Loading…
Prime editing efficiency prediction from pegRNA sequence, with every biochemical step of the editing mechanism modeled as its own learned rate.
No providers recorded yet. Browse all providers
Correcting the CFTR p.F508del mutation by prime editing took the Liu lab several years of iterative pegRNA optimization to reach 11% efficiency. The reason is combinatorial: every prime editing guide RNA carries a spacer, a reverse transcriptase template (RTT), a primer-binding site (PBS), and optionally a set of silent edits that shield the editing intermediate from mismatch repair (MMR), and each choice has dozens of plausible settings. Existing predictors compress that space into a single black-box regressor trained per experimental context, which leaves them good at rejecting bad pegRNAs but weak at ranking a field of already-decent ones.
OptiPrime instead writes the prime editing mechanism into the model's mathematical structure. Six genomic states, spanning the nicked target, 3′ flap, heteroduplex, and MutS-bound intermediates, are connected by transitions whose effective rates ("pseudorates") are each predicted by a separate small model that sees only the features biologically relevant to that step: prime-editor binding and nicking, flap synthesis, MMR activity, and heteroduplex resolution. Those rates populate a generator matrix whose system of ordinary differential equations is integrated forward to the experiment's endpoint. Because experimental covariates such as PE2 versus PE4 enter at the step they actually act on, one model can be trained jointly across many labs' datasets rather than fine-tuned per context.
OptiPrime was developed in David Liu's group at the Broad Institute, with collaborating groups contributing the primary-cell and mouse work, and published in Nature Biotechnology in August 2026 after a February 2026 preprint. In the catalog it sits beside crispAIPE, which predicts the same quantity from a single conventional network with calibrated uncertainty rather than a mechanistic decomposition.
The heteroduplex repair and MutS off-rate models share the "HetFormer", a transformer inspired by the Evoformer of AlphaFold 2 and pretrained on 64 million simulated heteroduplexes; binding, synthesis, and MMR rates come from a feature-driven MLP, a synthesis model, and a linear model. Training used 297,962 PE efficiencies across 40 experimental contexts, combining published screens with two new lentiviral pegRNA–target-site libraries: Lib-MMR (10,000 pegRNAs over 200 exonic sites) and Lib-CV (10,406 pegRNAs correcting 944 ClinVar pathogenic variants), each screened with PE2 and PE4 in HEK293T and HeLa cells for 74,769 efficiencies over 1,290 target sites. Under fivefold cross-validation stratified by protospacer, OptiPrime reached mean r = 0.693, ρ = 0.745, against ρ = 0.590 for PRIDICT2.0-K562 and ρ = 0.399 for DeepPrime-FT. The JAX/Flax implementation ships as a fixed five-model ensemble whose predictions are averaged, with the weights committed to the source repository under a noncommercial license; the screening data are deposited in the SRA.
The model is aimed at teams that need a working prime editing strategy quickly and can only synthesize a handful of guides. Assaying the eight top-ranked designs doubled CFTR p.F508del correction to 22%; three synthesized guides reached 37% correction of COL7A1 p.R185X in fibroblasts from individuals with recessive dystrophic epidermolysis bullosa, where six guides from other predictors stayed below 5%. In a mouse model of KIF1A-associated neurological disorder, 15 pegRNAs across two rounds identified a lead that, with a nicking guide and the PE6b editor, gave 64% correction in fibroblasts and above 40% editing in bulk brain cortex after dual-AAV9 delivery.
OptiPrime's broader claim is methodological: encoding a researcher-specified reaction mechanism as the model's computational graph buys generalization a monolithic regressor does not get, and makes intermediate quantities interpretable enough to test against experiment. The cost is that the model inherits the mechanism's gaps — protospacer quality still enters through a Cas9 nuclease score, which is where predictions fail most visibly, and chromatin context is absent because training data came from randomly integrated synthetic reporters. Distribution is narrow too: the weights exist only as files in a git repository under a noncommercial license, with no versioned checkpoint release.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.