Topology and structure of an engineered human cohesin complex bound to Pds5B

The cohesin subunits Smc1, Smc3 and Scc1 form large tripartite rings which mediate sister chromatid cohesion and chromatin structure. These are thought to entrap DNA with the help of the associated proteins SA1/2 and Pds5A/B. Structural information is available for parts of cohesin, but analyses of entire cohesin complexes are limited by their flexibility. Here we generated a more rigid ‘bonsai' cohesin by truncating the coiled coils of Smc1 and Smc3 and used single-particle electron microscopy, chemical crosslinking-mass spectrometry and in silico modelling to generate three-dimensional models of cohesin bound to Pds5B. The HEAT-repeat protein Pds5B forms a curved structure around the nucleotide-binding domains of Smc1 and Smc3 and bridges the Smc3-Scc1 and SA1-Scc1 interfaces. These results indicate that Pds5B forms an integral part of the cohesin ring by contacting all other cohesin subunits, a property that may reflect the complex role of Pds5 proteins in controlling cohesin–DNA interactions.

C ohesin is a protein complex that physically connects replicated DNA molecules until they are separated in anaphase [1][2][3] . This cohesion is essential for bi-orientation of sister chromatids on the spindle and is thus required for proper chromosome segregation. Cohesin also has important functions in mediating higher-order chromatin structure, gene regulation and DNA damage repair 4 and is related to other 'structural maintenance of chromosomes' (SMC) complexes such as condensin in both eukaryotes and bacteria 5 . Malfunctioning of cohesin can result in rare genetic diseases (cohesinopathies) 6 and is thought to be a major cause of trisomy 21 and spontaneous human abortions 7 . Mutations in cohesin genes have also been found in human cancers 8 , with the cohesin subunit gene STAG2 (also known as SA2) being one of only 12 human genes which are significantly mutated in 44 cancer types 9 . Understanding the structure of cohesin and how this is used to perform cohesin's various functions on DNA is therefore of great importance.
Biochemical experiments 10,11 and electron microscopy (EM) 12,13 revealed that cohesin complexes form ring-shaped structures. Experiments with budding yeast mini-chromosomes indicate that cohesin mediates cohesion by entrapping sister DNA strands inside this ring structure 14,15 . In human cells, cohesin complexes can be composed of up to seven subunits (Smc1, Smc3, Scc1, SA1 or SA2, Pds5A or Pds5B, Wapl, sororin). Only three of these, Smc1, Smc3 and Scc1 (also known as Rad21 or Mcd1), are required to generate molecular rings. These are formed at one end by hetero-dimerization of the long coiled coil subunits Smc1 and Smc3 via their hinge regions, and at the other end by the 'closing' (greek 'kleisin' 16 ) function of Scc1, which connects the nucleotide-binding domain (NBD) of Smc1 with the NBD-proximal coiled coil region of Smc3 (refs 13,14,17). The latter interface between Smc3 and Scc1 is thought to function as a regulatable 'DNA gate', which can be opened so that DNA can pass through 13,14,[18][19][20] . Gate opening requires Wapl (refs 21,22) and possibly its binding partner Pds5 (refs 23,24), which in vertebrates exists in two isoforms, Pds5A and Pds5B (ref. 25). Pds5 proteins bind to Scc1 directly, whereas Wapl interacts with Pds5 proteins and SA1 or SA2 (refs 13,23,24). Because Wapl can release cohesin from DNA, cohesin has to be stabilized on DNA once it has generated cohesive structures during DNA replication, so that cohesion can be maintained throughout G2-phase and early mitosis. This is achieved by acetylation of Smc3 (refs 23,26,27). In vertebrates, this modification leads to recruitment of sororin 28,29 , which inhibits Wapl and thereby prevents cohesin release from DNA 29 . In early mitosis, sororin bound to cohesin on chromosomes arms is inactivated by phosphorylation 30,31 , which together with SA2 phosphorylation 32 leads to cohesin release by Wapl and partial loss of cohesion between chromosome arms. At centromeres, sororin-bound cohesin is protected by protein phosphatase 2A bound to Sgo1 (ref. 33), and these Wapl-resistant complexes are removed from chromosomes in metaphase by the protease separase 34,35 . Like Wapl, separase opens the cohesin ring, but in this case by cleaving the Scc1 subunit.
Structural information has been obtained for the three ringforming subunit interfaces of cohesin (Smc1-Smc3, Smc1-Scc1, Scc1-Smc3) 14,17,36 , for parts of Scc1, SA2 and Wapl (refs 37-40) and structures have been predicted for the Huntingtin-Elongation factor 3-A subunit-TOR (HEAT) repeats of Pds5 proteins 13 . After submission of this manuscript, crystal structures of Pds5 proteins bound to fragments of Scc1 (refs 41,42) or Wapl and sororin 43 have also been reported. However, little is known about how Pds5 proteins interact with entire cohesin complexes. Addressing this question is important as Pds5 proteins have essential functions which are not understood at the mechanistic level. These include positive roles in establishment and/or maintenance of cohesion 25,[44][45][46][47][48] such as Smc3 acetylation [49][50][51] and recruitment of sororin 29 , but also negative effects on cohesion 52 such as release of cohesin from DNA 23,24 .
