Design of Highly Functional Genome Editors by Modelling CRISPR-Cas Sequences
First fully AI-designed CRISPR gene editor (OpenCRISPR-1) successfully edits human genome — 95% reduction in off-target edits vs SpCas9
Design of Highly Functional Genome Editors by Modelling CRISPR-Cas Sequences
Abstract
Profluent trained protein language models on the CRISPR-Cas Atlas — the largest curated CRISPR dataset assembled to date — to generate novel gene editors entirely from scratch. The resulting system, OpenCRISPR-1, is the first fully AI-designed gene editor (nuclease, guide RNA, and deaminase all generated by AI) to successfully edit the human genome. It matches or exceeds SpCas9 at on-target sites while drastically reducing off-target editing.
Key Contributions
- First fully AI-designed CRISPR system — every component (nuclease, gRNA, deaminase) is generated from scratch by LLMs, not engineered from natural variants.
- Equivalent on-target performance — 55.7% editing efficiency vs SpCas9's 48.3%.
- 95% reduction in off-target editing — 0.32% off-target activity vs SpCas9's 6.1%.
- Highly novel sequence space — OpenCRISPR-1 is 403 mutations away from SpCas9 and 182 mutations away from the nearest natural variant.
- Open release — the sequence was released publicly in April 2024 to enable broad research and commercial use.
Methodology
- Built the CRISPR-Cas Atlas: 5.1M CRISPR-Cas proteins (2.7x expansion of known natural diversity; 4.1x for Cas9 specifically), curated via systematic mining of 26 terabases of assembled genomes and metagenomes.
- Trained protein language models on the atlas, then generated 4M candidate sequences (4.8x diversity expansion vs natural proteins).
- For Cas9 variants specifically, trained on 238,917 natural Cas9 proteins.
- Trained separate models to generate guide RNAs and paired deaminases.
- Applied bioinformatic filtering to classify generated proteins by CRISPR-Cas family.
Results
| Metric | OpenCRISPR-1 | SpCas9 |
|---|---|---|
| On-target editing | 55.7% | 48.3% |
| Off-target editing | 0.32% | 6.1% |
| Mutations from SpCas9 | 403 | — |
| Mutations from nearest natural variant | 182 | — |
- Generalizable platform — the approach is not one-off; Profluent frames this as a proof point that LLMs trained on biological diversity can design functional proteins in many families.
- Democratization angle — open release pairs with Profluent's stated mission to lower cost/barrier for CRISPR therapeutics, agriculture, and research.
Limitations
- The Profluent materials do not emphasize limitations; peer-reviewed Nature paper likely includes discussion of delivery, immunogenicity, and clinical translation barriers that cannot be evaluated from press materials alone.
- AI-designed proteins may encounter novel safety/regulatory scrutiny: no prior clinical history for the specific sequences.
- Public sequence release creates dual-use concerns that the company has acknowledged but not fully resolved.
Full Content
Content summary pulled from Profluent's open press materials because the Nature paper URL sits behind a paywall/cookie wall. Full paper at DOI 10.1038/s41586-025-09298-z. Company also raised $106M in Nov 2025 on the back of this work.
Source: Design of Highly Functional Genome Editors by Modelling CRISPR-Cas Sequences, Ruffolo, Madani et al., Profluent / Nature, July 30 2025. Content extracted from Profluent press materials.