Multistate structures of the MLL1-WRAD complex bound to H2B-ubiquitinated nucleosome

We determined the structure of the human MLL1-WRAD complex (17) bound to a ubiquitinated nucleosome using single-particle cryo-EM. The Xenopus laevis nucleosome contained ubiquitin linked to K120 of histone H2B via a nonhydrolyzable dichloroacetone (DCA) linkage (see Materials and Methods and ref. 42). In addition, K4 of histone H3 was substituted with norleucine, a nonnatural amino acid which has been shown to bind more tightly to SET methyltransferases than the native lysine (43). Replacing H3K4 with norleucine was previously shown to drive tighter binding of the yeast homolog of MLL1-WRAD, COMPASS, to nucleosomes (44). The minimal fully active human MLL1 fragment containing the catalytic SET domain and spanning residues 3745 to 3969 was combined with full-length RbBP5, Ash2L, DPY30, and WDR5 as previously reported (17) to reconstitute MLL1-WRAD. The complex of MLL-WRAD bound to ubiquitinated nucleosomes was stabilized by cross-linking with glutaraldehyde (see Materials and Methods).

The MLL-WRAD complex was imaged with a Titan Krios equipped with a Gatan K3 direct electron detector (SI Appendix, Table S1). Our approach (see Materials and Methods) enabled us to resolve four distinct MLL1-WRAD states bound to the nucleosome at an overall resolution (Fourier Shell Correlation [FSC] 0.143 criterion) of 3.4 Å for state 1, 3.6 Å for state 2, 4.7 Å for state 3, and 4.3 Å for state 4 (SI Appendix, Fig. S1). The local resolutions of the MLL complexes are lower than that of the nucleosome in all four states, with the nucleosome at a ∼3.5- to 4.5-Å resolution, whereas MLL1-WRAD was resolved to a ∼5- to 9.5-Å resolution (SI Appendix, Fig. S2). Subunits from previously reported human MLL1-WRAD structures (38, 39) were placed in each map. State 4 fits well with a previously reported MLL1-WRAD structure (Protein Data Bank [PDB] ID 6KIU) (39), but also contained additional density, at a resolution of 7 Å to 9.5 Å, adjacent to the Ash2L SPRY domain that was not accounted for by the 6KIU model (Fig. 1 B and C). We hypothesized that this density might correspond to the DPY30 dimer and the more dynamic IDR and SPRY insertion regions of Ash2L (Fig. 1A), which were not resolved in previous reports (38, 39). The three additional MLL1-WRAD states identified in our study (SI Appendix, Fig. S1) and described in further detail below have not been previously reported.

To provide an experimental basis for molecular modeling of the extra density in state 4, we used cross-linking MS on the MLL1-WRAD complex to identify close contacts that could be used as restraints in model building. We cross-linked the MLL1-WRAD complex with BS3, a homobifunctional primary amine reactive cross-linker that has a theoretical maximum cross-linking distance constraint of 34 Å between Cα atoms of cross-linked lysine residues (45). Each MLL1-WRAD subunit contains seven or more lysine residues well distributed throughout the polypeptide, with one or more lysines found in each domain. Cross-linked MLL1-WRAD was digested with trypsin and analyzed by MS, and the resulting spectra were searched with the Plink2 cross-link search algorithm (46).

The final dataset comprised 167 lysine–lysine cross-links at a 2% false discovery rate (FDR) cutoff, with 40 interprotein cross-links and 127 intraprotein cross-links (Fig. 2). The diverse network of interprotein cross-links reflects the intertwined subunit architecture of the MLL1 complex. An exception is DPY30, which only has four cross-links to four residues in Ash2L, consistent with the position of DPY30 at the periphery of the MLL1-WRAD complex. Ash2L is the subunit with the largest number of cross-links, with 75 interprotein and intraprotein cross-links (Fig. 2 and SI Appendix, Fig. S3). There were 26 intersubunit cross-links for the MLL1-SET domain, reflecting this protein’s central position in the MLL1-WRAD complex. By contrast, there were far fewer cross-links to WDR5 and RbBP5, which are larger proteins and contain a greater number of lysines compared to the MLL1-SET domain. The small number of cross-links to these subunits could be due to the relative inaccessibility of lysines in the WD40 folds of both proteins, which reduces cross-linking efficiency (47). Alternatively, it is possible that there is greater mobility in these regions, which could similarly reduce the number of observed cross-links.

