Crystal Structures of Wolbachia CidA and CidB Reveal Determinants of Bacteria-induced Cytoplasmic Incompatibility and Rescue

Cytoplasmic incompatibility (CI) results when Wolbachia bacteria-infected male insects mate with uninfected females, leading to embryonic lethality. “Rescue” of viability occurs if the female harbors the same Wolbachia strain. CI is caused by linked pairs of Wolbachia genes called CI factors (CifA and CifB). The co-evolution of CifA-CifB pairs may account in part for the incompatibility patterns documented in insects infected with different Wolbachia strains, but the molecular mechanisms remain elusive. Here, we use X-ray crystallography and AlphaFold to analyze the CI factors from Wolbachia strain wMel called CidAwMel and CidBwMel. Substituting CidAwMel interface residues with those from CidAwPip (from strain wPip) enables the mutant protein to bind CidBwPip and rescue CidBwPip-induced yeast growth defects, supporting the importance of CifA-CifB interaction in CI rescue. Sequence divergence in CidAwPip and CidBwPip proteins affects their pairwise interactions, which may help explain the complex incompatibility patterns of mosquitoes infected with different wPip strains.

W olbachia pipientis is an intracellular bacterium infecting~40% of all terrestrial arthropod species and certain filarial nematodes [1][2][3] . They are primarily inherited maternally 4 . Their remarkable success in spreading in arthropod populations is in part driven by the ability of Wolbachia to increase the number of infected females 5 , often by acting as reproductive parasites that distort host sex ratios or reproductive outcomes 6 . The most common type of reproductive parasitism is cytoplasmic incompatibility (CI) 6 , which is a type of post-zygotic male sterility characterized by improper paternal chromatin condensation and separation in the first mitosis cycle [7][8][9] . CI causes Wolbachia-infected male insects to be sterile in matings with uninfected females but not similarly infected females (called 'rescue'), providing a selective advantage for infected females and thus driving Wolbachia into the host population 6,10 . Insect species infected by two or more different Wolbachia strains can be bidirectional incompatible [11][12][13] , which causes reproductive isolation of populations infected with different strains 11,14 .
Based on the outcomes of crosses between Wolbachia-infected insects, it has been proposed that CI is governed by a modificationrescue system 15,16 . Specifically, a modification activity from Wolbachia is postulated to modify sperm during spermatogenesis, and a rescue activity in the infected egg reverses or neutralizes the original sperm modification following fertilization. The existence of both unidirectional and bidirectional incompatibilities in infected insect populations has led to agreement that Wolbachia may carry multiple modification and rescue factors 12,13 , but the molecular identity of these factors remained unknown for nearly half a century [17][18][19][20] .
Recently, the key factors involved in CI induction and rescue were identified [21][22][23][24] . They are known as CI factors or Cifs, which are encoded by pairs of linked genes, cifA and cifB. Transgenic expression of both cifA and cifB genes, or cifB alone, in the germline of male Drosophila melanogaster or Anopheles gambiae, induces sterility that is highly similar to CI induced by Wolbachia 21,22,[25][26][27] . This embryonic lethality can be rescued by crossing transgenic males with Wolbachia-infected females or those with a transgenic cifA gene expressed in the germline [22][23][24] . The cif genes have diverged into at least five distinct phylogenetic groups, which are named types I-V 22,[28][29][30] . The degree of similarity and the presence/absence of cif gene homologs between Wolbachia strains correlates with known patterns of bidirectional incompatibility. For instance, Wolbachia strain wRi is able to rescue wMel-induced CI in same-species crosses, probably because these two strains share highly related type-I homologs of the cifA gene (99% amino acid identity); however, the reverse is not true: wMel cannot rescue wRi-induced CI, likely due to wRi also encoding a type II gene pair that is much more divergent 12,13,24 . In general, the cif genes from bidirectionally incompatible Wolbachia pairs are highly divergent, with only 29-68% amino-acid identity 22 . Recent analysis of transgenic cif gene expression in D. melanogaster is largely consistent with this view 31 .
The cifA and cifB genes are coevolving 22,30,32 . Protein-binding experiments and expression of cif genes in yeast have shown that CifA binds its cognate CifB specifically and rescues CifB-induced yeast growth defects 21,23 . Although the specific roles of cifA and cifB in CI induction and rescue are still not fully understood, the co-divergence of the cifA and cifB genes 22,32,33 has been proposed to be partially responsible for bidirectional incompatibility, possibly by modulating CifA-CifB binding 34,35 .
