Crystal structure and interaction studies of human DHTKD1 provide insight into a mitochondrial megacomplex in lysine catabolism

DHTKD1 is a lesser-studied E1 enzyme belonging to the family of 2-oxoacid dehydrogenases. DHTKD1, in complex with the E2 (dihydrolipoamide succinyltransferase, DLST) and E3 (lipoamide dehydrogenase, DLD) components, is implicated in lysine and tryptophan catabolism by catalysing the oxidative decarboxylation of 2-oxoadipate (2OA) in the mitochondria. Here, we solved the crystal structure of human DHTKD1 at 1.9 Å resolution in binary complex with the thiamine diphosphate (ThDP) cofactor. Our structure explains the evolutionary divergence of DHTKD1 from the well-characterized homologue 2-oxoglutarate (2OG) dehydrogenase, in its preference for the larger 2OA substrate than 2OG. Inherited DHTKD1 missense mutations cause the lysine metabolic condition 2-aminoadipic and 2-oxoadipic aciduria. Reconstruction of the missense variant proteins reveal their underlying molecular defects, which include protein destabilisation, disruption of protein-protein interactions, and alterations in the protein surface. We further generated a 5.0 Å reconstruction of the human DLST inner core by single-particle electron microscopy, revealing a 24-mer cubic architecture that serves as a scaffold for assembly of DHTKD1 and DLD. This structural study provides a starting point to develop small molecule DHTKD1 inhibitors for probing mitochondrial energy metabolism.


INTRODUCTION
species in the mitochondria via radical formation from a catalytic intermediate. DHTKD1 is increasingly recognized as essential for mitochondrial function and energy production [18,19].
While structural studies have been carried out for PDH and BCKDH E1 enzymes from various organisms across the phyla, including human [25,26], only prokaryotic OGDHs have been crystallised. These include the apo structure of E. coli OGDH (ecOGDH) [27], as well as various structures of Mycobacterium smegmatis OGDH (msOGDH) complexed with active site catalytic intermediates [28][29][30]. In this study, we report the crystal structure of human DHTKD1 and a model of the human DLST catalytic core determined by cryo-electron microscopy. We also characterised disease-causing variants of DHTKD1 that destabilise the protein, or disrupt the interaction with DLST. protein surface via different trajectories (Fig. 1d). Importantly, the hDHTKD1 linker packs against two loop regions that are longer than the equivalents in ecOGDH and msOGDH (Fig. 1e).
To a lesser degree, hDHTKD1 is also structurally homologous to the E1 enzymes of human PDH and BCKDH (also known as 2-oxoisovalerate dehydrogenase) (z 30, 3.0-3.5 Å, 16%, [25,26]), which are heterotetramers built from two copies of two subunits (Fig. 1f). This contrasts with DHTKD1, OGDH and presumably OGDHL which are homodimers. PDH and BCKDH form a more compact shape, lacking several surface insertions to the α / β core that are unique to the DHTKD1/OGDH/OGDHL subgroup. These include the helical bundle at the N-terminus, and the β -and helical hairpins (aa 527-565, 606-630) within the α / β 2 domain (Fig. 1f, red ribbons). PDH and BCKDH structures also contain K + binding sites that play a role in enzymatic regulation (Fig.   1f, green spheres). We did not observe any difference density that suggests metal binding in the equivalent region of DHTKD1. Metal-dependent regulation is also featured in mammalian OGDH enzymes [32,33], mediated by Ca 2+ binding motifs unique to the OGDH N-terminus and a region equivalent to the DHTKD1 Δ 3 [34]. Again, these motifs are not present in prokaryotic OGDHs, or DHTKD1.

