Integrated computational and Drosophila cancer model platform captures previously unappreciated chemicals perturbing a kinase network

Drosophila provides an inexpensive and quantitative platform for measuring whole animal drug response. A complementary approach is virtual screening, where chemical libraries can be efficiently screened against protein target(s). Here, we present a unique discovery platform integrating structure-based modeling with Drosophila biology and organic synthesis. We demonstrate this platform by developing chemicals targeting a Drosophila model of Medullary Thyroid Cancer (MTC) characterized by a transformation network activated by oncogenic dRetM955T. Structural models for kinases relevant to MTC were generated for virtual screening to identify unique preliminary hits that suppressed dRetM955T-induced transformation. We then combined features from our hits with those of known inhibitors to create a ‘hybrid’ molecule with improved suppression of dRetM955T transformation. Our platform provides a framework to efficiently explore novel kinase inhibitors outside of explored inhibitor chemical space that are effective in inhibiting cancer networks while minimizing whole body toxicity.

Introduction Protein kinases play a key role in cell signaling and disease networks and represent major therapeutic targets. The limited capacity to test large numbers of compounds to explore diverse chemical scaffolds, coupled with the difficulty in translating in vitro kinase inhibition into whole animal efficacy, has limited the chemical space of the known kinase inhibitors (KIs). As a result, obtaining optimal KIs with clinically relevant therapeutic activity has proven challenging despite extensive academic and industry effort.
To expand the number of kinase inhibitors, a variety of platforms have recently emerged as useful tools for compound screening. The fruit fly Drosophila melanogaster provides an inexpensive and efficient whole animal platform for cancer drug screening, capturing clinically relevant compounds [1][2][3]. For example, Drosophila was used to help validate vandetanib as a useful treatment for medullary thyroid cancer [4] (MTC). As a screening platform, Drosophila offers several advantages: First, flies and humans share similar kinome and kinase-driven signaling pathways [5], facilitating the use of flies to predict drug response in humans [1,6]. Second, the ease of breeding and the short (~10 day) life cycle of Drosophila makes it possible to carry out efficient moderate-throughput chemical screening in a whole animal system. Third, the screening readout provides a quantitative, animal-based measurement of structure-activity relationships (SAR) as well as information on the therapeutic potential or toxicity of the tested compounds: measurable parameters include survival and multiple phenotypic indicators that depend on kinase activity.
A key limitation of a Drosophila-based moderate-throughput screening platform is its inability to explore very large chemical libraries [7] such as the ZINC library, which has over 750 million purchasable compounds [8]. In contrast, structure-based virtual screening is a fast and inexpensive computational method that can screen large compound libraries, useful for identifying unique chemical probes [9]. If the structure of the protein is unknown, virtual screening can be performed against the homology models of the target based on experimentally determined structures. However, the automated construction of homology models-with sufficient accuracy for simultaneous virtual screening of multiple targets and the application of molecular docking to signaling networks-remains challenging in particular for highly dynamic targets such as kinases [10,11] and would benefit from a readily accessible whole animal platform. In this paper, we demonstrate how combining Drosophila and computational approaches provides a synergistic platform for lead compound discovery, combining the strengths of computational methods-which enable rational and rapid drug candidate selection-and a Drosophila animal model that enables fast and relevant biological readouts of tested compounds. We demonstrate the practicality of this approach using MTC as a test case.