To fit missing portions of the model to the extra density in our maps, we first utilized two DPY30 crystal structures (PDB 4RIQ and 6E2H) (33, 37) and fit them to the density map using the cross-linking data as a guide (see Materials and Methods). Both crystal structures contain DPY30 bound to the Ash2L C-terminal SDI motif, which protrudes from the Ash2L SPRY domain (33). This portion of human WRAD was either absent or poorly defined in previous structures of the MLL1 complex bound to nucleosomes (38, 39). By placing the SPRY domain of crystal structure 6E2H (33) in our map, the Ash2L SDI motif (residues 503 to 523) and DPY30 from the crystal structure were near the unallocated density (Fig. 3C). A simple rotation of the SDI-DPY30 portion of PDB 6E2H relative to its SPRY domain positions these domains in the density (Fig. 3 and SI Appendix, Fig. S4B). The 15-residue linker (residues 485 to 500) connecting the Ash2L SDI-DPY30 region and the Ash2L SPRY domain was manually rebuilt and then refined using MDFF (see Materials and Methods). The resulting position of DPY30 and the Ash2L SDI motif fits well into the active state map and satisfies a cross-link (SPRY-K299 to SDI-Linker-K493) (Fig. 2B), consistent with the conformational rearrangement modeled here.

The ability of the Ash2L linker and the SDI motif protruding from the SPRY domain to adopt multiple conformations had been observed in previous structures (33, 37) (Fig. 3 and SI Appendix, Fig. S4B). In our model, the partially bent SDI motif most closely resembles the helix conformation observed in the crystal structure of DPY30 alone bound to the SDI motif (4RIQ) (37) but differs from that previously reported for the cryo-EM structure of MLL1 bound to a nucleosome, 6PWV (Fig. 3B and SI Appendix, Fig. S4B). We note that orientation of the SDI motif in our structure is dictated by the positioning of the DPY30 dimer, which is bound to the SDI motif. Whereas the reported position of the SDI motif in 6PWV could fit in the density of our active state map (Fig. 3B and SI Appendix, Fig. S4B), the resulting position of the bound DPY30 dimer does not fit the density in our map (Fig. 3B and SI Appendix, Fig. S4C). The position of DPY30 in our model is consistent with observed cross-links between DPY30 residues K35 and K40, which are located in a disordered region, and surface lysine K406 and K440 of the Ash2L SPRY domain (Dataset S1).

Placement of DPY30 and the Ash2L SDI motif left an unassigned density of about 7- to 9.5-Å resolution adjacent to Ash2L that could potentially correspond to the IDR region (residues 200 to 300) and the SPRY insertion (residues 400 to 440) (Fig. 1A). There is no high-resolution structural information on the Ash2L IDR, and previously proposed models of this region based on the Ash2L yeast homolog, Bre2 (38), do not fit our EM map well, nor does a recently reported model generated by AlphaFold2 (48). The 40-residue insertion in the Ash2L SPRY domain has been shown to contact DNA (38, 39), and a homologous stretch of yeast Bre2 similarly extends toward nucleosomal DNA in COMPASS (49, 50). We used these observations, combined with cross-linking restraints and MDFF, to model the IDR and SPRY insertion domain in our map (Fig. 4B).