The DHTKD1 active site favours 2OA as substrate
Each DHTKD1 subunit in the crystal homodimer is bound with a ThDP cofactor (Fig. 1b, Compared to an apo E1 structure such as that of ecOGDH, our ThDP-bound DHTKD1 structure highlights 4 loop segments in the active site that undergo disorder-to-order transition during cofactor binding (Fig. 2a). Using nomenclature from ref [11] (Suppl. Fig. 4), these include: which contributes residues to engage with E2 enzyme for acyltransfer (e.g. His435). The conformations seen in our holo structure are similar to those of msOGDH structures bound with the post-decarboxylation cofactor conjugates [29] (Fig. 2b).
In one x-ray dataset, we observed omit map electron density at one active site of the homodimer that is not accounted for by the cofactor or any component of the crystallisation condition (Suppl. Fig. 5a). This density is adjacent to but disjoint from the ThDP cofactor, at a location partly overlapping the two conformations of post-decarboxylation intermediate seen in the msOGDH structures (PDB: 3ZHT, 3ZHU; [29]) (Fig. 2b). The size of this density feature can accommodate a C6 ligand such as 2-oxoadipate (2OA) without covalent linkage to ThDP. Although the observed ligand, likely co-purified with the protein, did not undergo enzymatic turnover, the keto carbon can be placed at 3.5 Å from the ThDP thiazolium C2 and hence compatible with the nucleophilic attack and subsequent decarboxylation (Suppl. Fig. 5b). While in good agreement with omit map, the 2OA model was not included in the deposited structure, in light of no further experimental evidence of its presence.
DHTKD1 and OGDH overlap to some extent in their in vitro reactivity towards 2OG and 2OA [13,35]. For example, soaking msOGDH crystals with 2OA and 2OG both yielded similar postdecarboxylation intermediates [29]. Nevertheless, hDHTKD1 turns over 2OA with 40-fold catalytic efficiency over 2OG [13]. The hDHTKD1 active site reveals several amino acids poised to interact with the substrate, but vary in sequence with OGDH orthologues. Two of them involve substitution to more polar residues i.e. Tyr190 (from Phe OGDH ) and Tyr370 (Phe OGDH ), while the other two to less bulky residues i.e Ser263 (Tyr OGDH ) and Asp707 (Glu OGDH ) (Suppl. Fig. 3).
Overlaying the two msOGDH post-decarboxylation intermediates (first and second conformers, sticks in Fig. 2b; [29]) onto the hDHTKD1 substrate pocket clearly explained how these substituted amino acids can stabilise catalytic intermediates generated from the longer 2OA substrate (Fig. 2c)

DHTKD1 preferentially interacts with 2OA in solution
We further explored the substrate preference of hDHTKD1 by mapping its metabolite interactome using MIDAS, a mass spectrometry-based equilibrium dialysis approach [36].

DHTKD1 and DLST form direct interactions
There is literature evidence that the E1 and E2 components of 2-oxoacid dehydrogenase complexes interact directly as a binary subcomplex, in the absence of E3. For some E1 enzymes such as OGDH and PDH, the N-terminus is known to be important for the direct interaction with E2 [38,39] and E3 [40], although this region is notably different for DHTKD1. For example, the hDHTKD1 precursor encodes a mere 50-aa segment before the first α -helix of the structure, while the hOGDH equivalent region is longer (121 aa) and contains two DLST-binding motifs [38] not preserved in DHTKD1 (Suppl. Fig. 1b). This suggests that the manner in which DHTKD1 and OGDH (E1) interact with DLST (E2) could be different.
Human DLST as a precursor protein is structurally composed of ( did not yield a stable three-way complex in SEC, as was the case shown for OGDHc previously [38].
Similar DHTKD1-DLST complex can also be formed by co-expression in the baculo Sf9 cells.
When expressed alone in Sf9, the hDLST 68-453 protein is highly prone to degradation, with a significant proportion fragmenting into two halves (Suppl. Fig. 7a). When the DHTKD1 and DLST proteins are co-expressed, hDHTKD1 45-919 co-purified in SEC together with both the hDLST 68-453 intact protein and the C-terminal fragment (containing the catalytic core), while the N-terminal fragment (containing the lipoyl domain and linker) was not part of this complex (Suppl. Fig. 7b). This suggests that the DLST N-terminal fragment alone is not sufficient to interact with DHTKD1, although this DLST region was previously mapped to be interacting with the binding motifs at the hOGDH-unique N-terminus [38]. Altogether, our data reinforce the notion that DHTKD1 and OGDH interact with DLST differently.

