Base and prime editing
Gentler gene editors change single DNA letters or rewrite short stretches without cutting both strands.
Open in the interactive tree →In David Liu's lab at the Broad Institute, base editors (2016) swap one DNA letter for another and prime editors (2019) can search-and-replace short stretches, without making a double-strand break. They are more precise and suit the point mutations behind many genetic diseases. They are the basis of the first personalized gene-editing treatment, given in 2025.
As of October 2026
In February 2025 a team from Penn and the Children's Hospital of Philadelphia gave baby KJ Muldoon, who has a life-threatening liver enzyme (CPS1) deficiency, a custom adenine base editor in lipid nanoparticles; the therapy was built within about six months of diagnosis, he received three doses by April 2025 and left hospital on 2 June 2025 (NEJM, May 2025). In March 2026 the team said it was discussing with the FDA a single umbrella trial of custom prime-editing therapies for seven urea cycle disorders. The next three results are early data from a few patients: Lilly's VERVE-102 (cholesterol) lowered LDL cholesterol by up to 62 percent in a company-reported phase 1 trial (May 2026), Beam's BEAM-302 has company-reported early human data for alpha-1 antitrypsin deficiency, and Prime Medicine's company-sponsored PM359 trial restored immune-cell function in two patients with chronic granulomatous disease (NEJM, December 2025). Only a small number of patients have been treated so far.
Open steps
- Delivery beyond the liver Medium AI leverageLipid nanoparticles mostly reach the liver; getting editors into muscle, brain, lung or blood stem cells safely is the main barrier for most diseases.
- Bystander and off-target edits High AI leverageBase editors change neighbouring letters and prime editors vary by site; predicting and measuring unintended edits must be good enough for approval.
- Smaller, better editors High AI leverageEditors are large and need two-part delivery; AI-designed smaller or PAM-flexible enzymes could fit standard vectors and reach more sites.
- One-patient therapies, made fast Medium AI leverageRoutine custom editors need standard nanoparticle recipes, fast safety assays and release testing that work for any guide, not six months of bespoke work.
Where AI could help
Medium AI leverage. AI improves guide and editor design, but delivery, safety follow-up and tiny patient numbers dominate.
- Predicting prime-editing guide efficiency by edit type and cell state
- Language-model design of new editor proteins and deaminases
- Predicting off-target and bystander edits before a patient is treated
- Faster design of custom guides for one-patient therapies
Shown so far
- In June 2024 Nature Biotechnology published PRIDICT2.0, a machine-learning model that predicts prime-editing guide efficiency for edit types up to 15 bp in different cell types. source
- In July 2025 Nature published OpenCRISPR-1, a gene editor designed with language models trained on CRISPR-Cas sequences, which edited the human genome (company-developed, Profluent). source