Structures and function of a tailoring oxidase in complex with a nonribosomal peptide synthetase module

Nonribosomal peptide synthetases (NRPSs) are large modular enzymes that synthesize secondary metabolites and natural product therapeutics. Most NRPS biosynthetic pathways include an NRPS and additional proteins that introduce chemical modifications before, during or after assembly-line synthesis. The bacillamide biosynthetic pathway is a common, three-protein system, with a decarboxylase that prepares an NRPS substrate, an NRPS, and an oxidase. Here, the pathway is reconstituted in vitro. The oxidase is shown to perform dehydrogenation of the thiazoline in the peptide intermediate while it is covalently attached to the NRPS, as the penultimate step in bacillamide D synthesis. Structural analysis of the oxidase reveals a dimeric, two-lobed architecture with a remnant RiPP recognition element and a dramatic wrapping loop. The oxidase forms a stable complex with the NRPS and dimerizes it. We visualized co-complexes of the oxidase bound to the elongation module of the NRPS using X-ray crystallography and cryo-EM. The three active sites (for adenylation, condensation/cyclization, and oxidation) form an elegant arc to facilitate substrate delivery. The structures enabled a proof-of-principle bioengineering experiment in which the BmdC oxidase domain is embedded into the NRPS. Nonribosomal peptide synthetases work with additional enzymes to synthesise secondary metabolites and therapeutics. Here, the authors explore bacillamide D synthesis and show the oxidase action is done while the intermediate is attached to the synthetase and replicate this with an oxidase bound synthetase for bioengineering applications.

N onribosomal peptides are a large group of natural products, bioactive secondary metabolites that are often useful to society as therapeutics and green chemicals 1,2 . They include important medicines such as immunosuppressants (rapamycin), antivirals (cyclosporin A), antibiotics (daptomycin), antifungals (caspofungin), and antitumors (bleomycin) [3][4][5] . Despite their modest size of~2-18 residues, nonribosomal peptides exhibit an impressive range of bioactivities because they can occupy a relatively large volume of chemical space. Nonribosomal peptides can contain many residues not typically found in proteins, such as D-, methylated, halogenated, and other nonproteogenic aminoacyl residues, aryl acyl residues, fatty acyl residues, and hydroxy acyl residues. Furthermore, nonribosomal peptides often contain heterocycles or are macrocyclic or branched. These topologies provide advantages such as pre-organization for binding targets or protease resistance [6][7][8][9][10] . The varied, powerful bioactivities and interesting chemical structures of nonribosomal peptides have led to a large number of studies on the total synthesis of the peptides and on biosynthesis and bioengineering of the enzymes which make them 11 .
Nonribosomal peptides are made in microbes by elegant biosynthetic megaenzymes called nonribosomal peptide synthetases (NRPSs) 12,13 . NRPSs are organized as assembly lines of repeating sets of domains. Each set of domains, known as a module, is responsible for adding one acyl monomer substrate, typically an amino acid residue, to the growing peptide chain 1 . A minimal NRPS elongation module contains three domains: the adenylation (A) domain selectively binds the monomer substrate, activates it by adenylation, and transfers it as an aminoacyl thioester to the prosthetic phosphopantetheine (ppant) moiety on the peptidyl carrier protein (PCP) domain. The PCP domain transports the covalently bound amino acid to the condensation (C) domain. The C domain catalyzes peptide bond formation between the aminoacyl residue on this PCP and the peptidyl moiety on the upstream PCP n-1 domain, elongating the peptide chain. The PCP domain, now with the newly elongated nascent peptide, translocates to the downstream module, where the peptide is further elongated and passed downstream in the next condensation reaction.
Variations and additions to basic NRPS biosynthesis contribute to the diversity of nonribosomal peptides. These variations can occur before, during, and after NRPS assembly-line synthesis. First, NRPS substrates can be produced from cellular metabolites by dedicated, pre-assembly-line synthetic enzymes 14,15 . Second, tailoring can occur during NRPS assembly-line synthesis, as an additional step in the catalytic cycle of a module. This can be encoded into the NRPS enzyme itself: the NRPS module can contain an alternative tailoring domain such as the heterocyclization (Cy) domain, which replaces the C domain and performs both condensation and heterocyclization [16][17][18] , or the NRPS module can include an additional tailoring domain such as a reductase, methylation, monooxygenases, oxidase or formylation domain, which acts on a biosynthetic intermediate tethered to the PCP domain [19][20][21] . Tailoring during NRPS assembly-line synthesis can also be catalyzed by separately encoded enzymes: Enzymes including halogenases, oxidases, hydroxylase, and P450s [22][23][24][25][26] can interact non-covalently with NRPSs and act on PCP-tethered intermediates. Third, after the nonribosomal peptide is released from the NRPS, one or more separately encoded enzymes can catalyze additional reactions to yield the final natural product 27 . Tailoring after release from the NRPS occurs in chloroeremomycin, vancomycin, and penicillin biosynthesis [28][29][30] .
A full biosynthetic pathway for a nonribosomal peptide can thus involve as few as one enzyme (an NRPS), or may require NRPSs as well as other enzymes for pre-, co-, and/or postassembly-line synthetic steps. Genes for such enzymes are often found together with the NRPS genes in the producer microbe, in a biosynthetic gene cluster (BGC) 31 . However, from the sequence of a BGC it is difficult to predict whether a separately encoded enzyme will act on an acyl substrate before assembly-line synthesis, on a PCP-tethered intermediate of NRPS assemblyline synthesis, or on a small molecule after release from the NRPS. The natures of these possible substrates are quite different, and an enzyme will act on only one of these species 24,32 .
