Structural basis for small molecule targeting of 
Doublecortin Like Kinase 1 DCLK1


 Doublecortin-like kinase 1 (DCLK1) is a bi-functional protein classified as a Microtubule-Associated Protein (MAP) and as a serine/threonine kinase that plays a critical role in regulating microtubule assembly. This understudied kinase is upregulated or mutated in a wide range of cancers. Knockdown studies have shown that DCLK1 is functionally important for tumour growth. However, the presence of tissue and development specific spliced DCLK1 isoforms and the lack of systematic evaluation of their biological function have challenged the development of effective strategies to understand the role of DCLK1 in oncogenesis. Recently, DCLK1-IN-1 was reported as a potent and selective DCLK1 kinase inhibitor, a powerful new tool to dissect DCLK1 biological functions. 
Here, we report the crystal structures of DCLK1 kinase domain in complex with two DCLK1-IN-1 precursors and DCLK-IN-1. Combined, our structural data analysis illuminates and rationalises the structure-activity relationship that informed development of DCLK1-IN-1 and provides the basis for DCLK1-IN-1 increased selectivity. We show that DCLK1-IN-1 induces a drastic conformational change of the N-lobe, which uncovered a new allosteric site. In addition, we demonstrate that DCLK1-IN-1 binds DCLK1 long isoforms with high affinity but does not prevent DCLK1 MAP function. Together, our work outlines the need for in-depth studies to rationally design of isoform-specific modulators and provides an invaluable structural platform to further the design of selective DCLK1 therapeutic agents.


Introduction
Doublecortin-like kinase 1 (DCLK1) is a large multi-domain bi-functional protein that belongs to both the protein kinase superfamily and the doublecortin (DCX) superfamily, within the microtubule associated protein (MAP) family. The tandem doublecortin domain (DCX1 and DCX2), located in the N-terminal region of DCLK1, drives its MAP function, while C-terminal region harbours an active serine/threonine kinase domain (Fig. 1a).
Beyond its established role in neurogenesis, DCLK1 has been identified as an intestinal and pancreatic stem cell maker 1-3 and growing evidence supports a role for DCLK1 in various malignancies. Many MAPs are known to play an important role in stabilization or destabilization of microtubules and changes in their expression levels are often associated with the development and progression of cancer 4,5 . Overexpression of DCLK1 has been reported in multiple cancers, including colon, pancreatic, renal cell carcinoma and rectal neuroendocrine tumours [6][7][8][9][10] . Aditionally, a relative high number of DCLK1 mutations have been identified in human gastric tumours 11,12 .
The regulation of DCLK1 is highly complex and takes place at many levels, however it is poorly understood. Multiple isoforms of DCLK1 exist, generated by alternative promoter usage and alternative splicing [13][14][15] . These isoforms differ drastically in their domain composition and hence in their biological function. The most relevant functional differences between these isoforms lies in the presence or abscence of the N-terminal tandem doublecortin domains that contribute to DCLK1's MAP function. Two main human isoforms have been reported in the literature, DCLK1-long (DCLK1-L), which contains the tandem doublecortin domains at the N-terminus and a kinase domain at the C-terminus and DCLK1-short (DCLK1-S), which lacks the tandem doublecortin domains and contains only the C-terminal kinase domain 14,16,17 . For each of the DCLK1-L and -S isoforms, two other splicing variants exist (a and b), which differ in the length and sequence of the C-terminal regulatory tail that immediately follows the kinase domain. Unfortunately, the nomenclature used to describe the various DCLK1 isoforms has not been consistent over the years, as recently highlighted in the Swiss-Prot and NCBI databases 8,17 , which has led to some discrepancy. Earlier studies in mice have demonstrated differential expresssion of spliced variants in embryonic and adult brains 16 .
