Integrated Proteomics-Based Physical and Functional Mapping of AXL Kinase Signaling Pathways and Inhibitors Define Its Role in Cell Migration

Overexpression of AXL, a member of the tumor-associated macrophage (TAM; TYRO3, AXL, and MER) family of receptor tyrosine kinases (RTK), is often associated with epithelial–mesenchymal transition (EMT), increased aggressiveness, and poor prognosis in a vast number of cancers (1–3). AXL overexpression has also been linked to the development of resistance to chemo-, immune- and targeted therapy (4–10), thus making it an attractive target for therapeutic intervention.

AXL is often reported to be overexpressed, both at mRNA and protein levels, in lung adenocarcinoma (2, 11, 12). However, AXL is rarely reported to be genetically amplified, fused, or mutated in cancers with an alteration frequency of approximately 3% or less in breast, lung, and head and neck squamous cell carcinoma as per data from The Cancer Genome Atlas (TCGA) and Clinical Proteomic Tumor Analysis Consortium (CPTAC; refs. 13, 14). Thereby, it is uncertain if AXL functions as a dominant oncogenic driver like other genomically altered RTKs, such as mutant EGFR or anaplastic lymphoma kinase (ALK). In addition, the mechanism of AXL kinase activation is complex and still poorly understood. AXL is activated by its cognate ligand, Gas6 (15). However, some studies suggest that AXL-GAS6 interaction may be constitutive and functional activation of downstream signaling is achieved only when an additional interaction with phosphatidylserine (PtdSer) on apoptotic bodies exists (16, 17). Recent studies have shown that AXL can act as a sensor for localized clusters of ligands wherein extracellular vesicles enriched in PtdSer accumulate Gas6 leading to an increase in local ligand concentration and activation of AXL receptor signaling (18). AXL can also be activated in a ligand-independent manner either via clustering of overexpressed protein (19) or by heterodimerization with other receptors such as EGFR and MET (7, 20, 21). Also, different types of cells (such as fibroblasts and bone-marrow progenitor cells) and conditions of stress (such as hypoxia and nutrient deprivation) in the tumor microenvironment can produce GAS6 and lead to microenvironment-mediated activation of AXL in cancers (22–25).

There are multiple kinds of AXL targeting agents that are in various stages of development, including ATP-competitive small molecule inhibitors, anti-AXL mAbs, drug-antibody conjugates, soluble receptors, and nucleotide aptamers (26, 27). Of these, small molecule AXL TKIs form the largest group and interestingly, most of these drugs that inhibit AXL were not designed against AXL but inhibit it as an off-target. Kinase inhibitors are in general known to have a wide array of targets in addition to the one that they were designed to inhibit, thereby exhibiting a range of specific and nonspecific phenotypic effects. There have been contradictory reports on whether AXL inhibition with TKIs is able to restore sensitivity to the targeted agent (28, 29). This could be because the mechanism for AXL-mediated drug resistance is not well understood or because of differences in model systems and/or AXL targeting strategies used for these studies. A study on kinases capable of bypassing EGFR dependence in EGFR-mutant non–small cell lung cancer (NSCLC) cells shows that while most of the common bypass kinases were associated with sustained activation of either phospho-AKT or both phospho-AKT and ERK, AXL was associated with sustained phospho-ERK activation only, thereby suggesting that AXL had different mechanisms to rescue EGFR mutations compared with other EGFR bypass genes (30). These studies reemphasize the importance of characterizing off-targets of purported specific kinase inhibitors as well as generating context-dependent signaling maps for AXL.

Given the complexities associated with AXL kinase activation and its context dependency, signaling downstream of AXL is also not well defined. Small-scale, focused studies in specific model systems have reported that GAS6/AXL can activate a wide array of pathways, including AKT, ERK, SRC, FAK/RAC, and JAK-STAT (31, 32). However, the pathways activated and hence the phenotypic response elicited by AXL signaling depends on cell/tissue type and disease state. Yet, no unbiased AXL protein complex or global phosphoproteomic analyses have been reported and it is important that such systems biology approaches be utilized to elucidate AXL's role in primary oncogenesis, metastasis, and resistance to therapy.