RET is a receptor tyrosine kinase associated with multiple roles in development and disease. The gain-of-function M918T mutation of RET (analogous to Drosophila M955T) activates multiple proliferation pathways and is directly associated with MTC pathogenesis [12,13]. Transgenic Drosophila expressing the analogous dRet M955T isoform show key aspects of transformation, including proliferation and aspects of metastasis [6,14]. Genetic modifier screens with dRet M955T flies led to the identification of multiple RET pathway genetic 'suppressors' and 'enhancers', loci that when reduced in activity improve or worsen the dRet M955T phenotype, respectively. These functional mediators of RET-dependent transformation include members of the Ras/ERK and PI3K pathways as well as regulators of metastasis such as SRC [6,15].
Oral administration of the FDA-approved multi-kinase inhibitor analogs sorafenib and regorafenib-along with additional structural analogs-partially rescued dRet M955T -induced transformation in Drosophila [1,15]. Sorafenib-class inhibitors are classified as 'type-II' KIs that bind the kinase domain in their inactive state [16], a conformational state regulated by the aspartate-phenylalanine-glycine (DFG)-motif ( Fig 1A) [17,18]. In the inactive, 'DFG-out' conformational state the directions of DFG-Asp and DFG-Phe 'flip', vacating the DFG-Phe pocket ('DFG-pocket') that modulates binding to type-II inhibitors. A key challenge of targeting kinases in the DFG-out conformation with structure-based virtual screening is that few kinase structures have been reported with the DFG-out conformation [19,20]. We recently developed DFGmodel [10], a computational method for modeling kinases in DFG-out conformations. This method informed the mechanism of clinically relevant multi-kinase inhibitors that target the MTC network [15].
In this study, we report the development of an integrated platform (Fig 2) that combines (i) computational modeling of kinases in their inactive state plus massive multi-target virtual screening with (ii) whole animal Drosophila assays to identify previously unappreciated chemicals that perturb RET-dependent transformation. This integrative platform combines the strengths of computational methods-including facilitating rational and rapid compound prioritization for experimental testing-and Drosophila models that provide a whole animal readout of compound efficacy. We leverage this integrated fly/computational modeling platform to The conformational state of protein kinases (e.g., KDR) including DFG-in (red) and DFG-out (blue) is determined by the DFG-motif. The DFG-pocket (cyan mesh) is unique to the DFG-out conformation. Sorafenib is shown in pink. Broken yellow lines indicate hydrogen bonds. (B) A scheme depicting the positive and negative effects of drug acting on genetic modifiers of medullary thyroid cancer in a Drosophila model. ptc-driven dRet M955T induces lethality during development. 'Suppressors' or 'enhancers' suppress or enhance, respectively, dRet M955T -induced disease phenotypes as revealed in genetic screening. A drug can suppress lethality by inhibiting the suppressors. It can also induce toxicity and/or worsen transformed phenotypes by inhibiting the enhancers, which results in enhanced lethality. create a novel 'hybrid' molecule with unique chemical structure and biological efficacy. Finally, we discuss the relevance of this approach to expedite the discovery of novel chemical scaffolds targeting disease networks.