The bacillamide biosynthetic pathway (Fig. 1a) is a threeprotein BGC that contains genes for BmdA, tryptophan decarboxylase which converts L-Trp into tryptamine (Tpm) 33 ; BmdB, an NRPS with domains A 1 -PCP 1 -Cy 2 -A 2 -PCP 2 -C 3 that include alternative tailoring domain Cy 2 , which installs a thiazoline ring in the peptide intermediate 16 ; and BmdC, which had not been studied but can be recognized by its sequence as an oxidase enzyme in the nitro-FMN reductase superfamily 34,35 . Mature bacillamides have all been reported to contain thiazole rings [36][37][38][39] , and thus BmdC appeared likely to oxidize the thiazoline ring to the thiazole seen in bacillamide D (Fig. 1a), but there is no indication of the timing of this oxidation.
Here, we describe studies on bacillamide synthesis that reveal unexpected biochemical aspects of the system, including stable dimerization induced by a tailoring protein, and an elegant NRPS-tailoring protein structure, previously undescribed for any nonribosomal peptide synthesis.

Results
In vitro reconstitution of the bacillamide D biosynthetic pathway. The bacillamide BGC is one of the best-represented pathways in genomic databases. Database searches return >400 clusters encoding putative proteins with high sequence identity to BmdA, BmdB, and BmdC, including~350 clusters which show >65% protein identity. The clusters are mostly found in Bacilli, though other Bacillaceae are represented, such as Laceyella and Thermoactinomyces species. Yuwen et al. showed B. atrophaeus BmdA is a pyridoxal-5'-phosphate (PLP)-dependent tryptophan decarboxylase 33 , and we showed Thermoactinomyces vulgaris F-5595 BmdB produces pro-bacillamide (AlaCys thiazoline Tpm) in NRPS assembly-line synthesis 16 (Fig. 1).
To perform in vitro reconstitution of the full bacillamide D biosynthetic pathway, we heterologously expressed T. vulgaris BmdA, BmdB and BmdC in E. coli individually and purified them to homogeneity ( Supplementary Fig. 1a). Reactions containing BmdA, BmdB and BmdC, PLP, ATP, alanine, cysteine, and tryptophan displayed robust production of a species of [M + H + ] = 315.14 ( Fig. 1b). This mass and comparison to synthetic standard confirmed this compound as AlaCys thiazole Tpm, i.e., bacillamide D (compound 1) (Supplementary Fig. 1b), indicating that the full bacillamide D pathway was reconstituted in vitro. Reactions which contained tryptamine in place of BmdA, PLP and tryptophan also showed robust bacillamide D production ( Supplementary Fig. 1e), indicating that BmdA can fulfill its role simply by producing tryptamine, and that interaction of BmdA with BmdB or BmdC is not required. Removal of BmdC from the reaction conditions gave a species with [M + H + ] = 317.15, identified as AlaCys thiazoline Tpm (pro-bacillamide, compound 2) 16 ( Supplementary Fig. 1f-h). These in vitro reconstitutions confirm BmdC as the oxidase involved in the bacillamide biosynthetic pathway (Fig. 1).
The structure of the oxidase BmdC. We performed X-ray crystallography of BmdC to gain a better understanding of its structure and function. Highly purified BmdC was subjected to sparse array crystallization. Initial yellow-colored crystals were obtained, consistent with the presence of a flavin cofactor. Iterative optimization gave crystals that yielded high-quality diffraction datasets (Supplementary Table 1). Molecular replacement using a nitro-FMN reductase domain from Anabaena variabilis (PDB ID 3eo7, unpublished) as a search model allowed structure determination at 2.7 Å resolution (Supplementary Table 1, PDB:7ly6). A single copy of BmdC is present in the asymmetric unit ( Fig. 2 and Supplementary Fig. 2), and the adjacent symmetrically related copy completes the biological dimer. Nitro-FMN reductase domain protein typically exists as dimers or pseudodimers 40 , and size-exclusion chromatography experiments reveal that BmdC is dimeric in solution (Supplementary Fig. 2e).
The structure of BmdC shows it to be a dimer of di-domain protomers ( Fig. 2 and Supplementary Fig. 2). The small N-terminal domain contains 96 residues and features a winged helix-turn-helix motif (wHTH) with a very small β-sheet. A DALI 41 structural similarity search reveals this domain to be similar to the precursor peptide recognition elements (RRE) in ribosomally synthesized post-translationally modified peptide (RiPP) processing enzymes with Z-scores ranging from 5.3 to 6.5 and rmsd values of 2.0-2.9 Å over 61-69 alpha carbons ( Fig. 2b) [42][43][44][45][46] . These RREs are often found in RiPP pathways for the production of linear azole-containing peptides, azolecontaining cyanobactins and thiopeptides 43 . The larger C-terminal (Ox) domain is comprised of residues 142-325 and folds into a six-stranded β-sheet sandwiched between α-helices on both sides, similarly to the flavodoxin-like fold in other nitro-FMN reductases 47 . BmdC contains a 45-residue loop segment between the RRE-like (RREL) and Ox domains, that wraps around the Ox domain of the other protomer. The loop contributes 1764.3 Å 2 of the total 5777.3 Å 2 per-protomer surface area buried by BmdC dimerization 48,49 . These very high values of buried surface area are consistent with our observation that BmdC exists as a dimer in solution. The overall structure of BmdC is most similar to ThcOx (PDB: 5lq4), an FMN-dependent oxidase involved in the production of the cyanobactin patellamide 50 , though ThcOx contains additional copies of the RRE and does not contain the wrapping loop ( Supplementary  Fig. 2c). McbC (PDB: 6grh), an oxidase involved in microcin B17 biosynthesis 42 , does not contain an N-terminal domain, but does have a similarly dramatic wrapping loop (Fig. 2c).