However, a lack of appropriate detection tools has rendered the study of the isoform specific expression patterns difficult. The detection of DCLK1 proteins in cells has been done using commercial antibodies that specifically target a C-terminal sequence that is not present in all isoforms 3,9,17 . Hence, the overall expression profile of DCLK1 isoforms in various tissues, the systematic evaluation of their biological function, and their relative contribution to tumorigenesis have been largely overlooked and remain to be investigated. Adding to this complexicity, recent studies have highlighted that DCLK1 undergoes epigenetic regulation. Hypermethylation of the DCLK1 5'(a) promoter in human colon adenocarcinomas (CRCs), resulted in the loss of DCLK1-L isoform and the usage of an alternate b-promoter drove the expression of DCLK1-S isoform [17][18][19] . In addition, the cleavage of DCLK1 by the cysteine protease calpain has been reported in neurons, resulting in the release of the kinase domain from the tandem doublecortin domains, a mechanism proposed to drive the relocalisation of the kinase domain to the nucleus 20 . Corroborating these data, the nuclear transcription factor, Jun dimerization protein 2 (JDP2), was recently identified as a substrate for the zebrafish DCLK1 kinase domain 21 . Despite these studies, the calpain-dependent proteolytic regulation of DCLK1 is unclear and the functional relevance of DCLK1 kinase domain localisation to the nucleus remains to be established.
Developing targeted strategies for such a protein with isoform-specific functions and varying expression levels is therefore particularly challenging. The therapeutic effect of DCLK1 knockdown or silencing has been widely demonstrated in various cancer models [22][23][24] , highlighting DCLK1 as an attractive target. The dual functions of DCLK1, as a MAP and as a kinase, may each contribute to tumorigenesis differently 25 . However, studies that demonstrate the direct impact of independently targeting DCLK1 kinase activity or its microtubule polymerization/stabilisation function and the potential effect on tumour growth are lacking. The lack of selective DCLK1 small molecule modulators that specifically target either DCLK1 kinase function or MAP function in isolation has precluded our understanding of the relative contribution of each function to oncogenicity. Most studies reported to date have commonly used as DCLK1 kinase inhibitors two compounds based on a benzopyrimido-diazipinone scaffold, the LRRK2 compound, LRRK2-IN-1 and the ERK5 compound, XMD8-92, as they both showed significant off target activity against DCLK1 kinase function 26,27 . However, their pan selectivity (targeting not only DCLK1) makes the published studies difficult to interpret with respect to DCLK1 kinase function. Recently, the Gray group has generated a bespoke highly selective DCLK1/DCLK2 inhibitor (DCLK1-IN-1) derived from this benzopyrimido-diazipinone series, as a way forward for dissecting the contribution of DCLK1 kinase activity on DCLK1 tumoregenesis activity 28 .
Here, we present the crystal structures of DCLK1 kinase domain in complex with DCLK1-IN-1 and the two critical intermediates. Combined, our structural data rationalises the structure-activity relationship (SAR) behind DCLK1-IN-1 and provides the basis for DCLK1-IN-1 increased selectivity over ERK5 and LRRK2 kinases. Interestingly, binding of DCLK1-IN-1 induced a significant shift in the position of the N-lobe, which results in an opening of the ATP binding site and importantly uncovers a new allosteric site. Differences in the binding kinetics of DCLK1-IN-1 corroborate our structural data that demonstrate that DCLK1-IN-1 adopts an intermediate binding mode compared to the two precursors which bind in a conventional Type I binding mode. Combined, our data show that DCLK1 selectivity can be achieved with compounds capable of inducing such a conformational change. In addition, we rigorously highlight the complexity of DCLK1 regulation and demonstrate that DCLK1-IN-1 binds to DCLK1 long isoform with high affinity but does not prevent DCLK1 MAP function. Such a result is in line with our data suggesting that kinase activity is required to prevent DCLK1 MAP function 25 . Taken together, our structural data presented here will undoubtedly provide a framework for the generation of highly selective and highly potent DCLK1 modulators and chemical probes (inhibitors or targeted protein degraders), for future therapeutic intervention.