The resulting model of Ash2L places the 40-residue SPRY insertion loop in the EM density at the periphery of the SPRY domain (Fig. 4C). The loop extends over 40 Å on the periphery of the SPRY domain to the nucleosomal DNA and contains an antiparallel beta sheet comprising residues 415 to 419 and 424 to 428 (Fig. 4C), whose homologous residues also form antiparallel beta strands in K. lactis Bre2 (51). This positions K419 and K421 near the DNA, consistent with the deleterious effect of substitution of these residues on DPY30-dependent methylation by MLL1 (52). The fit satisfies eight cross-links, including two intersubunit cross-links between Ash2L residue K437 and MLL1 residues K3870 and 3873 (Fig. 4C). As a next step, we placed IDR residues K207 and K244 in the unassigned EM density based on cross-links between these residues with multiple lysine residues within Ash2L, as well as to MLL1 (Fig. 4D). Specifically, Ash2L K207 cross-links to two adjacent IDR lysines (K205 and K218) and to MLL1 K3966 (Fig. 4 D and F), while Ash2L K244 cross-links to the SPRY-domain K311 and to MLL1 residue K3924 (Fig. 4D). Next, we included secondary structure based on Bre2 homology that predicts a beta strand comprising residues 247 to 250 that hydrogen bonds with the 424 to 428 strand in the SPRY insertion, forming a three-stranded beta sheet (Fig. 4B). Residues 241 to 246 were modeled as a loop and strand preceding the 247 to 250 beta strand based on sequence similarity to Bre2 (51). Based on secondary structure prediction (SI Appendix, Fig. S5A), we modeled Ash2L residues 200 to 210 as an α-helix and placed it in corresponding density (Fig. 4F). Notably, the proximity of the helix to DNA (Fig. 4A) is consistent with the observation that alanine substitutions of K205/K206/K207 markedly reduce MLL1 methyltransferase activity on nucleosomes (38). Finally, we fit Ash2L residues 253 to 278 to a short stretch of unassigned density extending to the periphery of the complex (Fig. 4 D and E), and in a position that satisfies the observed cross-link between K262 and K273 (Fig. 4E and SI Appendix, Fig. S4B). Given the limited resolution of the maps in this region (∼7- to 9.5-Å resolution), it is important to note that sidechains are included in the modeled regions solely for model completeness and cross-link validation, but are not experimentally determined based on map density.

The proximity of the Ash2L IDR and SPRY insertion region to nucleosomal DNA in our model (Fig. 4A) could account for the importance of DPY30 binding to MLL1-WRAD activity. Previous studies have shown that DPY30 greatly increases MLL1 methyltransferase activity on nucleosomes and that this increase in activity depends upon the Ash2L IDR (52). DPY30 forms multiple contacts with the IDR and the SDI helix, which help to stabilize folding of the IDR, as supported by NMR studies (52). This ordering of the IDR, in turn, would favor the predicted interactions with nucleosomal DNA, thus accounting for the importance of these residues to MLL1 activity on nucleosomes (38).

There was no apparent density corresponding to the Ash2L N-terminal PHD finger (residues 11 to 67), which has been shown to bind DNA (30, 31), or the atypical WH domain (residues 68 to 163). Previous cryo-EM studies of MLL complexes also did not locate these domains (38, 39). We therefore utilized the cross-links we identified involving the Ash2L N-terminal WH and PHD domains as restraints to guide docking of the crystal structure of these domains (PDB ID 3RSN) (30) to our active state model (Fig. 5A). Of the 22 cross-links we obtained (Dataset S1), the 12 cross-links between lysines within the WH and PHD finger domains were satisfied by the crystal structure. The additional 10 cross-links were used as restraints to dock the crystal structure on our MLL1-WRAD active state model. Eight of the cross-links connected to other regions of Ash2L (Fig. 5 B and C), and two cross-links connected to MLL1 (Fig. 5D) and RbBP5 (SI Appendix, Fig. S3B). In the highest-scoring model, all cross-links are satisfied except for one connecting K94 of the Ash2L WH and PHD finger domains to K112 of RbBP5 (SI Appendix, Fig. S3B), and the top four hits position the WH and PHD finger comparably, with a backbone RMSD of 0.336 ± 0.045 Å. The modeled position of the Ash2L WH and PHD finger domains (Fig. 5A) lies in a groove between MLL1 and the Ash2L SPRY domain. Importantly, the model places two lysine residues, K131 and K135, in contact with the sugar-phosphate backbone of the nucleosomal DNA (Fig. 5E), consistent with their importance to the binding of Ash2L to DNA (30). The location of Ash2L at the periphery of the nucleosome could facility interactions with transcription factors known to be recruited by this subunit (25). Given the positioning of the WH at the end of the DNA in the nucleosome core particle, it is possible that this domain could make additional contacts with DNA that extends beyond the minimal fragment associated with the nucleosome.