Target selection from fly genetic screen and structural analysis
In transgenic patched-GAL4;UAS-dRet M955T (ptc>dRet M955T ) flies, the ptc promoter drives expression of an oncogenic isoform of Drosophila Ret in multiple tissues; the result is lethality prior to adult eclosion [1,15]. We previously used this and similar fly MTC models in genetic screens to identify~40 kinases that mediate dRet M955T -mediated transformation [6,15]  To narrow this list, we prioritized candidate kinases based on two considerations: (i) pharmacological relevance as known mediators of RET signaling [6,21]; (ii) structural coverage, specifically kinases with known DFG-out structures or those that can be modeled with sufficient accuracy [10]. Atypical kinases (e.g., mTOR and eEF2K) and members of the RGC family were excluded as they have diverse sequence and structure features that limit our ability to generate accurate homology models. Applying these criteria to our genetic modifier list, we focused on targeting four key kinase targets: RET (receptor tyrosine kinase), SRC (cytoplasmic tyrosine kinase), BRAF (tyrosine kinase-like), and p70-S6K (AGC family).

Modeling kinases in DFG-out conformation
Description of the various conformations adopted by the kinases during activation and inhibition is needed for rationally designing novel, conformation-specific inhibitors. Therefore, our  approach was to perform massive structure-based virtual screens of purchasable compound libraries against multiple models with DFG-out conformation; our goal was to identify generic kinase inhibitors that may target one or more prioritized kinases but, importantly, demonstrate an effect on the disease pathway in the animal model.
The structure of two of the kinases identified in our dRet M955T genetic screens-BRAF and SRC-have been solved in the DFG-out conformation; the DFG-out structures of RET and p70-S6K have not been reported. We therefore generated DFG-out models using DFGmodel, a computational tool that generates homology models of kinase in DFG-out conformation through multiple-template modeling that samples a range of relevant conformations [10]. In our previous study, we tested and confirmed subsets of DFG-out models that enrich for known type-II inhibitors among a diverse set of non-type-II KIs found in the Protein Data Bank (PDB) with accuracy similar to or better than that obtained for experimentally determined structures [10]. For example, in a recent application of DFGmodel, models generated by this method were used in parallel with medicinal chemistry to optimize clinically relevant compounds that are based on the established kinase inhibitor sorafenib [15]. Conversely, in this study models generated by DFGmodel were used to develop compounds outside of the current kinase inhibitor chemical space.
To guide the identification of a 'generic' kinase inhibitor that can affect a disease pathway we first compared the DFG-out models of each kinase, identifying key similarities and differences in physicochemical properties among their inhibitor-binding sites. First, we noted that the prioritized targets RET, BRAF, p70-S6K, and SRC present negative electrostatic potential on the DFG-pocket surface, while many non-targets such as ERK have positive electrostatic potential ( Fig 3A, S2 Fig). This difference may partially explain the partial selectivity of type-II inhibitors (e.g., sorafenib) toward our prioritized targets while avoiding electrostatic positive kinases such as ERK. Second, RET and SRC have large DFG-pocket volumes (163 Å 3 , 196 Å 3 , respectively) and p70-S6K and BRAF have moderately large pockets (158 Å 3 , 136 Å 3 ). In contrast, ERK has a small DFG-pocket (113 Å 3 ; Fig 3B). We used this size difference to computationally select for kinases with larger DFG-pockets (e.g., RET, SRC) while excluding kinases with smaller DFG-pockets (e.g., ERK).

Virtual screening against MTC pathway
We performed virtual screening against multiple DFG-out models of MTC targets to identify putative small molecules that modulate the disease network ( Fig 2C). We docked a purchasable lead-like library from the ZINC database (2.2 millions compounds; [22]) against 10 DFG-out models for each kinase target, yielding over 88 million total docking poses. To combine the screening results, a two-step consensus approach was used. In the first step, the top scoring pose of compounds that ranked in the top 10% in 5 or more of the 10 models of each kinase were selected, resulting in approximately 2,000 compounds per kinase. In the second step, compounds that ranked in the top 25% in at least 3 of 4 targets were selected, resulting in 247 compounds. For comparison, sorafenib, an inhibitor that rescues ptc>dRet M955T flies, would rank eighth in this consensus docking result. From these consensus compounds, eight commercially available compounds were purchased to test their ability to rescue ptc>dRet M955T flies (S1 Table). These compounds were selected based on their interactions with key elements of the "ensemble" of targets' binding sites, with the emphasis on the conserved glutamate in αC-helix, the amide backbone of DFG-aspartate, and if present, the amide backbone of the hinge region (S3 Fig). Although the compounds are not predicted to bind optimally to each one of our targets, we hypothesized that these compounds may have an additive effect on the disease pathway, which could be improved with medicinal chemistry.