We had previously demonstrated the ability of LRRK2-IN-1 and XMD8-92 to bind to purified recombinant DCLK1 kinase domain (DCLK1-KD) using a thermal stability assay 25 . However, despite attempts to crystallise these inhibitors in complex with DCLK1 kinase domain (DCLK1-KD), we never obtained diffraction quality crystals, likely due to the weak affinity of these inhibitors. We therefore took advantage of the availability of more potent DCLK1 benzopyrimido-diazipinone scaffold inhibitors such as XMD8-85 to utilise for crystallography, so as to provide a structural framework to further guide the development of DCLK1 selective tools.
We first confirmed the ability of DCLK1-IN-1's precursor, XMD8-85, and subsequently DCLK1-IN-1, to bind to DCLK1-KD (residues 372-649) by carrying out thermal stability assays (Fig. 1b,   Supplementary Fig. 1). We used DCLK1-NEG, a structurally related negative control compound, which was predicted to have significantly reduced affinity for DCLK1, owing to incorporation of an additional methyl substituent at R3 (Fig. 1a) that would cause steric clash with residues located at the start of the activation loop and near the floor of the ATP binding site 25 . By performing a compound titration between 2.5 µM and 40 µM, we demonstrated an increased shift in the melting temperature of DCLK1-KD as the concentration of the compounds increased, confirming that both XMD8-85 and DCLK1-IN-1 bind DCLK1-KD. As expected, the DCLK1-NEG compound showed severely reduced binding to DCLK1-KD ( Fig. 1b and Supplementary Fig. 1).

Crystal structure of DCLK1-KD:XMD8-85:
To understand the structural basis by which XMD8-85 increased inhibition of DCLK1 as compared to XMD-8-92 and to elucidate its retained pan-activity against ERK5 and LRRK2, we solved the structure of DCLK1-KD (residues 372-649) in complex with XMD8-85 to 2.5 Å (Fig. 2a, Supplementary Table 1, Supplementary Fig. 2). The DCLK1-KD:XMD8-85 complex crystallised with two molecules in the asymmetric unit with a head to tail packing similar to DCLK1-KD:AMPPNP (5JZJ) and DCLK1-KD:NVP-TAE684 (5JZN) crystal structures ( Fig. 2a and b) that we solved previously 25 . However, the "face-to-face" arrangement previously seen in these structures, promoted by the packing of an extended aC helix from one molecule against the activation loop of the second molecule, is not observed in this new structure. In the DCLK1-KD:AMPPNP and DCLK1-KD:NVP-TAE684 structures, this unusual activation loop dimerization mode was found to be stabilised by the coordination of a sulfate molecule between Arg510 from the catalytic loop of one molecule and Thr546 from the activation loop of the other (sulfate occupying the position where a phosphate would be if the protein was phosphorylated on the activation loop) (Fig. 2b). In DCLK1-KD:XMD8-85 structure, the sulfate coordination is no longer maintained. Despite this, the activation loop is entirely visible and adopts an active conformation state (Fig. 2a). This active conformation state is further confirmed by the presence of the canonical salt-bridge interaction between the conserved aC helix glutamate (Glu436) and the invariant b3 strand lysine (Lys419) (Fig. 2c) . However, the extended aC helix conformation previously reported is not seen ( Fig. 2a and b) 25 . Instead, an additional turn in the aC helix is visible (Fig. 2c). Residues 424-430, that were previously involved in crystal packing in the NVP-TAE684 structure, could not be modelled, highlighting the inherent flexible property of the aC helix in DCLK1 to accommodate the binding of XMD8-85. Additionally, XMD8-85 induces a conformational change in the glycine loop and residues 399-400 located at the tip of the glycine loop also could not be modelled (Fig. 2c).
While the position of ring C of XMD8-85 aligns with ring B of NVP-TAE684, both ring A and B of XMD8-85 sit deeper into the backpocket within the ATP binding site (Fig. 2d). XMD8-85 is further stabilised by van der Waals (vdw) interactions between ring A and Gly532 (that preceeds the DFG motif) and between ring B and Val449 (that is located between aC helix and b5 strand) (Fig. 2c).