In this study, we aim to better characterize AXL TKIs, understand their effects on signaling and phenotype of cells, and develop assays to visualize active AXL signaling complexes. For this, we used multiple mass spectrometry (MS)-based proteomics approaches, such as proximity-dependent biotinylation (BioID) and phosphoproteomics to delineate AXL-mediated signaling in lung cancers. Our proteomic analyses provide insights into AXL interacting partners, changes in the phosphoproteome after pharmacologic AXL inhibition, and putative AXL pathway substrates, which can serve as a valuable resource for understanding AXL signal activation at a systems-level. Subsequent comparative analysis of different AXL TKIs using Western blotting, viability assays, and activity-based protein profiling (ABPP), identifies diverse target profiles and phenotypic effects for each inhibitor. In addition, treatment of our cell line models of acquired erlotinib resistance with EGFR and AXL TKI combination therapy demonstrate inability of pharmacologic inhibition of AXL to reverse resistance to erlotinib in these model systems. Notably, our MS and phenotypic assay data cumulatively show coupling of AXL to downstream PI3K/AKT signaling, but not MEK/ERK, and reveal a key role for AXL in mediating cell migration/invasion in a lung cancer cell line model. These results can guide the development of proximity ligation assays to detect active AXL signaling foci in situ, that can help identify patients likely to benefit from AXL targeted therapy.

Cells were provided by the Moffitt Lung Cancer Center of Excellence Cell Line Core (Tampa, FL). The H1299 AXL KO and PC9 AXL KO cells were developed in the Meyer lab at UCLA (Los Angeles, CA). Cells were confirmed to be free of Mycoplasma (PlasmoTest, Invivogen) and authenticated by short tandem repeat (STR) analysis (ACTG Inc.). RXDX106, bemcentinib, and cabozantinib were purchased from Selleckchem, dissolved in DMSO at a 10-mmol/L concentration, and diluted as necessary.

AXL variant 1 (from Meyer Lab at UCLA) was cloned into the pENTR TOPO vector using pENTR D-TOPO cloning kit (Invitrogen) as per manufacturer's protocol. AXL was then transferred from the entry clone to pSTV6-C-BirA*-FLAG Gateway destination vector (ref. 33; a gift from Gingras Lab, Lunenfeld-Tanenbaum Research Institute, Ontario, Canada) using LR-Gateway cloning, creating pSTV6-C-BirA*-FLAG-AXL lentiviral vector. Lentivirus packaging was done as described before (33). H1299 cells were infected with the produced virus and then supplemented with 2 μg/mL puromycin for selection 48 hours posttransfection to generate H1299 cells stably expressing pSTV6-C-BirA*-FLAG-AXL.

Cells were grown and expanded into one 15-cm petri dish per replicate for BioID proteomics experiments. Three biological replicates were used per sample. Samples for BioID proteomics analysis were prepared using methods described in (33). LC-MS/MS was carried out on the Thermo Q Exactive Plus quadrupole-orbitrap hybrid mass spectrometer. Peptides identified using Proteome Discoverer (2.2.0.388) and visualized using Scaffold (4.8.7). Putative AXL binding partners were identified by Automated Processing of SAINT Templated Layouts (APOSTL; ref. 34).

Proximity Ligation Assays (PLA) were performed as described (35). Briefly, cells were plated, fixed, and permeabilized in eight-well chamber slides. Nonspecific binding was blocked, followed by overnight incubation at 4°C with primary antibodies, rabbit anti-AXL (Cell Signaling Technology; 8661) and mouse anti-pY (pY-100; Cell Signaling Technology; 9411). Subsequent incubations with rabbit (+) and mouse (−) PLA probes and detection with Duolink In Situ Detection Reagents FarRed (Sigma) were carried out as per the manufacturer's protocol. Fluorescent images were acquired on a fully automated upright Zeiss Axio Imager Z.1 microscope at 40X magnification using Axiovision software suite (version 4.6, Carl Zeiss, Inc.).

H1299 cells were plated in six 15-cm dishes for each sample (in triplicates) and treated with RXDX-106. Phosphoproteomics samples were prepared using the PTMScan Kit (Cell Signaling Technology) as per the manufacturer's protocol. Phosphotyrosine peptides were enriched using the antibody beads [PTMScan Phospho-Tyrosine Rabbit mAb (P-Tyr-1000) Kit, Cell Signaling Technology; #8803) and analyzed with LC-MS/MS for label-free quantitation.