Among the five types of cif genes, only type I and certain type V cifB genes encode a deubiquitylase (DUB) domain 21 . (Hereafter, we will use "Cif" for the entire set of CI factors and "Cid" for the deubiquitylase family and primarily focus on the type I Cid protein family.) The cidA and cidB genes from Wolbachia strains wMel (hereafter cidA wMel and cidB wMel ) and wPip (hereafter cidA wPip and cidB wPip ) are among the most well studied cif genes (Fig. 1a). These Wolbachia strains are expected to be incompatible, as shown by crosses between trans-infected Aedes albopictus mosquitoes 36 . Unlike the wMel strain, Wolbachia wPip strains found in different populations of Culex pipiens are highly diversified. Many of them carry several polymorphic copies of the cidA wPip and cidB wPip genes, which may function as independent modification and rescue factors and associate with the diverse CI phenotypes in wPip-infected C. pipiens 32,37 . Thus, the Cid proteins can be used as a model to study how the diversification of cidA wPip and cidB wPip affects protein-protein interactions and how it is related to CI induction and rescue.
Here, we report crystal structures for CidA wMel and the DUB domain of CidB wMel and build a model of the CidA wMel -CidB wMel complex using AlphaFold-Multimer 38 , which is validated and confirmed by comparison to a crystal structure of a homologous CidA wPip -CidB wPip complex. Substituting the CidA wMel residues at the CidB-binding interfaces with those of CidA wPip enables it to bind CidB wPip and rescues CidB wPip -induced yeast growth defects, which suggests a role of cid gene co-evolution in modulating CidA-CidB binding specificity. Analysis using the crystal structures of the CidA wPip -CidB wPip complex further shows that sequence divergence in Wolbachia wPip may result in complex binding patterns between Cid wPip protein variants. These results provide a solid base for future investigations of CI mechanism and shed light on how co-evolution of CidA-CidB pairs modulates their interactions and bidirectional CI.

Results
Crystal structures of CidA wMel and CidB wMel DUB , and a model of CidB wMel ND1-ND2 . We determined the crystal structure of CidA wMel to 2.75 Å resolution (Supplementary Table 1). CidA w-Mel is made up largely of α-helices (Fig. 1b). Residues 111-154 and 158-165 are disordered and cannot be modeled, which divide CidA wMel into an N-terminal domain (NTD, residues 1 to 161) (see below for why we include these disordered residues as part of the N-terminal domain) and a C-terminal domain (CTD, residues 162 to 422) (Fig. 1b). The C-terminal domain folds into a twisted set of six HEAT repeats, a structural motif that primarily mediates protein-protein interactions 39 (Fig. 1b).
CidB wMel was expressed poorly in Escherichia coli, making direct structure determination difficult. We used a "divide-andconquer" strategy to obtain structural information for each of its domains separately. CidB wMel consists of a region, which is predicted to contain two PD-(D/E)XK (pseudo) nuclease domains (residue 1-751; hereafter CidB wMel ND1-ND2 , ND: nuclease domain) and a C-terminal deubiquitylase (DUB) domain. An active DUB domain is required for inducing a CI-like phenotype in transgenic D. melanogaster for cid operons 21 . We determined the crystal structure of the CidB wMel DUB domain to 1.85 Å resolution (Supplementary Table 1 and Fig. 1c). Similar to the structures of other proteases in the CE clan/Ulp1-like protease family 40,41 , the DUB core consists of a five-stranded β sheet flanked by α helices on both sides. His957, Asp976, and Cys1016 form the catalytic triad (Fig. 1c). Three variable regions (VRs), a constant region (CR) and a C-terminal accessory domain may account for the S1 ubiquitin-binding interface 41 ( Supplementary  Fig. 1a-d).
A model for CidB wMel ND1-ND2 (Fig. 1d) was built with the AlphaFold 42 program. Although the predicted structural model has an average pLDDT score as high as 83.190, which confirms its quality (pLDDT > 70 indicates the backbone prediction is correct 43 ), we wanted to validate this model further.
CidB wMel ND1-ND2 shares sequence homology with other CifB molecules 34 . Of note, the ND1-ND2 region of CidB wPip _I(b/2) (a CidB wMel ortholog from the Wolbachia strain wPip Tunis line 30 Supplementary Fig. 3a). Moreover, the residues that are varied among CidB homologs are mainly located at the surfaces of the protein ( Supplementary Fig. 4a), which should not affect protein folding. Thus, the model for CidB wMel ND1-ND2 is likely to be accurate. Since the cif genes from wMel have been well characterized, we will mainly focus on the structures and functional analysis of the CidA wMel -CidB wMel ND1-ND2 pair in the following sections. The results from the CidA wPip(Tunis) -CidB wPip(Tunis) system will be mentioned when necessary.
