OptiPrime for Prime Editing: How David R. Liu's AI Model Improves pegRNA Design and Gene Therapy Efficiency

AI-Driven Prime Editing Enters the Engineering Optimization Era

On February 20, 2026, the team led by David R. Liu published their latest study, “Mechanistic machine learning enables interpretable and generalizable prediction of prime editing outcomes,” introducing OptiPrime, a mechanistically driven machine learning framework.

OptiPrime shifts pegRNA optimization from empirical trial-and-error screening toward systematic, interpretable engineering design.

Unlike traditional black-box prediction tools such as DeepPrime or PRIDICT, OptiPrime explicitly models the biophysical processes underlying Prime Editing, treating it as a dynamic multi-step kinetic system. This approach enhances both prediction transparency and cross-context generalizability.

OptiPrime: Modeling Prime Editing as a Kinetic System

The OptiPrime architecture is built upon a detailed mechanistic decomposition of Prime Editing. The editing process is divided into multiple stages:
l Target binding
l nCas9 nick formation
l Reverse transcription
l Flap integration
l Heteroduplex formation
l DNA repair pathway resolution

A key innovation is the quantitative modeling of mammalian mismatch repair (MMR) interference. Since MMR frequently restores edited DNA back to wild-type sequences, it significantly impacts editing efficiency.

Figure 1. Design of prime editing (PE) strategies in a high-dimensional search space.

To address this, the team developed HetFormer, a Transformer-based module inspired by AlphaFold’s EvoFormer architecture. HetFormer models interactions between heteroduplex DNA and MutS complexes and was pre-trained on 64 million simulated heteroduplex sequences.

OptiPrime then parameterizes each sub-process using pseudo-rate constants and applies ordinary differential equations (ODEs) to simulate time evolution, ultimately integrating predicted editing yields.

Rather than producing a single predictive score, OptiPrime outputs full kinetic profiles. Researchers can directly identify bottlenecks—such as insufficient reverse transcription rates or excessive MMR off-rates—enabling rational pegRNA redesign.

Performance in Clinically Relevant Models

OptiPrime demonstrates strong performance, particularly in disease-relevant scenarios.

For the cystic fibrosis mutation CFTR p.F508del, associated with Cystic fibrosis, conventional optimization using enhanced epegRNA and PE2 achieved only 11% correction efficiency. In contrast, OptiPrime achieved 22% efficiency by evaluating just eight top-ranked candidate designs—while the top 16 candidates from competing models yielded below 1%.

Figure 2. OptiPrime accelerates the development of corrective strategies for pathogenic mutations in vivo.

In a mouse model targeting the Kif1a “leg dragger” mutation (linked to motor dysfunction), OptiPrime nominated eight silent mutation strategies and identified OP-5 as optimal. By fine-tuning RTT and PBS combinations, only 15 pegRNAs were required to reach 22% correction efficiency in heterozygous MEFs.

With the addition of non-paired sgRNA (nsgRNA) and the PE6 variant, editing efficiency increased to 64%. Delivery via dual AAV9 vectors carrying split PE6b validated successful in vivo correction in mouse cortex and spinal cord tissues four weeks post-injection.

Importantly, the workflow reduced optimization time to modular stages:
l Synthesis: 7–10 days
l Transfection: 3 days
l Sequencing: 1 day

This is substantially more efficient than traditional large-scale screening involving hundreds of pegRNAs.

Industry Implications:
From Empirical Tuning to Mechanistic Engineering

OptiPrime represents not merely an algorithmic upgrade, but a paradigm shift in pegRNA design logic.

Previously, pegRNA optimization relied heavily on empirical adjustments. Mechanistic modeling now enables engineering-style rational design, integrating kinetic simulation and data-driven learning.

However, limitations remain:
l Training data primarily derived from synthetic reporter systems, potentially overlooking endogenous chromatin effects
l Prediction accuracy may require calibration in highly heterogeneous tissues (e.g., tumors or neural systems)
l HetFormer pre-training relies on simulated datasets, which may introduce bias

Future integration of multi-omics datasets (e.g., ATAC-seq, ChIP-seq) and real-time feedback loops could further strengthen predictive robustness.

Overall, this study signals a broader transition of gene editing from artisanal methodology to scientific engineering. In the era of AI-driven genome engineering, OptiPrime serves not only as a tool, but as a computational framework—accelerating therapeutic development for rare diseases such as cystic fibrosis and Kif1a-associated neuropathies.

Upon formal publication and code release, OptiPrime is expected to stimulate open-source development and accelerate translation from diagnosis to therapy worldwide.

Original link: https://www.biorxiv.org/content/10.64898/2026.02.20.706353v1

Follow us on social media