PaperlayerView exact revision on arXiv

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

arXiv:2404.18021v2cs.AIcs.CLcs.HCq-bio.QM

TL;DR

CRISPR experiment design requires domain expertise, and general-purpose language models lack the precise, up-to-date knowledge needed for accurate biological design. CRISPR-GPT combines language models with domain knowledge and computational tools to support CRISPR experiment design. In a real-world application, sequencing showed consistently high rates of expected editing outcomes across four targeted genes.

  • Problem

    CRISPR experiment design requires domain expertise and trial-and-error, while general-purpose LLMs lack the precise, up-to-date knowledge needed for accurate biological experiment design.

    Source: §1 · p. 3, §1.1 · p. 3

  • Method

    CRISPR-GPT combines LLMs with domain-specific knowledge and computational tools in a modular agent to support CRISPR experiment design.

    Source: §1.2 · p. 4, §2.1 · p. 5

  • Results

    NGS showed consistently high rates of expected editing outcomes across all four targeted genes in the real-world application.

    Source: §3.3 · p. 15

  • Takeaways & Limitations

    The reported use case shows CRISPR-GPT providing computational guidance alongside researcher expertise during gene-editing experiment execution.

    Source: §3.3 · p. 15

    CRISPR-GPT cannot dynamically add or delete tasks during automatic execution, a capability the authors leave for future work.

    Source: LLM-planner automatically generates a… · p. 6

Abstract

from arXiv · show

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 combines specialized biological choices with practical validation requirements, while general-purpose LLMs can give inaccurate or incomplete guidance; CRISPR-GPT addresses this with a domain-informed agent and computational tools for designing experiments.

1 Introduction

  • CRISPR design requires selecting suitable systems and guides and planning validation, demanding domain expertise and knowledge of target-organ biology.CRISPR-Cas editing uses a guide RNA to recognize a target DNA sequence and a Cas enzyme to cut it.Source: §1 · pp. 2–3

1.1 General-purpose LLMs do not know how to design biological experiments.

  • General-purpose LLMs can hallucinate invalid gRNA sequences and omit practical design details, including materials, protocols, off-target effects, efficiency, and specificity.Their responses may also overwhelm users with irrelevant information, making practical guidance harder to identify.Source: §1.1 · p. 3

1.2 Overview of CRISPR-GPT

  • CRISPR-GPT combines an LLM agent, expert knowledge, literature, and computational tools to guide users through CRISPR experiment design.Its workflow covers system selection, gRNA design, delivery, off-target assessment, protocols, validation, and primer design.Source: §1.2 · p. 4
  • The agent also offers freestyle Q&A and off-target prediction for pre-designed guides, with safeguards addressing human-subject use, genetic-data privacy, and unintended consequences.Its stepwise design process uses chain-of-thought reasoning and state machines to support iterative refinement by novice users.Source: §1.2 · p. 4

2 Methods and Algorithms

CRISPR-GPT combines an LLM planner and agent with state-machine task execution and tool access to guide gene-editing workflows. Its methods include expert rating of design responses and human-agent wet-lab validation.

2.1 Large Language Model

  • CRISPR-GPT has four core modules: an LLM planner, Tool provider, Task executor, and user-facing LLM Agent.The modules are identified in Figure 2.Source: §2.1 · p. 5

Task Executor operates as state machines, providing robust subgoal decomposition and progress control.

  • The Task Executor implements 22 tasks as state machines that decompose work into sub-goals and use defined transitions to track progress.Four predefined Meta-Tasks support complete gene-editing pipelines; task state machines can be chained, including tasks customized by the planner.Source: Task Executor operates as state… · pp. 5–6

Tool Provider connects Task Executor with external APIs

  • The Tool provider wraps external APIs in task-specific instructions and responses, connecting the executor to web search, Primer3, and external data resources.Resources include guide RNA libraries, research papers, and experimental protocols; the system guides human and LLM users in using tools [46] [49] [23] [37].Source: Tool Provider connects Task Executor… · p. 6

LLM-planner automatically generates a list of tasks based on the user’s request

  • Using ReAct prompting [53], the LLM planner decomposes user requests into dependency-respecting task sequences and chains their state machines.It draws on LLM knowledge and written task descriptions and dependencies; dynamically adding or deleting tasks during automatic execution is not allowed and is left for future work.Source: LLM-planner automatically generates a… · p. 6

