CRISPR-GPT
Four-agent LLM planner for CRISPR-Cas gene-editing experiments, spanning 22 individual tasks across knockout, base, prime, and epigenetic editing.
| Affiliation | Stanford University (Cong lab), Princeton, UC Berkeley, Google DeepMind (Qu et al.) |
| First introduced | 2024-04 (arXiv v1); peer-reviewed version in Nat. Biomed. Eng. 2026-02 |
| Lifecycle stages | Experiment design, Analysis |
| Autonomy level | Assistive (human-in-the-loop; user-proxy agent operates autonomously but user oversight is encouraged) |
| Domain focus | Biology (CRISPR-Cas gene editing — knockout, base editing, prime editing, epigenetic editing) |
| Availability | Closed (no full code release pending US regulatory clarity on AI in biology); welcome page on GitHub |
Approach
Multi-agent LLM system with four roles:
- Planner — decomposes user requests into a chain of state-machine tasks.
- Task Executor — manages workflow.
- User-Proxy — mediates user interaction.
- Tool Providers — wrap external tools, databases, and web search via APIs.
The system implements 22 individual gene-editing tasks (sgRNA design, off-target prediction, delivery selection, protocol drafting, validation assay design) across Meta, Auto, and QA modes.
Validation
Real-world case study of non-expert researchers using CRISPR-GPT to plan and execute gene-editing experiments from scratch, as reported in the Nature Biomedical Engineering paper.
Notable results
First LLM agent system reported to span the full CRISPR experimental-design workflow across four editing modalities. Demonstrated to help non-experts plan and execute real gene-editing experiments.
Primary paper
Other references
Code
Welcome page — full codebase withheld pending regulatory clarity.