In addition to the active state MLL1-WRAD complex described above, we were able to resolve three additional states in our dataset (Fig. 6 and SI Appendix, Fig. S1 and Table S3). All four states have well-resolved density for RbBP5-WD40 (∼5- to 6-Å resolution), whereas other MLL1-WRAD subunits are less well resolved but seem to be in similar positions as compared to the active state model. In state 1 (Fig. 6A), only the RbBP5-WD40 domain shows strong density on the nucleosome. There is no density corresponding to the remainder of RbBP5 or to the methyltransferase subunit, MLL1. In state 2 (Fig. 6B), WDR5 is also well resolved (∼6- to 7-Å resolution), in addition to the strong RbBP5-WD40 density, while the remaining MLL1-WRAD density is poorly resolved. State 2 contains additional strong density contiguous with the DNA at SHL −6 to SHL −7 (Fig. 6B) that may correspond to a slight peeling away of the DNA from the histone core, although the density is not sufficiently well resolved to rule out alternative models that place additional protein in this location. State 3 contains equally well resolved density for RbBP5-WD40 and WDR5 as well as better-resolved MLL1-SET domain density (∼7- to 8-Å resolution) and low-resolution Ash2L density (∼7- to 9.5-Å resolution) (Fig. 6C). Interestingly, in this state, the DNA density at SHL −6 to SHL −7 is very weak, as is the MLL1 density as it approaches the nucleosome disk.

A comparison of the four states reveals little change in the docking of the MLL1-WRAD complex on the nucleosome, except for state 2, where the density indicates that WDR5 is somewhat tilted toward RbBP5 as compared to states 3 and 4 (SI Appendix, Fig. S6). This altered position could be a result of the absence or high mobility of the MLL1-SET domain, although further experiments would be needed to validate this hypothesis. Overall, we speculate that the four resolved states may represent an apparent progressive assembly of subunits from state 1, which contains RbBP5 and partial density for WDR5, through state 4 (the active state), which contains all five subunits (Fig. 6D). While it is not possible to distinguish ordered assembly from simple heterogeneity of states bound to the nucleosome, one possibility is that RbBP5 binding to the ubiquitin conjugated to histone H2B-K120 facilitates stepwise assembly of the full MLL1-WRAD complex on nucleosomes. While the RbBP5-WD40 domain contacts ubiquitin, the flexible C-terminal loops of RbBP5 (residues 330 to 475) contact all MLL1-WRAD subunits except DPY30. We speculate that the RbBP5 C-terminal residues could play a role in the assembly of MLL1’s active state on the nucleosome. Consistent with the model for stepwise assembly of the complex, a hydrodynamic analysis of the MLL1-WRAD complex identified multiple assembly states in the absence of nucleosomes, which included several different substates in addition to the full five-component complex (53).

A striking feature of reported structures of MLL1 complexes (38, 39, 52, 54) is their highly variable positioning on the nucleosome, in contrast with the conserved spatial relationship among the five subunits of the MLL1-WRAD complex. These orientations of the complexes on the nucleosome can be categorized according to the superhelical positions of RbBP5 and Ash2L/DPY30, which define the two ends of the complex. The positions of these subunits in the present and in previous studies are listed in SI Appendix, Table S4, and a subset are depicted in Fig. 7.

A comparison of structures suggests a role for the ubiquitin conjugated to H2B-K120 in positioning the MLL1-WRAD complex on the nucleosome. Monoubiquitination of H2B has been shown to increase the catalytic efficiency of MLL1-WRAD on nucleosomes by twofold in vitro (39). The active state MLL1-WRAD complex reported here (Fig. 1B and 6D) is docked on the H2B-ubiquitinated nucleosome in essentially the same orientation as a previously reported structure of MLL1-WRAD bound to H2B-ubiquitinated nucleosome [6KIV (39)], with RbBP5 at SHL −2.5 and Ash2L at SHL −7 (Fig. 7A). The MLL3-WRAD complex (6KIW), which contains a related methyltransferase subunit and the same four adapter proteins, is oriented in a similar manner on H2B-ubiquitinated nucleosome (39). By contrast, structures of MLL1 complexes bound to unmodified nucleosomes (38, 39, 54) show highly variable positioning of the MLL1 complex. Two structures of MLL1-WRAD bound to unmodified nucleosomes, 6PWV (38) and 6KIZ (39), have the Ash2L subunit positioned around SHL −7, as in the active complex, but with MLL1-WRAD markedly pivoted by about 35° (Fig. 7 A and B). This rotation places RbBP5 in contact with nucleosomal DNA at SHL −1.5, in addition to reducing the contact area with the nucleosome core. A subset of these particles adopted the active orientation (6KIW), which suggested that MLL1-WRAD could sample both positions in the absence of ubiquitin (39). A similar heterogeneity in positions resembling the active and peripheral complexes was also observed in a recent study of MLL1-WRAD bound to unmodified nucleosomes [7MBM and 7MBN (54)]. Taken together, these observations suggest that the interaction with ubiquitin favors the docking arrangement of the active complex (Fig. 7A).