Testing candidates in ptc>dRet M955T fly viability assay
Transgenic ptc>dRet M955T flies express the oncogenic Drosophila dRet M955T isoform in several tissues in the developing fly, leading to aspects of transformation of dRet M955T tissues [6,14]. As a result, ptc>dRet M955T flies exhibited 0% adult viability when cultured at 25˚C, providing a quantitative 'rescue-from-lethality' assay to test drug efficacy [1,15]. Compounds were fed at the highest accessible concentrations (see Experimental Procedures). We used sorafenib as a positive control, as it previously demonstrated the highest level of rescue among FDA-approved KIs in ptc>dRet M955T flies [15]. Similar to our previous results, feeding ptc>dRet M955T larvae sorafenib (200 μM) improved overall viability to 3-4% adult survival (P < 0.05).
The docking poses of 1-1 and 1-2 resemble the proposed docking pose of 1 (Fig 4B), which has a typical DFG-out-specific, type-II KI binding pose and is predicted to occupy the DFGpocket with its terminal phenyl moiety. The 1,2,4-triazole moiety, resembles the urea moiety found in sorafenib (Fig 1A), forms favorable hydrogen bonds with the side chain of the conserved αC-helix glutamate residue and the backbone amide of the DFG-Aspartate. In addition, this series of compounds is smaller and shorter (MW < 360) than the fully developed type-II KIs (MW > 450) such as sorafenib, as they lack an optimized moiety that interacts with the hinge region of the ligand-binding site (Fig 4C). Compound 1-2 is structurally different from 1 and 1-1 and was less effective in rescuing ptc>dRet M955T flies, even though it was tested at higher concentrations (Fig 4A). While 1 and 1-1 have an (1H)-1,2,4-triazole moiety, 1-2 has an 1,2,4-oxadiazol-5-amine moiety where the (1H)-nitrogen is replaced by an oxygen. This modification distinguishes 1-2 from 1 and 1-1 in their interaction preference: 1-2 loses a hydrogen bond donor due to the nitrogen-to-oxygen substitution, while the electronegative oxygen introduces an unfavorable electrostatic repulsion to the carboxylate sidechain of the conserved αC-helix glutamate (Fig 4C, 1-2 insert).
We examined the kinase inhibition profile (DiscoverX) of 1 against a subset of the human protein kinome (Table 1). At 50 μM, 1 did not appreciably inhibit SRC, BRAF, or S6K1, while it demonstrated weak activity against wild-type RET and moderate activity against the oncogenic isoform RET M918T . Of note, 1 inhibited other cancer-related targets such as FLT3 (Table 1), which activates the Ras/ERK signaling pathway [23].
1 also showed activity against aspects of transformation and metastasis in the fly. In the mature larva, the ptc promoter is active in epithelial cells in a stripe pattern in the midline of the developing wing epithelium (Fig 4C; wing disc). ptc-driven dRet activates multiple signaling pathways, promoting proliferation, epithelial-to-mesenchymal transition (EMT), and invasion of dRet M955T -expressing cells beyond the ptc domain [14] (Fig 4C). Similar to sorafenib, oral administration of 1 blocked the invasion of dRet M955T -expressing cells into the surrounding wing epithelium (Fig 4B).
At lower dosage (25 μM), compound 2 weakly rescued ptc>dRet M955T flies (1%; P < 0.05) (Fig 5A). Unlike 1, 2 did not act synergistically with sorafenib. This difference was confirmed by the kinase inhibition profile of 2 (Table 2), in which it has stronger inhibition of RET and RET M918T , but loses the inhibition of FLT3, two key differences between the kinase inhibition profiles of 1 and 2.
Through a chemical similarity search of the ZINC database, we identified five compounds that share the 1H-indole-2-carboxamide moiety with docking poses similar to that of 2 (Fig 5B  and 5C; S2 Table), and confirmed all five analogs increased viability of ptc>dRet M955T flies (Fig Integrated computational and Drosophila cancer model platform for targeting kinase network 5A) albeit with weak efficacy (some with a P-value above 0.05). At low dose (10 μM), 2-1 showed improved efficacy in rescuing ptc>dRet M955T flies relative to 2 and displayed similar efficacy as sorafenib at 200 μM. However, 2-1 showed poor solubility, limiting its usefulness as a lead compound. 2-3 was also more efficacious than 2 and displayed better solubility in both DMSO and water than 2-1; it also has the N-phenylacetamide moiety as a linker group, a common linker feature found in type-II KIs such as imatinib. Compound 2-3 displayed a different inhibition profile than 1 and 2 ( Table 2): it strongly inhibits FLT3 and PDGFRB, weakly inhibits RET and RET M918T , and does not inhibit SRC.