CidB wMel ND1-ND2 consists of two NDs; ND1 comprises residues 1-391 and ND2 comprises residues 392-751 (Fig. 1d). The ND1 and ND2 domains of CidB wMel interact through an extensive interface burying a surface area of~2200 Å 2 . This interface is highly similar to that of CidB wPip(Tunis) (Supplementary Fig. 3b-c) and the residues involving in ND1-ND2 interaction are highly conserved among CidB homologs (Supplementary Fig. 3d-f). Interestingly, in addition to the (inactive) PD-(D/E)XK modules predicted by sequence homology analysis (residues 277-375 and 607-730) 21,28-30 , the structural model of CidB wMel shows that each ND has an additional more divergent PD-(D/E)XK module (residues 30-137 and 431-523) ( Supplementary Fig. 5). None of these modules present the key canonical catalytic residues E-D-E-K 23 (the putative corresponding residues are P33, K77, S92, S94 in module I, K279, Y311, N330, T332 in module II, F435, K475, I491, D493 in module III and K609, G663, V680, G682 in module IV based on structural analysis ( Supplementary Fig. 5)). These modules may play structural roles or support an alternative catalytic activity.    CidA wMel binds CidB wMel through a large distinct interface. We built a model for the CidA wMel -CidB wMel ND1-ND2 complex using AlphaFold-Multimer 38 (Fig. 2b). The predicted end-to-end structure has a ranking score (TM-Score) as high as 0.8568 with a ptm score (intra-chain quality) of 0.86436 and an iptm score (interface score) of 0.85498. This indicates our prediction is highly reliable both for the protein themselves and the interface between them. The obtained end-to-end model was further optimized by Molecular Dynamics (MD) simulations. Two independent trajectories were performed and both achieved equilibrium within the first 100 ns and were then trapped in optima in the next 100 ns. The Root-Mean-Square Deviation (RMSD) for one of the simulation trajectories was recorded ( Supplementary Fig. 6a). Trajectory cluster analysis was performed and the most stable binding conformation was extracted from the largest cluster of the trajectory as the final binding complex of CidA wMel -CidB wMel ND1-ND2 . The complex also predicts many hydrogen bonds and salt bridges between CidA wMel and CidB wMel ND1-ND2 at the three interfaces ( Fig. 2c-e) and only 0.57% of the residues are Ramachandran outliers ( Supplementary  Fig. 6b), which demonstrated the side chains of the predicted model have been reliably predicted. Therefore, we believe that the model of the CidB wMel ND1-ND2 and its complex with CidA wMel is reasonable.
The interface between CidA wMel and CidB wMel ND1-ND2 can be divided into three regions (Fig. 2b); each region mainly involves a pair of structural motifs. At the first region (Interface I), the helices consisting of the residues 93-158 of CidA wMel interact with a loop in CidB wMel ND1-ND2 (residues 448-462) through a network of hydrogen bonds and salt bridges (Fig. 2c). Interestingly, the region of CidA wMel which is disordered in the absence of CidB wMel (residue 111-154 and 158-165) is modeled to form a helix bundle upon complex formation (Figs. 1b and 2c), consistent with its role in mediating CidA wMel -CidB wMel ND1-ND2 interaction. These residues are part of the N-terminal domain. Interface II is formed by the helices at the N-terminus of CidA wMel (residues 2-60) and helices from both NDs of CidB wMel ND1-ND2 (residues 337-353 and 395-418) (Fig. 2d). Interface III involves multiple HEAT repeats in CidA wMel and a cross-cutting helix (residues 241-264) in CidB wMel ND1-ND2 that is bolstered by two β strands and their connecting loop (residues 365-387) to stabilize the interaction (Fig. 2e). Interestingly, although the model of CidA wMel -CidB wMel ND1-ND2 complex is very similar to the structure of CidA wPip(Tunis) -CidB wPip(Tunis) ND1-ND2 (Fig. 2a, b), the residues involving in interaction at the three interfaces are very different between the two complexes ( Fig. 2c-h), which could help explain their cognate-specific binding. (The residues at the interface are well resolved in CidA wPip(Tunis) -CidB wPip(Tunis) ND1-ND2 with good electron density ( Supplementary Fig. 7)).