The interaction of XMD8-85 with DCLK1-KD results in two hydrogen bonds interactions within the ATP binding site (Fig. 2e). These all occur through interaction of XMD8-85 within the hinge region, involving a donor/acceptor hydrogen bond pair with the backbone of Val468. The amine group (N2) in the linker between rings D and C of XMD8-85 acts a hydrogen bond donor for the backbone carbonyl oxygen of Val468, and the amine group (N5) in ring C of XMD8-85 acts as an acceptor for the backbone amide -NH of Val468. Additionally, the hinge region residues Gly471 and Lys469 further stabilise ring D conformation through main chain interactions. Residue Leu518 located on b7 at the floor of the ATP binding pocket also stabilises rings A, B and C through hydrophobic/vdw interactions. Notably, the amide carbonyl oxygen of diazepine ring B participates in a network of water-mediated hydrogen bond interaction with the invariant Lys419 and aC Glu436 that play a critical role in nucleotide binding ( Fig. 2f and 2g). The amide N-methyl subsitutent at position R4 on diazepine ring B confers significant affinity for DCLK1. This R4 N-methyl group is able to form favourable vdw interactions with the gatekeeper residue Met465 (on b5) located at the back of the ATP binding site as well as Ala417, Val449 and Glu466 (Fig. 2e). Removal of the N-methyl group was demonstrated to result in a loss of affinity towards DCLK1, ERK5 or BRD4, but led to a modest reduction in affinity for LRRK2 28 . As for DCLK1, LRRK2 has a gatekeeper methionine residue, suggesting that for LRRK2 this residue is not a significant contributor to the observed affinity 38 .
Interestingly, the salt bridge interaction between residue Glu415 (b3 strand) and Lys469 from the hinge region is maintained, acting as anchor point to maintain the structural integrity of this interface (Fig. 2e). The structure also revealed other features of XMD8-85 that likely account for its improved DCLK1 affinity relative to XMD9-92. In particular, the ortho methoxy substituent in ring D in XMD8-85 is likely better accommodated in DCLK1 than the bulkier ortho ethoxy substituent of XMD8-92; secondly, protonation of the para N-methylpiperazine of XMD8-92 at physiological pH would enable an additional favourable electrostatic interaction with Asp475 of DCLK1, which isn't possible for the ortho 4-hydroxypiperidine subsituent of XMD8-92 28 . However, the DCLK1-KD:XMD8-85 crystal structure also revealed aspects of sub-optimal shape complementarity of XMD8-85 for the DCLK1 ATP binding site and additional cavities near the back pocket that might be exploited to improve selectivity towards DCLK1 ( Fig. 2g and 2h). Overall, our DCLK1-KD:XMD8-85 crystal structure clearly indicated that the N-methyl group in the R4 position is orientated towards the gatekeeper residue Met 465, such that modifications in this position might be used to improve specific selectivity for DCLK1, in particular relative to other kinases with a differing gatekeeper residue.  (Fig. 2h). In addition, considering the conformational variability of the aC helix observed in the DCLK1-KD:XMD-8-85 structure as compared to ether the DCLK1-KD:AMP-PN or DCLK1-KD:NVP-TAE684 structures ( Fig. 2a and 2b), it was conceivable that these modifications might induce DCLK1 kinase domain to adopt an inactive conformation.

R4 modification provides increased selectivity towards
As illustrated by compound FMF-03-055-1, introduction of a N-ethyl substitution at position R4 was well tolerated by DCLK1 and in fact resulted in 5-fold enhanced potency for DCLK1 as compared to XMD8-85 28,37 . This modification also reduced off-target binding to both ERK5 and BRD4, but only to a limited extent for LRRK2 28 . LRRK2 (humanized ROCCO4), has identical residues to DCLK1 within the ATP binding pocket, including the gatekeeper methionine, which could explain this observed retention of potency with this modification (PDB:4YZM 38 ).
