GPT-5 Achieves 79-Fold Efficiency Boost in Molecular Cloning Through Autonomous Lab Iterations

OpenAI's latest research indicates that its GPT-5 model has autonomously optimized molecular cloning processes, leading to a 79-fold increase in experimental efficiency. The AI system, operating with "zero decision intervention," iterated through five rounds of experimentation to develop a unique solution dubbed RAPF. This development marks AI's entry into physical laboratory work beyond simulation, with all data verified through sequencing and replicated on automated robotic platforms.

Optimizing Molecular Cloning

The 79-fold efficiency improvement primarily stems from GPT-5's optimization of two critical steps in molecular cloning: assembly and transformation. Previous efforts often addressed these steps in isolation, but GPT-5 developed an integrated approach.

For the transformation step, GPT-5 introduced the T7 optimization method. This method challenges conventional laboratory practices by proposing a more aggressive cell handling technique. Typically, competent cells used for DNA uptake are handled delicately to prevent damage. However, GPT-5's approach involved centrifuging cells, removing half the liquid to concentrate them, and then adding DNA. This method significantly increased the collision frequency between cells and DNA, resulting in a more than 30-fold improvement in transformation efficiency during verification. This outcome suggests that the AI's strategy was based on fundamental physical probabilities rather than traditional experimental intuition.

In the enzymatic reaction phase, GPT-5 designed the RAPF-HiFi scheme. This scheme integrates two proteins, gp32 and RecA, which are not commonly used together in cloning protocols. The gp32 protein functions to straighten tangled single-stranded DNA, reducing interference, while the RecA protein guides homologous searches to precisely align matching DNA fragments. This combined action led to an approximately 2.6-fold increase in assembly efficiency compared to the baseline HiFi method. The core mechanism involves gp32 stabilizing and straightening single-stranded DNA, followed by RecA driving homologous search and pairing. Auxiliary proteins are then detached through heating, with polymerase filling gaps and ligase sealing nicks to complete the assembly.

The combination of these two optimizations—a 2.6x improvement in assembly and a 30.4x improvement in transformation—resulted in a multiplicative end-to-end efficiency gain of 79 times. This enhancement not only boosts efficiency but also alters experimental strategy, potentially allowing for successful cloning in a single experiment where dozens might have been needed previously.

AI's Iterative Evolution

GPT-5's solution was not derived from existing databases but evolved through five rounds of "experiment-feedback-optimization" cycles. Early theoretical solutions often failed in real-world experiments. However, GPT-5 demonstrated strong online learning capabilities. Researchers provided the model with data from failed experiments, which it used to analyze reasons for failure and adjust its approach in subsequent rounds.

By the third and fourth rounds, GPT-5 began to exhibit advanced task planning. It autonomously proposed the introduction of new tools, specifically the gp32 and RecA proteins. The model also specified the precise order and temperature switching points for these proteins: initial reaction at 50°C, cooling to 37°C for RecA activity, and then reheating to 50°C to deactivate auxiliary proteins before polymerase completes the final steps. This level of detail indicates advanced multi-dimensional logical reasoning.

Throughout these five iterative rounds, human scientists were responsible only for execution and data recording, without influencing the solution design or directional judgment. This suggests that the key mechanisms and path selections originated from the model's own reasoning, rather than human guidance. Within a controlled experimental framework, the AI independently proposed, tested, and corrected new experimental paths, exploring optimization routes not explicitly provided by humans.

Bridging AI and Physical Action

GPT-5's intelligence was translated into physical action through an automated system driven by Robot on Rails. To eliminate human "feel" from the experimental process and verify the AI solution's universality, OpenAI utilized robots. GPT-5 generated experimental protocols in natural language, which were then converted into robotic action commands via a "human-to-robot" large model intermediary layer. For example, a command like "centrifuge at 3000g for 5 minutes at 4°C" enabled the robot to autonomously locate the centrifuge, balance tubes, and execute the action.

The system incorporates advanced visual recognition to navigate dynamic laboratory environments, allowing robots to locate well plates and centrifuge tubes in real-time. This enables accurate aspiration even if reagents are slightly misplaced. Experimental results showed that robots executing GPT-5's optimized R8 scheme achieved consistent relative improvement trends compared to human expert operations. While absolute colony counts were lower for robots due to hardware limitations, the robots replicated the human performance trend in terms of optimization improvement factors relative to the baseline scheme.

OpenAI's report also addresses biosafety, emphasizing that the experiment was strictly confined to "benign systems" under controlled conditions. This caution underscores the potential power of AI to accelerate scientific discovery while also amplifying risks. The ability of AI to independently complete the full closed loop from hypothesis to verification and optimization suggests that future scientific discovery rates may be less constrained by human physical limits or experimental intuition. This shift could free scientists from repetitive tasks, allowing them to focus on more foundational research questions.