Protein engineering has long been constrained by a brutal arithmetic: the space of possible protein sequences is astronomically large, functional solutions are rare, and laboratory experiments are expensive. For a specific class of enzymes called non-ribosomal peptide synthetases (NRPSs), those constraints are even more severe. These molecular machines produce many of the antibiotics, immunosuppressants, and anticancer agents used in clinical medicine today. A study published in Nature Communications now demonstrates that generative AI models can design entirely new protein domains for these enzymes from scratch, with some designs outperforming the natural versions they were built to replace.
The Enzyme That Refuses to Be Engineered
NRPSs are not ordinary enzymes. They operate as biological assembly lines, with multiple domains working in sequence to build complex peptide molecules step by step. Each domain has a specific job: selecting a building block, carrying an intermediate, forming a chemical bond. The problem is that these domains do not work in isolation. They communicate through transient, state-dependent interfaces, meaning the contacts between domains shift depending on what stage of the catalytic cycle the enzyme is in.
This dynamic quality is precisely what makes NRPSs so difficult to engineer. Swapping one domain for another, even a structurally similar one, often disrupts the delicate coordination between partners. Activity drops, or disappears entirely. Evolutionary analysis reinforces this picture: natural recombination events in NRPSs tend to preserve specific interdomain contacts rather than enabling free domain exchange. The enzyme’s function is embedded not just in individual domains but in the relationships between them.
Previous engineering frameworks, including an approach called eXchange Unit Thiolation (XUT) that defines fusion sites designed to maintain structural compatibility, have enabled some functional chimeras. Yet even with such guidance, engineered junctions frequently compromise activity. The core difficulty is that productive interdomain communication cannot be read from sequence patterns alone. It depends on context-specific interface features and the dynamics of the whole assembly.
What Generative AI Did Differently
The research team integrated three pretrained generative models: ESM3, a sequence-based foundation model; EvoDiff, a diffusion-based approach; and ProteinMPNN, which conditions design on protein structure. Rather than using these tools in isolation, the team coupled them with iterative design-build-test-learn (DBTL) cycles, using experimental results from each round to guide the next.
The target was the thiolation (T) domain, the carrier component of the NRPS module that shuttles chemical intermediates between catalytic partners. T-domains are small but occupy a critical position: they must interact productively with both upstream and downstream domains across multiple catalytic states. Designing one that works in a non-native context is, by the field’s own assessment, a demanding problem.
The team generated 76 de novo T-domain sequences and tested them across 578 recombinant NRPS variants in living cells, spanning minimal constructs, full-length assembly lines, and hybrid architectures. The results were notable. AI-designed T-domains supported peptide biosynthesis across all tested architectures. The best-performing designs increased product yields by up to approximately threefold compared to NRPSs carrying the native T-domain.
One representative design was characterized in biochemical detail. It showed improved soluble expression, better refolding behavior, and a melting temperature 12 degrees Celsius higher than the native domain, indicating substantially greater thermal stability. Molecular dynamics simulations of this design revealed that its global fold was preserved, but its interdomain contact networks were reshaped in a state-dependent manner. In other words, the AI-designed domain did not simply mimic the natural one. It found a different solution that still satisfied the functional requirements of the interface environment.
The team also found that functional compatibility was strongly context-dependent, particularly sensitive to the identity of the downstream partner domain. This is not a limitation of the approach so much as a confirmation of what makes NRPSs hard: the T-domain does not operate in a vacuum, and designs that work in one architectural context may not transfer directly to another.
Why This Matters Beyond the Laboratory
The significance of this work extends in two directions. The first is practical. NRPSs produce compounds that are difficult or impossible to synthesize by conventional chemistry. Engineering them to produce new variants, or to produce known compounds more efficiently, has been a long-standing goal in biotechnology and drug discovery. Demonstrating that generative AI can design functional carrier domains that outperform natural sequences is a meaningful step toward that goal.
The second direction is conceptual. This study addresses a question that has been open in the field: whether generative protein design can handle highly dynamic, multidomain systems where function depends on coordinated conformational changes rather than static structural features. The answer, at least for this class of enzymes, appears to be yes, provided the design process is coupled with systematic experimental feedback and guided by an understanding of the interface constraints involved.
AI here does not replace the biochemist’s judgment about what matters in an enzyme. It augments the team’s capacity to explore sequence space at a scale and speed that would be impractical through conventional methods alone.
In Short
Researchers used three generative AI models, combined with iterative experimental testing, to design 76 new thiolation domains for non-ribosomal peptide synthetases. Across 578 tested variants, AI-designed domains supported biosynthesis in multiple architectures, with the best designs increasing product yields by up to approximately threefold over the native sequence. One design showed a 12-degree Celsius improvement in thermal stability. The study establishes that generative design can navigate the dynamic, context-dependent interface constraints that have historically made NRPS engineering so difficult, opening a more systematic route to reprogramming these clinically important biosynthetic machines.
Based on reporting from Nature: Machine Learning.