Source-linked AI summary
CRISPR-GPT for Agentic Automation of Gene-editing Experiments
Yuanhao Qu, Kaixuan Huang, Ming Yin, Kanghong Zhan, Dyllan Liu, Di Yin, Henry C. Cousins, William A. Johnson, Xiaotong Wang, Mihir Shah, Russ B. Altman, Denny Zhou, Mengdi Wang, Le Cong
TL;DR
CRISPR gene-editing experiment design requires specialized knowledge, while general-purpose LLMs can generate inaccurate or incomplete biological guidance. CRISPR-GPT combines LLM reasoning with domain knowledge, external tools, and structured task execution to support design workflows, and the paper reports expert evaluation and real-world collaborative validation. Its automatic planner remains bounded by fixed task structures during execution.
Problem
CRISPR experiment design requires specialized, up-to-date biological knowledge, but general-purpose LLMs can hallucinate guide sequences and omit essential experimental details.
Method
CRISPR-GPT combines an LLM agent with domain knowledge, retrieval, external tools, and structured task workflows for system selection, guide design, delivery, protocols, and validation.
Results
CRISPR-GPT supported diverse gene-editing scenarios, automated design tasks, and a real-world experiment in which its guidance covered the workflow from system selection through sequencing validation.
Takeaways & Limitations
CRISPR-GPT demonstrates the potential of domain-augmented LLM agents to assist non-experts with complex CRISPR experiment design through human-agent collaboration.
Takeaways & Limitations
The planner cannot dynamically add or delete tasks or state machines during automatic execution, which the authors leave for future work.
Abstract
from arXiv · showhide
The introduction of genome engineering technology has transformed biomedical research, making it possible to make precise changes to genetic information. However, creating an efficient gene-editing system requires a deep understanding of CRISPR technology, and the complex experimental systems under investigation. While Large Language Models (LLMs) have shown promise in various tasks, they often lack specific knowledge and struggle to accurately solve biological design problems. In this work, we introduce CRISPR-GPT, an LLM agent augmented with domain knowledge and external tools to automate and enhance the design process of CRISPR-based gene-editing experiments. CRISPR-GPT leverages the reasoning ability of LLMs to facilitate the process of selecting CRISPR systems, designing guide RNAs, recommending cellular delivery methods, drafting protocols, and designing validation experiments to confirm editing outcomes. We showcase the potential of CRISPR-GPT for assisting non-expert researchers with gene-editing experiments from scratch and validate the agent's effectiveness in a real-world use case. Furthermore, we explore the ethical and regulatory considerations associated with automated gene-editing design, highlighting the need for responsible and transparent use of these tools. Our work aims to bridge the gap between beginner biological researchers and CRISPR genome engineering techniques, and demonstrate the potential of LLM agents in facilitating complex biological discovery tasks. The published version of this draft is available at https://www.nature.com/articles/s41551-025-01463-z.
1 Introduction
CRISPR experiment design demands specialized biological knowledge, and general-purpose LLMs can produce inaccurate, incomplete, or distracting guidance. CRISPR-GPT addresses this gap by combining LLM reasoning with domain knowledge, computational tools, and structured workflows for gene-editing design.
- 1 Introduction: CRISPR experiments require selecting editing systems, designing guide sequences, and validating outcomes, often demanding substantial domain expertise and trial-and-error.The relevant biology includes guide RNA targeting and Cas-enzyme activity at specific DNA sequences.
- 1 Introduction: General-purpose LLMs lack precise, up-to-date domain knowledge and may hallucinate guide RNA sequences that do not match known genomic regions.Such errors can mislead researchers and waste resources if sequences are not vetted against reference databases.
- 1 Introduction: General-purpose responses often omit practical experimental details, including materials, protocols, off-target considerations, and guide efficiency or specificity.These omissions can leave researchers, especially newcomers, unprepared to execute experiments.