To better understand how the ethyl substitution at position R4 provided increased potency towards DCLK1 and increased selectivity over ERK5, we solved the crystal structure of DCLK1-Cter:FMF-03-055-1 (residues 372-686) to 3.1Å resolution ( (Fig. 3b). In addition to enhancing binding to DCLK1, the ethyl substitution also significantly decreased off-target binding to ERK5. This could be attributed to the presence in ERK5 of a leucine instead of a methionine as the gate keeper residue (Leu137). While a leucine residue is shorter, it has more constrained side chain conformations compared to a methionine. Analysis of ERK5:XMD8-92 (PDB 5BYY) structure indeed suggests that ethyl group substitution at position R4 would result in a clash with Leu137 in its current rotamer conformation ( Fig. 3c). Our structural data therefore provides the rationale for the increased affinity of FMF-03-055-1 towards DCLK1 and its loss of affinity towards ERK5. Conversely, the higher affinity of XMD8-92 towards ERK5 compared to DCLK1 can be attributed by the collapse of the N-towards the C-lobe resulting in a tight binding of this compound (XMD8-92 IC50 for ERK5 is 98nM (Fig.   3d) 34 .

DCLK1-IN-1 selectively binds DCLK1-KD:
While the R4 ethyl group in FMF-03-055-1 enhanced affinity towards DCLK1 and decreased off-target affinity for ERK5, FMF-03-055-1 still showed appreciable binding to LRRK2 and BRD4 28 . To further probe the R4 position, an electronegative trifluorethyl group was introduced instead of hydrophobic ethyl, which led to the compound known as DCLK1-IN-1 37 . Interestingly, DCLK1-IN-1 showed modest reduction in affinity for DCLK1 compared to FMF-03-055-1 but simultaneously a significant improvement in selectivity against ERK5, LRRK2 and BRD4 37 . To better understand the effect of this change, we crystallised DCLK1-KD in complex with DCLK1-IN-1 to 3.1Å resolution (Fig. 4a and 4b). The DCLK1-KD:DCLK1-IN-1 complex crystallised in a similar head to tail conformation as seen for the DCLK1-KD:XMD8-85 and DCLK1-Cter:FMF-03-055-1 structures (Fig. 2a, 3a and 4a). The ATP binding site shows a clear unbiased electron density for DCLK1-IN-1 ( Supplementary Fig. 2). Overall, the substitution of the methyl or ethyl group at position R4 with the trifluoroethyl group does not affect the conformation of the Met465 side chain nor the conformation of the aC helix or the A-loop (Fig. 4c). However, a striking feature of the DCLK1-KD:DCLK1-IN-1 structure, compared to DCLK1-KD:XMD8-85 and DCLK1-KD:FMF-03-055-1, is a significant opening of the ATP binding site, whereby both the glycine loop and the b3 strand undergo an upward shift of 5Å to accommodate the bulkier trifluoroethyl group (Fig. 4c and 4d). As a likely consequence of this effect due to binding of DCLK1-IN-1, no electron density was observed for residues 397-401 from the glycine loop, highlighting its high flexibility. However, the specific orientation adopted by the trifluoroethyl group allows additional contacts with residues Ala417, Val404, Ala417 and Met465 (Fig. 4b). Significantly, the presence of the trifluoroethyl group, which also causes a significant shift in the position of the b strands that make the N-term lobe (Fig. 4c), disrupts the salt-bridge interaction between the invariant lysine (Lys419) of b3 strand with the conserved glutamate (Glu436) of the aC helix (Fig. 4c).
