepPrime, and ablation experiments showed that removing any part of its mechanistic architecture hurt accuracy, confirming the design matters.

Because the model’s internal rates correspond to real biological processes, OptiPrime can also be adapted to predict editing strategies it was never trained on. The team used its learned rates to predict PE3 outcomes (editing with an extra nicking guide) and twinPE outcomes (using two pegRNAs), with useful accuracy.

The practical payoff came in therapeutic settings. OptiPrime-nominated guides outperformed expert-designed and existing-model guides for correcting mutations in CFTR (cystic fibrosis) and COL7A1 (in fibroblasts from patients with recessive dystrophic epidermolysis bullosa), as well as for installing an orthogonal IL-2 receptor variant in primary human T cells. Most striking, the team used OptiPrime to develop a corrective prime editing strategy for the “leg dragger” mouse model of KIF1A-associated neurological disorder in just four weeks and with only 15 pegRNAs. After delivery to newborn mice by AAV, editing in the brain reached above 40% average in bulk cortex and more than 70% in transduced cells.

The study has caveats: OptiPrime was trained largely on synthetic integrated reporter data, so it does not yet incorporate chromatin context, and the mouse results, while promising, are a preclinical demonstration in one disease model. Still, the team’s webserver (https://optipri.me/) makes the tool available to the community, and future iterations may add chromatin features and smarter optimization loops to shrink the search space even further.

Reference: Hsu, A., Chen, P.J., Li, A.H. et al. Mechanistic machine learning for prediction of prime editing outcomes. Nature Biotechnology (2026). https://doi.org/10.1038/s41587-026-03261-7