GPT-5 improves a cloning protocol 79-fold

GPT-5 increased the efficiency of a DNA cloning protocol 79-fold in a laboratory, according to an evaluation conducted by OpenAI with biosecurity company Red Queen Bio. The model proposed changes, analyzed the results from each round and adjusted its next recommendations without human guidance on what to test.

The result does not mean GPT-5 can run a laboratory on its own. Scientists still had to carry out the experiments and upload the data. But it does show something important: AI can help improve real biological processes, not just summarize papers or suggest hypotheses.

What GPT-5 did in the lab

The team used a controlled molecular biology system based on Gibson assembly, a standard technique for joining DNA fragments. The goal was straightforward to measure: produce more bacterial colonies from a fixed amount of DNA.

In each round, GPT-5 proposed between eight and ten different reactions. The researchers tested them, compared the results with the reference protocol and sent the most promising data into the next round.

The model eventually proposed a combination of two proteins that, according to OpenAI, had not previously been used as a functional method of this kind:

• RecA, a protein that helps find compatible DNA sequences.
• Gp32, a protein that binds to single-stranded DNA and helps prevent it from tangling.

The combination was named RAPF-HiFi. In initial tests, this new procedure increased efficiency by more than 2.5 times. The total 79-fold increase also included improvements to the transformation phase, the step in which DNA enters bacterial cells.

Where the biggest improvement came from

The most effective modification in the transformation step was less sophisticated than expected. The model proposed concentrating the cells before adding the DNA and removing some of the liquid in which they were suspended. This variant outperformed the reference procedure by 30 times in the final validation.

The full 79-fold improvement was confirmed with experimental replicates. Even so, it is not a universal figure. The result applies to the specific system used in the study, including its reagents, cells and laboratory conditions.

Some changes that appeared to work very well in the first round also had smaller effects when repeated with greater DNA dilution. That highlights a common problem in science: a striking result must hold up in further testing before it can be considered robust.

The model reasoned and learned from the results

The evaluation aimed to measure more than the ability to follow instructions. GPT-5 had to propose modifications, explain possible mechanisms and change its strategy based on the data obtained.

The RecA and Gp32 case is relevant because the model did not simply add two ingredients. It also proposed combining them with different incubation timings, so that each protein acted during a different phase of the reaction. The researchers later found that removing either one reduced performance, suggesting that both contribute to the improvement.

OpenAI describes this as evidence of experimental biological reasoning. However, the work does not show that the model understands biology like a human researcher or that it can generalize the solution to other experiments.

They also tested a laboratory robot

Red Queen Bio and Robot on Rails used a robotic system capable of turning written instructions into physical actions. The robot could move liquids, handle tubes, apply heat and prepare plates, while a camera identified the laboratory equipment.

When running cloning protocols, the robot achieved relative improvements similar to those of the manual experiments. But it produced around ten times fewer colonies in absolute terms, a difference the researchers attribute to factors such as liquid-handling precision, temperature control and small details of manual manipulation.

What this changes for research

For a laboratory, an improvement like this could mean less wasted material, more experiments on the same budget and faster research cycles. In areas such as protein engineering or genetic test design, where many variants are analyzed, improving the efficiency of a basic technique can have a considerable practical impact.

The main limitation is that the system still operates within a supervised loop: the AI proposes, people carry out the work and the data goes back to the model. The researchers also used a benign biological system and a narrowly defined task to reduce biosecurity risks.

The next step will be to determine whether these methods maintain their performance with other protocols and whether models can better balance exploring new ideas with optimizing those that already work. The most important signal from this study is not that AI has replaced the scientist, but that it is starting to become a tool capable of designing and improving physical experiments with help from its own results.