The flow through from the immunoprecipitation of phosphotyrosine peptides was labeled using TMT 10-plex reagents following the manufacturer's recommendation (TMT10plex Isobaric Label Reagent Set, Thermo Fisher Scientific), enriched using IMAC and used for global phosphoproteomics (pSTY).

A nanoflow ultra-high-performance liquid chromatography (HPLC; RSLCnano, Thermo Fisher Scientific) coupled to a quadrupole-orbitrap hybrid mass spectrometer (Q Exactive Plus, Thermo Fisher Scientific) was used for LC-MS/MS analysis. MaxQuant 1.5.2.8 (36) was used for peptide identification and reporter ion quantification. Data were normalized and analyzed for differential expression between different treatment conditions.

Scratch wound cell migration and invasion assays were carried out per the recommended protocol for the IncuCyte system (Essen Bioscience).

H1299 and PC9 cell pellets, in triplicate for each sample, were harvested in ice-cold PBS, lysed, and the ATP-binding proteome labeled according to the Pierce Kinase Enrichment Kits and ActivX Probes (Thermo Fisher Scientific). Lysates were incubated with 20 μmol/L RXDX106, cabozantinib, bemcentinib or DMSO for 15 minutes and the desthiobiotin-ATP probe was added for the competition reaction at a final concentration of 5 μmol/L for 15 minutes, followed by reduction, alkylation, trypsin digestion, and pulldown of desthiobiotinylated peptides. LC-MS/MS data were acquired, and peptides identified using MaxQuant. Data were normalized and analyzed for differential expression between different drug treatments.

Cell lines models for acquired resistance to first-generation EGFR TKI, erlotinib, were derived by dose escalation treatment of EGFR-mutant cell lines with erlotinib as described before (37). The cell line models for perister-derived erlotinib resistance were generated by passing PC9 cells through a persister bottleneck as described before (38). Briefly, parental PC9 cells were exposed to 2.5 μmol/L erlotinib for approximately 6 weeks until few surviving single cells were observed. After passing through a senescent period, single cells emerged to form colonies. These colonies were isolated, passaged to new plates, and maintained in 1 μmol/L erlotinib for further studies.

DNA and RNA were extracted simultaneously using QIAGEN AllPrep DNA/RNA mini kit as per manufacturer's protocol. From extracted RNA, RNA sequencing (RNA-seq) libraries were prepared using the NuGen FFPE RNA-Seq Multiplex System (Tecan US, Inc.). Whole-exome sequencing (WES) was performed on extracted genomic DNA using the Agilent SureSelect XT Clinical Research Exome kit (Agilent Technologies, Inc.).

Gene expression data were normalized and differential expression between experimental groups were evaluated using DEseq2 (39). RNA-seq data were filtered, and normalized genes were subjected to single sample gene set enrichment analysis (ssGSEA) using the cancer hallmarks (40). The ssGSEA parameters used were sample.norm.type: rank, weight: 0.75, statistic: area.under.RES, output.score.type NES, nperm: 1000, min.overlap: 5, correl.type: z.score, and run.parallel: TRUE. ssGSEA results were visualized using the Complex Heatmap R package (41). Gene sets with FDR > 0.05 were denoted with an "X".

For detailed experimental procedures, see the Supplementary Materials and Methods.

The supportive data for the MS analyses have been provided as Supplementary Tables S1–S5. The MS proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE partner repository with the dataset identifiers PXD023911 (BioID), PXD023913 (ABPP), and PXD023923 (Phosphoproteomics; ref. 62). The whole-exome and RNA-seq data have been deposited in NBCI Sequence Read Archive (SRA; PRJNA790835) and Gene Expression Omnibus (GEO; GSE191185) respectively.