Insight into complex assembly from cryo-EM and SAXS studies
To provide a structural context for the DHTKD1-DLST interactions, we attempted single particle cryo-electron microscopy (cryo-EM) on the reconstituted binary complexes co-expressed in E.  [42] and bovine BCKDH E2 [43]. Upon close inspection, we observed in the E. coli co-expressed complex more heterogenous particles, some of which reveal extra density emanating from the cubic core to approximately 10-20 Å (Suppl. Fig. 8a).
This was not the case in the Sf9 co-expressed complex, where particles are more homogenous and contain only cubic cages (Suppl. Fig. 8b). To date, we managed to collect a dataset of 33072 particles from the latter sample, and generated a 3D reconstruction at 5 Å resolution (Suppl. Fig.   9). This shows 24 DLST C-terminal catalytic domains assembled as eight trimer building blocks into a cubic cage with 432 symmetry (Fig. 3f). Our EM map allows tracing of a humanised DLST model (aa 219-453 of human DLST) based on the E. coli structure (PDB 1E2O; 60% identity) [41].
Considering the sequence conservation, the catalytic cores of E. coli and human DLST display essentially identical topology and symmetry along 2-, 3-and 4-fold axes (Fig. 3f). In this assembly, all 24 C-terminal catalytic domains have their first residue (aa 219) exposed to the surface of the core (Fig. 3f, inset), presumably projecting the adjacent inter-domain linker outwards from the core in order to deliver the N-terminal lipoyl domain for engagement with E1 and E3. We reasoned that the additional density protruding from the core in some of our cryo-EM grids (Suppl. Fig. 8a)  are clustered in the α /β2 domain (Fig. 1a,c). From >100 DHTKD1 and OGDH orthologues surveyed, the aa 715 position is invariantly Arg, while aa positions 305 and 777 are also highly conserved (82% and 93% respectively) (Fig. 4a). None of these residues directly affect the conserved catalytic machinery common to the 2-oxoacid dehydrogenase family.
To understand their putative biochemical defects, we reconstructed the 7 DHTKD1 missense mutations recombinantly. All hDHTKD1 45-919 variants were expressed as soluble protein to similar level as wild-type (wt), with the exception of p.L234G and p.S777P which showed significantly lower yields and propensity to degradation, suggesting these variant proteins are misfolded, compared to wt (Fig. 4b, Suppl. Fig. 11). Leu234 is located at the protein centre ~20 Å from cofactor site (Fig. 4e), and the p.L234G change introduces a smaller side-chain thereby leaving a cavity at the hydrophobic core (Fig. 4f). Ser777 is partially exposed to the protein surface (Fig. 4e), and the p.S777P change introduces a proline side-chain that likely disrupts hydrogen bonds with neighbouring residues (Fig. 4g).
However, p.P773L exhibited a significantly reduced melting temperature (ΔTm = -5.2 °C), suggesting that while expressed as soluble protein this variant is thermally more labile than wt (Fig. 4c). Pro773 forms a bend for the surface exposed loop which connects the α/β2 and α/β3 domains and also harbours the abovementioned Ser777. Replacing Pro773 with Leu likely alters the structural integrity of this loop (Fig. 4h) and could affect protein folding. The observation of two destabilising mutations within this one loop region strongly implicates its importance in the functioning of DHTKD1.
Arg715 is located at the two-fold axis of the homodimer, and together with Arg712 forms a salt bridge network with Asp677 of the opposite subunit (Fig. 4h) Fig. 12). Nevertheless, when assayed in our co-expression and affinity pull-down, the hDHTKD1 45-919 p.R715C variant has significantly reduced ability to bind hDLST 68-453 directly ( Fig. 4d, Suppl. Fig. 13), compared to wt (Fig. 3b). Therefore the effect of p.R715C substitution could be transmitted from the dimer interface to engagement with E2 through loop 3. As control, hDHTKD1 45-919 bearing the p.R455Q or p.P773L substitution (both located at protein exterior) interacts with hDLST 68-453 to similar extent as wt.
Our data did not reveal any discernible defects on protein stability or interaction with DLST in vitro for the p.Q305H, p.R455Q and p.G729R variants, the latter two being found in the majority of reported cases of 2-aminoadipic and 2-oxoadipic aciduria. These results imply that additional functions or unknown binding partners could be involved in the OADHc. Future efforts can be focussed on studying their in vivo impact using patient-relevant cells.