- 1 Introduction: CRISPR-GPT combines an LLM-powered planning agent with expert knowledge, literature, and computational toolkits to automate gene-editing experiment design.Its workflow covers CRISPR-system selection, guide RNA design, delivery approach, off-target prediction, and protocol recommendations.
- 1 Introduction: CRISPR-GPT also provides freestyle Q&A and off-target prediction functions to support users during experimental design.These auxiliary modes address ad hoc questions and analysis of pre-designed guide RNAs.
- 1 Introduction: The system includes safeguards for human applications, including use restrictions, genetic-information privacy measures, and alerts about unintended consequences.The stated safeguards are framed as measures for responsible use of automated gene-editing design.
2 Methods and Algorithms
CRISPR-GPT implements gene-editing design as modular, state-machine-based interaction between an LLM planner, external tools, a task executor, and the user. The system decomposes requests into dependent tasks while retaining user oversight, and it was evaluated through expert ratings and a real-world collaborative experiment.
- System architecture: CRISPR-GPT comprises an LLM planner, tool provider, task executor, and LLM agent that interfaces with users.The tool provider connects external APIs, tools, libraries, and documents to the workflow.
- Task execution: The system implements 22 tasks as state machines, with each state representing a sub-goal and defined transition logic guiding multi-round textual interaction.Four predefined meta-tasks support complete gene-editing pipelines, while customized task chains can be generated for user requests.
- Task decomposition: The LLM planner decomposes user requests into dependent tasks using task descriptions, dependencies, and internal model knowledge.Selected task state machines are chained to fulfill the user’s request.
- Limitations: The planner does not dynamically add or delete new tasks or state machines during automatic execution.The authors identify dynamic task modification as future work, retaining fixed task structures for robustness.
- State-machine interaction: Each interaction state provides instructions and required inputs, receives user responses, may call external APIs, and then transitions to the next state.The LLM agent interprets state information, user requests, interaction history, and tool results to select actions.
- Human oversight: Users monitor the workflow and can correct the LLM agent or interact manually, keeping user input central to experiment planning and design.The paper frames this arrangement as collaboration between human expertise and artificial intelligence.
- Evaluation and validation: The system was evaluated by 12 CRISPR and gene-editing experts and used in a human-agent collaboration for real-world knockout experiments.The reported biological workflow included cell culture, crRNA cloning, genomic DNA extraction, PCR, and sequencing.
3 Results
CRISPR-GPT combines domain knowledge, retrieval, external tools, and interactive modules to support gene-editing experiment design. Expert evaluation and a wet-lab knockout demonstration showed broad assistance across design tasks, while complex and rare cases remained difficult.
- 3 Results: CRISPR-GPT combines LLM reasoning, domain knowledge, retrieval techniques, and external tools to support diverse gene-editing experimental designs.Supported scenarios include single-gene knockout, base editing, prime editing, and epigenetic editing.
- 3.1 CRISPR-GPT assists researchers with gene-editing experimental design through three modules.: The agent provides Meta, Auto, and Q&A modes for guided pipelines, customized task planning, and CRISPR-related questions during experiment design.Meta Mode uses predefined pipelines, Auto Mode decomposes user requests into dependent tasks, and Q&A Mode provides interactive information support.
- 3.1 CRISPR-GPT assists researchers with gene-editing experimental design through three modules.: Meta Mode organizes 22 gene-editing design tasks across four CRISPR-based systems, including system selection, delivery recommendations, sgRNA design, off-target prediction, protocols, and validation planning.The workflow guides users through task-specific decisions using predefined pipelines and external tools.
- 3.2 CRISPR-GPT outperforms general LLMs in gene-editing design tasks through human expert evaluations.: Expert evaluation found significantly higher accuracy for CRISPR-GPT than general LLM agents across MetaMode, AutoMode, and QAMode, with stronger completeness and conciseness patterns overall.General LLM agents performed better on Q&A completeness, whereas CRISPR-GPT was designed to provide concise answers.
