A preprint describes 25 rounds using PRAXIS, a system that designs and tests proteins autonomously

A preprint describes 25 rounds using PRAXIS, a system that designs and tests proteins autonomously

In a preprint dated 14 August 2026, the authors describe PRAXIS. Software agents select protein variants, and a robotic laboratory assembles the DNA encoding those variants and measures the activity of the resulting proteins. The results of each experiment update a shared computational model, which the agents use to select the next batch of variants.

Protein engineering involves developing proteins with specific properties, which requires deciding which protein variant to test next. The number of possible variants is too large to test them all. PRAXIS structures this choice as a series of experiments in which each result informs the next experiment.

The authors tested the system on GH1 family glycoside hydrolases, enzymes that break down sugar-containing molecules. They divided the sequences of six natural GH1 enzymes into eight segments and recombined them while preserving the overall protein structure. These segments can produce about 1,7 million chimeras, which are proteins assembled from fragments of the original enzymes.

Three agents searched for variants with activity and selectivity toward glucose, xylose, or mannose. Each agent selected its next batch independently, but the results of every experiment entered a shared dataset and updated a single model. An experiment targeting one sugar could therefore change the variants selected next by the other agents.

Over approximately one month, the system completed 25 rounds. In each round, the laboratory assembled the DNA of the selected variants, used it to produce proteins, and measured their activity with fluorescent substrates, which are molecules that generate a measurable light signal. Each measurement determined the next experiment.

The best variants were then retested manually alongside the original enzymes. The authors obtained several variants with altered selectivity toward the target sugars. The three most active variants shared a sequence fragment. Subsequent testing showed that these proteins had higher expression, meaning that the laboratory produced more protein.