Likewise, little is known about cohesin's overall architecture, which so far has only been visualized by rotary shadowing EM 12,13 . These studies analysed cohesin complexes composed of Smc1, Smc3, Scc1 and either SA1 or SA2, isolated either from HeLa cells or Xenopus eggs 12 or assembled from recombinant human subunits 13 . Although ring-shaped cohesin complexes could be seen in these experiments, they also revealed a high degree of structural flexibility, observed in the 50-nm-long coiled coil regions of Smc1 and Smc3. Because such flexibility is incompatible with single-particle EM techniques based on image averaging, we generated structurally more rigid 'bonsai' cohesin by truncating the coiled coils of Smc1 and Smc3, following an approach previously used for the kinetochore Ndc80 complex 53 . We obtained low-resolution three-dimensional (3D) structures of bonsai trimers (Smc1 B , Smc3 B , Scc1) bound to either SA1, or Pds5B or both, and used these together with crosslinking-mass spectrometry to analyse how SA1 and Pds5B interact with cohesin. The results of these experiments imply that Pds5B is not simply associated with the cohesin ring only via binding to Scc1, but forms an integral part of the ring by forming contacts also with Smc1, Smc3 and SA1. This structural information may help to explain the numerous functions that have been attributed to Pds5 proteins.

Results
Analysis of full-length human cohesin by EM. We observed previously that human recombinant cohesin complexes containing full-length Smc1, Smc3, Scc1 and SA1 (500 kDa) can be visualized by rotary shadowing EM as ring-shaped complexes of variable shape 13 , closely resembling cohesin purified from human or amphibian cells and analysed by the same technique 12 . To obtain insight into how Pds5 proteins interact with cohesin, we bound recombinant purified human Pds5B to cohesin tetramers prepared as above and analysed these complexes by SDS-PAGE and silver staining (Fig. 1a) and rotary shadowing EM ( Fig. 1b and Supplementary Fig. 1). Compared with cohesin tetramers (Fig. 1c), an additional density was observed in Pds5B-containing complexes in close vicinity to the NBDs of Smc1 and Smc3. Otherwise these complexes showed the ring conformation characteristic for cohesin tetramers, often with kinks in the coiled coil regions of Smc1 and Smc3. These observations are consistent with Pds5B binding to Scc1, as previously observed in biochemical experiments 13,23,24 , and indicate that Pds5B does not cause conformational changes in cohesin detectable by rotary shadowing EM.
Because rotary shadowed specimens cannot be used for determining 3D structures, we tested if cohesin complexes could be analysed by negative staining EM. For this purpose, we purified tetrameric complexes composed of Smc1, Smc3, Scc1 and SA1 by glycerol density gradient centrifugation and stained them with uranyl formate (Fig. 1d and Supplementary Fig. 2). Also under these conditions structures were seen that are reminiscent of cohesin's 'open' ring conformation, although the thin coiled coils of Smc1 and Smc3 (diameter 1 nm) are more difficult to visualize by negative staining than by rotary shadowing. Unexpectedly, however, negatively stained cohesin complexes often had a different, rod-shaped conformation in which the coiled coils of Smc1 and Smc3 appeared connected (Fig. 1d,e and Supplementary Fig. 2), resembling a conformation also observed for bacterial SMC complexes and human condensin 12,54 . In some cases, these rods were not straight but kinked, possibly reflecting the kinks that have also been observed by rotary shadowing in the ring-like conformation of cohesin 12,13 . Together with the globular domain visible at one end of the rods, presumably representing the NBDs, Scc1 and SA1, these kinked rod-shaped complexes resemble the symbols typically used for musical notes. More work will be required to understand why rotary shadowed and negatively stained cohesin complexes appear in different conformations and which of these reflects the structure of cohesin under physiological conditions, but we note that the existence of rod-shaped cohesin conformations may be biologically relevant as crosslinking-mass spectrometry performed with cohesin in solution revealed numerous contacts between the coiled coils of Smc1 and Smc3 (refs 13,55), and because crystallographic analysis of bacterial condensin has unequivocally shown that the coiled coils in this complex can exist in direct juxtaposition 56 .