Interestingly, this opening creates a new shallow pocket, which is in our structure occupied by extra electron density that could be attributed to a fragment of polyethylene glycol (PEG), a component of the crystallisation condition (Fig. 4c). Significantly, this shallow pocket, surrounded by the aC helix, the activation loop DFG motif and the b3 strand lysine (Lys419), represents a novel DCLK1 allosteric site, unseen in other DCLK1 structures solved to date ( Fig. 4c and 4d). A similar allosteric site has been reported in ERK5 when bound to an allosteric inhibitor ( Supplementary Fig. 4) 34 .
Together, our structural data demonstrates that in contrast to XMD8-85 and FMF-03-055-1, which adopt a conventional type I binding mode ( Fig. 2c and 3b), DCLK1-IN-1 adopts a type 1.5 binding mode with DCLK1 by inducing a large conformational change in the N-lobe and a creating a new allosteric site ( Fig. 4c and 4d). These changes can account for the dramatically increased DCLK1-IN-1 selectivity for DCLK1 against ERK5, LRRK2 and BRD4 37 . Several studies have shown the beneficial effect of DCLK1 knockdown on tumor growth 10,24,[39][40][41] Considering the dual biological function of DCLK1 and the presence of multiple isoforms ( Fig. 5a and Table 1), understanding the role of DCLK1 in cancer development and progression is likely to require mechanistic studies that pinpoint and dissect the expression pattern and the contribution of each DCLK1 isoform. Unfortunately, several of the studies published to date do not provide such comprehensive expression data. To begin to dissect the underlying mechanisms that link DCLK1 to oncogenesis and evaluate the use of DCLK1-selective inhibitors, we summarise those studies which investigate the specific role of distinct DCLK1 isoforms. In colorectal cancer (CRC), DCLK1-S isoforms, which lacks the microtubule binding domain, are reportedly overexpressed in several codon 12 KRAS mutant cells including G12D 42 . Likewise, human colon adenocarcinomas (hCRCs), overexpress DCLK1-S from an alternate (b)-promoter of DCLK1, due to the epigenetic silencing of the 5'(a)-promoter, which results in the specific loss of the DCLK1-L isoform. In addition, high expression of DCLK1-S is also associated with a significantly worse overall survival rates 17 .
Therefore, DCLK1 kinase inhibition, using an inhibitor such as DCLK1-IN-1, may be suitable to further study the role of DCLK1-S in these cancer models.
In non-small cell lung cancer (NSCLC), DCLK1-S isoform is predominantly over-expressed in cell lines H460 and A549, whereas DCLK1-L isoform is predominantly expressed in H1299 cells 39 .
However, the specific functions of each of the DCLK1 isoforms and their regulation of NSCLC remains unclear and requires futher study. Similarly, in renal clear cell carcinoma (RCC), the most common type of kidney cancer, both DCLK1-S and DCLK1-L isoforms are over-expressed (Table   1). Tumors expressing DCLK1-L isoform are enriched in cancer stem cell markers, however, tumors expressing DCLK1-S do not show such enrichment. While DCLK1-L has been shown to drive molecular and functional stemness 43 , the role of DCLK1-S in RCC remains unclear and will require further clarification.
In addition to the differential expression profiles observed for each isoform in various cancers, the expression levels of a particular isoform can vary between the early and late stages of a particular cancer type. For example, the levels of DCLK1-L in the serum of patients with pancreatic ductal adenocarcinoma (PDAC) significantly varies between early and advanced stages 44 . Interestingly, DCLK1-IN-1 seems ineffective on PDAC cells lines, PATU-8988T and PATU-8902, and was only effective on organoid samples isolated from an early stage of the disease but not from the advanced stage, despite all expressing DCLK1-S isoforms 37 . Unfortunately, this study did not specifically look at differences in levels of DCLK1-L, nor have the differences in DCLK1-S expression levels between the early and late stages of pancreatic cancer been reported. Thus, it is possible that either a lower DCLK1-S expression, or an altered DCLK1-L expression, in the late stage of pancreatic cancer could explain the lack of DCLK1-IN-1 activity. Together, these studies highlight an urgent need for a greater understanding of DCLK1 isoforms expression patterns, their differential expression levels during the various stages of cancer, as well as their epigenetic regulation in order to design the best therapeutical strategy.

