Cryo-EM Structure-guided Selection of Computed Ligand Poses to Enhance Potency in MTA-synergic Inhibition of Human Protein Arginine Methyltransferase 5


 The potential of using cryo-electron microscopic (cryo-EM) structures of 2.5-4.0 Å resolutions for structure-based drug design was proposed recently, but is yet to be materialized. Here we show that a 3.1 Å cryo-EM structure of protein arginine methyltransferase 5 (PRMT5) is sufficient to guide the selection of computed poses of a bound inhibitor and its redesign for much higher potency. PRMT5 is an oncogenic target and its multiple inhibitors are in clinical trials for various cancer types. However, all these PRMT5 inhibitors manifest negative cooperativity with a metabolic co-factor analog --- 2-methylthioadenosine (MTA), which is accumulated substantially in cancer patients carrying defective MTA phosphorylase (MTAP). To achieve MTA-synergetic inhibition, we obtained a pharmacophore from virtual screen and synthesized a specific inhibitor (11-2F). Cryo-EM structures of the 11-2F/MTA-bound human PRMT5: MEP50 complex and its apo form together showed that the inhibitor binding in the catalytic pocket causes a shift of the cofactor-binding site by 1.5 – 2.0 Å, disfavoring cofactor-binding and resulting in positive cooperativity between 11-2F and MTA. Coarse-grained and full-atomistic MD simulations of the ligands in their binding pockets were performed to compare computed poses of 11-2F and its redesigned analogs. Three new analogs were predicted to have much better potency. One of them, after synthesis, was ~4 fold more efficient in PRMT5 inhibition in the presence of MTA than 11-2F itself. Computational analysis also suggests strong subtype specificity of 11-2F among PRMTs. These data demonstrate the feasibility of using cryo-EM structures of near-atomic resolutions and computational analysis of ligand poses for better small molecule therapeutics.