Generation and characterization of 'bonsai' cohesin. The visualization of cohesin by negative staining revealed a high degree of flexibility of full-length cohesin, as seen by rotary shadowing, precluding structure determination by single-particle analyses. We therefore engineered 'bonsai' versions of Smc1 and Smc3 that largely lack their coiled coil regions ( Fig. 2a and Supplementary Fig. 3), inspired by the previous finding that truncating the coiled coils of the Ndc80 complex enabled its structure determination 53 . Bonsai SMCs (Smc1 B and Smc3 B ; approximately half the size of their full-length equivalents) were co-expressed with Scc1 in Baculovirus-infected insect cells and purified as stoichiometric trimers (Fig. 2b). Importantly, these assemblies still contain the hinge domains with which Smc1 and Smc3 interact with low-nanomolar affinity 36 . Bonsai trimers containing wild-type NBDs, but not those containing a mutation in the ATP-binding region of Smc3 B (K38A), were able to hydrolyse ATP, indicating that the coiled coil truncations do not prevent NBD interactions in bonsai cohesin (Fig. 2c,d). The ATPase rate of bonsai cohesin is comparable to that of full-length cohesin (data not shown). Differential scanning fluorimetry showed that the presence of the slowly hydrolyzable ATPanalogue ATPyS, as well as a near physiological pH improved the stability of bonsai cohesin ( Supplementary Fig. 4). Samples for single-particle EM were prepared accordingly.
A 3D model of tetrameric bonsai cohesin containing SA1. To obtain insight into cohesin's structure, we co-expressed the bonsai trimers characterized above with SA1 and purified tetrameric complexes. Recombinant Pds5B could be bound to these complexes, resulting in four different complexes for further structural analysis (Fig. 3a). Pds5B did not bind Smc1 B -Smc3 B dimers, indicating that Scc1 is required for the interaction of Pds5B with bonsai cohesin, as observed for full-length complexes 13 ( Supplementary Fig. 5). Complexes were purified by glycerol density gradient centrifugation, either in the absence of glutaraldehyde to enable sample analysis and quality control by SDS-PAGE and silver staining ( Fig. 3b and Supplementary Fig. 6) or following the GraFix method 57 (that is, in the presence of low glutaraldehyde concentrations to preserve structural integrity of cohesin) for negative staining EM ( Fig. 3c and Supplementary  Fig. 7). The presence of Pds5B increased the sedimentation of  these complexes, implying that many of them interacted with Pds5B stably enough to allow their purification ( Supplementary  Fig. 6). We used single-particle analysis to calculate models of tetrameric bonsai cohesin (Smc1 B , Smc3 B , Scc1 and SA1; 367 kDa) and pentameric bonsai cohesin (Smc1 B , Smc3 B , Scc1, SA1 and Pds5B; 533 kDa) (Supplementary Fig. 7 and Fig. 3d, respectively). Calculations were performed independently to prevent any model bias. This resulted in two models with comparable features: the alleged front side contains three fragments ('top', 'bottom left', 'bottom right') with comparable dimensions that are separated by regions with lower density. The fourth fragment (termed 'clasp') is considerably larger and is positioned at the back ( Fig. 3d and Supplementary Fig. 7). An additional density (termed 'bridge') in the pentamer was observed near the NBDs of Smc1 B and Smc3 B , possibly revealing the position of Pds5B (Fig. 3d). How we interpret the structures with respect to the location of subunits and domains will be described below.
Computational sorting of the bonsai pentamer data set for the presence of this putative Pds5B density identified a fraction of  images where this density was missing ( Supplementary Fig. 8).
Further processing of this subset (10,582 particle images) resulted in a 3D model with the characteristic top, clasp, bottom left and bottom right densities (Fig. 3e), resembling the independently calculated 3D model of tetrameric cohesin (Smc1 B , Smc3 B , Scc1, SA1) ( Supplementary Fig. 7). The model was determined at a resolution of 28 Å (FSC 0.5 criterion) ( Supplementary Fig. 9). We conclude that the pentameric bonsai cohesin sample contained a Pds5B-bound and a Pds5B-free population, presumably because Pds5B had dissociated from some cohesin complexes during sample preparation. This is consistent with the observation that Pds5B was present in two 'peaks' in glycerol density gradient fractions, only one of which also contained cohesin (Fig. 3b), and with results from fluorescence recovery after photobleaching (FRAP) experiments which indicate that Pds5 proteins interact dynamically with cohesin in vivo 58 . In both 3D models (Fig. 3d, Fig. 10). This top region represents B15% of the total density of the Smc1 B -Smc3 B -Scc1-SA1 model (367 kDa), corresponding well to the molecular mass of the hinge domain (49 kDa). We therefore attribute the top region to the Smc1-Smc3 hinge ( Supplementary Fig. 10) and the bottom densities to the NBD of Smc3 bound to the N terminus of Scc1 (Smc3 NBD -Scc1 N ) and to the NBD of Smc1 bound to the C terminus of Scc1 (Smc1 NBD -Scc1 C ; Fig. 3e). According to this interpretation, the elongated density that connects the hinge and the NBD domains represents the residual coiled coil regions of Smc1 B and Smc3 B . For reasons that are explained below, we assume that the clasp represents SA1 bound to the central region of Scc1 (Scc1 middle ). At the present resolution, we cannot distinguish which bottom density represents Smc3 NBD -Scc1 N and Smc1 NBD -Scc1 C . The 10-nm distance between the bottom densities implies that ATPyS did not trap bonsai cohesin in a NBD engaged, despite its stabilizing effect on the complex.