Improving efficacy through compound hybridization
Interestingly, the chemical scaffolds of our newly identified active compounds are not associated with inhibition of protein kinases based on an analysis with SEA search [24], which relates ligand chemical similarity of ligands to protein pharmacology. Nevertheless, they provided rescue of ptc>dRet M955T flies at a level similar to sorafenib and regorafenib [15]. The docking poses of these compounds suggest a less-than-optimal interaction with the hinge region of protein kinases, a common feature of most KIs. Furthermore, the relatively low molecular weight (~350 g/mol) of these lead-like compounds provides a window for conducting lead optimization with medicinal chemistry. Hence, we sought to improve the efficacy of our computationally derived leads by installing a hinge-binding moiety found in known type-II KIs such as sorafenib.
To select the optimal position on our initial hit to conduct fragment exchange with known type-II kinase inhibitors (sorafenib and AD80 [1]), we took into consideration (i) the docking poses and phenotypic results of these compounds and (ii) the synthetic accessibility and the novelty of the putative hybrid compounds, even if they do not dock optimally to our intended kinase targets. We focused on 2/2-3 due to: 1) their 1H-indole moiety docks uniquely into the DFG-pocket and with the potential to interact with the αC-helix glutamate ( Fig 5C); 2) their 1H-indole-2-carboxamide moiety resembles the urea linker that is commonly found in type-II KIs such as sorafenib (Fig 6A; blue box); 3) the N-phenylcarboxamide moiety of 2-3 is a common linker between the hinge-binding and the DFG-pocket moieties of type-II KIs, e.g. imatinib (Fig 6A; grey box), while the N-(piperidin-4-yl)carboxamide moiety of 2 is not a common linker; 4) the docking pose of 2/3's 1H-indole moiety overlaps with the trifluoromethylphenyl moiety of sorafenib/AD80. Based on these four criteria, we chose to perform a fragment exchange at the carboxamide position to combine the 1H-indole-2-carboxamide moiety of 2/2-3 with the hinge-binding moiety of sorafenib and of AD80, a multi-kinase inhibitor that has shown promise in MTC treatment [1], to create 3 and 4, respectively (Fig 6B). Oral administration of 3 and 4 to ptc>dRet M955T flies demonstrated that the efficacy of 4 was low with only 3% rescue, while 3 demonstrated strongly improved efficacy with 15% rescue (Fig 6C; P < 0.05); this level of rescue was significantly higher than the parent compound 2/2-3 or sorafenib. Additionally, 3 suppressed the invasion/migration of dRet M955T -expressing cells in the wing epithelium (Fig 6D), further confirming its efficacy against dRet M955T -mediated phenotypes. The kinase inhibition profile of 3 ( Table 3) resembles that of the parent compound 2-3 ( Table 2) with at least two notable exceptions: 3 inhibits CSF1R, PDGFRB, and FLT3, receptor tyrosine kinases and orthologs of Drosophila Pvr that activate the Ras/ERK signaling pathway [25] and play key roles in SRC activation and tumor progression; 3 inhibits Aurora kinases AURKB and AURKC (Drosophila ortholog aurA or aurB). Of note, although 4 did not improve the viability of ptc>dRet M955T flies, it shares chemical similarity to several known type-I½ kinase inhibitors that have the common adenine moiety and a related indole moiety. This group of inhibitors was shown to inhibit other related kinases, increasing our confidence in the relevance of this chemical space for kinase pathway modulation [26].

Integrated discovery pipeline
This study demonstrates the utility of an integrated platform that combines Drosophila genetics, computational structural biology, and synthetic chemistry to enrich for the discovery of useful chemical tools in an established Drosophila MTC model (Fig 2). We have previously shown that Drosophila can provide a unique entry point for drug development by capturing subtle structural changes in lead compounds that are often missed by cellular or biochemical assays [15]. Here we refine this approach by iteratively combining experimental testing with computational modeling. A key strength of the integrated approach is its ability to rapidly derive candidates from a large, purchasable chemical library via virtual screening to test chemically unique compounds with our fly models in a cost-effective manner. This platform allowed us to quickly confirm the in situ relevance of active chemotypes through iterations of computational modeling, synthetic chemistry, and phenotypic testing in the fly. We expect this integrated pipeline is generally applicable to kinase networks associated with other diseases [7].

DFG-out modeling approach
DFGmodel is a recently developed computational tool that generates models of kinases in their inactive, DFG-out conformation for rational design of type-II KIs [10]. In a recent study, models generated by DFGmodel were used to guide the optimization of the drug sorafenib to better target a new disease space [15]. Here, we demonstrate a successful application of DFGmodel to explore compounds that are not appreciated as kinase inhibitors. For each kinase target, DFGmodel uses multiple experimentally determined structures as modeling templates and generates multiple homology models. Thus, this method samples a large fraction of the DFG-out conformational space during the model construction, which enables us to account for the flexibility of the binding site during virtual screening [27]. Notably, DFG-out models capture key features that are important for protein-ligand interactions in multiple kinases simultaneously, providing a framework for rationalizing activity of known inhibitors and developing unique KIs that target a signaling pathway. For example, our results suggest that the electrostatic potential within the DFG-pocket is a key feature for inhibitor selectivity: ERK has an inverse electrostatic potential in the DFG-pocket than that of our target kinases RET and BRAF (Fig 3), which may explain the insensitivity of ERK toward inhibitors such as sorafenib.