INTRODUCTION
Single particle cryo-electron microscopy (cryo-EM) has undergone a so-called "resolution revolution" and has produced hundreds of cryo-EM structures of 2.5 -4.5 Å resolutions in the past 10 or so years [1][2][3][4][5][6][7][8] . Structures of many macromolecules complexed with small molecule agonists, antagonist, or modulators have been published in near-atomic resolutions 2,9 . Although single particle cryo-EM often lags behind crystallography in resolutions, it was expected to open a new phase in structure-based drug design, a process that has been successful with high-resolution crystal structures and fragment-based drug design. Although truly atomic resolutions (~1.2 Å) for high-symmetry complexes were achieved recently 10,11 , most of cryo-EM structures of low-or no-symmetry complexes are resolved to 2.5 to 4.5 Å. With such developments, the potential of cryo-EM structures for structure-based drug design became quite promising as proposed by multiple groups, and is tantalizing for many important biological or pharmacological targets; but it is still yet to be fully materialized, in part because most cryo-EM structures (>80%) are not resolved to 1.0 -2.5 Å, where crystal structure-based drug design has been successfully 2 . With development of advanced computational tools for analysis of ligand configurations (poses) 12,13 , it became interesting to test the combination of virtual screening, molecular docking and energy minimization of ligands with near-atomic resolution cryo-EM structures to enhance ligand potency. We started this direction in 2015 and selected human protein arginine methyltransferase 5 (PRMT5) as our first target.
As the only type II PRMT enzyme well studied so far, PRMT5 is responsible for catalyzing symmetric arginine dimethylation of histone proteins (e.g., H4R3, H3R8) 36 and a wide array of nonhistone proteins, such as p53 and NF-kB 37,38 . It utilizes S-adenosyl-L-methionine (SAM) as a cofactor to donate a methyl group and catalyzes formation of an ω-N G -monomethyl arginine (MMA), yielding an S-adenosyl-L-homocysteine (SAH) as a by-product. A second SAM molecule is consumed to generate ω-N G ,N' G -symmetric dimethylarginine (SDMA) as the final product in a distributive fashion [39][40][41] . The enzymatic activity of PRMT5 requires a crucial partner named methylosome associated protein 50 (MEP50) 36,42 , which dramatically augments its activity via increased capability of substrate recognition and presentation 43,44 . Structural studies of PRMT5 from C. elegans, X. laevis, and H. sapiens 39,43,45 all revealed a conserved triosephosphate isomerase (TIM) barrel at the N-terminal half, a Rossmann-fold domain for cofactor binding and a βsandwich domain at C-terminal part for substrate binding and dimerization. MEP50, a 7-WD40 repeat β-propeller protein, binds PRMT5 via contacts with the TIM barrel domain, constituting a hetero-octameric complex with four PRMT5:MEP50 heterodimers arranged in a head-to-tail fashion of an apparent D2 symmetry 39 . The octamer may form larger complexes with other binding partners, such as COPR5 46 , SWI/SNF 47 , pICln 48 , Riok1 49 , Menin 50 , etc., in order to methylate a broad spectrum of substrates 39 . Suppressing the elevated PRMT5 activity specifically in cancer cells thus constitutes a potential treatment.
Currently, all PRMT5-targeting small molecule inhibitors fall into three classes based on their sites of action. The first class is cofactor-site competitive inhibitors 51 , which are SAM analogs whose ribose and adenine moieties are strongly favored by the Rossmann-fold domain of PRMT5. The second class is substrate-site competitive inhibitors with high potency and selectivity 52 . The third class is allosteric modulators, which interfere indirectly with PRMT5's canonical binding sites 53 .
Many PRMT5 inhibitors were discovered and designed in the past years 54 . Two molecules, JNJ-64619178 of the first class and GSK-3326595 of the second class, have entered clinic trials for multiple cancer types since 2018 51 . A new molecule, MRTX9768 (from a conference report), had some advantages than these two in preclinical studies (unpublished) 55 . Moreover, a nucleosidebased covalent inhibitor could attack a unique cysteine residue (Cys449) to form a covalent adduct in the SAM-binding site 56 , but it faces significant difficulty in achieving high specificity among SAM-binding enzymes. On the other hand, although class 2 inhibitors work differently from the class 1, their function relies on SAM or its analogs 57 , but may not work well in the presence of 5'-methylthioadenosine (MTA). MTAP-/-cancer cells accumulate MTA by >10 folds in cytosol.
MTA competes with SAM and reduces the efficacy of many SAM-dependent class 2 inhibitors.
For the broad spectrum MTAP-/-cancers, class 2 inhibitors working synergistically with MTA are desired to suppress or kill them specifically.
Available structural data are insufficient for understanding potential interactions between a substrate-site inhibitor and MTA. The crystal structure (PDB: 3UA4) of the apo C. elegans PRMT5: MEP50 complex determined a decade ago differs significantly from that of the human complex 45 , whereas the apo human complex has been resistant to crystallization. The first co-crystal structure of the human complex with a SAM analog A9145C and a histone H4 substrate peptide (PDB: 4GQB) reported in 2012 was a milestone and triggered competition in crystal structure-based drug design. 39 However, all published crystal structures of human PRMT5 were obtained in complex with one or more ligands 37 and did not reveal the basis for their lack of synergy with MTA.
The need for PMRT5 inhibitors that work more potently with elevated MTA levels (called MTAsynergistic inhibitors) in cancer cells asks for a new strategy. Although chemical screen against the PRMT5/MTA is a conventional strategy that may provide a practical, yet costly, approach (unpublished data from Mirati Therapeutics) 55 , we asked whether the combination of near-atomic resolution cryo-EM structures with computational analysis of compound poses might constitute a different and less expensive approach that would allow us to search and select the right (or most probable) pose of a target compound and use it to guide the design of high-potency binders. The first cryo-EM structure of PRMT5:MEP50 in complex with a cofactor analog (dehydro-sinefungin) was solved at a 3.7 Å resolution in 2018 58 , which would be insufficient. We reasoned that structures of near 3.0 Å resolutions may have sufficient structural constraints for large ligands and allow proper docking and selection among possible poses optimized by computational analysis 2,9 .
The top candidate(s) will be accurate enough for structure-based redesign and reselection, ultimately leading to high potency binders. In this paper, we applied this strategy to a new PRMT5 inhibitor initially discovered by virtual screen, and were able to determine its pose with high accuracy, unravel the chemical basis underlying its synergy with MTA, design and reselect new compounds, and synthesize one of the designed compounds to confirm its higher potency. Our results suggest that this non-crystallographic strategy works, and can be applied more broadly for cryo-EM structure-based molecular therapeutics.