The electron density maps showed very strong density for flavin mononucleotide (FMN) (Fig. 2d). Because the cofactor was co-purified with BmdC, we confirmed its identity through heat denaturation followed by mass spectrometry, which showed its [M + H + ] of 457.1, characteristic of FMN and ruling out FAD ( Supplementary Fig. 2d). As with other dimeric flavodoxins, BmdC binds FMN in an active site between its protomers (Fig. 2e) Arg146 and Ser147) and the catalytic isoalloxazine rings (hydrogen bonding of Arg149 with FMN N1 and O2, Asn236 with N3 and O4, plus Gly294 with N5; analogous to interactions observed in flavin reductase P 51 ). The second protomer contributes backbone interactions through residues 182-186. The isoalloxazine ring system appears planar, but ultra-highresolution would be needed to definitively show planarity 51 . The conserved active-site Tyr260 points toward the FMN in each protomer. The analogous tyrosine in indigoidine synthetase has been proposed as a general base for oxidation, and BmdC is likely to use the same catalytic mechanism 52 . There is room for the thiazole ring to bind between this Tyr260 residue and the catalytic FMN N5 atom; a glycine from the crystallization buffer is observed occupying this site.
BmdC acts during BmdB assembly-line synthesis and induces BmdB dimerization. Inspection of the bacillamide BGC and of the BmdC structure does not clearly indicate whether oxidation by BmdC occurs during assembly-line synthesis or after small molecule release from the NRPS. A reaction mixture of 1 μM BmdC and 2 mM pro-bacillamide showed no production of bacillamide D (Fig. 3a), suggesting BmdC does not act on free pro-bacillamide. To test directly whether the oxidation occurs with pro-bacillamide or with AlaCys thiazoline -S-BmdB as the substrate for BmdC, we designed dialysis experiments. Using dialysis membrane which allows small molecules but not proteins to pass through, we saw that bacillamide D production occurs only when BmdB and BmdC are found on the same side of the dialysis membrane, not when they are separated (Fig. 3a-c). When BmdB and BmdC are separated, only pro-bacillamide is detected. These results suggest that BmdC acts during NRPS assembly-line synthesis and show that it must physically interact with BmdB for its function.
To investigate the interaction between the oxidase BmdC and the NRPS BmdB, we performed size-exclusion chromatography titration experiments. Mixing increasing amounts of BmdC with fixed amounts of BmdB led to a distinct peak shift corresponding to a BmdB-C complex (Fig. 3d). Interestingly, the BmdB-C complex elutes from the column substantially earlier than expected for a dimer of BmdC binding to a single NRPS. Rather, it corresponds tõ 650 kDa, consistent with a complex comprised of two molecules of BmdB and a dimer of BmdC (605 kDa). This dimerization is unusual, as NRPS systems usually act as monomers 53 .
Although dimerization in NRPSs is rare, polyketide (PKS) systems are typically dimers [54][55][56][57] . Complementation experiments, where each protomer of a dimer has a different inactivating mutation (for example in the conserved serines of the carrier proteins (CP) or the active-site cysteines in ketosynthase domains), have been used to show that the growing polyketide passed from one protomer to the other during synthesis [58][59][60] . We created mutants of BmdB that prevent pantetheinylation on PCP 1 (BmdB_S792A) and PCP 2 (BmdB_S1820A), respectively (Supplementary Fig. 3a, b). Neither BmdB_S792A nor BmdB_S1820A produced pro-bacillamide. Crucially, mixing BmdC with BmdB_S792A and BmdB_S1820A did not lead to substantial bacillamide production either, indicating that the two BmdB mutants could not complement each other in the context of a BmdC-induced dimer, and that it is unlikely the nascent peptide is passed between protomers.
We next compared rates of peptide synthesis of probacillamide and bacillamide D by monomeric and dimeric BmdB(-BmdC). The reconstitution experiments described above show similar total production of pro-bacillamide by BmdB and bacillamide D by BmdB-BmdC, but that BmdB-BmdC proceeds with around twice the initial rate (4.6 min −1 vs 2.5 min −1 ; Supplementary Fig. 3a-f). The rate difference is likely because of a preference for a thiazole intermediate over a thiazoline intermediate for condensation with tryptamine at the C3 domain of BmdB. To discern more directly whether dimerization and/or oxidation has impact on the overall rate of synthesis, we created an FMN-free BmdC. BmdC double mutant R146E S292F does not co-purify with FMN, but still induced dimerization of BmdB. Peptide synthesis reactions showed that BmdC_R146E-S292Finduced dimeric BmdB produces pro-bacillamide at the same rate as (monomeric) BmdB alone ( Supplementary Fig. 3f). Therefore, dimerization does not appear to impart a catalytic advantage to bacillamide synthetase.
To identify the domain of BmdB to which BmdC binds, we made a series of BmdB truncation constructs and tested their ability to bind BmdC by gel filtration. This showed BmdC binds the A domain of the second module of BmdB (BmdB M2 ) ( Fig. 3e and Supplementary Fig. 4). We confirmed this interaction by isothermal titration calorimetry (Fig. 3f).
Structural investigations of the BmdB M2 -BmdC complex. The complex of BmdB M2 -BmdC was subjected to structural investigation by X-ray crystallography and cryo-electron microscopy. BmdB M2 was expressed, purified, modified with various substrate analogues, bound with BmdC, and the complex purified by gel filtration prior to crystallization screening and/or grid preparation. We present three resulting structures: a BmdB M2 -BmdC complex structure with a cysteine-vinylsulfonamide adenylate 61,62 , determined by X-ray crystallography at 3.8 Å resolution (PDB:7ly7), a BmdB M2 -BmdC complex structure determined by cryo-EM at 3.8 Å resolution (PDB:7ly4), and an in situ proteolyzed complex of two-thirds of the BmdB A 2 domain with the Ox domain of BmdC, determined by X-ray crystallography at 2.5 Å resolution (PDB:7ly5) (Fig. 4, Supplementary Fig. 5, Supplementary Fig. 6, and Supplementary Tables 1-3).