The presence of the DCX domains does not prevent DCLK1-IN-1 binding to the kinase domain:
To begin to unravel if a selective DCLK1 kinase inhibitor can target DCLK1 long isoforms, we first sought to determine if the presence of the N-terminal doublecortin domains and additional N-and Cregulatory elements impacted DCLK1-IN-1 binding to the DCLK1 kinase domain by thermal shift assay (TSA). We generated two DCLK1-L (Uniprot O15075-2) constructs: DCLK1-FL1D (residues 50-686) and DCLK1-FL2D (residues 1-700) (Fig. 5a). Both constructs were generated as WT suggesting binding ( Fig. 5b and Supplementary Fig. 5).

DCLK1-IN-1 does not inhibit DCLK1 MAP function: Having established that DCLK1-IN-1 binds
to DCLK1-L isoform with a similar affinity to the DCLK1 kinase domain, we next wanted to show if DCLK1-IN-1 binding had an impact on DCLK1 MAP function. We and others have previously shown that purified DCLK1 promotes tubulin polymerisation in vitro 25,46 . We additinonally showed that DCLK1 kinase domain activity regulates its tubulin polymerization activity 25 . Using a fluorescent based polymerisation assay, we previously demonstrated that phosphorylated DCLK1 full-length no longer polymerises tubulin and DCLK1's tubulin polymerisation activity could be restored upon phosphatase treatment or when using a catalytically dead mutant protein, DCLK1-D511N. We again took advantage of this mutant to determine if binding of DCLK1-IN-1 to the kinase domain had an impact on DCLK1 tubulin polymerisation activity using microtubule polymerisation and pelleting assays, followed by analysis by SDS-PAGE of the supernatant and pellet fractions. We first performed a phos-tag gel analysis to confirm the difference in phosphorylation between phosphorylated DCLK1-FL1D WT and DCLK1-FL1D D511N (Supplementary Fig. 7). We next tested tubulin polymerisaton in the presence of phosphorylated DCLK1-FL1D WT and the corresponding catalytically dead mutant D511N. As expected, DCLK1-FL1D WT does not polymerise tubulin (Fig. 5d, lanes 1-2) while the catalytically dead mutant D511N does (Fig. 5d, lanes   3-4), as seen by the increased presence of tubulin (and DCLK1 D511N) in the pelleted fraction. This is consistent with our previous observation using tubulin polymerisaton assays based on fluorescencebased dye (Cytoskeleton Inc) 25 . We next tested tubulin polymerisation in the presence of 20 µM and 40 µM DCLK1-IN-1, concentrations which showed binding to DCLK1 based on our TSA data (Fig.   5b). The presence of DCLK1-IN-1 at 20 µM (Fig. 5d, lanes 5 and 7) or 40 µM (Fig. 5d, lanes 8-9) does not affect tubulin polymerisation capability of DCLK1-FL1D D511N compared to the control condition without the compound (Fig. 5d, lanes 3-4). We confirmed that the addition of 40 µM DCLK1-IN-1 does not cause DCLK1-FL1D D511N to aggregate and pellet upon centrifugation (Fig.5d, lanes 12-13). We also confirmed that DCLK1-IN-1 alone does not promote tubulin polymerisaton in the absence of DCLK1-FL1D D511N, (Fig. 5d, lanes 10 -11). As negative control, we tested tubulin polymerisation in the presence of nocodazole, a well-characterised microtubule destabliser. As expected, the presence of 40 µM nocodazole inhibits tubulin polymerisation (Fig. 5d,   lanes 14-15), while the presence of 40 µM of DCLK1-NEG compound has no impact on tubulin polymerisation (Fig. 5d, lanes 17-20). Nocodazole has been suggested to target both kinases and microtubules 47 . We therefore confirmed whether nocodazole was capable of binding DCLK1 using thermal stability assays ( Supplementary Fig. 7 microtubule binding through phosphorylation. Interestingly, our previous structural characterisation of presumed pathological mutations found in DCLK1-L in the context of gastric cancers, indicated that mutations occurring within the kinase domain would lead to a kinase disfunction. Together these data point towards a role for the kinase domain in regulating DCLK1 microtubule binding affinity and consequently its tubulin polymerisation activity. From these observations, it is clear that DCLK1isoform specific targeting strategies will be required to account for their distinct biological functions. Evidence is accumulating on the benefit in reducing DCLK1 expression levels to reduce tumorigenesis. Our structures will provide the basis to develop a chemically induced degradation strategy, such as proteolysis-targeting chimeras (PROTAC) 54,55 , in which a DCLK1 selective kinasebinding moiety might provide the necessary handle to recruit DCLK1 for degradation, thereby also dismantling DCLK1 MAP function.