Virtual screening leads to selective inhibitors of PRMT5
Pharmacophores differing from a reported inhibitor (EPZ015666) the Cambridge database were used to screen against PRMT5 and identify a new pharmacophore with comparatively better binding efficiency. The resulting pharmacophore was modified manually based on the crystal structure of the catalytic pocket to increase non-covalent interactions between the inhibitor and the residues lining the binding pocket, and was synthesized in the lab (Fig. 1A). The manual optimization of the molecular design was guided by structure-activity relationship (SAR) and docking of the resulting compound in the PRMT5 catalytic pocket. Compounds with a better binding network were synthesized and tested for in vitro inhibition. Recombinant enzyme (PRMT5:MEP50) was expressed on sf9 cells and purified to biochemical homogeneity (Supplementary Fig. S1). The compounds at varying concentrations were incubated with 100 nM enzyme, 2.0 µM H4-histone peptide substrates and 10 µM SAM. The inhibition of enzyme activity (methylation of H4 peptide) was monitored by measuring luminescence from the MTase-Glow detection reagent. One of the compounds, 11-2F, inhibited PRMT5 strongly (IC50 = 0.73 ± 0.2 µM). Surface plasma resonance (SPR) studies detected that 11-2F binds to the PRMT5:MEP50 complex with an apparent affinity of 13.6 µM without MTA, but its affinity increased drastically to ~82 nM in the presence of 25 µM MTA ( Fig. 1C). How to increase the potency of 11-2F further is a challenging question. Because the 11-2F-bound PRMT5: MEP50 complex could not be crystalized well, it made a good candidate for obtaining near-atomic resolution cryo-EM-structures and testing the proposed drug-design strategy by selecting compound poses from computation analysis.  Table 1). The estimated resolution of the cryo-EM map is in good agreement with the level of visible structural details (Figures 2A-D). Local resolutions estimated by ResMap vary in 2.4 -4.1 Å (Fig. 2B). After an X-ray structure (PDB: 6CKC ) was modeled and fitted into the cryo-EM map by real-space refinement ( Fig. 2A), there were clear structural differences between the cryo-EM-map-based model (green) and the crystal structure (orange; Fig. 2C). Fig. 2D shows the clearly resolved densities corresponding to side chains within two neighboring α-helixes. More differences occur within the MEP50 domain at the periphery of the complex (Fig. 2C), where all β-sheets were well resolved ( Fig. 2A). Because MEP50 will not play a significant role in our structure-based analysis of the compound, we will not discuss these structural differences further.
To our satisfaction, the 3.1 Å cryo-EM map clearly shows the density expected for an inhibitor that binds inside the catalytic site of PRMT5 (Fig. 2E, where the ligand model is not optimized yet). The fact that the shape of the density can be accounted for by the inhibitor (stick model in Fig. 2E) suggests that the cryo-EM map at ~3.0 Å probably be suitable for modeling more accurately the binding pose of the compound. In addition, the cryo-EM map shows clear density for MTA, which is smaller than 11-2F (Fig. 3A), but accounts for the density in the cofactor-binding pocket quite well. These structural features in the cryo-EM Coulombic potential map triggered us to test if it is feasible to perform accurate modeling by molecular docking and energy minimization, pose selection, and structure-based design of 11-2F at ~3.0 Å cryo-EM resolutions.