Pds5B curves around the NBDs of Smc1 and Smc3. To analyse the structure of bonsai cohesin bound to Pds5B, 10,416 images of pentamers were used to generate a 3D model with a determined resolution of 35 Å (FSC 0.5) ( Fig. 3d and Supplementary Fig. 9). Compared with the model of tetrameric bonsai cohesin without Pds5B ( Fig. 3e and Supplementary Fig. 7), a prominent additional density (bridge) is evident below the bottom left and bottom right densities (Fig. 3d, Fig. 4a,c and Supplementary Fig. 8, Supplementary Fig. 10). This density reappeared when reconstructions were calculated using a reference model from which this density had been manually removed ( Supplementary Fig. 8), confirming that it represents a density that is specifically present in cohesin pentamers. We therefore assume that this density represents Pds5B. Consistent with recently determined structures of Pds5 from three different organisms [41][42][43] , the Pds5B density has an elongated shape with an overall length of 17 nm.
SA1 and parts of Scc1 form cohesin's backside clasp. If the density assignment of Pds5B, the hinge domain, and the NBDs is correct, the clasp-like density that resides behind the residual coiled coils of Smc1 and Smc3 corresponds to SA1 (146 kDa) and Scc1 middle . To test this, we analysed the structure of bonsai cohesin complexes lacking SA1. We were unable to obtain 3D models of bonsai trimers (Smc1 B -Smc3 B -Scc1; 222 kDa), possibly because of the relatively small size and pseudo-symmetrical structure of this complex. However, we were able to analyse images of these trimers when they were bound to Pds5B (387 kDa). Processing was again performed de novo to avoid reference bias of previous models. A 35 Å (FSC 0.5) model was calculated using images of 4,027 particles ( Fig. 3f and Supplementary Fig. 9). Consistent with the assignment described above, these complexes also contained a top region (corresponding to the Smc1-Smc3 hinge), the bottom densities (representing Smc3 NBD -Scc1 N and Smc1 NBD -Scc1 C ) and the curved bridge density assigned to Pds5B. In contrast, less density was observed in the region which in SA1-containing complexes is occupied by the clasp-like structure, although a smaller density was still observed there (Figs 3f and 4a,c and Supplementary  Figs 10 and 11). The majority of the clasp-like density is thus formed by SA1, whereas the central region of Scc1, and possibly also fragments of the residual coiled coils and Pds5B, contribute to the clasp fragment that remains present in the absence of SA1. This interpretation attributes a central position to Scc1 middle , consistent with Scc1 bridging the NBDs of Smc1 and Smc3 and recruiting both SA1 and Pds5B to cohesin. Since parts of SA1 and Pds5 are in direct proximity and both proteins bind Scc1, it is likely that their exact positions and conformations are interdependent. This could account for differences between the various complexes, such as change in the position of Pds5B or a twist of the top region that is observed in the absence of SA1 ( Supplementary Fig. 10). However, the biological relevance of such effects remains unclear, as there is no evidence for the existence of cohesin complexes lacking SA1 (or SA2) in vivo.
Proximity map and topology of bonsai cohesin bound to Pds5B.
To further validate and interpret our 3D models, we used chemical crosslinking followed by mass spectrometry and generated a map of distance constraints with 102 inter-and 244 intramolecular crosslinks ( Fig. 4b and Supplementary Figs 13 Table 1). The proximity of these residues (all numbered as in full-length human cohesin) is further supported by an extensive network of intramolecular crosslinks (Supplementary Fig. 13 and Supplementary Table 2) and consistent with the attribution of the residual coiled coil regions to the solventaccessible X-shaped density of our model. Pds5B K282 crosslinks extensively with multiple of these residues, indicating that it points towards the residual coiled coils of bonsai cohesin.
Crosslinks between lysines 185, 188, 1,034 and 1,038 of Smc3 B and lysines 50, 72 and 86 of Scc1 N (green, Fig. 4b) are in agreement with a crystal structure of the yeast Smc3-Scc1 N interface 14 and reveal that the pairing of Smc3 with Scc1 N is retained in the residual coiled coil fragment of Smc3 B . Crosslinks between Scc1 317-406 and SA1 (gold, Fig. 4b) indicate extensive contacts between Scc1 middle and SA1. Spatial constraints imposed by these crosslinks are largely consistent with the structure of human SA2 in complex with a Scc1 281-420 fragment 38 , suggesting that SA1 and SA2 bind Scc1 in a comparable manner.