BmdB M2 -BmdC complex is a very elongated dimeric structure in which the two BmdB modules are kept far apart, and in which the active sites accessed by PCP 2 form a catalytic arc (Fig. 4). The center of the complex is BmdC in the same dimeric configuration described above. The area of the BmdC Ox domain most distal from the dimeric interface (residues 223-231 and 303-306) makes a small interface with one end of the A domain (residues 1321-1323, 1468, 1512-1513), burying 491 Å 2 of surface area per protein. This positions the two BmdB M2 modules in a plane perpendicular to the BmdC rotational axis, stretching out at~45°t o the core BmdC dimer interface. The configuration is extremely oblong, with the BmdB M2 :BmdC complex having overall dimensions of 65 Å by 65 Å by 280 Å and placing the Cy 2 domains on either side of the massive complex. There is some flexibility in this configuration, with the relative Ox:A 2 orientation varying by~11°a cross the three structures. The Cy 2 :A 2 interface is largely constant and the distal N-lobe of Cy 2 substantially weaker in both EM and crystallography maps, indicative of both N-lobe:C-lobe flexibility 63,64 and overall variation in the periphery of the complex. Furthermore, though both BmdB modules of the BmdB M2 -BmdC complex are visible in EM maps, the highest quality map was obtained by performing refinement with a mask that enveloped the BmdC dimer and a single BmdB M2 protomer, because the modest flexibility of the Ox:A 2 interface does break the overall twofold symmetry of the full complex. Notably, the cryo-EM map was calculated using a sample in which the PCP domain was not locked into a single conformation, and as a result, the PCP and A sub domains are not visible, despite attempts at focused classification around PCP-binding sites. The EM maps do confirm the overall conformation and mode of dimerization observed by crystallography. In the BmdB M2 -BmdC crystal structure, the cysteine-vinylsulfonamide-adenylate biases the PCP toward the thiolation state 62,65 , and PCP 2 is visible in this position ( Fig. 4a) with the domain-domain interaction dominated by the A domain helix of residues 1520-1532 and the PCP helix of residues 1820-1833. Surprisingly, A 2sub is disordered, despite no visible impediment to its binding in its classic thiolation position 66 , suggesting A 2sub may not be required to help position PCP 2 for aminoacylation.
In the catalytic cycle of module 2, PCP 2 must visit A 2 for thiolation/aminoacylation to ligate the cysteine substrate (forming Cys-S-PCP 2 ), then the Cy 2 domain for condensation and heterocyclization (forming AlaCys thiazoline -S-PCP 2 ) and the BmdC Ox domain for oxidation (forming AlaCys thiazole -S-PCP 2 ). The positions of the active sites are marked out by: the present BmdB-C crystal structure (step 1-thiolation), previous Cy 2 and LgrA PCP 1 -C 2 containing structures 16,17,67 ) (steps 2 and 3condensation and heterocyclization), and the position of FMN in BmdC (step 4-oxidation). These three active sites form an inplane arc, with each opening facing the inside of the concave surface (Fig. 4c). This decreases the distance PCP 2 is required to travel to transport its intermediates.
The proteolyzed complex provides a high-resolution view of the core BmdB-BmdC interface (Fig. 4d). The BmdB-BmdC binding interface is quite small, burying 491 Å 2 of surface area per molecule in the full proteins. Of this, 423 Å 2 is present in the truncated complex 48,49 . Isothermal calorimetry binding experiments between A 2 and dimeric BmdC reveal a K d value of 75 nM (Fig. 3f) 68 . The interaction features a network of hydrogen bonding, including two central salt bridges, between BmdC Asp225and BmdB Arg1513 and between BmdC Asp231 and BmdB Arg1512, plus side-chain-backbone and backbonebackbone contacts (Fig. 4c). Single point mutation of A 2 (R1512D), A 2 (R1513D), BmdC(D225R), and BmdC(D231R) each leads to disruption of the BmdB-BmdC complex, as observed by gel filtration (Fig. 4e). Peptide synthesis assays with BmdC(D225R) and wildtype BmdB decreased oxidation by~50%, which confirms the importance of this BmdC-BmdB_A 2 interaction. It is not clear whether the remnant oxidation is because BmdC(D225R) retains some affinity to A 2 of BmdB (albeit not sufficient to maintain the complex through gel filtration), or is caused by the transient binding of PCP 2 to BmdC.
Embedding the BmdC oxidase domain into BmdB. NRPSembedded Ox domains are thought to have arisen from the genetic insertion of genes for stand-alone Ox enzymes into NRPS genes. That the Ox domain of BmdC has homology with NRPSembedded Ox domains 47,52,69 led us to ask whether we could replicate evolution and embed BmdC Ox into BmdB to turn it into an embedded tailoring domain. Searching the NCBI database of nonredundant proteins returned 40 proteins which include both RREL and Ox domains contained within a larger NRPS or PKS system. Thirteen showed similar insertion points within NPRS A domains and featured an average sequence identity of~45% to BmdC. Sequence alignments with BmdB A 2 indicate that the embedded Ox domains are inserted at a point corresponding to that between A 2 loop residues Ser1534 and His1535 (Fig. 5). Notably, this loop is adjacent to the helix that includes BmdB Arg1512 and Arg1513 of the BmdB-BmdC interface. We, therefore, created fusion genes in which BmdC, flanked by linkers observed in embedded RREL-Ox didomains from Paenibacillus tianmuensis (LinkPt) or from Bacillus pseudomycoides (LinkBp), is inserted into full-length BmdB after residue 1534. By inserting the BmdC Ox domain without the RREL domain were we able to achieve protein expression and purification. In vitro assays show that these constructs, BmdB-BmdC(Ox)LinkBp and BmdB-BmdC(Ox)LinkPt, are able to produce total bacillamide at 50% and 25% efficiency compared to native BmdB-BmdC. Although the total bacillamide produced by BmdB-BmdC(Ox) LinkPt is lower, it is more efficient at oxidation, with half the total bacillamide produced being pro-bacillamide and half bacillamide D (Fig. 5). BmdB-BmdC(Ox)_LinkBp oxidizes~30% of the bacillamide it produces.