Feasibility of pose selection at the binding sites
As a positive control for the proposed strategy, we first analyzed the MTA-binding site because we have a crystal structure to check the quality of our results. But cryo-EM maps at 3.0 -4.0 Å resolutions may have significant uncertainty for small ligands 2 . We first asked whether the cryo-EM density corresponding to MTA is sufficiently good to distinguish its right g pose among various possible ones. Different poses of MTA can be generated by software packages for molecular docking, and be ranked relative to each other by minimized binding energy. If the selection of the binding pose based on the cryo-EM map is accurate, the resulted pose is expected to be very similar, if not identical, to that determined by X-ray crystallography at a higher reported resolution.
We first generated 25 million random poses of MTA and used AUTODOCK to introduce rotational freedom around all rotatable bonds and find the one of the lowest binding energy in each run 12,59,60 . AUTODOCK clustered these energy-minimized poses for 2,000 runs internally based on a RMSD threshold (Supplementary Fig. S3-1; Supplementary Table S3). The top pose in the 1 st cluster represented closely 89% of the 2,000 poses from random starting configurations, suggesting that it was heavily favored. The mean binding energy of the first cluster is significantly lower than the second one. Structural comparison found that the top pose of the first cluster differs significantly from the ones from other clusters (Supplementary Fig. S3-1). When the top 3 poses from the first cluster were compared with the cryo-EM density, the top one (colored green in  Table S5) based on the known X-ray structural models, and found that the top pose for each agrees well with the respective X-ray model, except minor differences at flexible tail regions ( Fig. 3D and 3E). These comparisons support our general strategy and argue strongly that the predictions from ligand docking and energy minimization in AUTODOCK are relatively accurate for ring-containing compounds, and can be further improved when cryo-EM densities of sufficient resolution are available to constrain them.  Table S5) was compared with the cryo-EM map (Fig. 3F), and then refined against the density by all atomistic molecular dynamics calculations (Figs 3G & 3H). It is very close to the final refined model (yellow vs. green, Fig. 3G) with only one rotation of the quinoline group by 180 degrees after refinement , suggesting that the cryo-EM map of 3.1 Å contains sufficient structural features of the ligand for selecting and refining the top pose generated by computational analysis (Fig. 3H), leading to an accurate model of the ligand, even though individual atoms in two multimember rings of the quinoline are resolved in the cryo-EM maps.
As a control, we also tested whether different software packages would generate similar or the same top poses for the same compound against the same structural model of the binding pocket. for their respective energy minimization processes. Our data so far showcase that the poses generated from computational analysis can be selected and refined against a cryo-EM density of the ligands at ~3.1Å resolution, producing a ligand-binding model that is fairly accurate.

Structure of the apo human PRMT5:MEP50 complex
Comparison of the 11-2F/MAT-bound cryo-EM structure with the known X-ray structures containing other ligands suggests that the flexible loop deduced from the structure of the C. elegans PRMT5 might be induced into an ordered state and contribute to the compound-binding pocket from the periphery. The flexible loop thus makes an important part for the 11-2F binding pocket, and could be used for structure-based drug design. Moreover, a structure of the apo complex would help verify that the densities assigned to MTA and 11-2F are real. Using a similar procedure, we obtained a 3.2 Å cryo-EM structure of the human PRMT5:MEP50 complex without any ligands ( Supplementary Fig. S4, and Fig. 4A). Parameters for data processing and molecular modeling are given in Supplementary Table 1. The local resolutions of the map vary (Fig. 4B), and the structural model fits the density well after real-space refinement (Fig. 4C), except the flexible loop region (Fig. 4E). As expected, no clear density in the two binding pockets for MTA and the substrate are visible even at a lower threshold level. There is very low density corresponding to most parts of the flexible loop, whereas the density corresponding to the loop was fairly strong in the map of the 11-2F bound complex (yellow, Fig. 4F). In the atomic model for the apo complex, residues 292-294, 304-307, and 312-329 were thus omitted. The modeling of the leftover residues also harbors a high level of uncertainty.
Since in the apo state, both MTA and 11-2F-binding pockets are empty, we compared the volume changes of the two pockets between the two cryo-EM structures. The MTA-binding pocket in the apo state is roughly 28% smaller in the estimated volume than that in the MTA/11-2F bound state.
The putative substrate-binding pocket in the apo state is larger because the flexible loop is disordered, leaving an open end. Such differences tell two important points. 1) MTA-binding induces a change in the binding pocket, probably due to induced fit. 2) The substrate-binding pocket in the apo state has a large volume so that a substrate or an inhibitor (11-2F) has significant freedom in testing different poses before becoming securely bound with the flexible loop making part of the pocket (Fig. 4F).