Identification of biologically active compounds
Although used in the clinics for MTC, sorafenib (and its close analog regorafenib) show limited efficacy in MTC patients; this poor activity is mirrored in the ptc>dRet M955T fly model, which was rescued 3-4% at 200 μM [15]. Despite considerable academic and industry effort, the known chemical space of kinase inhibitors is limited [7]. For example, sorafenib and Integrated computational and Drosophila cancer model platform for targeting kinase network regorafenib differ in only one non-hydrogen atom. Through structure-based virtual screening against multiple kinase targets in the MTC pathway, we discovered chemically unique compounds (S2 Table) with an ability to rescue ptc>dRet M955T viability that is similar to sorafenib, an FDA-approved KI (Figs 4 and 5). Importantly, our data indicates that these compounds act through key cancer networks. For example, compounds 1, 2, 2-3 and 3 all have shown the ability to suppress invasion of ptc>dRet M955T cells in the wing epithelium. Previous work demonstrated that dRet M955T -mediated wing cell invasion is controlled by SRC [15,28], which acts by controlling E-cadherin and Matrix Metalloproteases (MMPs). Of note, 1, 2, 2-3 and 3 each demonstrated significant activity against orthologs of Drosophila Pvr, a key regulator of Src: all show significant activity against human FLT3, while 3 shows additional activity against Pvr orthologs CSF1R and PDGFRB. In addition to being orthologs of Pvr, FLT3, CSF1R, and PDGFRB similarly can activate SRC [29]. We propose that this activity against regulators of SRC account for the abil- ity of 1, 2, 2-3 and 3 to suppress invasion, a key first step in tumor metastasis. Other activities, for example, 3's inhibition of Aurora kinases-required for proliferation during tumor progression [30]-likely also contributes. Indeed, AURK inhibitors are known to be active against MTC [31,32] and synergy between AURKs and FLT3 is currently being explored clinically through a number of dual-AURKB/FLT3 inhibitors [33,34].

Recombination of building blocks for future inhibitors
Although the new tool compounds 1 and 2 are not themselves sufficiently potent to serve as therapeutic leads, they reveal diverse fragment-like pharmacophores that serve as starting points for an exploration of new chemical space. These pharmacophores can be further optimized by combining with well-developed chemotypes that are known to interact with kinase binding sites (e.g., the hinge binding region) to form more efficacious chemical probes [35]; this provides a key second step towards building effective compounds. For example, 2 and 2-3 include an 1H-indole moiety capable of occupying the DFG-pocket of protein kinases from different families and a carboxamide group commonly found in type-II KIs (Fig 6A). Guided by the docking poses of these compounds, the 1H-indole-2-carboxamide group was combined with an optimized hinge-binding moiety from sorafenib, to form a significantly more efficacious compound (i.e., 3). As indicated in the kinase inhibition profile of 3 (Table 3), it shares part of the target set of its constituents 2 and 2-3.
Despite its promise, our approach has several limitations. The computational modeling does not take into account conformational changes modulated by inter-or intra-molecular interactions between the kinase domain and binding partners (e.g., SH2/SH3 domains, scaffolding proteins), as well as the differential propensity among kinase domains to adopt DFGout states [36]. This weakness is partly mitigated through careful template selection for model building as well as docking of the small molecules to multiple models representing different conformers. Second, although kinases are closely conserved between humans and Drosophila, fly models have some differences with human disease networks including lacking an adaptive immune system. They lack some relevant target organs (e.g., thyroid, breast, prostate, pancreas); in these cases we focused on developing eye and wing epithelia, which provide 'generic' polarized epithelia that can give biological and pharmacological results in Drosophila that have proven translatable to mammals as we previously reported [1,15,[37][38][39]. Despite these limitations, flies offer a useful genetic and pharmacological toolkit which can facilitate drug development for cancer as we show in this study. Future studies will include testing the compounds discovered in this study on mammalian models including human MTC cell lines and mouse xenografts [15].
In summary, we demonstrate the potential of combining chemical modeling with Drosophila genetics to rapidly and efficiently explore novel chemical space. This provides an accessible and cost-effective platform that can be applied to a broad palette of diseases that can be modeled in Drosophila. Combining the strengths of these two high-throughput approaches opens the opportunity to develop novel tool and lead compounds that are effective in the context of the whole animal.