Discussion
Tailoring domains and proteins play an important part in allowing NRPS products to occupy such a large volume of chemical space 28,70 . These domains and proteins are presumably evolutionarily co-opted to function with the NRPS as their genes have been physically taken into the BGC or spliced into the NRPS gene itself. The presence of a RiPP precursor peptide recognitionlike element (RRE) as the N-terminal domain of BmdC raises the possibility that BmdC was co-opted from a RiPP cluster into the bacillamide BGC. The RRE binds its RiPP precursor peptide as a beta-strand along the edge of its β-sheet. This β-sheet is present in the RRE-like domain of BmdC, and overlaying RRE:peptide complexes shows this beta-strand would point the peptide towards the FMN at the BmdC active site ( Fig. 2 and Supplementary Fig. 2b). The heterocyclic peptidyl substrate (Ala-Cys thiazoline ) contains no leader and is far too short to bind both the RRE-like strand and the active site. In addition, the PCP contains no β strands, suggesting this RRE feature is likely an interesting evolutionary relic left over from coopting an RiPP oxidase, rather than a functional element in bacillamide synthesis.
The biochemical (Figs. 3 and 4) and structural (Fig. 4) results suggest that oxidation during bacillamide synthesis requires substrate delivery to the Ox domain by PCP 2 . It is likely that Ox:PCP 2 protein-protein interactions are required for productive substrate binding to the Ox active site. The BmdC structures show the active site of BmdC to be fairly wide and shallow, hinting that a small molecule such as pro-bacillamide may not be able to make extensive binding interactions.
Bacillamide synthetase BmdB exists as a BmdC-induced dimer. As mentioned, dimerization is an uncommon feature in exclusively NRPS systems 53,65 , but is common in polyketide and fatty acid synthases 57,58,71 . One known example of a dimeric NRPS is in the vibriobactin biosynthesis pathway, where NRPS subunit VibF is dimerized through a catalytically inactive C domain 72 . Similar to the situation with BmdB, this dimerization is not necessary for function and does not significantly increase product formation 72 . Very recently, an elegant structure of a Cy-A-PCP construct of FmoA3 showed it has a head-to-tail homodimer architecture with a massive dimerization interface along the back side of the Cy and A domains 73 . The Cy-A domain interface observed in BmdB is similar to that seen in FmoA3 73 and in other C-A-containing structures 67,[74][75][76] . The BmdB M2 -BmdC complex structures (Fig. 4) clearly show why dimerization through BmdC is neutral for bacillamide biosynthesis and why BmdB mutants do not complement each other: The BmdB-BmdC binding interaction holds the BmdB protomers very far apart, such that they are unlikely to interfere with each other. We suggest that dimerization of BmdB through BmdC is a product of the dimeric nature of BmdC, which is common in oxidases 40,51,77,78 . The twofold symmetry of BmdC presents two BmdB binding sites that can be occupied simultaneously, and because dimerization does not impart a catalytic disadvantage, there is no evolutionary pressure to eliminate it.
The relatively stable interface between BmdB A 2 and BmdC means that the flexibility of the BmdB M2 -BmdC interaction is modest. There is clearly some flexibility visible in the EM dataset, so it is more straightforward to refine the region of the map consisting of a single BmdB M2 and a dimer of BmdC, but the BmdB M2 are in the same general orientation relative to BmdC. This feature of the dimerized module contrasts markedly with the dimodular NRPS linear gramicidin synthetase subunit A, where the first and second modules assume many different relative orientations 67 . The difference is explained by the fact that the two BmdB M2 modules are bound to each other through a fairly consistent interaction with the Ox domain of BmdC, whereas modules 1 and 2 of LgrA are tethered by a very flexible linker.
Many separately encoded tailoring enzymes that act during assembly-line synthesis have been described and characterized biochemically [23][24][25] . However, only systems with P450 tailoring 29,79 and prolyl oxidation 80 have been structurally characterized, and they feature transient NRPS: tailoring enzymes interactions, rather than the stable tailoring complex reported here. Indeed, although Bmp3 80 and BmdC are both flavin-dependent oxidases of PCPbound substrates, little other commonality is observed: Bmp3 has a completely different fold from BmdC, uses FAD, not FMN, is tetrameric, and needs to interact only with the type II PCP Bmp1, not with any other NRPS component. The two studies show similar oxidation function can be achieved in very different ways by different NRPS systems.
Oxidation during NRPS assembly-line synthesis can be catalyzed by separately encoded enzymes (in a stable complex like BmdB-BmdC, or a transient complex as in Bmp1:3) and by embedded domains. In addition to the RREL-Ox domains inserted into modules described above, famous examples of thiazole oxidation by embedded domains include those in the biosynthesis of bleomycin, epothilone, and indigoidine 69,81-86 . Bleomycin and epothilone are made by hybrid NRPS-PKS systems, so dimerization of the synthetases is not unexpected. The Ox domains are inserted within the A domain of EpoB and after the PCP domain of BlmIII respectively, and likely contribute to dimerization. As in EpoB, the Ox domain of indigoidine synthetase is inserted between A domain motifs A8 and A9, a common position for tailoring domains: The embedded methyltransferase in TioS is in this position and makes an apparently rigid interface at its insertion site within the A sub subdomain 87 . Interestingly, the inserted Ox sequence in indigoidine synthetase (IndC) is larger than that in EpoB (Supplementary Fig. 7b). Careful analysis and structure prediction using the Robetta web server 88 reveal the insertion is an Ox pseudodimer, suggesting IndC would not dimerize through the inserted pseudodimeric Ox domain. Indeed, gel filtration of purified IndC shows it is a monomer (Fig. 5). We suggest all indigoidine synthetases will have a pseudodimeric Ox which greatly resembles a BmdC dimer fused to their A domain (Supplementary Fig. 7c).