We took an unbiased multipronged systems level proteomics approach, consisting of BioID interaction mapping as well as tyrosine and global phosphoproteomics, to annotate AXL protein complexes and investigate AXL signaling in lung cancers (Fig. 1A). We hypothesize that the identified interacting and signaling proteins and their functions would guide the understanding of AXL's role in these cells. For this, H1299 NSCLC cells were chosen as they are known to have high levels of AXL expression and activation and are more likely to give us pertinent interacting proteins and mechanisms. To identify AXL interacting proteins, we used the BioID system which utilizes a proximity-dependent labeling strategy. For this, the bait protein is fused to a mutant form of biotin ligase (BirA*) that biotinylates interacting protein within a 10-nm radius of the protein of interest, which can then be enriched and identified by MS (33). This system allows for harsher lysis conditions over other affinity-based MS approaches because proximal proteins are now marked by covalent biotin modifications and preservation of protein complexes is not necessary for identification.

H1299 cells expressing pSTV6-BirA*-FLAG-AXL were created. To confirm expression of the transgene, immunofluorescence staining with anti-FLAG antibody was done on parental H1299 and induced H1299-BirA*-FLAG-AXL cells. FLAG staining, with distinct membrane localization, was observed only in the cells that were transduced with the transgene and induced with doxycycline and biotin (Fig. 1B). H1299-BirA*-FLAG-AXL-Puro cells were then either treated with biotin only (negative control, no expression of BirA*-FLAG-AXL), doxycycline only (negative control, BirA*-FLAG-AXL expressed but no biotinylation), and both biotin and doxycycline (expression and biotinylation of BirA*-FLAG-AXL). Treated cells were lysed, processed, and prepared for MS analysis (Fig. 1A). The data was then analyzed using APOSTL (34) wherein the proteins found in the negative controls were subtracted to identify 119 AXL interacting proteins. The results show that the biotinylated bait protein, AXL, was only detected in the sample that was induced with both biotin and doxycycline, as expected (Fig. 1C). Other proteins identified as putative AXL interactors included known partner proteins, such as EGFR and ITGB1, as well other interesting proteins known to be involved in kinase biology (such as ERBB2IP, GAB1, STAT3, PTPN11, and SCRIB; Fig. 1C). To identify signaling pathways and functional modules that are associated with the AXL interactome, these proteins were further subject to STRING-based network and Kyoto Encyclopedia of Genes and Genomes (KEGG)- and gene ontology (GO)-based pathway analyses. STRING network analysis with high confidence resulted in an undirected network with 119 nodes (proteins) and 150 edges (interactions) that was imported and visualized in Cytoscape (Fig. 1C). The AXL interacting proteins could be classified into adherens/tight junction (consisting of ITGB1, BAIAP2, YES1, LLG1, EGFR, EZR, SCRIB, CTNND1, TJP1, DLG1, MLLT4, OCLN, PVRL2), cell junction assembly/organization (consisting of CTTN, FERMT2, ITGA2, ITGA6, GJC1, DSG2, NUMBL, FLNA, NUMB, PVRL2, OCLN, MLLT4, DLG1, TJP1, CTNND1), and protein localization to cell periphery (consisting of EPB41L2, ITGB1, EGFR, DLG1, EZR, SCRIB, NUMB, FLNA, PALM, VAMP3, CAV1, EPHA2, PICALM, FLOT1) clusters (Fig. 1C). We identified adherens/tight junction, proteoglycans in cancer, cell junction assembly/organization, and protein localization to the cell periphery as the top enriched functional modules from KEGG and GO analyses (Fig. 1C and D), thereby suggesting that AXL has a major role to play in regulating cell adhesion and migration in these cells.

In addition to profiling AXL interactors by BioID, we performed tyrosine (pY) and global (pSTY) phosphoproteomics in H1299 cells treated with an AXL kinase inhibitor to understand signaling downstream of AXL kinase activation. For this, a potent TAM inhibitor, RXDX106, was used. We first treated H1299 cells with RXDX106 for varying amounts of time to confirm its ability to inhibit AXL phosphorylation by Western blotting (Fig. 2A). Our data shows that 100 nmol/L of RXDX106 potently inhibited AXL phosphorylation at Tyr⁷⁰² within 30 minutes. This was accompanied by a concurrent increase in AXL and a decrease in phosphorylation of downstream AKT (both at Ser⁴⁷³ and Thr³⁰⁸), but not ERK (Fig. 2A). We also show that siRNA-mediated silencing of AXL expression triggers a decrease in downstream pAKT without altering the pERK levels (Fig. 2A). This suggests that AXL activation is driving downstream PI3 kinase/AKT signaling events in this model. As an alternate method to assess changes in AXL tyrosine phosphorylation and activation, that can potentially to be translated to clinic as a diagnostic platform, we developed PLAs to detect activated AXL:pY signaling complexes in these cells. H1299 and Calu1 cells, both of which have detectable levels of phosphorylated AXL Tyr⁷⁰² (Fig. 6C), have high basal AXL:pY PLA foci that are abrogated by treatment with AXL TKI, RXDX106 (Fig. 2B). As expected, the HCC827 cells, which lack pAXL, do not show significant labeling in this PLA (Fig. 2B).

