GPT-5 improves cloning and accelerates biological research | Keryc
GPT-5 is starting to show that artificial intelligence can do more than summarize papers: it can propose concrete changes to lab protocols that scientists can test. Does that sound like science fiction to you? Not so much.
In a controlled experiment, the model suggested improvements that boosted the efficiency of a molecular cloning protocol by 79-fold, and it did so working with real experimental data.
What they did
A team at OpenAI partnered with the startup Red Queen Bio to see how a frontier model, GPT-5, behaves when asked to optimize a wet lab experiment. They picked a simple, well-defined system: cloning two fragments using the green fluorescent protein gene (GFP) and the plasmid pUC19, aiming to increase the number of sequence-verified colonies.
The rules were clear: the model proposed changes, scientists ran the experiments and uploaded results. Prompting was fixed, with minimal human intervention, so improvements could be attributed directly to the model’s ability to reason and propose new ideas.
What they found
After several iterative rounds, GPT-5 achieved a cumulative improvement in cloning efficiency of 79x compared to the initial HiFi/Gibson-based protocol. What does 79x mean? With the same input DNA amount, they got 79 times more verified clones.
Improvements came from two fronts:
A new enzymatic procedure proposed by the model, called RecA-Assisted Pair-and-Finish HiFi Assembly (RAPF-HiFi), which incorporates two proteins: RecA (E. coli recombinase) and gp32 (T4 phage single-stranded DNA-binding protein). That combo helped DNA ends find and pair more effectively.
A simple but effective tweak to the transformation step: concentrate the cells (spin down, remove half the volume and resuspend at 4°C) before adding DNA, which increased transformation efficiency by more than 30x in the final validation.
How RAPF-HiFi works (explained simply)
The core idea of RAPF-HiFi is to help DNA pieces find the right partner before the usual enzymes seal them. In plain terms:
gp32 acts by cleaning and untangling single-stranded DNA tails, preventing secondary structures that block pairing.
RecA searches for and pairs complementary strands, guiding the correct alignment.
Then the temperature is raised so both helpers come off and the assembly enzymes (polymerase and ligase) finish the job.
Control experiments showed that removing either protein reduced the benefit, suggesting both are needed for the proposed mechanism.
Experimental process and robotics
To scale and speed testing, they built a robotic system (Robot on Rails) that turns natural-language protocols into robotic actions, recognizes materials with computer vision, and plans movements. The robot ran the standard protocol and an AI-modified version in parallel.
The robot reproduced the relative improvements seen by humans, but produced absolute colony counts about 10 times lower, revealing clear areas for improvement: liquid handling, temperature control, and the subtle manual care used with competent cells.
What limitations and precautions exist
Not everything is ready for general use. The authors emphasize that:
Results come from a benign, highly controlled system; effects may be specific to the experimental setup.
The experiment was designed with explicit constraints and biosecurity reviews. The goal was to measure capabilities, not to deploy a method without oversight.
Although prompting was fixed to measure the model’s autonomy, that choice limited fine optimization of promising ideas. A better balance between exploration and exploitation could yield larger gains.
These tests also help evaluate risks: if advanced models can propose practical changes to lab methods, they must be accompanied by safeguards and preparedness frameworks.
What does this tell us about AI in science?
First: AI can contribute practical, mechanical ideas, not just literature summaries. Here it suggested an unusual enzyme combination (RecA + gp32) and simple operational tweaks that worked. Surprised that something obvious to a scientist might be new? Sometimes the obvious appears when you explore combinations outside the usual human routine.
Second: human-AI interaction remains essential. Scientists ran experiments, interpreted results and kept safety controls. AI speeds iteration, but doesn’t replace people.
Third: automation can multiply pace, though it needs fine-tuning. Robots did well on relative comparisons, but there’s still craft in the lab that’s hard to replicate.
Final reflection
This work doesn’t announce an instant revolution, but it points to a practical direction: models like GPT-5 can propose and learn from real experiments, helping reduce time and cost in research. At the same time, those advances call for responsibility: risk assessments, experimental limits and technical and operational safeguards.
Science advances by small, cumulative iterations. Here we see an iteration that came from an AI: useful, concrete and open to improvement. Can you imagine the pace of discovery with AI working shoulder to shoulder with well-regulated researchers and robots? The possibility is underway, but it must be handled carefully.