- 3.2 CRISPR-GPT outperforms general LLMs in gene-editing design tasks through human expert evaluations.: CRISPR-GPT showed difficulties with more complex gene-editing scenarios and rare biological cases, motivating future updates to domain knowledge and external tool sets.This limitation was reported alongside the overall expert-evaluation results.
- 3.3 CRISPR-GPT demonstrates its efficacy through real-world application.: In a human-AI knockout experiment in A375 cells, CRISPR-GPT supported system selection, delivery, guide design, protocols, culture, transduction, and validation across four target genes.The experiment used Meta Mode, selected AsCas12a and lentiviral transduction, and reported consistently high expected editing outcomes from NGS.
4 Safety and Ethical Concerns
CRISPR-GPT incorporates safeguards for human genome editing and privacy. These include mandatory organism checks, warnings and confirmations for human targets, and filtering of potentially identifiable genome sequences.
- Human genome-editing targets trigger warnings, a link to an international moratorium, and confirmation that users understand the risks and guidelines.
- CRISPR-GPT follows healthcare data-privacy rules and treats genome sequences of up to 20 bp as non-identifying.
- The system does not store identifiable long genome sequences that could reveal patient information.
- A prompt filter detects genome sequences of ≥20 bp and asks users to remove them before sending prompts to external LLMs.
5 Discussion
CRISPR-GPT combines domain knowledge, external tools, and modular workflows to support CRISPR experiment design, including context-specific guidance and free-form interaction. Its current limitation is that it cannot generate complete constructs or vectors from natural-language input.
- CRISPR-GPT supports CRISPR system selection, guide RNA design, delivery planning, protocol drafting, and validation-strategy design.
- Unlike chemical workflows, biological experiments require guidance tailored to living cells and organisms’ dynamic experimental context.
- Free-style prompting and ad hoc Q&A let researchers obtain contextualized guidance for unexpected experimental challenges.
- CRISPR-GPT can design individual guides and primers but cannot yet generate complete constructs or vectors from natural-language input.
- Integration with laboratory automation and robotics is proposed as a future direction for connecting computational design with physical execution.
A Failure Cases of GPT-4
The failure-case section identifies GPT-4 as the language model used in the reported examples.
- The experiments use GPT-4 version gpt-4-0613.
A.1 Examples of inaccurate information
The examples show GPT-4 producing inaccurate or insufficient CRISPR design guidance across base-editor selection, delivery, validation, guide RNA design, and primer design.
- GPT-4 recommends ABEs, CBEs, and prime editing for broad EGFR mutation introduction, but the evaluation notes that it omits expanded-window or non-specific base editors and gives irrelevant recommendations.
- GPT-4 recommends lipofection as the common delivery method for CRISPR knockout systems in HepG2 cells, although the evaluation identifies this as inaccurate for a hard-to-transfect line.
- The EGFR-knockout validation response proposes genotypic and phenotypic assays, while the evaluation specifically flags qRT-PCR as unreliable for assessing single-guide knockout edits.
- GPT-4 supplies EGFR-targeting sgRNA sequences that the evaluation reports do not align with the human EGFR sequence.
- GPT-4 provides TGFBR1 CRISPRa sgRNA sequences and recommends RNA, protein, phenotype, and off-target validation, but the supplied sequences are reported not to align with human TGFBR1.
- GPT-4 proposes EGFR knockout PCR primers from a genomic coordinate, while the evaluation reports that the supplied primer sequences do not align with human TGFBR1 sequences.
A.2 Examples of incomplete information:
The examples show that generic GPT responses to CRISPR knockout requests can omit key design, protocol, and implementation details needed for experimental execution. The documented omissions include CRISPR-system selection, usable sgRNA sequences, cloning and transfection protocols, reagent quantities, and detailed verification procedures.
- The suggested workflow includes gRNA design, cloning gRNA and Cas9 into a plasmid, and plasmid delivery into A549 cells by lipofection.The passage mentions CRISPOR or Benchling for guide design and Lipofectamine 3000 for transfection.