For the phosphoproteomics profiling, we treated H1299 cells either with DMSO or RXDX106 for 0.5 or 4 hours. Cells were then lysed, trypsin digested, and processed for MS (Fig. 1A). In the global phosphoproteomics dataset, we identified 13,638 unique phosphopeptides, which correspond to 3,679 unique phosphoproteins. In the pY dataset, 634 unique phosphopeptides, corresponding to 329 unique phosphoproteins, were identified. We focused on the phosphopeptides that are significantly altered by AXL inhibitor, RXDX106. A phosphopeptide was determined to be differentially expressed between two conditions if |log₂ ratio| ≥ approximately 0.585 (1.5-fold change) and P value < 0.05. This filtering analysis gave us 148 unique peptides (mapped to 105 unique proteins) from the pY dataset and 5,117 unique peptides (mapped to 2,032 unique proteins) from the pSTY dataset.

To gain a global view of signaling networks regulated by AXL kinase, we generated an experimentally consistent literature network. For this, differentially expressed phosphopeptides from the pSTY and pY datasets were combined and assigned to five broad pattern categories as shown in Supplementary Fig. S1A. Selected differentially expressed phosphoproteins were then searched for activating or inhibiting transcriptional and phosphorylation literature interactions with MetaCore followed by in-house filtering for experimental consistency and visualization to generate the network shown in Supplementary Fig. S1A.

This experimentally consistent literature network provides a global view of the signaling networks regulated by AXL. As expected, phosphorylation of TAM kinases, including AXL, were downregulated as they are known targets of RXDX106 (42). The network shows decreased phosphorylation of several proteins, including kinases such as PAK2 and PAK4 and other proteins such as CTTN, KRT18, NUMB, that are involved in cell adhesion, migration, and actin reorganization. Protein kinases, such as ERBB2, CDK2, MTOR, PRKCD, and several MAPKs (MAPK1, 3, 7), had increased phosphorylation as per the network. Increased MYC, ETS1, and JUN phosphorylation were also shown as prominent nodes with high connectivity (Supplementary Fig. S1A). We identified signaling by Rho GTPases, which are known to regulate cytoskeletal and cell adhesion dynamics, as the top enriched pathway from MsigDB (40) canonical pathways (CP) analysis (Supplementary Fig. S1A).

We also took a more focused look at the changes in only tyrosine phosphorylation induced by RXDX106 (Fig. 2C). This analysis revealed affected pY sites from proteins with diverse functions and revealed both known and novel interactors of AXL. Tyrosine phosphorylation of proteins involved in cell adhesion/migration/invasion (such as TJP1, CTNND1), actin remodeling (such as CTTN, EPB41), phosphoinositide signaling (such as INPPL1, GAB1, GAB2), GTP-mediated signaling (such as RASA1, RIN1, ARHGAP35), as well as kinases (such as IGF1R, EGFR, EPHB4, LYN, FER, MAPK1/3), receptors (such as ITGB1/4), adaptor (such as LAT2, CRKL, SHB, GAB1/2, ERBIN), and scaffold (such as FLOT1) proteins, are modulated by AXL inhibition, indicating that these cellular processes may be important for propagating AXL signaling and function downstream (Fig. 2C). Anchoring junction and signaling by receptor tyrosine kinases were identified as top enriched functional modules from MsigDB GO and CP analyses, respectively (Supplementary Fig. S1B).