The in-plane catalytic arc of BmdB M2 :C (Fig. 4c) is reminiscent of catalytic chambers in polyketide and fatty acid synthases 89,90 . Fig. 4 Co-complex structures of BmdB module 2 and BmdC. a Overall structure of the biological dimer of BmdB M2 -BmdC solved by X-ray crystallography. The asymmetric unit contains one protomer each of BmdB and BmdC with symmetry mates completing the biological dimers. b Electron microscopy of the BmdB M2 -BmdC. Top: An EM map of the full dimeric BmdB M2 -BmdC (gray) was calculated at 4.2 Å and showed some variability in the relative positions of the BmdB M2 . Imposing a mask around the BmdC dimer and one copy of BmdB produced higher resolution maps, at 3.8 Å resolution, into which BmdC and BmdB Cy2A2 could be modeled. Bottom: Superimposition of BmdB M2 -BmdC structures determined by X-ray crystallography and by cryo-EM show some modest differences in the BmdC:BmdB interaction angle. c Proposed positions BmdB PCP2 assumes around the catalytic arc during the synthetic cycle in module 2. The position of PCP2 for thiolation (aminoacylation) (1) is observed in this study, the position for condensation and heterocyclization (2/3) is based on the condensation state of LgrA (PDB:6mfz) 67  The arc seems to facilitate elegantly the transport function of PCP through the synthetic cycle of module 2 (Fig. 4c), while the BmdB:C geometry keeps each BmdB protomer separate. We believe the BmdB-C complex described here is the only nontransient tailoring complex characterized to date. It remains to be seen whether all such tailoring complexes, which provide a myriad of chemical modifications, have evolved equally elegant architectural solutions.

Methods
General reagents. All commercial reagents purchased for this study are listed in Supplementary Table 4 16 and BmdC (WP_022737638.1; https://www.ncbi.nlm.nih.gov/ protein/WP_022737638.1?report=genpept), cloned from genomic DNA obtained from the Agricultural Research Service Culture Collection. Constructs were cloned into a pET21-derived vector containing an N-terminal TEV-cleavable calmodulinbinding peptide tag and a C-terminal TEV-cleavable octa-histidine tag (pBacTRev), or into a pET21-derived vector containing an N-terminal TEV-cleavable octahistidine tag and a C-terminal TEV-cleavable calmodulin-binding peptide tag (pBacT). All PCR primer sequences, sources of amplified DNA, parental plasmids, restriction sites used, and resulting plasmid names for cloning in this study are listed in Supplementary Table 5. For all site-directed mutagenesis steps, plasmids and primer sequences are listed in Supplementary Table 6.
BmdC was heterologously expressed with growth, harvest, and lysis as described for BmdA. Clarified lysate was applied to a 5 ml HiTrap IMAC FF column charged with Ni 2+ , and BmdC eluted with IMAC buffer B. Fractions containing BmdC were pooled and applied to a 20-ml CBP column equilibrated with CBP buffer A, and eluted with CBP buffer B. Pooled fractions were dialyzed overnight against GF buffer A plus 2 mM βME, at which time affinity tags were cleaved with N-His-TEV protease (1 mg per 20 mg of BmdC) at 4°C. The cleaved BmdC sample was passed again through the HiTrap IMAC FF and CBP columns, the flowthrough was concentrated and applied to a Superdex200 16/60 column equilibrated with GF buffer A (50 mM Tris-HCl pH 7.5, 200 mM NaCl) plus 2 mM βME.
BmdB was expressed and purified similarly to described 16 : protein was expressed in BL21(Bap1) E. coli 91 , grown in LB media containing 40 μg/ml kanamycin at 37°C before being induced, harvested, and purified in the same manner as BmdC.
Bacillamide synthesis assays. Bacillamide synthesis reaction conditions were adapted from Duerfahrt et al. 92 . Reaction containing 1 μM BmdA, 1 μM BmdB, 1 μM BmdC, 1 mM L-alanine, 1 mM L-cysteine, 2 mM ATP, 1 mM pyridoxal-5'phosphate (PLP), 2 mM FMN, 1 mM L-tryptophan, 50 mM HEPES pH 7.5, 100 mM NaCl, 0.5 mM TCEP and 10 mM MgCl 2 in a final volume of 50 μL were incubated for 2 h at 37 o C. Accompanying reactions lack components/substitute components as indicated in Supplementary Fig. 1 (e.g., no ATP controls; 2 mM tryptamine in place of PLP, tryptophan, and BmdA; no BmdC and FMN reaction). Assays depicted in Supplementary Fig. 1i were performed with 2 mM of probacillamide, 2 mM ATP, 1 μM BmdC, 50 mM HEPES pH 7.5, 100 mM NaCl, 0.5 mM TCEP, and 10 mM MgCl 2 in the same volume. Reactions were stopped with 50 μL of 4:1 n-butanol:chloroform, lyophilized to dryness and resuspended in 50 μL of 10% methanol. Five microliters of sample were applied to a ZORBAX Extend-C18 (Agilent) column, analyzed with LC-ESI-MS and eluted with LCmethod 1 (Supplementary Table 7). ESI-MS was performed with an in-line Bruker amaZon speed ETD ion-trap mass spectrometer ( Fig. 1 and Supplementary Fig. 1). The pro-bacillamide and bacillamide D standards were purchased from Zamboni Chem Solutions. In vitro bacillamide reactions including BmdC_D225R were performed the same way (Fig. 4f). When sets of experiments included quantification of bacillamide D and pro-bacillamide production (Figs. 4f and 5c), production was quantified using areas under the curves of extracted ion chromatograms (EICs). Bacillamide D was quantified using the area under the EIC for m/z = 315.1 across the time in which bacillamide peaks elute. Because some samples did not have full separation between peaks for bacillamide D and probacillamide, quantification of pro-bacillamide was performed by integrating the area under the EIC for m/z = 317.1 and subtracting 6% of the area under the EIC for m/z = 315.1, to compensate for the m/z = 317.1 contribution from isotopic distribution of heavier isotopes within bacillamide D.