In conclusion, the DCLK1-KD crystal structures presented here in the presence of benzopyrimidodiazipinone scaffolds have revealed a remarkable plasticity of the DCLK1 ATP binding pocket and additionally identified a novel allosteric site not seen previously. The existence of distinct DCLK1 isoform signatures in cancer progression and development calls for in-depth studies to rationally design isoform-specific modulators. Our study now establishes a structural framework as a platform to guide the design and development of novel isoform-specific modulators as therapeutic agents.

DNA Constructs
Human DCLK1 (Uniprot O15075-2) constructs for protein expression were cloned into a modified pCOLD vector encoding an N-terminal tobacco etch virus protease cleavage site and an 8XHis tag (Takara). The constructs for the catalytically dead mutant was designed using primers containing the mutation, and PCR products were cloned in pCOLD vector. All constructs were verified using Sanger Sequencing (Micromon).

Thermal Shift Stability Assay
Thermal shift stability assays were performed as described previously 45  inhibitor structures were solved by the molecular replacement method using PHASER in CCP4, with DCLK1-KD structure (PDB 5JZJ) as a model. The structures were refined using Phenix.refine. 58 Model building was carried out using COOT 59 . The ligand was modelled using the Grade Web Server . The overall structure were validated using MOLPROBITY. All molecular graphics representations were created using PyMOL (The PyMOL Molecular Graphics System, Version v1.8.0.3, Schrödinger LLC).

Surface Plasmon Resonance
Immobilisation. Proteins were immobilised at 25°C using standard amine coupling to a CM5 chip

Phos-Tag gels
Phos-tag TM pre-cast gel anaylsis of phosphorylated and non-phophorylated samples was done according to manufacturer's instruction (FUJIFILM Wako Pure Chemical Corporation). DCLK1 WT, DCLK1 WT treated with lamda phosphatase and DCLK1 D511N at 1 µg or 0.5 µg were run on a SuperSep TM Phos-tag TM 12.5% SDS-PAGE pre-cast gel. Electrophoresis was carried out at 150 volts for 60 min. The gel was stained using InstantBlue® Coomassie Protein Stain (Expedeon). A concevtional 12.5% SDS-PAGE analysis with molecular weight markers was done in parallel to ensure the proteins for the phos-tag analysis were not degraded.

Tubulin Polymerization and Pelleting Assay
Tubulin polymerization assays were performed in buffer containing 80 mM PIPES at pH 6.9, 2.0 mM MgCl2, 0.5 mM EGTA, 2.5 mM β,γ-Methyleneguanosine 5′-triphosphate, and 10 µM tubulin and 5 µM DCLK1 for one hour at 37°C. For tubulin polymerisation in the presence of inhibitors, various concentrations of the inhibitors were first incubated with DCLK1 and allowed to stand on ice for 5min before adding this mix to the tubulin polymerisation reaction. Following the tubulin polymerisation, the reaction mix was pelleted at 20,000 g for 20min and the supernatant and the pellet fractions were run on a 12% SDS gel. The data shown are representative of two independent experiments.