Structural basis for synergy between MTA and 11-2F
Our data in Figs 1B and 1C suggest positive cooperativity between MTA and 11-2F. Because it would be ideal to maintain this cooperativity when 11-2F is redesigned, we would like to understand its structural underpinnings. To do that, we aligned the MTA/11-2F-bound cryo-EM structural model with the MTA/H4-bound X-ray model in their N-terminal TIM barrel domains (bottom part in Fig 5A), and then compare the two catalytic domains. It is obvious that the whole cofactorbinding pocket is pushed upwards by ~2.0 Å when 11-2F is present in the substrate-binding site (Fig. 5B). Published data showed that MTA and H4 peptide had no positive cooperativity, suggesting that the shift of the catalytic domain (Fig. 5B) is probably the root cause for the positive co-operativity between 11-2F and MTA. This synergistic mechanism for MTA and 11-2F appears different from the positive cooperativity between SAM and H4 peptide (Fig. 5B) or between SAM and JNJ inhibitor (EPZ015666 in Fig.   5E) because the longer tail of SAM or its analogs (e.g. LLY283 in Fig. 5D) does not favor the shift of the co-factor binding pocket. Alternatively, the physical shift for MTA-binding (Fig. 5C) does not favor the binding of the H4-peptide substrate or the JNJ inhibitor so that they do not show positive cooperativity with MTA. From this line of thinking, a guiding principle for redesigning 11-2F would be to preserve the interactions between the quinoline of 11-2F and residues in the binding pocket, including Glu435, Glu444, Phe327, Trp579, etc., in order to retain the positive cooperativity with MTA.

Predicted subtype specificity of 11-2F among PRMTs
With the top poses predicted by computational analysis being very close to the final binding pose refined against the cryo-EM density (Fig. 3G), it was tempting to ask whether the same analysis of 11-2F among available structural models of six other PRMT proteins would reveal something unique for PRMT5 (Fig. 6). We first identified key residues coordinating 11-2F in the substratebinding pocket of the PRMT5 that are conserved among PRMTs (Supplementary Fig. S5 PRMT6 and PRMT7 clearly cannot accommodate the quinoline ring in the right position for the two catalytic Glu residues to interact. The PRMT2 is pretty poor because the alkylated indole ring of 11-2F has very limited interactions with its binding pocket. The main reason appears that the PRMT2 substrate-binding pocket is fairly shallow and cannot accommodate the two parts of 11-2F completely, making its docking energy fairly high (Fig. 6H). PRMT4 is probably the only one that might have a relatively good binding affinity because its Glu266 interacts with the 2-amine on the quinoline ring and its Tyr154 hydroxyl interacts with the N-atom inside the quinoline ring.
Other interactions next to the alkylated indole ring help stabilize the tail part of 11-2F (Fig. 6D).
The resulted docking energy in PRMT4 agrees with the predicted interactions in the binding pocket (Fig. 6H). These analyses predict that 11-2F is able to differentiate PRMT5 from other PRMTs due to the chemical differences among their binding pockets. It will be interesting to test if mutations in the binding pockets of PRMT4 can enhance 11-2F-binding.

Structure-based design of 11-2F for higher potency
The above analysis highlighted three principles that could be considered to enhance the binding affinity of 11-2F analogs to PRMT5. 1) It is important to retain the quinoline ring backbone to maintain the positive cooperativity with MTA. 2) The pi-stacking interactions of Trp579 and Phe327 with the quinoline ring could be enhanced for keeping the inhibitor properly oriented, which could be achieved by introducing small groups (-F, or -CH3, or -NH3 + ) to the ring. 3) The alkylated indole ring in the tail part of 11-2F is relatively flexible, and could be stabilized by introducing H-bonds or electrostatic interactions with the binding pocket, especially with the residues on the flexible loop. We used these principles to guide the design of dozens of different 11-2F derivatives, and utilized computational analysis to predict their most stable poses before ranking them and selecting the most potent ones. Three 11-2F analogs predicted to have higher potency introduce more interactions with the binding pockets ( Table 1; Supplementary Figs S6 & S7;   Supplementary Tables S6 & S7). From the constraints in the cryo-EM model (Fig. 6E), the quinazoline-corresponding parts of these compounds are in almost the same place and orientation as the quinoline of 11-2F. The predicted binding affinity for 11-9F is ~18 nM, and ~1.0 nM for HWIem2104 and 2109 in a similar pose (Table 1, right column). In the binding pocket, the 2,4di-NH2-quinazoline of 11-9F forms H-bonds with E444, E435, and S439, and its aromatic ring is sandwiched between the sidechains of Phe327 and Trp579. The backbone amino group of F580 also contributes to its binding. When 11-9F was synthesized and assayed, it inhibits PRMT5 enzyme activity 4-5 folds more potently than 11-2F does (Fig. 7E vs. Fig. 1A; more details in a manuscript being prepared by X. Yang. et al. ), in good agreements with predictions of the docking analysis and energy minimization (Fig. 7D).
Similarly, the other two compounds (HWIem2104 and HWIem2109) were picked out of different virtual designs, and their top poses are presented in Fig. 7B and 7C, showing the extra interactions of their tail portions with the residues on the flexible loop (Residues 292 -329) when the added 4-member ring interacts with Thr323 hydroxyl and the carbonyl group in the middle linker region interacts with Ser310. For HWIem2109, AUTODOCK Vina predicted its top pose close to that of 11-2F (Fig. 3F vs. 7C; Supplementary Fig. S6 and Supplementary Table S6). It contains an -NH-in the 6-member alkyl ring fused to the indole ring, which introduces an extra H-bond with Ser 310, albeit it was predicted to have the same potency as HWIem2104 (Fig. 7D). Given that the quinazoline ring in both compounds is constrained as 11-2F, we expect that the high potency for HWIem2104 and HWIdm2109 predicted by computational analysis is very likely a good indicator of successful increase in potency, which still awaits testing after chemical synthesis. Expectedly the results will probably be even better with cryo-EM maps of 2.0 -3.0 Å in resolution, whereby individual atoms of certain multi-member rings in ligands and some of the water molecules at the binding pockets will become recognizable.
The strategies we tested above appear to work well for the redesign of 11-2F based on accurate modeling of the protein-ligand interactions and the computational analysis to select three different ways out of virtual modifications to improve binding affinity. One of them, 11-9F, did show significantly better potency in enzyme inhibition assays, close to the predicted enhancement based on relative binding energy (Table 1; Fig. 7E). Because of the preserved head part (quinazoline) of the compounds, HWIem2104 and HWIem2109 are very probably going to follow the predictions and show even higher potency than 11-9F, even though they still need to be synthesized for experimental tests.