BmdC crystallography. Crystals of BmdC formed with the vapor diffusion, sittingwell technique at 22°C with a protein concentration of 5 mg/ml and a crystallization solution of 35% PEG1500 (wt/vol), 0.1 M SPG buffer (0.148% (w/vol) succinic acid, 0.604% (w/vol) sodium dihydrogen phosphate monohydrate, 0.328% (w/vol) glycine, pH 9.0). Crystals were cryoprotected by the addition of 10% ethylene glycol, then looped and flash cooled in liquid nitrogen before data collection. Diffraction data were collected at the CLS 08-ID beamline of the CMCF at the Canadian Light Source (λ = 0.979 Å), in Saskatoon, Canada. The data were indexed in the space group I2 1 3 with iMosflm 93 and scaled with AIMLESS in CCP4 94 . Initial phases were obtained by molecular replacement using the program MORDA in CCP4 95 and search models homologous to the Ox domain, including 3EO7 (unpublished). The RREL portion was modeled using AutoBuild in PHENIX 96 . Geometry restraints for FMN were obtained with SKETCHER in CCP4 95 . The overall structure was then iteratively rebuilt in COOT 97 and refined in PHENIX 96 to produce the final structure (Supplementary Table 1). There is one molecule in the asymmetric unit, and a symmetry mate forms the biological dimer of BmdC.
Cofactor identification. BmdC was heat-denatured at 95°C for 15 min and pelleted by centrifugation at 21,100×g (15,000 RPM with FA-45-24-11 rotor). The supernatant was filtered through a 10000 Da Amicon filtration device and the flowthrough used for HPLC 98 : Samples were applied on the ZORBAX Extend-C18 column and eluted with LC-method 2 (Supplementary Table 7). The cofactor eluted at~10 min and was analyzed by direct injection and ESI-MS with an amaZon speed ETD ion-trap mass spectrometer. The standards of FAD and FMN used were purchased from Sigma-Aldrich.
Assays with flavin-free BmdC. Each reaction of 500 μL had 50 μL samples removed at 0, 1, 2, 5, 10, 15, 30, 60, and 120 min. The reaction contained 50 mM HEPES pH 7.5, 10 mM MgCl 2 , 100 mM NaCl, 2 mM DTT, 1 μM enzyme(s), 1 mM L-alanine, 1 mM L-cysteine, 2 mM ATP, 10 mM tryptamine, plus 2 mM of FMN for the wildtype BmdC and BmdB reaction. FMN was not added to the BmdC_ R146E_S292F because this mutant retains some affinity to FMN, but the 2 mM of FMN was confirmed not to promote oxidation (Supplementary Fig. 1). Reactions were stopped with 50 μL 4:1 n-butanol:chloroform before lyophilizing overnight and resuspending in 50 μL of 10% methanol before LC-MS analysis. The sample was centrifuged at 21,100×g (15,000 RPM with FA-45-24-11 rotor) before applying 10 μL of the supernatant to a ZORBAX Extend-C18 (Agilent) column and analyzing with LC-method 3. Crystallization of the proteolyzed BmdB-BmdC complex. The proteolyzed BmdB-BmdC complex is a result of unexplained proteolysis of both BmdB-A 2 PCP 2 and BmdC during crystallization. BmdB_A 2 PCP 2 was expressed and purified as described for BmdB_A 2 , but adding an anion exchange step before gel filtration, using a Mono Q HR 16/10 column equilibrated in a Q buffer A and a gradient to Q buffer B over 100 mL. The eluted sample was concentrated and applied to a Superdex200 column equilibrated with GF buffer A with 1 mM TCEP. BmdB_A 2 PCP 2 was post-translationally modified using AlaCys thiazole -amino-CoA, converted 99 from AlaCys thiazole -amino-ppant (kindly gifted by Jon Patteson and Dr. Bo Li-University of North Carolina Chapel Hill), but the proteolysis removed the PCP 2 domain from the crystallized complex. A 2 PCP 2 was incubated with a twofold molar excess of BmdC and applied to Superdex200 Increase 10/300 column with GF buffer B (50 mM Tris-HCl pH 7.5, 150 mM NaCl) with 1 mM TCEP. Crystals were formed by mixing 4.5 mg/ml of complex with a crystallization solution of 50 mM Tris-HCl pH 7.5, 160 mM KCl, and 21% PEG3350 (wt/vol), using sitting-drop vapor diffusion at 22°C. The crystals appeared after~12 days and only appear in 24-well sitting drops with drop ratios of 2 μL of protein and 1 μL of mother liquor. Crystals were cryoprotected by dipping the crystal in a solution containing 160 mM KCl, 21% PEG3350 (wt/vol) and 20% ethylene glycol before flash vitrification in liquid nitrogen. Diffraction data were collected using the 24-ID-E beamline of the NE-CAT at the Advanced Photon Source (APS) in Argonne, Illinois. The data were indexed and scaled with HKL2000 100 in space group H32. The phases were obtained by molecular replacement in PHENIX 96 using one chain of BmdC and a homology model of the BmdB A 2 domain generated by SWISS-MODEL 101 as search models. The overall structure was then iteratively rebuilt with COOT 97 and refined with PHENIX 96 . There is one molecule in the asymmetric unit and a symmetry mate forms the biological dimer of the proteolyzed complex of BmdB and BmdC.
