Mechanistic machine learning for prediction of prime editing outcomes

In plain English

Rather than predict one editing score as a black box, OptiPrime models steps in the prime-editing process and uses the learned rates to estimate outcomes across PE3 and twin prime editing. The authors report prospective demonstrations in primary human and mouse cells and an in vivo correction of a pathogenic mutation in a mouse model of KIF1A-associated neurological disease.

What they found

• The model learned determinants associated with mammalian mismatch repair and nominated silent edits intended to evade that pathway.
• The authors report prospective editing improvements in primary cells and in vivo correction in mouse brain.

The catch

• The abstract does not provide the full benchmark composition or error distribution.
• Prospective molecular correction in selected models does not establish clinical safety or efficacy.
• Performance outside the tested editors, cell types, and sequence contexts remains uncertain.

Methods at a glance

Study type

Mechanism-based machine-learning model development with prospective experimental validation

Biological system

Prime editing in mammalian cells and mouse brain

Organism

Human cells and mice

Dataset

Prime-editing outcome datasets spanning multiple editing configurations

Sample size

Not reported

Model architecture

Mechanism-based machine-learning model with learned pseudorates for prime-editing steps

Training objective

Predict prime-editing efficiency and outcome distributions

Code

Training and modeling code is available under a noncommercial license at https://github.com/alvin-hsu/optiprime-src; a noncommercial webserver is available at https://optipri.me/.

Data

The article links raw sequencing data in SRA BioProject PRJNA1314411 and processed data in its supplementary information.

Tasks and metrics

Tasks: prime-editing efficiency prediction; PE3 outcome prediction; twin prime-editing outcome prediction; silent-edit design

Metrics: predictive accuracy; measured editing efficiency; editing outcome distribution

Baselines: state-of-the-art prime-editing outcome predictors reported by the authors

Principal findings

• The authors report state-of-the-art predictive accuracy across evaluated prime-editing settings.
• Prospective designs were tested in primary human and mouse cells.
• A selected design corrected a pathogenic mutation in the brain of a KIF1A disease mouse model.

What remains unresolved

External validation: Prospective experiments included primary human and mouse cells and an in vivo mouse-brain demonstration.

• The abstract does not disclose the complete benchmark composition or subgroup performance.
• Experimental demonstrations cover selected editing systems and biological contexts.
• In vivo molecular correction does not establish therapeutic safety or efficacy.