LLM-Agent automatically interacts with the Task Executor based on the user’s meta request

  • The LLM Agent interprets each state and selects the next action using its instructions, the user’s request, interaction history, and external-tool results.Users monitor the process and can correct errors, override the agent, or interact with the Task Executor themselves.Source: LLM-Agent automatically interacts with… · pp. 6–8

2.2 Human evaluation

  • Twelve CRISPR and gene-editing experts rated responses from three modes on a 1-to-5 scale, comparing them with ChatGPT 3.5 and ChatGPT 4.0 (gpt-4-0613).The same criteria were used for all modes and comparison outputs; evaluation rubrics are detailed in Appendix C.Source: §2.2 · p. 8

2.3 Biological experiment and wet-lab validation

  • An independent scientist unfamiliar with gene-editing experiments used CRISPR-GPT with the ChatGPTv4 API to facilitate knockout gene-editing experiments in a cancer research project.In A375 cells, researchers cloned four crRNAs targeting TGFBR1, SNAI1, BAX, and BCL2L1 into a Cas12a-expressing backbone, established stable lines by lentiviral transduction, and sequenced targeted loci after seven days.Source: §2.3 · p. 8

3 Results

CRISPR-GPT combines LLM reasoning, domain knowledge, retrieval, and external tools to support diverse gene-editing scenarios through three interactive modes. Expert evaluations and a real-world knockout experiment assess its design assistance.

3 Results

  • CRISPR-GPT supports knockout, base editing without double-strand breaks, prime editing for insertions, deletions, or replacements, and CRISPRa/i epigenetic editing.The system combines LLM reasoning, domain knowledge, retrieval techniques, and external tools for these gene-editing scenarios.Source: §3 · pp. 8–9

3.1 CRISPR-GPT assists researchers with gene-editing experimental design through three modules.

  • CRISPR-GPT’s three modules provide predefined pipelines, request-driven task planning, and CRISPR-focused question answering during experimental design.Meta Mode supports newcomers, Auto Mode adapts to user requests, and Q&A Mode answers queries throughout the process.Source: §3.1 · p. 9

3.1.1 Meta Mode

  • Meta Mode guides users through 22 design tasks spanning four CRISPR-based systems, from system and delivery selection through guide design, protocols, and validation.It uses predefined pipelines, published protocols, a multi-species guide database, off-target instructions, and external tools such as Primer3.Source: §3.1.1 · p. 10, Table 1 · p. 11

3.1.2 Auto Mode

  • Auto Mode uses an LLM planner to break requests into dependent tasks, autofill relevant fields from the request, and explain its design choices.It supports 13 unique gene-editing design tasks without relying on predefined meta-task pipelines.Source: §3.1.2 · p. 10

3.1.3 Q&A Mode

  • Q&A Mode uses a knowledge base and expert-selected document retrieval to answer CRISPR and gene-editing questions during Meta Mode and Auto Mode.Users can, for example, ask for information about a selected system such as Cas12a.Source: §3.1.3 · p. 10

3.2 CRISPR-GPT outperforms general LLMs in gene-editing design tasks through human expert evaluations.

  • Expert evaluations found significantly improved overall performance over general LLM agents, with higher accuracy across all three modes and stronger conciseness.CRISPR-GPT also scored better on completeness in design tasks, but ChatGPT 3.5 and 4.0 scored higher on completeness in Q&A mode, reflecting a completeness–conciseness tradeoff.Source: §3.2 · pp. 12–13, Figure 6 · p. 13

3.3 CRISPR-GPT demonstrates its efficacy through real-world application.

  • Consistent high rates of expected editing outcomes across four targeted genes demonstrated CRISPR-GPT’s assistance in a human A375 knockout experiment.The workflow used Meta Mode, AsCas12a and lentiviral delivery, and followed agent-provided protocols through guide design and NGS validation; the researchers also noted ethical considerations raised by CRISPR-GPT.Source: §3.3 · pp. 13–15

4 Safety and Ethical Concerns

The section identifies safety and ethical risks from AI-guided genome editing and describes safeguards against human heritable editing and genome-data exposure.