Pulldown experiments and yeast growth assays were used to validate these structural models. As a model system, it has been demonstrated that expression of cidB genes in the yeast Saccharomyces cerevisiae inhibits cell growth, but growth is at least partially restored if the cognate cidA gene is coexpressed, resembling CI induction and rescue in insects 21,23 . Substituting the CidA wPip(Tunis) residues at its CidB-binding interface with the corresponding ones from CidA wMel abolished its ability to bind CidB wPip(Tunis) ND1-ND2 and neutralize the CidB wPip(Tunis)induced growth defect in yeast ( Supplementary Fig. 8). Thus, the structure of the CidA wPip(Tunis) -CidB wPip(Tunis) ND1-ND2 complex accurately captures the CidA-CidB interaction mode and is biologically relevant. However, due to the poor expression of CidB wMel ND1-ND2 in E. coli and the low toxicity of CidB wMel in yeast, we could not directly validate the model of the complex by pulldown or yeast growth analysis. To test the reliability of this model, we rationally designed a set of mutations based on the model to demonstrate that the interaction mode is conserved between the wPip and wMel CidA-CidB pairs (see below).
CidA-CidB binding specificity is determined by residues at their interfaces. Until now, only CifA and CifB proteins expressed from the same operon have been shown to interact 21,23,35 . The specific interactions between CifA and CifB have been proposed to play a role in CI induction and/or rescue, depending on different CI models. We substituted multiple residues at the binding interface of CidA wMel with the corresponding region of CidA wPip(Tunis) (Fig. 3a-c) and investigated if the chimeric construct (named CidA wMel (ST)) could now bind CidB wPip(Tunis) ND1-ND2 . Their interaction, while detectable, was weak, probably because the residues substituted are not the major binding determinants for interaction between CidA wPip(Tunis) and CidB wPip(Tunis) ND1-ND2 (Fig. 3d). Interestingly, CidA wMel (ST) could bind a closely related CidB wPip ND1-ND2 from Wolbachia strain wPip(Pel) (hereafter CidB wPip(Pel) ND1-ND2 ) to an extent similar to the binding of the cognate CidA wPip(Pel) (Fig. 3d, Fig. 9). We used CidB wPip(Pel) for further biochemical investigation.
Agreeing with the pulldown experiments, yeast growth defects induced by CidB wPip(Pel) could be rescued by both CidA wMel (ST) and CidA wPip(Pel) , but not CidA wMel (Fig. 3e) even though the expression levels for CidA wMel and CidA wMel (ST) were similar ( Supplementary Fig. 10). These data clearly show the importance of the interfacial residues (Fig. 3e) in determining binding specificity. They also support the accuracy of our CidA wMel -CidB wMel ND1-ND2 model. To locate the critical regions that determine binding specificity, the substituted interfacial residues of CidA wMel (ST) were divided into nine groups/regions based on their location in the primary sequence and three-dimensional structure. Each region was individually reverted back to the original sequence in CidA wMel (Fig. 3b,  ND1-ND2 (Fig. 3d). They also failed to rescue CidB wPip(Pel) -induced yeast growth defects (Fig. 3f). (These CidA variants were expressed at similar levels. Interestingly, however, co-expression of CidA wMel variants with CidB wPip(Pel) in yeast did strongly enhance levels of the latter protein but only if they were able to bind to it. (Supplementary Fig. 10)) Thus, these specific residues at the binding interface are particularly important in determining the binding specificity of CidA wMel .  Table 1 and Fig. 3g). The structure of the complex was very similar to that of CidA wPip(Tunis) -CidB wPip(Tunis)  Fig. 11; electron density for representative residues shown in Supplementary Fig. 7) and explained why certain CidA wMel (ST) variants show reduced binding to CidB wPip(Pel) ND1-ND2 . For example, at region 4, residues R114 and K121 of CidA wMel (ST) interact electrostatically with residues E460 and D459 of CidB wPip(Pel) ND1-ND2 , respectively (Fig. 3h). In CidA wMel (ST-4), residues 114 and 121 were reverted back to the corresponding residues of CidA wMel , which are both glutamines (Fig. 3c). These residues are not charged and may lead to the reduced binding of CidA wMel (ST-4) to CidB wPip(Pel) ND1-ND2 (Fig. 3d). Similarly, at region 7, residues R288 and D299 of CidA wMel (ST) have charge complementarity with D246 and R252 of CidB wPip(Pel) ND1-ND2 , respectively (Fig. 3i). In the CidA w-Mel (ST-7) revertant, residues 288 and 299 are glutamine and asparagine, respectively (Fig. 3c), which can no longer form salt bridges with D246 and R252 of CidB wPip(Pel) ND1-ND2 . Finally, at region 9, residue R393 of CidA wMel (ST) interacts with residues E257 and D261 of CidB wPip(Pel) ND1-ND2 through salt bridges (Fig. 3i). In CidA wMel (ST-9), residue 393 is glutamic acid (Fig. 3c), which cannot bind to E257 and D261 of CidB wPip(Pel) ND1-ND2 due to electrostatic repulsion. The crystal structure clearly explains the reduced binding of CidA wMel (ST-4), CidA wMel (ST-7) and CidA wMel (ST-9) to CidB wPip(Pel) ND1-ND2 (Fig. 3d), which closely parallels their loss of function in yeast growth rescue assays (Fig. 3e, f).