The DPY30 subunit may also play a role in positioning MLL1 complexes on nucleosomes lacking H2B-K120Ub. Structural studies of a four-protein MLL1-WRA complex lacking both DPY30 and the Ash2L SDI helix (6W5I, 6W5M, and 6W5N) (52) showed the complex to bind unmodified nucleosomes in multiple orientations (SI Appendix, Table S4). While one class (6W5N) resembles that of the five-protein MLL1-WRAD complex bound to unmodified nucleosomes (6PWV), with Ash2L at SHL −7, the remaining classes bind in another orientation, with Ash2L at SHL −4.5 (6W5I and 6W5M) (Fig. 7C). This observation is consistent with a role for DPY30 in favoring binding between SHL −6 and SHL −7, as observed in structures of the complete MLL1-WRAD complex containing DPY30 (6KIV, 6KIX, 6KIZ, 6PWV, and this work). DPY30 likely exerts its effect through its interactions with the Ash2L IDR, which undergoes conformational changes in the presence of DPY30, as reflected in changes in NMR spectra (52). The proximity of the Ash2L IDR and SPRY insertion region to nucleosomal DNA in our model (Fig. 5), as well as in a previously proposed model (38), could account for the coupling of DPY30 binding to ordering of this region of Ash2L.

The MLL1-WRAD complex was reconstituted from proteins expressed in Escherichia coli (17, 62).

A human MLL1 construct comprising residues 3745 to 3969, as well as full-length human WDR5, RbBP5, and ASH2L proteins, were individually expressed in E. coli (Rosetta II, Novagen) and purified as described previously (62). DPY30 was expressed and reconstituted with the other four proteins as described in ref. 17.

Nucleosome core particles containing histone H3 with norleucine in place of lysine 4 as well as H2B ubiquitinated at K120 via a nonhydrolyzable DCA linkage were prepared as described (49). Briefly, histones H2A, H4, and the Widom 601 DNA were expressed and purified as described in ref. 63. Histone H3 containing norleucine in place of lysine 4 was expressed and purified as described in ref. 44. Ubiquitinated H2B was generated by expressing and purifying H2B K120C and ubiquitin K79C, which were cross-linked with DCA as described in ref. 42. A complete protocol for generating ubiquitinated nucleosomes can be found in ref. 64.

MLL1-WRAD complex was buffer exchanged into a cross-linking compatible buffer (50 mM Hepes pH 8.0, 200 mM KCl, 5 mM MgCl2, 0.1 mM (ethylenedinitrilo)tetraacetic acid, 5% glycerol, 1 mM tris(2-carboxyethyl)phosphine [TCEP]). Twenty micrograms of the buffer-exchanged complex was cross-linked with 2 mM and 4 mM BS3 for 4 h at 4 °C and frozen at −20 °C until further processing for liquid chromatography (LC)-MS/MS as described in refs. 65 and 66 with minor modifications. Briefly, cross-linked complexes were thawed at room temperature, and an equal volume of trifluoroethanol was added and incubated at 60 °C for 30 min to denature. The proteins were then reduced by the addition of 5 mM TCEP for 30 min at 37 °C followed by alkylation with iodoacetamide at a 10-mM final concentration for 30 min in the dark at room temperature. The sample was diluted 10-fold with 20 mM triethanolamine, and then digested with 2 μg of trypsin (Promega) overnight at 37 °C. The peptides were further purified on Sep-Pak C18 cartridge (Waters), dried, and resuspended in 5% acetonitrile/0.1% trifluoroacetic acid solution and then analyzed by LC-MS/MS. BS3-cross-linked peptides were analyzed on a Thermo Scientific Orbitrap Elite and Lumos with higher energy collision dissociation (HCD) fragmentation and serial MS events that included one FTMS1 event at 30,000 resolution followed by 10 FTMS2 events at 15,000 resolution. Other instrument settings included MS mass range greater than 1,800; m/z value as masses enabled; charge-state rejection: +1, +2, and unassigned charges; monoisotopic precursor selection enabled; dynamic exclusion enabled: repeat count 1, exclusion list size 500, exclusion duration 30 s; HCD normalized collision energy 35%, isolation width 3 Da, minimum signal count 5,000; and FTMS MSn AGC target 50,000. The RAW files were converted to mgf files and analyzed by the cross-link database–searching algorithm pLink2 under default settings: 1) up to three missed cleavages, 2) differential oxidation modification on methionine (+15.9949 Da), 3) differential modification on the peptide N-terminal glutamate residues (−18.0106 Da) or N-terminal glutamine residues (−17.0265 Da), 4) static modification on cysteines (+57.0215 Da), and 5) 2% FDR. All possible tryptic peptide pairs within 20 ppm of the precursor mass are used for cross-linked peptide searches. The cross-linked peptides were considered confidently identified if at least four consecutive b or y ions for each peptide were observed. Cross-links used for this study are in Dataset S1.