Crystallization experiments with BmdB M2 -BmdC complex. BmdB M2 was heterologously expressed for structure determination from pBacT_BmdB_M2_struct in BL21(EntD-) E. coli 102 in LB-kan media. BmdB M2 was purified as described for BmdC, but adding an anion exchange step before gel filtration, using a Mono Q HR 16/10 column equilibrated in a Q buffer A (50 mM Tris-HCl pH 7.5, 2 mM βME) and a gradient to Q buffer B (a Q buffer A plus 1 M NaCl) over 150 mL. The eluted sample was concentrated and applied to a Superdex200 column equilibrated with GF buffer A with 1 mM TCEP.
The purified BmdB M2 protein was first modified with ppant from coenzyme A (BioShop Canada Inc.) with 1 μM SFP, 10 mM MgCl 2 , and a twofold molar excess of coenzyme A for 1 h. To remove the excess SFP and coenzyme A, the sample was applied to a Superdex 200 Increase 10/300 column equilibrated with GF buffer B with 1 mM TCEP. Cysteine-vinylsulfonamide adenylate inhibitor (Zamboni Chem Solutions) was used to lock PCP 2 in the thiolation state in a reaction with 10 mM MgCl 2 and a threefold molar excess of the inhibitor at room temperature overnight. This reaction was then applied to the Superdex 200 Increase 10/300 column equilibrated with the same buffer to separate the excess inhibitor. BmdB M2 was incubated with twofold molar excess of purified BmdC and applied to the Superdex 200 Increase 10/300 column equilibrated with GF buffer B with 1 mM TCEP. BmdB M2 :BmdC was crystallized by sitting-drop vapor diffusion after 7 days at 22°C using 2 μL of 5.6 mg/ml protein complex and 1 μL of crystallization solution (0.2 M sodium citrate, 14% PEG3350, 5.6 mg/ml, 0.08 M guanidine hydrochloride and 0.1 M Bis Tris propane pH 6.1). Cryoprotection was accomplished by increasing the amount of MPD, using 2% steps, to a final concentration of 10%, before flash cooling in liquid nitrogen. Diffraction data from this crystal was collected at the CLS 08-ID beamline of the CMCF at the Canadian Light Source (λ = 0.979 Å), in Saskatoon, Canada. The data was indexed with iMosflm 93 to P3 and scaled with AIMLESS to P6 5 22 in CCP4 94 . Initial phases were obtained by molecular replacement in PHENIX 96 using one copy of BmdC, one copy of the A 2 domain fragment solved of the proteolyzed complex structure and one copy of BmdB_Cy 2 (PDB: 5t3e) as search models. Geometry restraints for the cysteinevinylsulfonamide adenylate inhibitor were generated with a the combined used of SKETCHER in CCP4 94 and REEL in PHENIX 96 . A homology model of the PCP 2 domain was generated using SWISS-MODEL 101 and manually placed in the electron density. The overall structure was then iteratively rebuilt with COOT 97 and refined in PHENIX 96 to lead to the final structure described in Supplementary Table 1 (PDB:7ly7). There is one molecule in the asymmetric unit and a symmetry mate forms the biological dimer of this complex.
IndC purification. The gene encoding IndC 69,103 from Streptomyces chromofuscus (AFV27434.1), which produces indigoidine and consists of the domains A-Ox-PCP-Te, was codon-optimized and synthesized by Bio Basic Corp. and subcloned in the vector pBacT, resulting in the vector pBacT_SC_IndC. IndC was expressed and purified as described for BmdB_S792A.
EM structural determination and analysis. BmdB-BmdC complex formation for EM experiments was performed as described above, with pantetheine or Ala-Cys thiazole -amino-ppant attached to PCP 2 . Lacey carbon grids (200 mesh, SPI supplies) were glow discharged for 15 s at 15 mA. Sample (3.5 μl) of were added to the grid surface, blotted for 2 s, and rapidly plunged into liquid ethane, using a Vitrobot (Mark IV, FEI Company). The frozen grids were imaged at 300 kV in a Titan Krios, equipped with a Gatan K3 and Bioquantum (Gatan Inc.) using SerialEM 104 . Each micrograph was acquired in counting mode and consisted of 40 frames to a total fluence of 109 e − /Å 2 with a pixel size of 0.855 Å.
All processing steps were carried out in cryoSPARC V2.15 software. All 8140 movies were motion-corrected using patch motion correction. Following patch CTF estimation, blob picking was performed using an elliptical template of size 50 × 300 Å. Only particles with local NCC values between 192 and 521 were kept, resulting in 3,156,357 particles that were extracted using a box size of 512 pixels and binned by 4. Four rounds of 2D classification were performed to remove debris and broken particles, and the remaining 453,440 particles were extracted again using a box size of 640 pixels and downsampled to 480 pixels, resulting in a pixel size of 1.14 Å. Three ab initio models were generated and used for four rounds of 3D heterogenous refinement as to bad particles. The final particle stack, containing 117,491 particles, was used for nonuniform refinement, resulting in a 4.2 Å reconstruction. The map displayed a full dimer complex, with one BmdB molecule showing much better signal than the other. Next, a mask containing this arm and the entire oxidase dimer was created with Chimera and used for local refinement as nonuniform refinement, resulting in the final 3.8 Å map.
Reporting summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
The crystallographic data (Supplementary Table 1