Discussion
Here, we have presented structural, biochemical and functional studies on the Wolbachia CI factors CidA wMel and CidB wMel , providing a solid foundation for investigation of the molecular functions of the Cid proteins in CI. These proteins are used here as a model to investigate the evolution of CidA-CidB interactions and their biological impact. Our structure-based mutagenesis has led to a clear demonstration of noncognate CifA-CifB binding, which suggests how sequence changes specifically at the interfaces between these proteins could lead to the complex incompatibility relationships that can be observed in the wild 32 .
The Cid systems from Wolbachia strains wMel and wPip are evolutionarily related, but these strains reside in different clades or supergroups (supergroups A and B, respectively). The two strains are expected to be incompatible based on crosses with trans-infected Ae. albopictus mosquitoes 36 . Here, we show that the CidA-CidB pairs from these Wolbachia strains share the same interaction mode but have different residues at their binding interfaces. These residues are critical for determining their binding specificity. CidA wMel (ST), a rationally designed chimeric construct containing the scaffold of CidA wMel and the interfacial residues of CidA wPip , binds CidB wPip(Pel) and rescues CidB w-Pip(Pel) -induced growth defects in yeast. It proves that CidA variants from different Wolbachia strains use the same pathway to carry out their function.
To further evaluate the conservation and evolutionary relationship of CidA and CidB found in different Wolbachia strains, the conservation scores of CidA and CidB residues were calculated by the ConSurf server and mapped onto their structures [45][46][47] . The residues in the core of the proteins tend to be highly conserved. In contrast, the residues located at the surface are more variable (Supplementary Fig. 4a-b). Many of the varied residues are located at the CidA-CidB binding interfaces (Supplementary Fig. 4c). (Although several key residues of CidA at interface III are identical among its homologs, their interaction with CidB cannot be maintained due to residue variations at interface III of CidB.) This agrees with previous hypotheses that the CidA and CidB interfaces are co-evolving 32,35 , which could modulate binding between CidA and CidB proteins and lead to incompatibility.
Wolbachia strains infecting C. pipiens usually contain more than one copy of the cidA and cidB genes 32 , which has been proposed to be responsible for the very complex incompatible patterns documented in C. pipiens infected with these strains. We showed here that the natural sequence variations in CidA wPip and CidB wPip proteins affects their binding specificity. The expression of multiple CidA wPip and CidB wPip variants in one Wolbachia strain could lead to a mix-and-match type of binding and may be one reason for the very complex crossing results. The presence of cidB_IV(2) variants was associated with the incompatible phenotype of mosquitoes infected with group IV wPip when crossing with those infected with wPip from other groups 32,37 . Indeed, the CidB_IV(2) variants have unique residues at interface I (including residues 451, 452, 460, 461 and 462) compared to the CidB of the other groups. These differences may impair the binding of the CidB_IV(2) variants to the CidA variants of the other groups and lead to incompatibility.