To form stable MLL1-WRAD–nucleosome complexes, 100 nM of modified nucleosome was mixed with 5× WRAD in EM buffer (20 mM Hepes, 300 mM NaCl, 2 mM TCEP, 0.1 µM Zinc chloride) and incubated for 1 h on ice. The sample mix was then cross-linked using fresh 0.05% (wt/vol) glutaraldehyde for 1 h on ice and quenched with 100 mM Tris for 2 h. The sample was concentrated to 0.5 mg/mL and directly applied to freshly glow-discharged 2/2 Cu Quantifoil grids in a Vitrobot Mark IV (Thermo Fisher Scientific) at 4 °C and 100% humidity. The sample was immediately blotted (3-s blot time) and flash frozen in liquid ethane.

All data were acquired at the Johns Hopkins Beckman Cryo EM Center on a Thermo Fisher Titan Krios G3 equipped with a Gatan K3 direct electron detector at a magnification of 21,500 in superresolution counting mode, corresponding to a pixel size of 0.529 Å/pix. A total of 5,091 movies were recorded using Serial-EM (67) using a varying negative defocus of 1.0 µm to 2.8 µm and recording 40 frames at 1.5 e/Ų per frame at 60 e/Ų total dose. Data acquisition was monitored and frequently evaluated using cisTEM (68). Movie stacks were aligned and down-scaled to a pixel size of 1.058 Å/pix (bin 1) using MotionCor2 (69), and contrast transfer function (CTF) correction was performed using Ctffind4 (70). The full dataset was then manually inspected, and 4,804 movie stacks were selected for further processing in Relion 3.0 (71).

Initial particle picking yielded 2,511,854 particles that were subjected to two-dimensional (2D) and 3D classification using threefold downscaled particle images (bin3) for faster processing speed and stronger signal-to-noise ratio, which removed junk particles and yielded 608,972 “good” particles that displayed MLL1 complex bound to the nucleosome. All “good” particles were reextracted at 1.058 Å/pix (bin1), consensus refined, and subjected to masked 3D classification using an Ash2L DNA interface mask, which yielded six unique classes displaying differences in Ash2L density and DNA topology(SI Appendix, Fig. S1). The 103,849 particles were subjected to one round of CTF refinement followed by Bayesian polishing as implemented in Relion-3.0. This procedure yielded a reconstruction after refinement of 3.33-Å resolution based on the FSC 0.143 (72) criterion termed state 1 (SI Appendix, Figs. S1 and S2), which was sharpened using an automatically calculated B factor of −70 Ų. Particles grouped in classes 2 and 5 from the initial masked classification were combined after confirming similar classes by visual inspection and reclassified into four classes (SI Appendix, Fig. S1). The resulting best class containing 134,528 particles showed flexible Ash2L and MLL1 density and density stretches associated with the DNA between SHL −6 to SHL −7 and was termed state 2. Particles were subjected to one round of CTF refinement followed by Bayesian polishing, yielding a 3.55-Å resolution reconstruction (according to the FSC 0.143 criterion) that was sharpened using an automatically calculated B factor of −86 Ų. The 56,526 particles grouped in class 3 from the initial masked classification were CTF corrected, particle polished, and subsequently refined to a final resolution of 4.65 Å (0.143 criterion). This reconstruction was sharpened using an automatically calculated B factor of −117 Ų (SI Appendix, Fig. S2) and termed state 3. The 66,449 particles grouped in class 4 from the initial masked classification were CTF refined, particle polished, and further refined to a final resolution of 3.95 Å (based on the 0.143 criterion). This class contained highly resolved nucleosome density but more poorly resolved MLL1 complex and Ash2L density. To further deblur the Ash2L components in this state, an additional 3D classification using only local searches was performed (SI Appendix, Fig. S1). This process yielded an overall better-resolved class (25,525 particles) that refined to 4.25 Å after postprocessing and was sharpened using a B factor of −75 Ų. This map (state 4) was used to guide Ash2L modeling as described below. We note that efforts to identify additional states using cryoDRGN (73) did not yield useful results.

