AI Joins The CRISPR Chat: AI Gene Editing Revolution!

Do you want to chat with your experiments? Well, I guess now you can, kind of! Researchers introduced CRISPR-GPT, a chat-based AI collaborator for gene editing!

AI Joins the CRISPR Chat

CRISPR: Powerful, but Tricky

These days, it feels like CRISPR-Cas is everywhere!

From research to therapy, gene editing is transforming biological research and medicine. And nowadays, there are even advanced versions:

• CRISPRa/i (activation and interference)

• Base editing

• Prime editing

CRISPR is now a powerful and versatile tool for genetic modification. And it’s getting out of the lab! We’re starting to see real-world impact, with the first approved therapies (for sickle cell disease and beta-thalassemia) and plant engineering for sustainable agriculture.

But starting a gene editing experiment is no easy feat, even with software and tools available! There is no solution to help you from idea to result. This often means putting together system selection, gRNA design, off-target evaluation, delivery method, and data analysis, all on your own.

An end-to-end solution would make CRISPR gene editing even more accessible and enable discoveries!

AI to the Rescue

Enter today’s paper. The authors created CRISPR-GPT, a large language model (LLM)-based tool for human-AI collaboration to design and execute CRISPR experiments. From start to finish!

But why LLMs?

LLMs are the base of chatbots such as ChatGPT, Claude, and more. They have great language skills, contain a ton of knowledge, and can be connected to external tools to make them more efficient and increase their problem-solving capabilities.

In science, they are starting to help solve problems. For example, in chemistry:

• ChemCrow uses LLMs augmented with tools to solve chemistry-related tasks, such as the synthesis of paracetamol

• Coscientist integrates LLMs and automation to synthesize compounds (Google has a virtual collaborator to brainstorm research ideas, also called “AI co-scientist. I found it funny!)

And similar tools have been applied to math and coding (I think we all tried this!).

But biological tasks bring new problems! Standard LLMs don’t have in-depth knowledge of biology, and bio experiments are challenging in unique ways:

• Living systems are variable

• Biological data are noisy!

• Plus, biological tools are highly specialized, and skills in one might not translate to another.

So, the authors decided to bring the power of LLMs to gene editing!

Building an AI Co-Pilot

CRISPR-GPT covers 4 gene editing modalities (CRISPR knockout, base/prime editing, and CRISPRa/i) with 22 experimental tasks. It has 3 user modes:

• Meta mode: For researchers without gene editing experience. This is an end-to-end protocol, with the user always in the loop.

• Auto mode: For more advanced researchers. There is no predefined task order, and you can request what you need.

• Q&A mode: You can ask your gene editing questions!

The system is built as four cooperating LLM agents. An agent is an AI system using LLMs to perform tasks autonomously. In this case:

• LLM Planner: Creates the entire workflow as a chain of smaller tasks from the user’s request. An example of a task chain: Cas selection → Delivery method → gRNA design → Validation → Data analysis.

• Task Executor: Executes the task decisions: pick Cas, select delivery, design gRNAs, etc. It also calls external tools and creates full pipelines!

• LLM User-proxy: Talks with the human user, asking clarifying questions, or fills in missing inputs and integrates results from tools.

• Tools Providers: Wrapped tool interfaces for Google search, Scholar, bioinformatics tools, etc.

In short, you have the Planner producing workflows, the Executor running them, Tools providing computation and literature, and the User-proxy mediating human input.

The idea is to have reproducible automation with human oversight.

Key Modules: From Idea to Reality

The authors created and tested the core modules using their newly made Gene-editing bench benchmark.

• gRNA design: One of the hardest parts of working with CRISPR. CRISPR-GPT uses precomputed tables and LLM reasoning to suggest the right guide for the experiment.

• Delivery selection: Another critical step for CRISPR success. The agent classifies the biological system (cell lines, primary cells, in vivo, etc.), searches the literature, and ranks candidate delivery methods.

• Q&A and domain fine-tuning: The authors tried to make the Q&A mode “think” like a scientist. They fine-tuned a LLM model on 11 years of an open CRISPR Google Group to teach it “how scientists discuss troubleshooting”. This fine-tuned model outperformed baseline LLMs!