Catalytic mechanism of PRMT5 and the MTA-inhibitor cooperativity
Understanding of the catalytic mechanism of PRMT5 may also be helpful in development of potent inhibitors. So far, all well-characterized PRMT5 inhibitors in published studies require the nucleoside component of the cofactor SAM, indicating that they are either SAM analogs or SAMdependent molecules. Our work demonstrated that the cofactor binding induces a conformational change of the loop region (residues 292-329) by forming a small helix α1 (residues 320-329) which stabilizes the nucleoside via Tyr324, followed by the formation of a "lid" (residues 314-319) and a substrate-binding tunnel (residues 295-313), where Phe327 sandwiches the substrate ring against Trp579. These provide a physical connection to achieve positive cooperativity between MTA in the cofactor binding site and the inhibitor at the substrate-binding site. Foreseeably, different SAM analogs other than MTA might be developed to enhance the potency of the compound inhibitor, especially for those MTAP+/+ cancer cells 51 . Given that the MTA and 11-2F binding sites are next to each other, a chimera compound harboring the key groups of the two may be prepared for the same purpose.
Our work demonstrated the feasibility of the proposed strategy in using cryo-EM structures of ~3.0 Å resolutions and computational analysis of compound poses for drug design. It also led to the development of a novel class of substrate-competitive inhibitors that preferentially bind the PRMT5:MEP50 complex with MTA accumulation and may be used to pharmacologically exploit the PRMT5-related vulnerability in MTAP-/-cancer cells. We expect that the same or similar strategy be applicable to other cryo-EM maps of similar resolutions for drug design, which will expand our experimental capacity in developing new molecular therapeutics.

AUTHOR INFORMATION
Corresponding Authors ‡These authors contributed equally.

Notes
The authors declare no competing financial interest. The manuscript was written by W.Z. G.Y. and Q.X.J, and revised by Q.X.J. with input from all authors. All authors have given approval to the final version of the manuscript.

Data availability
Cryo-EM density maps of the apo and 11-2F-bound forms of human PRMT5:MEP50 have been deposited in the EM data bank under the accession codes of EMD-20764 and ***, respectively.
Atomic coordinates for the molecular models have been deposited in the protein data bank under accession code PDB ID 6UGH and ****, respectively.
Acknowledgements are available online.

Methods and Materials (details available online):
Expression, purification, and characterization of PRMT5:MEP50 complex.
Preparation of purified PRMT5:MEP50 complex was described in literature 39 . Details are available in the Supplementary Information. Protein samples were concentrated to 12 mg/mL in a buffer containing 10 mM HEPES at pH 7.5, 150 mM NaCl, 10% (vol/vol) glycerol and 2.0 mM DTT.