The crystal structure of CidB wMel DUB provides a solid structural basis for further investigation of the molecular targets and function of CidB deubiquitylase family proteins. Given that bacteria do not Fig. 3 Mutagenesis with binding and yeast growth assays reveal how residues at the three interfaces determine CidA binding specificity. a, b CidA wMel (ST) is a chimera with the body of CidA wMel (pink) and interfacial residues from CidA wPip(Tunis) . The locations of the mutated residues are shown in orange on the CidA wMel structure. c The substituted residues in CidA wMel (ST) are divided into nine regions (R) and reversed back to those of CidA wMel , individually, to create CidA wMel (ST-1) through CidA wMel (ST-9). d CidA wMel (ST) does not bind wild-type CidB wMel but binds CidB wPip(Pel) to a similar extent as CidA wPip(Pel) . Regions 4, 7, and 9 play important roles in binding. The experiment was repeated three times independently with similar results obtained. One representative is shown. e CidA wMel (ST) is able to rescue yeast from CidB wPip(Pel) -induced lethality.  21,40 . Indeed, CidB proteins have been proposed to target nuclear-protein import and protamine-histone exchange factors 48 . CidB wMel DUB has three variable regions and a C-terminal region different from other bacterial proteases in the CE-clan/Ulp1-like protease family, which could be responsible for its substrate preference for ubiquitin over other ubiquitin-like molecules and for Lys63-linked ubiquitin chains 21,40,41 .
In different models that try to account for how Cif proteins induce and rescue CI, CifA-CifB interaction is suggested to play different roles. In one, which is named the "2-by-1" model, CifB is suggested to have an ancillary role by modulating CifA stability or activity in the male germline 25 . By contrast, in the "toxin- Fig. 4 Sequence variations modulate interactions between natural CidA wPip and CidB wPip alleles. a Sequence alignment shows that residues at certain positions are different among the CidA wPip and CidB wPip variants. The residues that are the same among the variants are not shown. The varied residues which are located at the binding interfaces are boxed. The numbers above sequence alignment were based on the sequences of CidA and CidB from wPip Tunis. b The locations of the varied residues are shown as spheres on the structure of the CidA wPip(Tunis) -CidB wPip(Tunis) ND1-ND2 complex. c The varied residues which are located at the binding interfaces are colored and labeled. Other residues at the binding interfaces are shown in gray. d, e His-tagged CidA variants were used to pull down GST-tagged CidB variants. The proteins were detected by Coomassie Blue staining. These experiments were repeated three times independently with similar results obtained. One representative is shown. ND nuclease domain. Source data for panels (d) and (e) are provided as a Source Data file.
antidote" model, specific CifA-CifB interaction in the egg has been proposed to be essential for CifA to rescue CifB-induced CI 32,35 . Two major differences between these models are where CifA and CifB interact (in male or female) and whether the interaction is involved in CI induction or rescue. Our structures of CidA-CidB complexes and attempts to modulate their binding specificity provide a way to rationally design CifA and CifB mutants with desired binding attributes. Analysis of these variants will help distinguish between these and other CI models.
For in vitro pull-down experiments, CidA variants were the same as those used for crystallization. A pET28a-GST vector was made where the GST coding sequence was inserted between restriction sites NcoI and BamHI on pET28a. The CidB wPip(Tunis) ND1-ND2 and CidB wPip(Pel) ND1-ND2 genes were cloned into the pET28a-GST expression vector between restriction sites EcoRI and XhoI so that the proteins expressed had an N-terminal GST-tag.
For yeast growth analysis, DNA fragments were subcloned from E. coli vectors by restriction digest or PCR amplification and ligated into yeast expression vectors. The 2-micron plasmid pRS425GAL1 (LEU2) utilizing a GAL1 promoter was used for galactose-induced CifA expression in yeast and the low-copy vector pRS416GAL1 (URA3) was used for galactose-induced CifB expression.
Primers used to generate CifA or CifB expression plasmids are summarized in Supplementary Table 2. All plasmids were verified by sequencing (Sangon Biotech, China).
Protein expression and purification. All proteins were expressed in E. coli (BL21 (DE3) strain). Briefly, E. coli transformed with an expressing plasmid was cultured in Luria broth (LB) at 37°C to an optical density (OD 600 ) of 0.6. Overexpression of the recombinant proteins was induced by adding isopropyl-β-Dthiogalactopyranoside (IPTG) to a final concentration of 0.5 mM at 16°C for 16-18 h.
The harvested bacteria overexpressing CidA wMel (or CidA variants) were resuspended in a lysis buffer (20 mM Tris-HCl, pH 8.0, 300 mM NaCl, 10 mM imidazole, 10% glycerol) and lysed via a high-pressure homogenizer at 4°C. The lysate was centrifugated at 26500 × g for 30 min at 4°C. After centrifugation, the supernatant was loaded onto a Ni-NTA column (GE Healthcare, USA). The column was washed using a lysis buffer supplemented with 50 mM imidazole and eluted using a lysis buffer supplemented with 500 mM imidazole. The eluted protein was diluted with buffer A (20 mM Tris-HCl, pH 8.0, 5 mM DTT) and further purified by anion-exchange chromatography (Hi Trap Q HP 5 mL, GE Healthcare, USA), using a linear gradient of 0%-40% mixture of buffer A and buffer B (20 mM Tris-HCl, pH 8.0, 1 M NaCl, pH 8.0, 5 mM DTT). Finally, the protein was purified by gel-filtration chromatography (Superdex 200 10/300 GL, GE Healthcare, USA), using a buffer containing 20 mM Tris-HCl, pH 8.0, 150 mM NaCl, 5 mM DTT.