Real-World Demonstrations

But all of this would be useless without real-world proof. So, the authors ran two real lab demonstrations:

1. Quadruple gene knockout in lung adenocarcinoma cells: A junior researcher (with no prior gene editing experience) used CRISPR-GPT to design and execute lentiviral CRISPR knockouts of 4 genes involved in tumour survival and metastasis. CRISPR-GPT proposed the enzyme, the lentiviral delivery, the gRNA design, the cloning protocol, and the NGS validation pipeline. NGS analysis (also done by CRISPR-GPT) showed high editing efficiencies, and functional assays confirmed biologically relevant phenotypes.

2. CRISPR activation in melanoma cells: Another inexperienced junior researcher used CRISPR-GPT to guide CRISPR activation of 2 genes involved in cancer immunotherapy. The system designed gRNAs, delivery, and validation steps. Functional assays showed high activation for the genes.

And both worked on the first attempt!

Safety, Ethics, and Limitations

Here is where it gets tricky. CRISPR brings obvious security and ethical risks: human germline editing, pathogen misuse, bio-weapons, maybe even more.

So, the authors included safety layers. You always have to declare the organism to edit, and get a warning for human-related work. Plus, you get automatically stopped for germline edit or dangerous pathogens. For privacy, they prevent external LLMs from receiving genomic sequences longer than 20 nt, that could be used to identify people.

There are still many limitations to LLM systems in biology:

• Data dependence: The systems needs high-quality domain data: protocols, curated literature, and forum threads. Collecting these data is expensive, time-consuming and far from perfect.

• Coverage: Many resources are data-limited, for example, the gRNA table only supports human and mouse targets.

• Hallucinations/rare cases: Human oversight is needed to avoid hallucinations or mistakes with rare biological cases or complex setups.

Impact and the Future of Work

CRISPR-GPT is a (powerful) prototype for AI-assisted biology. The authors see this as a way to accelerate reproducible experiment design, democratize gene editing for less experienced users, and reduce trial-and-error in wet labs. And there is huge potential to couple CRISPR-GPT with robots to create truly automated labs!

Systems like these open lots of questions for the future of the scientist role. We always had powerful tools at our disposal, but these are on a different level. They are not going to “replace” scientists; that’s silly, but what is going to happen?

Will we have smaller labs because there is less need for people? Bigger labs/universities because the AI “collaborators” are super expensive? Will money be used more wisely, because AI is more efficient, or less, because AI is biased? Is AI going to simply become a new colleague in a growing pie, or the opposite?

Lots of questions, and I have no answers! Just something I’ve been thinking about.

But in the meantime, CRISPR-GPT seems like a great tool! Go here, read the article, and try it out!

But if you made it this far, thank you! Are you ready to use CRISPR-GPT? Have you tried to do gene editing before? What was your experience? Reply and let me know!

More Room:

• Cryo-EM Goes Small: Don’t worry, the machines are still huge. But cryo-EM has trouble imaging small proteins (<100 kDa). This study tackles that. It presents Di-Gembodies, modular nanobody-based scaffolds that enhance cryo-EM imaging of small proteins. By covalently dimerizing nanobodies, they stabilize and display one or two proteins, improving particle alignment and resolution. Validated on targets as small as 14 kDa, this method offers a flexible and scalable platform for protein structure determination.

• AFM+DNA Origami+AI: Yep, finally more AI for DNA origami! This study introduces AFMSPA, an AI-assisted tool for automated, high-throughput AFM analysis of DNA origami nanostructures. Using a neural network with transfer learning, AFMSPA automatically detects, classifies, and measures structural features from raw AFM images with >95% accuracy. Initially supporting four 2D shapes, it can be easily adapted to other designs, providing a scalable and efficient solution for statistical analysis in DNA nanotechnology.

• Water-Fearing DNA Origami: DNA generally loves water, but hydrophobic interactions could be used with normal Watson-Crick base pairing. This study introduces a hydrophobicity-guided strategy for the programmable assembly of DNA nanostructures. Researchers designed small hydrophobic molecules conjugated to single-stranded DNA and integrated them into specific sites on DNA tiles and origami structures. This enables controlled hierarchical assembly driven by hydrophobic interactions to create complex superstructures.

• Share Plenty of Room with founders or builders

  I help biotech and deep tech companies transform complex technologies into engaging content that builds credibility with investors, partners, and potential hires. Let’s chat!

• Know someone who’d love this?
  Pass it on! Sharing is the easiest way to support the newsletter and spark new ideas in your circle.

• Got a tip, paper, or topic you want me to cover?
  I’d love to hear from you! Just reply to this email or reach out on social.