Enzymatic inhibition assay
Enzymatic inhibition activity of inhibitors was determined by a MTase-Glo™ methyltransferase assay (Promega Corporation, V7602). Details are presented in the Supplementary Information online.

Surface plasmon resonance (SPR) binding study
Binding affinity measurements were conducted in a Reichert2SPR system (Ametek) at 25 °C. Details in experimental setup, data collection and analysis are available online.

Cryo-EM grid preparation, data collection and analysis
Details are available online. Grids were prepared in a Vitrobot. Data were collected at both Florida State University (NIH SEM4 consortium) and National Cancer Institute's National Cryo-EM Facility (NCEF). A general protocol was used for data analysis with details specified in the Supplementary Information.

3D model building and refinement
The crystal structure of an inhibitor-bound (LLY-283) form of PRMT5:MEP50 dimer (PDB: 6CKC) was used as the initial molecular model and docked into the cryo-EM map in Chimera.
The model was subjected to real-space refinement in PHENIX 65 with secondary structure and geometry restraints, and manually adjusted in COOT 66             µM MTA was added into the running buffer to saturate the cofactor-binding pocket of immobilized PRMT5:MEP50 protein before the injection of analyte. The dissociation time for 11-2F binding in MTA containing running buffer was elongated to be 3 min to achieve complete dissociation.
Sensorgram data was processed using TraceDrawer software to calculate the equilibrium dissociation constant KD, association rate kon and dissociation rate koff values.

Cryo-EM grid preparation and imaging
The frozen protein sample was thawed and diluted by

Cryo-EM data processing for APO form (PRMT5:MEP50)
The movie stacks were motion-corrected, dose-weighted and binned by a factor of 2 using for each movie were estimated by CTFFind4 3 . Relion3.0 4 was utilized for particle picking, extraction, classification and refinement. LoG-based particle picking mode was utilized to select an initial set of particles as templates for the following template-based auto-picking procedure.
The auto-picked 597,002 particles were subjected to three rounds of 2D classification to remove junk particles, yielding a sub-set of 151,554 particles in 3D classification. An initial de novo 3D model was generated in Relion3.0. Two rounds of 3D classification were performed to identify distinct conformational states or sub-stoichiometric assemblies of the PRMT5:MEP50 complexes.
Unbinned particles from good 3D classes exhibiting 1:1 stoichiometry and high-resolution structural features were re-extracted. A cleaned dataset of 113,466 particles were auto-refined in Relion3.0, yielding a 3.9 Å resolution structure based on the gold-standard FSC (Fourier shell correlation) 5 with the 0.143 criterion with D2 symmetry imposed. The particle stack was then subjected to 2 iterations of CTF refinement and beam tilt correction, resulting in a 3.8 Å resolution structure. The resolution was improved to 3.6 Å by applying a soft mask around the entire protein complex during 3D auto-refinement. Bayesian polishing implemented in Relion3.0 and another iteration of CTF refinement further improved the final resolution to 3.4 Å. The refined particles were extracted and further refined in the cisTEM 6 which slightly improved the resolution to 3.17 Å. The final map was sharpened with a soft mask with an automatically calculated B-factor of -79.7 Å 2 . The local resolution of the apo form of PRMT5:MEP50 was assessed using ResMap 7,8 and colored in Chimera 9 .
Cryo-EM data processing for compound bound form (PRMT5:MEP50:11-2F/MTA) The movie stacks were motion-corrected, dose-weighted and binned by a factor of 2 using

3D model building and refinement
The crystal structure of an inhibitor-bound (LLY-283) form of PRMT5:MEP50 dimer (PDB: 6CKC) was used as the initial model and docked into the density map using chimera. The model was subjected to real-space refinement using PHENIX 10 with secondary structure and geometry restraints and manually adjusted in COOT 11 . A molecular dynamics (MD)-based optimization was performed using ISOLDE 12 with the technical assistance by Dr. Tristan Croll (CCPEM) 13 .
Overfitting and overinterpretation of the model were monitored by refining the model against one of the two independent half-maps and testing the refined model against the other map. The final structure was assessed in MolProbity and optimized to minimize clashes 14 . Cryo-EM data collection and modeling statistics are summarized in Supplementary Table 1 for the apo and inhibitor bound PRMT5 complexes.

Molecular docking analysis
Molecular docking was performed mainly by using AutoDock 4.