The study focused on so-called non-ribosomal peptide synthetases, or NRPS for short. These large enzyme complexes function like molecular assembly lines. They produce a wide variety of biologically active natural compounds in microorganisms, including substances that have served as the basis for the development of numerous antibiotics, immunosuppressants, and cancer drugs. The problem is that a product can only be formed if the individual components interact precisely at the right moment. This is why NRPS are particularly difficult to modify in a targeted manner.
The researchers combined several pre-trained AI models with experimental test series to develop new so-called thiolation domains (T-domains). These building blocks act like movable transport arms within the NRPS: they pick up molecular intermediates and pass them on to the next processing station. In total, the team used AI to design 76 new T-domains, produced hundreds of different NRPS variants, and tested them in living cells.
The results showed that several of the AI-designed building blocks were functional. Some even enabled the formation of products in artificially combined enzyme systems and increased the yield up to three times compared to natural variants. In addition, at least one of the designed proteins proved to be more soluble and thermally more stable than the natural counterpart. Higher yields are not only important for future production. They also make it easier, even in the early stages of research, to obtain new natural compounds in sufficient quantities to determine their structure and investigate their effects more closely. A key factor was that the AI did not design the new building blocks in isolation, but rather tailored them to their immediate surroundings within the respective enzyme system. This is because whether such a building block functions depends heavily on its ability to interact with its neighboring regions.
“Our results show that artificial intelligence can be used not only to design individual proteins, but also to specifically redesign key building blocks of complex biological production systems,” says Bozhüyük. “AI does not replace experiments, but it helps us select promising candidates more quickly from among many possible variants.”
“The study shows that AI can significantly advance the development of new active compounds based on natural products,” says Prof. Rolf Müller, Scientific Director of HIPS and head of the department “Microbial Natural Products”. “If we can adapt such biosynthetic systems in a more targeted manner in the future, this will open up new possibilities for developing urgently needed drugs more efficiently and thus making them available for clinical applications more quickly.”
Bülbül EF, Bang S, George K, Bianchi G, Raj P, Chung S, Pauline V, Hochstrasser R, Minas HA, Elgaher WAM, Kany AM, Hirsch AKH, Schmitt S, Heinz DW, Kalinina OV, Klakow D, Bozhüyük KAJ.
Generative AI designs functional thiolation domains for reprogramming non-ribosomal peptide synthetases.
Nat Commun. 2026 Sep 22;17(1):10084. doi: 10.1038/s41467-026-77963-6