- Generic GPT guidance for knocking out EGFR in A549 cells omits CRISPR-system selection, sgRNA sequences, cloning and transfection protocols, and detailed verification procedures.The response also repeats verification information without providing the needed details.
- The 293FT transfection answer describes cell culture, plasmid preparation, and Lipofectamine-based delivery but does not specify the quantities of cells, plasmids, or reagent.
A.2.1 Example 3:
The example asks how to evaluate CRISPRa activation of human EGFR in 293FT cells. The proposed answer names qRT-PCR and western blot, but the accompanying critique notes that it lacks procedural details for applying either method.
- CRISPRa activation efficiency for human EGFR in 293FT cells is assessed using qRT-PCR for EGFR mRNA and western blot for EGFR protein expression.The proposed interpretation is increased EGFR mRNA or protein expression relative to a control group.
- The evaluation response does not explain in detail how to perform the experiment with each named method.
B Prompt Formats
The prompt formats organize CRISPR-GPT around task decomposition, dependency-aware execution, and JSON-formatted agent responses. They specify task families for knockout, base editing, prime editing, CRISPRa/CRISPRi, and off-target prediction, then provide the agent with state and tool context for each inquiry.
- B.1 The prompt format of the task decomposition.: The task-decomposition prompt instructs the LLM planner to generate executable tasks while respecting dependencies and including prerequisite tasks.The response is required to be parsable by JSON.
- B.1 The prompt format of the task decomposition.: The task description table records implemented tasks and their dependencies for the planner to consult.
- B.1 The prompt format of the task decomposition.: Knockout planning separates Cas-system selection, delivery selection, guideRNA design, protocol selection, and Sanger or NGS validation-primer design.GuideRNA design depends on Cas-system selection, while protocol selection depends on both system and delivery choices.
- B.1 The prompt format of the task decomposition.: Base-editing planning includes base-editor selection, guideRNA design, delivery selection, protocol selection, and Sanger or NGS mutation-sequencing primer design.GuideRNA design follows base-editor selection, and protocol selection follows delivery selection.
- B.1 The prompt format of the task decomposition.: Prime-editing planning includes prime-editor selection, delivery selection, pegRNA design, protocol selection, and Sanger or NGS mutation-sequencing primer design.pegRNA design depends on prime-editor selection, and protocol selection depends on delivery selection.
- B.1 The prompt format of the task decomposition.: CRISPRa/CRISPRi planning covers activation or repression-system selection, delivery selection, guideRNA design, protocol selection, and qPCR-primer design.GuideRNA design follows activation or repression-system selection, while protocol selection follows delivery selection.
- B.1 The prompt format of the task decomposition.: Off-target planning uses CRISPRitz for off-target search or prediction as a standalone task.
- B.2 The format of the prompt for LLM-agent.: The LLM-agent prompt combines system state, interaction history, and external-tool results with the user’s meta request, then requests direct JSON-formatted answers.It instructs the agent to answer directly, select an option for multiple-choice inquiries, and defer gene-sequence requests to manual control.
C The Rubrics for Human Evaluations
The human-evaluation rubric compares CRISPR-GPT with ChatGPT versions across accuracy, reasoning, completeness, and conciseness. The rubric defines ordered performance levels for assessing factual correctness, justification, coverage, and response economy.
- CRISPR-GPT is evaluated against ChatGPT v3.5 and v4 on accuracy, reasoning, completeness, and conciseness.The evaluation uses expert-defined rubrics.
- The rubric also assesses conciseness alongside accuracy, reasoning, and completeness.
- Accuracy ranges from poor answers with multiple factual errors to excellent answers that are accurate and free of substantive errors.
- Reasoning is scored by the clarity, logic, and support provided for claims and suggestions.
- Completeness measures whether an answer supplies the critical information needed for a thorough CRISPR design or action plan.