1,019-1,028, 1,071-1,087 and 1,168-1,177 are all in close proximity in bonsai cohesin. The distance between these regions in the structure of the equivalent SA2-Scc1 middle (ref. 38) exceeds the spatial constraints of our crosslinker by almost 100 Å. However, the putative proximity of these SA1 regions when bound to cohesin is consistent with the clasp-like conformation of SA1 in our model and with their positions near the flexible regions that connect HEAT repeats in SA2 (ref. 38), indicating that these crosslinks may also reflect physiologically relevant interactions.
The majority of inter-and intramolecular crosslinks that involve Pds5B were observed in its C-terminal part. Residues in this region of Pds5B (845, 925, 1,213, 1,219, 1,244, 1,344 and  1,397) are proximal to Scc1 (335, 387, 527) and SA1 (48,49,52,57,453, 626, 910 and 1,086), but also to residues in the NBD of Smc3 B (157, 1,190 and 1,194) (dark blue, Fig. 4b). This suggests that the charged C-terminal tail of Pds5B, which is predicted to be unstructured and was previously shown to be dispensable for binding to cohesin 13 , is flexible and resides close to both the SA1-Scc1 middle region and to the NBD of Smc3 B .
While Pds5B K282 crosslinks with the residual coiled coils, more N-terminal residues of Pds5B (25, 36, 74, 115 and 121) seem to reside near residues 26 and 114 of Smc3 B and near the Smc3 B -Scc1 N DNA exit gate. This proximity is further supported by mutations in this region of Pds5 in budding yeast that were identified to functionally cooperate with the acetylation of Smc3 in controlling the release of cohesin from DNA 23 . Since one end of the elongated Pds5B density contacts an NBD density in our model, we attribute this density to the N-terminal region of Pds5B, the NBD of Smc3 B and Scc1 N. and speculate that more C-terminal HEAT repeats of Pds5B curve around the NBD region towards the SA1-Scc1 middle interaction interface (Fig. 4d). A comparison of the recently solved structure of human Pds5B 43 , reported after we obtained and submitted the EM structure of bonsai cohesin bound to Pds5B, supports our assignment of Pds5B's N-and C-terminal regions of Pds5B. Interestingly, Pds5B seems to be curved to a higher degree when bound to cohesin than in isolation (Fig. 4d). Based on our proximity map and the assignment of the N-and C-terminal regions of Pds5, we propose that the density defined as 'bottom right' contains Smc3 NDB -Scc1 N whereas Smc1 NBD -Scc1 C account for the density defined as 'bottom left'. Although Pds5B curves around the NBDs of Smc1 B and Smc3 B , the presence of Pds5B did not alter the ATPase activity of Smc1 B -Smc3 B -Scc1 under our assay conditions ( Supplementary Fig. 12).
In silico analysis of Smc1 and Smc3. Because little structural information is available about the coiled coils of Smc1 and Smc3, we analysed these regions of cohesin by in silico modelling. Previous work revealed that the coiled coils are formed intramolecularly by two anti-parallel strands 11 and that their length of almost 50 nm, as well as the position of several coiled coilinterrupting regions are conserved [59][60][61] . Compared with other coiled coil-containing proteins, the sequence of cohesin is remarkably conserved among metazoans but not among all eukaryotes 62 . Here we analysed the sequences of Smc1 and Smc3 and predicted interruptions based on sequence alignments, coiled coil prediction algorithms, available high-resolution structural information 14,17,36 and proximity information from intramolecular crosslinks 13 (Fig. 5). As expected, residues that are absolutely conserved in metazoans, insects, plants and fungi mainly reside in the proximity of cohesin's NBDs. A number of highly conserved residues and regions are also found in the hinge domains of Smc1 and Smc3. Within the coiled coils, Smc3 166-210 and Smc3 976-1,025 are well-conserved, indicating the importance of the interaction between Smc3 and Scc1 N . Micrographs of fulllength cohesin showed characteristic coiled coil disruptions in the hinge-proximal halves of the coiled coils of Smc1 and Smc3 ( Fig. 1b and Supplementary Fig. 1) 12,13 . Based on our analysis, we expect that the non-coiled coil regions Smc3 783-789 and Smc3 375-392 account for a kink in the Smc3 coiled coil region. Our analysis also highlights the conservation of the non-coiled coil region Smc1 781-797 and of residues that flank the non-coiled coil region Smc1 933-988 . Interestingly, Smc1 933-988 , as well as Smc3 1,065-1,083 , a remarkably non-conserved region within the NBD of Smc3, have previously been identified in proteomic experiments to contain numerous phosphorylation sites 63 , implying that some functions of cohesin might be regulated via post-translational modifications in these regions. Nonsense mutations and small in-frame deletions of 40 residues in Smc1 and Smc3 that have been identified in Cornelia de Lange Syndrome patients showed no obvious bias to conserved residues or regions (grey arrows, Fig. 5).