4.1 Mitigation of the risk of human heritable editing

  • For human-organ targets, CRISPR-GPT prevents bypassing organism selection and presents a warning, moratorium link, and read-and-understand confirmation before design proceeds.The safeguard addresses concerns about germline and embryo editing, including enhancement of normal traits; the paper notes such editing is illegal in the United States and many other countries.Source: §4.1 · p. 15

4.2 Protection of user genome data privacy

  • To protect genome privacy, CRISPR-GPT follows HIPAA, stores no identifiable long genome sequences, and flags prompts containing ≥20 bp of A/T/G/C/U for user deletion before external LLM transmission.The paper states that DNA segments up to 20 bp are considered safe and not able to identify a person.Source: §4.2 · pp. 15–16

5 Discussion

CRISPR-GPT combines LLMs, domain knowledge, external tools, and modular task execution to support CRISPR experiment design, while offering interactive Q&A and step-by-step guidance. Its current inability to generate complete constructs or vectors points to future development, including integration with laboratory automation.

Discussion

  • CRISPR-GPT supports CRISPR system selection, guide RNA design, protocol drafting, and validation planning through integrated domain knowledge, tools, and modular task execution.The paper says this workflow can accelerate design, mitigate errors and oversights, and enhance research quality and reproducibility.Source: §5 · p. 16
  • Unlike agents focused on well-defined chemical protocols, CRISPR-GPT provides step-by-step guidance tailored to biological experiments with living cells and organisms.The discussion highlights the need to account for the dynamic nature of living systems.Source: §5 · p. 16
  • Researchers can use free-style prompts and ad hoc Q&A to get contextualized guidance and adapt their approach to unexpected experimental challenges.These capabilities enable a more natural and intuitive interaction with CRISPR-GPT.Source: §5 · p. 16
  • CRISPR-GPT can design individual components such as guide RNAs and primers, but cannot yet generate complete constructs or vectors from natural-language input.The authors suggest integrating modular gene-editing design advances such as FragMID as a future direction.Source: §5 · p. 16
  • Integration with automated laboratory platforms and robotics could connect computational design to physical execution, enabling end-to-end automated experiments with less manual intervention.The authors anticipate this could accelerate discovery.Source: §5 · p. 16

A Failure Cases of GPT-4

The study uses GPT-4 version gpt-4-0613.

A Failure Cases of GPT-4

A.1 Examples of inaccurate information

Examples 1–6 illustrate inaccuracies in GPT-4’s CRISPR guidance, including omitted or unsuitable recommendations and sequences reported not to match the stated human gene targets.

Example 1:

  • GPT-4 omitted the specific expanded-window and non-specific base editors requested for broad EGFR mutagenesis, instead discussing ABEs, CBEs, and prime editing.The passage identifies that discussion as irrelevant to the requested editors.Source: Example 1 · p. 21

Example 2:

  • For hard-to-transfect HepG2 cells, GPT-4 inaccurately recommended lipofection as the top delivery choice for CRISPR knockout.The example notes that HepG2 is a well-known hard-to-transfect cell line.Source: Example 2 · p. 21

Example 3:

  • GPT-4 recommended qRT-PCR to assess EGFR knockout, although the example identifies it as unreliable for edits generated with a single guide RNA.The response also proposed genotypic, protein, and functional assays, but the stated error concerns qRT-PCR as knockout validation.Source: Example 3 · p. 22

Example 4:

  • The EGFR-targeting sgRNA sequences GPT-4 supplied did not align to human EGFR sequences.Source: Example 4 · p. 22

Example 5:

  • The TGFBR1-targeting sgRNA sequences GPT-4 supplied did not align to human TGFBR1 sequences.Source: Example 5 · p. 23

Example 6:

  • GPT-4 proposed primers using an EGFR target location, while the example’s note says the primers did not align to human TGFBR1 sequences.The note names TGFBR1 although the prompt asks for primers to verify EGFR knockout.Source: Example 6 · p. 23

A.2 Examples of incomplete information:

These examples illustrate that GPT-4's generic CRISPR guidance left key design and execution details unspecified, from selecting the system and providing sgRNAs to reagent quantities.

Example 1:

  • For EGFR knockout in A549 cells, GPT-4 outlined generic gRNA design, cloning, lipofection, indel and protein verification, functional assays, controls, and replication, but omitted system selection, sgRNA sequences, and actionable cloning/transfection protocols.Source: Example 1 · pp. 23–24

Example 2:

  • For CRISPR knockout transfection in 293FT cells, GPT-4 gave generic Lipofectamine steps and timing but omitted the cell, plasmid, and reagent quantities needed to execute the experiment.Source: Example 2 · p. 24

A.2.1 Example 3:

For evaluating CRISPRa activation of human EGFR in 293FT cells, the response proposes expression, functional, transcriptomic, and off-target assays, while noting that it does not explain how to perform them.