DFG-out models
Models of kinase targets (human RET, SRC, BRAF, S6K1) in the DFG-out conformation were generated using DFGmodel [10]. Briefly, the method takes a DFG-in structure or the sequence of the protein kinase as input. DFG-model relies on a manually curated alignment between the target kinase and multiple structures representing unique DFG-out conformations. It calls on the structure-based sequence alignment function of T_COFFEE/Expresso [40] v11.00.8 to perform sequence alignment of the kinase catalytic domain to the templates, followed by the multi-template function of MODELLER [41] v9.14 to generate 50 homology models covering a range of conformations. For each kinase 10 DFG-out models with largest binding site volume, as calculated by POVME [42] v2.0, were used for further study. These ensembles of DFG-out models of our targets BRAF, p70-S6K, RET, and SRC have been evaluated and confirmed to enrich known type-II inhibitors over non-ligands derived from kinase-inhibitor complexes found in PDB in our previous study [10]. The area-under-curve (AUC) of virtual screening performance of our targets BRAF, RET, and SRC DFG-out models are 87.7, 82.8, and 76.8, respectively, which correspond to at least 4-to 5-fold increase in the enrichment accuracy over randomly selected ligands in a known sample set [43][44][45]. This ensemble of models provides an approximation of the binding site flexibility, as well as optimizes the binding site for protein-ligand complementarity and structure-based virtual screening [11,27,43].

Virtual screening
Initial virtual screening utilized the ZINC12 [22] "available now" lead-like chemical library (downloaded in 2013, 2.2 million compounds). A maximum of 300 conformers for each compound were generated with OMEGA. Each conformer was docked with FRED using the default settings [40] [41]. The top scoring pose was used for further analysis. For each of the four targeted kinases (RET, BRAF, SRC, p70-S6K), the ensemble of 10 DFG-out models was used for screening. Each compound was docked against 10 models of each kinase target, resulting in 10 docking poses per each kinase-compound pair. The results were filtered by selecting compounds that rank in the top 10% in at least five models per kinase-compound pair, represented by the best scoring pose. This filtering process was done for all four targets. Next, compounds that scored in the top 25% in at least 3 of 4 the kinase ensemble models were collated into a final set of 247 compounds. These consensus best-scoring ligands, representing 0.0112% of the library, were visually inspected to remove molecules with energetically unfavorable or strained conformations, or with reactive functional groups that may interfere with assays [46], which are commonly observed in large virtual screenings. Eight compounds, all top-100 ranking among 247 candidates (S1 Table), were selected based on their interactions with the models (DFG-pocket occupancy, hydrogen-bond to conserved amino acids, etc) and chemical novelty and were purchased for Drosophila viability screening . Analogs 1-1, 1-2, 2-1 to 2-5, and others (S2 Table) were identified based on the structure of compounds 1 and 2 through the chemical similarity search function available in ZINC15 [8] and SciFinder using the default setting and Tanimoto coefficient above 70% similarity. These compounds are commercially available through vendors such as ChemBridge and Enamine.

Chemical methods
For synthetic procedures and characterization data related to compounds 1, 3, and 4, please see supplementary materials.

Kinase profiling of compounds
Kinase inhibition profile of the compounds was assessed at 50 μM through commercially available kinase profiling services (DiscoverX).
Stock solutions of the test compounds were created by dissolving the compound in DMSO at the maximum concentration. The stock solutions were diluted by 1,000-fold or more and mixed with semi-defined fly medium (Bloomington Drosophila Stock Center) to make druginfused food (0.1% final DMSO concentration; maximum tolerable dose in flies). Approximately 100 ptc>dRet M955T embryos alongside with wild-type (+;+/SM5 tubgal80 -TM6B) flies were raised until adulthood on drug-infused food for 13 days at 25˚C. The numbers of empty pupal cases (P in S1B Fig) and that of surviving adults (A) were used to determine percentage of viability, while their body size, which is affected by food intake, temperature, and humidity, were compared to vehicle-treated groups to standardize the experimental conditions. For toxicity test, the viability of the wild-type flies raised alongside with the ptc>dRet M955T flies in the same vials was examined. None of the tested compounds show detrimental effect in the control flies, showing > 90% viability at any doses tested for the compounds equivalent to vehicle.

Wing discs cell migration/invasion assays
Third-instar ptc>dRet M955T larvae were dissected, and developing wing discs were collected, fixed with 4% paraformaldehyde in PBS, and whole-mounted. At least 10 wing discs were analyzed for each treatment. Invasive GFP-labeled dRet M955T -expressing cells were visualized by their green pseudocolor under a confocal microscope. The apical and the virtual z-series views of the wing disc were examined to identify abnormal tissue growth and dRet M955T -expressing cells migrating beyond the ptc domain boundary.