Crystal structures of human ubiquitin (1UBQ), human WDR5^31-334 (3EG6), human MLL1^3814-3969-Ash2L^285-504-RbBP5^330-375 (5F6L), DPY30^50-96 (6E2H), human RbBP5^11-329, and X. laevis nucleosome (6KIU) were subjected to rigid-body fitting in the state 4 map using Chimera (74). These combined fitted structures allowed us to assign most of the density of the state 4 map, except for a region of continuous density (Fig. 1C) which was judged to correspond to Ash2L and DPY30 based on the similarity of the complex to solved structures of yeast homologs (49–51). The Ash2L model was computed through iterative rounds of cross-linking-based homology modeling utilizing available structures and manual model building, followed by MDFF using our state 4 map. Specifically, we use the previously resolved SPRY domain (287 to 504) from the MLL1^SET-ASH2L^SPRY-RbBP5^330-375 complex (PDB ID 5F6L) (36) and added residues 398 to 440 from a recently reported Ash2L homology model (38). All MDFF simulations were done in NAMD2.13 (75). The CHARMM36 force field (76, 77) was used for both protein and DNA. All simulations were carried out at 300 K in vacuum with a scaling factor of 1.0. To avoid potential structural artifacts that can arise from MDFF, secondary structure, chirality, and cis-peptide restraints were applied. For targeted refinement of Ash2L and DPY30, a 500 kcal/mol/Å harmonic restraint was applied on all backbone atoms except for Ash2L and DPY30. The cross-correlation coefficient was computed using the MDFF package implemented in VMD 1.9.3 (78). We then manually optimized the models fit to our experimental density, which generated the initial Ash2L model. We assigned adjacent helical tube density to DPY30 and performed automated fitting of the Ash2L-SDI/DPY30 dimer crystal structure (PDB 6E2H). In addition, our density allowed us to fully build the C-terminal helix of Ash2L (487 to 528). Ash2L residues 199 to 286 were built using MODELER (79) using cross-link restraints and underwent iterative rounds of manual model building guided by the EM density. The model was further refined through a 3.0-ns MDFF simulation (41). To improve the overall fit of Ash2L and DPY30, the state 4 map was filtered to 6.0 Å and was used as the template map. After refinement with MDFF, the cross-correlation coefficient between the overall structure and template map improved from 0.79 to 0.83. To correct any rotamer outliers that may occur from MDFF refinement, the model of the full complex underwent 1,000 iterations of minimization with secondary structure restraints using the Phenix geometry minimization module. Following, the model was iteratively refined with Phenix real-space refinement. Final statistics for the state 4 model are shown in SI Appendix, Table S1.

For the state 2 model, the positioning of WDR5, RbBP5^11-323, ubiquitin, and the nucleosome in state 4 were used as starting initial positions. WDR5 was then subjected to rigid-body fitting into the state 2 map using the Fit in Map function in ChimeraX. The entire structure was then subject to real-space refinement in Phenix. Final statistics for the state 2 model are show in SI Appendix, Table S2.

Docking of the human Ash2L WH–PHD domain (PDB 3RSN) guided by cross-linking data was carried out with HADDOCK 2.4 (80). Cross-links were used as unambiguous restraints between the Ash2L N-terminal residues and the remainder of the MLL1-WRAD complex. In HADDOCK, 5,000 rigid-body calculations were first carried out, and the top 1,000 structures underwent semiflexible refinement, resulting in 200 water-refined structures. Clusters were defined using fraction of common contacts with a cutoff of 0.60. All 200 structures were identified to a single cluster with a final HADDOCK score of −213.7 ± 1.4.