Discussion
It is now clear that structural and mechanistic insight into how cohesin interacts with DNA will be essential for understanding the 3D organization, cohesion and repair mechanisms of eukaryotic genomes and their correct spatiotemporal expression patterns. Understanding cohesin functions will likewise be of great importance for revealing the molecular cause of several human diseases and syndromes, some as widespread as trisomy 21 and spontaneous abortions 7 . Great progress has been made in obtaining structural insight into cohesin subunit interactions and domains 14,17,[36][37][38][39][40][41][42][43] , but the analysis of entire cohesin complexes has been limited by their inherent flexibility. To overcome this limitation, we have generated and structurally analysed engineered human cohesin. Our ability to obtain 3D models of these 'bonsai' complexes indicates that they are structurally more rigid than full-length cohesin. Further optimization of their design might pave the way for high-resolution single-particle EM, and similar approaches could be used for other SMC complexes.
Interestingly, our 3D models revealed that the Scc1-binding subunit SA1, which in rotary shadowed specimens can be seen attached to the ring and lying either inside or outside of it 13 , is located 'behind' Smc1 B and Smc3 B in the 3D model of bonsai cohesin, in close vicinity to their hinge and NBDs. The vicinity of SA1 to the NBDs may well be of functional relevance as previous observations have implicated orthologues and paralogues of SA1 in release of cohesin from DNA, which is thought to occur via dissociation of Scc1 from Smc3 (refs 13,14,18-20). In budding yeast, the SA1 orthologue Scc3 is required for release of cohesin from DNA 23 , and in human cells mitotic phosphorylation of the SA1 paralogue SA2 is necessary for Wapl-mediated release of cohesin from chromosome arms 29,32 . The relevance of SA1's proximity to the hinge is more difficult to interpret as it could have simply been induced by truncation of the coiled coils in bonsai cohesin. However, fluorescence resonance energy transfer (FRET) assays have detected interactions between the hinge and NBD-proximal subunits (Scc1 and Pds5) 64 , and by crosslinkingmass spectrometry we could also detect interactions of SA1 with hinge-proximal coiled coil regions in full-length cohesin 13 . SA1-hinge interactions may therefore also occur in wild-type cohesin and could regulate its functions.
The position in which SA1 and its binding partner Scc1 are typically assumed to be located, that is, directly between or next to the NBDs (depending on whether these are ATP bound and thus associated with each other or not), is occupied in our 3D model of bonsai cohesin by Pds5B. Consistent with the existence of numerous HEAT repeats in Pds5B, the density that represents Pds5B has an elongated and curved appearance. Interestingly, the location of this density is not only consistent with it binding to Scc1 (ref. 24), but the Pds5B density also directly contacts SA1 and Scc1 middle , as well as the NBDs of Smc1 B and Smc3 B . These contacts may be relevant for the observed positive and negative roles of Pds5 proteins in cohesion and raise the interesting possibility that these subunits could have direct effects on cohesin's ATP-binding and hydrolysing activities, which has been implicated in ring opening and cohesin loading. Our observation that Pds5B alone does not detectably alter the ATPase activity of cohesin does not exclude this possibility as the presence of other molecules (Wapl, DNA, the cohesin-loading complex) or posttranslational modifications may be required to reveal such an effect.
Direct physical interactions between Pds5B and Scc1, SA1 and Smc1 B and Smc3 B were also detected by crosslinking-mass spectrometry. This proximity map implies that the unstructured C-terminal tail of Pds5B contacts both SA1 and Scc1 in the region where SA1 clasps around Scc1 middle , despite the fact that this part of Pds5B is not essential for cohesin binding 13 . In budding yeast, point mutations in the N-terminal region of Pds5 prevent release of cohesin from DNA 23 . Interestingly, the corresponding N-terminal part of Pds5B appears to be closer to the Smc3-Scc1 N interface, raising the possibility that this part of Pds5 proteins regulates cohesin-DNA interactions either by influencing opening and closure of the DNA gate between Smc3 and Scc1, or-as discussed above-by modulating cohesin's ATPase activity. In support of the former possibility, it has been reported recently that Pds5B can inhibit closure of cohesin's exit gate in vitro 43 . The proximity between Pds5B and the NBD of Smc3 could also explain why Pds5 proteins are required for Smc3 acetylation [49][50][51] . In any of these cases, our observation that Pds5B occupies a similar bridge ('kleisin')-like position between the NBDs of Smc1 and Smc3 as Scc1 implies that Pds5 proteins may control cohesin-DNA interactions by directly regulating cohesin ring opening.