To identify new putative substrates of the AXL signaling pathway, we integrated the phosphoproteomics data with the proximity network. We found 23 peptides from the pY dataset that overlapped with the BioID data and used their peptide sequences to build a Position-Specific-Scoring-Matrix (PSSM) motif (Fig. 3D). Transition matrices such as PAM20 are frequently used to calculate the similarity between two given sequences (43). Thereby, we computed the mean PAM250 similarity scores between each peptide sequence of the global phosphoproteomics dataset and the 23 sequences forming the AXL pathway motif to rank all peptides by its similarity to the AXL pathway motif. This analysis identifies CBL, DCBLD2, LAMTOR1, ARHGAP35, LAT2, RASA1, ABI1, RIN1, and PTK2 as some of the top putative AXL pathway substrates (Fig. 2D).

Overall, our data supports our earlier observation that AXL inhibition deregulates phosphorylation of proteins involved in PI3K/AKT pathway and suggests that AXL has an important role to play in cell migration signaling pathways in these cells.

To functionally validate the effect of AXL inhibition on the migration and viability of cells, we performed scratch wound cell migration and invasion assays as well as cell viability assays. H1299 and Calu1 NSCLC cells were treated either with DMSO, the indicated concentrations of RXDX106, or AXL siRNA. The movement/migration of cells into the scratch wound was monitored by imaging the scratch wound every 3 hours using the IncuCyte system until the wound was completely closed. Our data shows that inhibition of AXL with RXDX106 or siRNA reduced the wound closure rate, as measured by relative wound density, in a dose-dependent manner in both H1299 and Calu1 cells (Fig. 2E and F). This implies that cells are moving into the wound slower in the presence of an AXL inhibitor or siRNA, both in migration (where the scratch is not overlaid with Matrigel; Fig. 2E) as well as in invasion (where the cells are moving through the Matrigel laid over the scratch; Fig. 2F) assays, thereby supporting the effect of AXL inhibition on components of the cell adhesion/migration/invasion pathways seen in our phosphoproteomics data. To assess the effect of AXL inhibition on cell viability, H1299, Calu1, H2009, H460, and HCC827 cells, each of whom express varying levels of phosphorylated and total AXL, were treated with increasing doses of RXDX106 for 72 hours. We observed that 1 μmol/L RXDX106 reduced the viability of these cells only by approximately 10%, except for H2009 whose viability was reduced by approximately 50% (Fig. 2G). Reduction of viability of cells was only observed at high doses of 5 and 10 μmol/L RXDX106 in all cell lines studied, irrespective of their level of AXL expression. Interestingly, for H1299 and Calu1, cell viability was not altered at drug doses of 100 nmol/L or 1 μmol/L (Fig. 2G) that were shown to inhibit phosphorylation of AXL (Fig. 2A and B) and migration/invasion of cells (Fig. 2E and F). Collectively, these data provide a map for AXL protein-complexes and signaling partners and evidence for AXL's role in mediating cell migration/invasion.