A.2.1 Example 3:

  • The response recommends qRT-PCR and Western blot to measure EGFR mRNA and protein increases, respectively, and advises confirming activation with at least two independent methods.It also calls for appropriate controls, replicates, and proper normalization of gene-expression measurements.Source: §A.2.1 · p. 25
  • Flow cytometry or immunofluorescence can quantify or visualize EGFR protein at single-cell resolution, while cellular assays can assess functional outcomes such as proliferation or wound healing.Source: §A.2.1 · p. 25
  • RNA-seq can assess EGFR overexpression and downstream transcriptome-wide effects, while whole-genome sequencing, Digenome-seq, GUIDE-seq, or CIRCLE-seq can evaluate potential off-target activation.The answer does not provide detailed instructions for performing any of the recommended methods.Source: §A.2.1 · p. 25

B Prompt Formats

The prompts direct the planner to break user requests into dependent CRISPR tasks and require the agent to make decisions from the current interaction state. They also prescribe structured responses and a manual handoff when users request gene sequences.

B.1 The prompt format of the task decomposition.

  • The planner decomposes requests using the task-description table, respects task dependencies, and returns a JSON-formatted task list.It is instructed to consider whether the user can jump directly to a task or must complete prerequisite tasks first.Source: §B.1 · pp. 25–26
  • Knockout tasks include Cas-system and delivery selection, guideRNA design and protocol selection, plus primer design for Sanger or NGS sequencing.GuideRNA design depends on Cas-system selection; protocol selection depends on both Cas-system and delivery selection.Source: §B.1 · p. 26
  • Base-editing tasks cover editor selection, guideRNA design, delivery, protocol selection, and primer design for Sanger or NGS sequencing.GuideRNA design depends on editor selection, while protocol selection depends on delivery selection.Source: §B.1 · p. 26
  • Prime-editing tasks cover system and delivery selection, pegRNA design, protocol selection, and primer design for Sanger or NGS sequencing.pegRNA design depends on prime-editing system selection, and protocol selection depends on delivery selection.Source: §B.1 · p. 27
  • CRISPRa/CRISPRi tasks include activation or repression system selection, delivery selection, guideRNA design, protocol selection, and qPCR primer design.GuideRNA design depends on system selection, and protocol selection depends on delivery selection.Source: §B.1 · p. 27

B.2 The format of the prompt for LLM-agent.

  • The agent prompt combines the user’s meta request with the current state, interaction history, and potentially external-tool or library results to guide decisions.It asks the agent to answer the tool’s current inquiry on the user’s behalf in JSON format.Source: §B.2 · p. 27
  • The agent is told not to ask follow-up questions, to choose directly for multiple-choice inquiries, and to answer “I don’t know” to gene-sequence requests so the user can take manual control.Responses follow a JSON structure with “Thoughts” and “Answer” fields.Source: §B.2 · p. 28

C The Rubrics for Human Evaluations

The rubric evaluates CRISPR-GPT against ChatGPT v3.5 and v4 on accuracy, reasoning, completeness, and conciseness.

Accuracy

  • Accuracy scores whether answers correctly reflect CRISPR technology and research, from multiple factual errors to complete accuracy.Intermediate scores distinguish significant inaccuracies that could flaw experimental design from minor errors or negligible oversights.Source: C The Rubrics for Human Evaluations · p. 28

Reasoning

  • Reasoning scores whether answers provide clear, logical, and well-supported rationales, from flawed or absent logic to insightful explanations.The rubric also distinguishes weak rationales from generally solid reasoning that could be better supported.Source: C The Rubrics for Human Evaluations · pp. 28–29

Completeness

  • Completeness scores whether answers cover the information needed for understanding and action, from missing critical details to a full experimental-design suite.Intermediate ratings reflect coverage of necessary points, with gaps in important or nuanced aspects of CRISPR design.Source: C The Rubrics for Human Evaluations · p. 29

Conciseness

  • Conciseness scores whether answers communicate relevant information efficiently, from verbose and extraneous to concise and directly relevant.Intermediate ratings allow some unnecessary detail while requiring the answer to remain clear and understandable.Source: C The Rubrics for Human Evaluations · p. 29

Related papers

Loading 2404.18021v2…