2026-09-25
ESM3, ProteinMPNN and EvoDiff generate 76 NRPS T-domains (578 variants). Best full-length titer is 257% of wild type; AI-2 melts 12°C higher and tracks the downstream partner.
Nonribosomal peptide synthetases (NRPSs) are modular assembly lines behind many clinical peptide antibiotics, immunosuppressants, and anticancer agents. Engineers have long wanted to swap modules and substrates to make new molecules. Yields often collapse, because the line is not a rigid kit of parts. The thiolation (T) domain, also called the peptidyl carrier protein (PCP), cycles between adenylation (A) and condensation (C) partners. Those interfaces are transient and state-dependent. Evolution-guided XUT fusion sites preserve some compatibility, yet chimeras still introduce non-cognate contacts. The question here is whether generative protein models can write new-to-nature T domains that still work inside a specified A–T–C neighborhood.
The pipeline is conditional generation plus design–build–test–learn (DBTL). During sampling the flanking A and C sequences (and structure when available) stay in context, the T segment is masked, and candidates are forced to condition on the target junction. Three generators run in parallel: the sequence model ESM3, the diffusion model EvoDiff, and structure-conditioned ProteinMPNN.
The assay is a type S split of the GameXPeptide synthetase GxpS. Subunit 2 alone reports linear tripeptides for prescreening; co-expression with subunit 1 reports titer of the main cyclic product cyclo(vLflL). Golden Gate swaps candidate T sequences into a sacB cassette. Three DBTL rounds: round 1 hand-picks 11 sequences from ESM3 and ProteinMPNN; round 2 adds EvoDiff plus filters (FFxxGGxS motif, ≥50% identity to wild type, ESM2 perplexity cap); round 3 trains surrogate models on earlier activity labels for active learning. In total, 76 designs and 578 recombinant NRPS variants were tested in vivo, spanning minimal, full-length, repositioned, and hybrid lines.
In the minimal assay, all four round-1 ESM3 designs produced peptide; AI-2 reached about 115% of wild type. Only two of seven ProteinMPNN designs were detectable. On full-length GxpS, all eight high-yield designs made product 6. AI-27 reached 257% of wild type (7.4 mg L⁻¹ in succinate XPP at 20 mL) and AI-38 reached 144% (4.1 mg L⁻¹). Wild type under the same condition is about 2.9 mg L⁻¹; in LB it climbs to about 50 mg L⁻¹ and rankings shift. The abstract’s “up to 3-fold” maps onto the best full-length gain.
Representative design AI-2 expresses more in the soluble fraction, refolds to a native SEC peak, and melts near 52°C versus about 40°C for refolded wild-type T3, a 12°C gap. Molecular dynamics keeps global fold stability but rearranges state-dependent interdomain contacts. The authors treat the MD as hypothesis-generating; they did not mutate contacts to prove causation.
Hybrids lock the interpretation. Designs AI-67–AI-76, generated for an LCL-type downstream C domain, made no product on a C/E-type partner and recovered activity with an LCL-type partner. Sequence identity to wild type barely predicts function. Mean pLDDT shows no significant correlation with titer across 66 designed T domains.
This is a working demonstration that generative design can touch dynamic megasynthases, with AI as the method and a biosynthetic assembly line as the object. For NRPS engineers the operational takeaway is specific: treat downstream partner class as an explicit design variable and condition T-domain generation on the junction, rather than screening chimeras after the fact. Gains such as AI-27’s 2.5-fold titer bump are scaffold- and media-specific, not universal parts. Success is defined as peptide biosynthesis, then binned by relative output. The paper does not claim the generators learned to optimize interfaces.
Function is junction-conditioned. Top designs do not transfer cleanly across hybrid architectures, so each new line may need its own round of optimization. LC–MS labels mix interface effects with expression, growth, and media; changing broth can reorder variants, which injects noise into DBTL. MD contact maps are not mechanistic proof. Biophysical work is concentrated on AI-2, not a large panel. The models are not shown to optimize interfaces explicitly; they sample under a context constraint and sometimes hit functional sequence. Tests stay near GxpS and related hybrids; phylogenetically distant NRPS systems remain untried.