Before engineering bonsai cohesin complexes, we also explored if full-length cohesin could be used for 3D structure determination by negative staining EM. Unexpectedly, under these conditions cohesin was often observed in a closed rod-shaped conformation in which the coiled coils of Smc1 and Smc3 were juxtaposed, as opposed to their clear separation in the open ring conformation that is typically seen by rotary shadowing 12,13 . We suspect that this rod conformation is not simply a sample preparation or staining artefact, as our previous crosslinkingmass spectrometry experiments of full-length cohesin in solution had also detected numerous close contacts between the coiled coils, and because similar rod-shaped conformations have NBDs CC Hinge S886 S1065 S1067 S1074 S1077 S1081 S1083  previously been observed for eukaryotic and bacterial condensin complexes 12,54,56 . Interestingly, for the Smc-ScpAB complex in Bacillus subtilis it has recently been observed that ATP-dependent binding of DNA to the hinge can change the closed rod conformation to an open ring conformation 56 , and this change has been proposed to be required for topological entrapment of DNA inside all SMC complexes 56,65 . Our observation that human cohesin can exist in a closed rod conformation supports this hypothesis. Protein purification and characterization. Baculoviruses were generated and used to express proteins in sf9 cells 67 . All constructs were purified using tandem metal-chelate-affinity and FLAG-affinity purification as described 13 . In brief, cell lysates were generated by dounce homogenization and centrifugation in buffer A (25 mM KH 2 PO 4 pH 7.5, 500 mM KCl, 5% v/v glycerol, 2 mM MgCl 2 ) supplemented with 20 mM imidazole, 20 mM b-Mercaptoethanol, 0.05% v/v Tween-20, 0.5 mg ml À 1 PMSF and complete protease inhibitor (Roche). Cleared lysates were incubated with Ni-NTA agarose beads (Qiagen) for 90 min. Beads were washed with buffer A þ 20 mM imidazole, 20 mM b-Mercaptoethanol and 0.05% v/v Tween-20 and eluted with buffer A þ 0,01% v/v Tween-20 and 200 mM imidazole. Eluates were incubated with anti-FLAG M2 agarose beads (Sigma) and washed with buffer B (25 mM Hepes pH 7.5, 150 mM NaCl, 5% v/v glycerol, 2 mM MgCl 2 ) þ 0.01% v/v Tween-20. Complexes were eluted in buffer B þ 0.5 mg ml À 1 FLAG peptide þ 183 mM ATPgS (Jena Biosciences).

Methods
Lysates of sf9 cells expressing Pds5B 10HIS were generated by dounce homogenization and centrifugation in buffer C (25 mM Tris pH 7.5, 200 mM NaCl, 5% v/v glycerol) supplemented with 20 mM imidazole, 20 mM b-Mercaptoethanol, 0.05% v/v Tween-20, 0.5 mg ml À 1 PMSF and complete protease inhibitor. Cleared lysates were incubated with Ni-NTA agarose beads for 90 min. Beads were washed with buffer C þ 20 mM imidazole, 20 mM b-Mercaptoethanol and 0.05% v/v Tween-20 and eluted with buffer C þ 0.01% v/v Tween-20 and 200 mM imidazole. Imidazole was removed by subsequent dilution in buffer C followed by concentration or by further purification by size-exclusion chromatography on a Superdex 200 10/300 or 16/600 column (GE Healthcare). Freeze-thawed Pds5B was bound to FLAG-immobilized cohesin by incubation at a large molar excess. BSA was added at 0.5 mg ml À 1 to promote specific binding and monitor sufficient washing between Pds5B binding and elution from the FLAG beads. The Pds5B sample that was used to compare binding to bonsai dimers, trimers and tetramers ( Supplementary Fig. 5) contained a very substoichiomteric amount of Wapl.
Purified protein complexes were analysed by SDS-PAGE on 7.5% or 4-12% polyacrylamide Tris-Glycine gels followed silver staining or coomassie staining. Antibodies against Smc1 (Bethyl laboratories), Smc3 (ref. 13) and Pds5B 25 were used for analysis by immunoblotting. Detailed descriptions of ATPase assays, differential scanning fluorimetry and crosslinking followed by mass spectrometry are included in the Supplementary Methods.
Low-angle metal coating and EM. Glycerol spraying, low-angle metal coating and EM of full-length cohesin complexes was performed as described 13,68 . In brief, FLAG-peptide eluates were diluted twofold with spraying buffer (200 mM ammonium acetate and 55% glycerol), sprayed onto a freshly cleaved mica and mounted in a MED020 high-vacuum evaporator (Baltec). After vacuum drying, a Platinum layer of 1-nm and a 7-nm Carbon support layer were subsequently evaporated onto the rotating specimen at respective angles of 5-6°and 45-60°. Pt/C replicas were released from the mica on water, captured by freshly glowdischarged 400-mesh Pd/Cu grids (Plano), plasma cleaned and visualized by transmission EM using a LaB 6 equipped Tecnai 20 (FEI) operated at an accelerating voltage of 80 kV. Micrographs were recorded at a nominal magnification of 50,000 and a resolution of 2.18 Å per pixel. Well-spread cohesin particles with distinguishable coiled coils were selected from raw micrographs and analysed by ImageJ.