The CidB wMel DUB protein was purified following a similar procedure, using affinity, anion-exchange and size exclusion chromatography. CidB wMel DUB has a GST-tag. Instead of using Ni-NTA affinity chromatography, GST affinity chromatography was used. The harvested bacteria overexpressing CidB wMel DUB were resuspended in a lysis buffer (20 mM Tris-HCl, pH 8.0, 150 mM NaCl, 5% glycerol) and lysed via a high-pressure homogenizer at 4°C. The lysate was centrifugated at 26500 × g for 30 min at 4°C. After centrifugation, the supernatant was loaded onto a glutathione-Sepharose column (GE Healthcare, USA). The GST tag was removed by incubating the loaded column with PreScission protease overnight at 4°C. The CidB wMel DUB protein was further purified through anionexchange and size exclusion chromatography using the same columns and buffers as described above.
Selenomethionine (SeMet)-derivatized proteins were expressed by bacteria growing in M9 SeMet medium supplemented with 100 mg/L L-selenomethionine. The purification procedure of SetMet-derivated proteins was the same as mentioned above for the native.
Crystallization, data collection, and structure determination. All crystals were grown by the microbatch-under-oil method unless otherwise specified 49 . CidA wMel was crystallized at 16°C by mixing 1 µL protein (5 mg/mL) with 1 µL crystallization buffer containing 0.2 M Sodium phosphate monobasic monohydrate, 20% w/v Polyethylene glycol 3350, pH 4.7. The crystals were cryoprotected by Parabar 10312 (previously known as Paratone oil, Hampton Research, USA). X-ray diffraction data were collected on beamline BL18U1 at the Shanghai Synchrotron Radiation Facility at 100 K and at a wavelength of 0.97852 Å. Data integration and scaling were performed using HKL3000 50 . The structure was determined by SeMet single-wavelength anomalous dispersion (SAD) method with the AutoSol program in PHENIX 51 . The CidA wMel model was initially built by the Autobuild program in PHENIX and subsequently subjected to iterative cycles of manual building in Coot 52 and refinement in PHENIX.
CidB wMel DUB was crystallized by hanging drop method at 18°C, with the crystallization reservoir solution containing 10 mM Nickel (II) Chloride hexahydrate, 100 mM TRIS pH 8.5 and 20% w/v Polyethylene Glycol Monomethyl Ether 2000. The crystals were directly flash frozen in liquid nitrogen using reservoir solution supplemented with 10% glycerol as cryoprotectant. X-ray diffraction data were collected at beamline BL17U1 at the Shanghai Synchrotron Radiation Facility. The structure was determined by SeMet SAD method as described above.
Data collection and structure refinement statistics are summarized in Supplementary Table 1. All Molecular graphics were created using UCSF ChimeraX 54 .
In vitro pull-down experiments. E. coli BL21(DE3) expressing either His-tagged CidA mutants or GST-tagged CidB ND1-ND2 variants were harvested by centrifugation and resuspended separately in a buffer containing 20 mM Tris-HCl, pH 8.0, 300 mM NaCl, 1 mM PMSF and 5% glycerol. The expression levels of CidA variants were approximated by the BCA Protein Assay (Tiangen Biotech (Beijing) Co., Ltd., China). Equal amount of CidA variants were mixed with a large and fixed volume of CidB ND1-ND2 variants to ensure that CidB ND1-ND2 variants were in excess to CidA. The mixture was co-lysed by sonication. Cleared lysates were incubated with 50 μL Ni-NTA resin at 4°C for 1 h. The resin was washed eight times with 800 μL buffer containing 20 mM Tris-HCl pH 8.0, 200 mM NaCl, 0.01% Tween-20, 50 mM imidazole each time. Proteins were eluted by adding 3 resin volumes of 20 mM Tris-HCl pH 8.0, 200 mM NaCl, 300 mM imidazole. The samples were analyzed using the SDS-PAGE and Coomassie stain.