CONCLUSION
DHTKD1 is emerging as a key player in mitochondrial metabolism through its influence in lysine metabolism, energy production, and ROS balance. The structure of DHTKD1 presented here provides the first template for the rational design of DHTKD1 small molecule modulators, to probe the enzyme's role in these mitochondrial functions and the associated disease states.
DHTKD1 exhibits key structural differences from others E1 enzymes particularly OGDH, providing a molecular basis for the subtle difference in substrate specificity and protein-protein interaction despite their close homology. These features would likely be exploited by the DHTKD1-specific modulators.
We have reconstituted the DHTKD1-DLST complex in vitro, and demonstrated for the first time that complex formation is disrupted in some disease-causing variants, likely via indirect (e.g. destabilising DHTKD1 to reduce its steady state level) or direct (e.g. altering the binding interface of DHTKD1) mechanisms. These data underscore the importance of DHTKD1

Expression and purification of human DHTKD1 and DLST
Site-directed mutations were constructed using the QuikChange mutagenesis kit (Stratagene) and and processed using the CCP4 program suite [46]. hDHTKD1  crystallized in the primitive space group P1 with two molecules in the asymmetric unit. The structure was solved by molecular replacement using the program PHASER [47] and the E. coli OGDH structure (PDB code 2JGD) as search model. The structure was refined using PHENIX [48], followed by iterative cycles of model building in COOT [49]. Statistics for data collection and refinement are summarized in Table 1.

DHTKD1 enzyme assay
The enzymatic activity assay was performed in a buffer containing 35  SAXS data were processed and analyzed using the ATSAS program package [52] and Scatter (http://www.bioisis.net/scatter). The radius of gyration Rg and forward scattering I(0) were calculated by Guinier approximation. The maximum particle dimension Dmax and P(r) function were evaluated using the program GNOM [53].

Solution Analysis
Analytical gel filtration was performed on a Superdex 200 Increase 10/300 GL column or T m , were determined as described [54] and final graphs were generated using GraphPad Prism (v.7; Graph-Pad Software). Assays were carried out in technical triplicate.

MIDAS protein-metabolite screening
Protein-metabolite interaction screening using an updated MIDAS platform was performed similar to Orsak, et al [36]. FIA-MS spectra collected from MIDAS protein-metabolite screening was qualitatively and quantitatively processed in SCIEX OS 1.5 software to determine relative metabolite abundance by integrating the mean area under the curve (AUC) across technical triplicates.

EM data collection and processing
A total of 619 dose-fractioned movies were beam-induced motion correction using MotionCor2 [55] with the dose-weighting option. CTF parameters were determined by ctffind4.1 [56]. The DLST particles were auto-picked using Relion 3.0 [57]. A total of 165739 particles were then extracted with a box size of 344 pixels rescaled to 172 pixels (2x binning). Particles were filtered for homogeneity using rounds of 2D, 3D classification with Relion 3.0 [57]. 33,072 particles where re-extracted un-binned and submitted to 3D auto-refinement with O (octahedral) symmetry imposed resulting in a 5 Å map after post processing (Suppl. Fig. 10a). 3D refinements were started from a 50 Å low-pass filtered version of a 3D class map from previous step. Global resolution was estimated by applying a soft mask around the protein complex density and based on the gold-standard (two halves of data refined independently) FSC = 0.143 criterion (Suppl. Fig. 10b). Local resolution was calculated with relion 3.0 [57] (Suppl. Fig. 10c).

EM Model building and refinement
To fit a template to the final map the E.coli DLST orthologue structure (PDB 1SCZ) was used.
The sequence was humanized and residues truncated to the alpha carbon using Chainsaw [58].
The oligomeric structure was docked into the O symmetry full EM density map using Molrep followed by one round of refinement using refmac5 through CCP-EM [59]. Fig. 5 and Suppl. Fig.   10 were made using Chimera [60].