Density gradients. FLAG-peptide eluates containing cohesin (100-200 ml) were applied on continuous 15-40% (v/v) glycerol gradients (25 mM Hepes pH 7.5, 150 mM NaCl, 2 mM MgCl 2 183 mM ATPgS) with a buffered 10% glycerol cushion (100 ml). When indicated, mild stabilization was accomplished by a modified GraFix protocol 57 with 0.1% (v/v) glutaraldehyde in the 40% glycerol fraction and 1 mM N-(p-Maleimidophenyl)isocyanate (PMPI) in DMSO in the 15% glycerol fraction. Centrifugation was performed in a TH660 (Sorvall) or a SW60Ti (Beckmann) rotor at 45,000 r.p.m. for 16 hours. To reduce sample heterogeneity, the centrifuge temperature was lowered from 4°C to À 10°C briefly after the maximum speed was reached. Samples shown in Fig. 1d, Supplementary Figs 2, 6 and 7 were sedimented at 4°C. Fractions (150 ml) were collected manually from top-to-bottom by pipetting or semi-automatically from bottom-to-top with a needle coupled to a P1 pump and a Frac-920 fraction collector (GE Heatlhcare). Aspartate pH 7.9 was added at a final concentration of 10 mM to collected GraFix fractions to quench the crosslinking.
Negative staining and EM. Protein complexes were adsorbed to a continuous carbon film and stained with freshly prepared uranyl formate (2% w/v in doubledistilled water) on Quantifoil R 3.5/1 grids. Bonsai tetramer images were recorded at a Titan Krios (FEI) operated at 80 kV equipped with a 4K4K CCD detector (Eagle detector; magnification: 37,000; twofold binning, estimated 1.2 Å per pixel). All other images were recorded with a CM200 (Philips) FEG microscope operated at 160 kV and equipped with a 4K Â 4K CCD camera (Tietz Video Systems; magnification: 88,000; twofold binning; 2.51 Å per pixel). The defocus was adjusted manually between 1 and 3 mm.
Image processing. Particles were selected semi-automatically or automatically using cowPicker (in-house developed software planned for release to the wider community in the near future; B. Busche and H.S., unpublished data) yielding different sets of images for each complex (bonsai tetramer, B52,000 particle images; Pds5-bound bonsai tetramer, B113,000 particle images; Pds5-bound bonsai trimer, B82,000 particle images). Initial CTF correction was performed using cowEyes (M. Lüttich and H.S., unpublished data) and non-particle images and images with unsatisfying contrast and CTF features were removed (with remaining B33,500 particle images for bonsai tetramer; B50,000 particle images for Pds5-bound bonsai tetramer and B34,000 particle images for Pds5-bound bonsai trimer). Images were two times downsampled, filtered, normalized and used for 2D classification. Initial models were obtained using a resampling approach of 2D class averages with a subsequent 3D MSA and classification as described 69,70 or by a maximum-likelihood approach implemented in Simple Prime 71 . Initial models were refined using in-house developed cowEyes and 3D classification was performed in Relion 72 . Density validation was performed by removing the density in UCSF Chimera 73 and supervised classification 74 . Different densities of the obtained 3D models were cross-validated by 3D classification in Relion or by projection matching with distinct input 3D models and distinct classified subsets of images ( Supplementary Fig. 8). Reference-free recovery of densities was judged as sign of authenticity of these densities. Particle images were sorted for the presence (Pds5B-bound bonsai tetramer, B19,000 particle images; Pds5-bound bonsai trimer, B8,000 particle images) or absence (Pds5B-free bonsai tetramers, B31,000 particle images) of Pds5 and used to generate improved 3D models (bonsai tetramer 3D model, 10,582 particle images; Pds5B-bound bonsai trimer 3D model, 4,027 particle images; Pds5B-bound bonsai tetramer 3D model, 10,416 particle images). Output models of Relion refinement or 3D classification were subjected to projection matching to obtain 2D class averages for comparison to reprojections. By difference mapping, the main differences between the distinct complexes ( þ / À Pds5B and þ / À SA1) were identified and assigned to Pds5B and SA1, respectively. To do so, 3D models were aligned, equally downfiltered, normalized and subtracted. To define the absolute handedness, a random conical tilt approach 75 on negatively stained sample and cryo prepared sample was tested but failed possibly due to particle heterogeneity, pseudo symmetry and particle size. By comparison and fitting with the available X-ray models, we propose the handedness of the 3D models as illustrated.
Data availability. Density maps were deposited in the EM Data Bank under the accession codes EMD-4029 ( þ SA1, þ Pds5B), EMD-4030 ( þ SA1) and EMD-4031 ( þ Pds5B). Raw mass spectrometry data are available upon request. The software suite used for particle picking and analysis is still in the development phase and will be released in the near future. The authors declare that all other data supporting the findings of this study are available within the article and its Supplementary Information Files.