Yeast growth assays. Growth was analyzed in the BY4741 strain as previously described 55 . Briefly, yeast cultures were grown overnight at 30°C in Yeast Extract Peptone Dextrose (YPD) media or synthetically defined (SD) raffinose media lacking uracil, leucine or both. Yeast were pelleted by centrifugation, washed with sterile water, and spotted in six-fold serial dilution from an initial OD 600 0.2 concentration on solid minimal synthetic media containing either 2% galactose or glucose and lacking either uracil, leucine, or both. Plates were placed at 30 or 36°C for 2 to 3 days.
Western blot analysis. For immunoblotting, co-expression culture in raffinose minimal medium (SD) lacking uracil and leucine were diluted to 0.2 OD 600 in galactose (inducing) minimal medium (SD) lacking uracil and leucine, kept 12-16 h at 30°C until reaching 0.8-1.0 OD 600 at which point the equivalent of 2.5 OD 600 units of cells were harvested, washed and resuspended in 1 mL dH 2 O followed by the addition of 200 μL dH 2 O and 200 μL 0.2 M NaOH, incubated for 5 min at room temperature. Cells were vortexed intermittently for 20 s and pelleted at 10,000 x g, 1 min. Pellets were stored at −80°C for at least 15 min, resuspended in 100 μL 1 x SDS-PAGE sample buffer and 4% β-mercaptoethanol then heated at 95°C for 3 min, centrifuged and 20 μL supernatant were loaded in the 10% SDS-PAGE gel and transferred to PVDF Immobilon-P transfer membranes (0.45 μM pore size) (Sigma-Aldrich) under 70 V, 2.5 h used for immunoblot analyses. Antibodies used for immunoblotting were as following: mouse anti-FLAG M2 (Sigma, 1:10,000), secondary antibody used was sheep anti-mouse NXA931V (GE Healthcare, 1:5,000); and mouse anti-PGK1 (yeast phosphoglycerate kinase; Abcam, 1:10,000), secondary antibody used was sheep anti-mouse NXA931V (GE Healthcare, 1:10,000). All immunoblot analyses used 5% milk for blocking. All serial dilution and Western blot data are representative of at least two biological experiments. Proteins were visualized by HRP-based chemiluminescence 56 .
AlphaFold modeling. In this study, AlphaFold was used to predict the monomer structure of CidB wMel ND1-ND2 , and AlphaFold-Multimer was used to predict the binding complex of CidA wMel -CidB wMel ND1-ND2 with multiple sequence alignments (MSA) set as the all genetics database used at CASP14. The prediction of complexes was run twice with different random seeds and 10 models were obtained. Beginning with visual inspection, four of them were selected to perform protein structural quality check for the side chain conformations using prime module of Schrödinger2021-3. Eventually, the one complex with the highest quality score was selected for further optimization with subsequent MD simulations.
Molecular dynamics (MD) simulations. MD simulations were performed by using Desmond package of Schrödinger2021-3 57 using the OPLS4 58 force field. The binding model of CidA wMel -CidB wMel ND1-ND2 obtained in the last step was explicitly solvated with TIP3P 59 water molecules under cubic periodic boundary conditions for a 15 Å buffer region. The overlapping water molecules are deleted and 0.15 M KCl is added, and the systems were neutralized by adding K + as counter ions. Brownian motion simulation was used to relax these systems into local energy minimum states separately. An ensemble (NPT) was then applied to maintain the constant temperature (300 K) and pressure (1.01325 bar) of the systems, and the simulations were started with different random initial velocities. The results were visually analyzed by using Maestro graphical interfaces and the RMSD was calculated based on C-alpha atoms. Produced trajectories were clustered using the Desmond trajectory cluster analysis panel. Finally, the most energetically stable binding complex from the largest cluster of conformation in MD trajectory was selected and minimized again with backbone constraints using the prime module of Schrödinger2021-3. The Ramachandran plot of the eventual model was generated with Schrödinger2021-3.
Reporting summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
The data that support this study are available from the corresponding authors upon reasonable request. The coordinates for crystal structures of CidA wMel , CidB wMel DUB , CidA wPip(Tunis) -CidB wPip(Tunis) ND1-ND2 complex, and CidA wMel (ST)-CidB wPip(Pel) ND1-ND2 complex have been deposited in the Protein Data Bank (PDB), with the accession codes 7FIT, 7FIU, 7FIV, and 7FIW, respectively. Source data are provided with this paper.