In our data, doses of AXL inhibitor RXDX106 that reduced AXL phosphorylation did not significantly alter viability of the lung cancer cells tested. However, there are reports in literature of other AXL inhibitors that reduce cell viability as well as reports of divergent effects of different AXL TKIs on cell viability (28, 44). These discrepancies could be attributed to the polypharmacology effect of kinase inhibitors wherein these drugs display a wide spectrum of targets in addition to the specific target they were designed against (45). To understand the differences in phenotypic effects between different AXL inhibitors, we compared three AXL inhibitors, RXDX-106 (a TAM inhibitor; ref. 42), cabozantinib (a multi-kinase inhibitor that inhibits AXL; ref. 46), and bemcentinib (an AXL inhibitor currently in clinical trials; ref. 47) in short-term viability assays, long-term colony formation assays, as well as their ability to modulate signaling downstream of AXL (Fig. 3). For these studies, along with H1299 cells, EGFR mutation–driven PC9 NSCLC cells were also used, since we were interested in subsequently looking at the role of AXL in the context of ERFR TKI resistance. To differentiate AXL-dependent effects from AXL-independent ones, we also used H1299 and PC9 cells where AXL has been knocked out by CRISPR (AXL KO) in these studies. In addition to AXL inhibitors, PC9 cells were also treated with erlotinib as a control drug, since its ability to potently inhibit viability, colony formation, and signaling in EGFR-mutant PC9 cells is well characterized. We observe that bemcentinib is more effective in reducing the viability of cells compared with other AXL inhibitors in both H1299 (Fig. 3A and B) and PC9 cells (Fig. 3F and G). This difference is more evident in the colony-forming assay, where 1 μmol/L bemcentinib completely inhibits growth in both H1299 and PC9 cells. Interestingly, bemcentinib shows more robust decrease in cell viability and colony-forming assays than the other AXL inhibitors even in H1299 and PC9 AXL KO cells (Fig. 3D, E, I, and J), thereby suggesting that this effect on viability of bemcentinib is likely AXL-independent. The effect of each of these AXL inhibitors on downstream signaling was also assessed by treating H1299 and PC9 cells with the indicated concentration of drugs for an hour (Fig. 3C and H). RXDX106 and cabozantinib both reduced AXL phosphorylation at Tyr⁷⁰² concomitant with a reduction in downstream AKT phosphorylation in H1299 cells. Bemcentinib was observed to cause a decrease in downstream AKT phosphorylation without inducing a robust inhibition of AXL pTyr⁷⁰² (Fig. 3C), which is consistent with previous reports of bemcentinib inhibiting AXL phosphorylation at tyrosine residues other than Y702 (47). In PC9 cells that express AXL but not detectable pAXL, RXDX106 and cabozantinib do not alter downstream pAKT or pERK (Fig. 3H). Interestingly, treatment with bencentinib decreases pAKT and completely inhibits pERK in PC9 as well as PC9 AXL KO cells at 1 μmol/L, thereby suggesting that this effect on signaling is AXL-independent (Fig. 3H and K). Treatment with erlotinib inhibits pEGFR Tyr¹⁰⁶⁸, pAKT, and pERK as expected in EGFR-driven PC9 cells (Fig. 3H).

To compare the kinase-binding targets between these three AXL inhibitors that exhibit diverse phenotypic and signaling responses, we performed an ABPP experiment in both H1299 and PC9 cells as shown in Fig. 4A. ATP-ABPP is a chemical proteomic tool that uses a kinase active-site directed competitive ATP probe to enrich and identify kinase binding targets and off-targets of a drug based on its activity (48, 49). Identified proteins from the ABPP experiment were labeled to be differentially expressed between conditions using 1.5-fold and P < 0.05 cutoffs. Kinases were our primary focus, so we filtered the differentially expressed proteins to select for protein kinases that were present in the DMSO control sample but significantly reduced in the drug treated samples. Kinases that were present in all DMSO samples but absent in all drug-treated samples, were manually selected to be differentially expressed since statistics could not be derived for these proteins. We observed that each drug had a unique kinase binding profile as illustrated in the kinome trees (Fig. 4B and E). Despite this heterogeneity, 10 and 8 kinases, including AXL, were common targets of all 3 drugs in H1299 and PC9 cells respectively (Fig. 4B, D, E, and G). Interestingly, there is a high degree of concordance between the kinase targets of a drug in the two different cell lines tested, with approximately 60% of the kinases being similar for bemcentinib and RXDX106 and approximately 65% for cabozantinib (Fig. 4C and F). Our data also showed that bemcentinib binds more kinases compared with the other two AXL inhibitors, thereby resulting in perturbations of the ATP-binding proteome that were distinctly different (Fig. 4B, C, E, and F). Proteins altered uniquely by bemcentinib belonged to the CK1, AGC, and CAMK kinase families (Fig. 4B and E) and when subjected to pathway enrichment analysis using ENRICHR (50), the topmost enriched pathway was found to be MAPK signaling.

Overall, here we compare the phenotypic and signaling effects of three commonly used AXL inhibitors and identify the spectrum of kinase targets for each of them. Our data suggests that of the three AXL TKIs characterized here, bemcentinib has the most off-targets and that viability effects of bemcentinib do not require AXL. The promiscuity of bemcentinib may help explain the unique phenotypic and signaling perturbations induced by this drug, including its AXL-independent effects.

Given the recent reports on the role of AXL in mediating resistance to EGFR-targeted therapy (7, 8, 10), we used two types of cell line models of erlotinib resistance to investigate the effect of AXL inhibitors on EGFR TKI resistance. For our first model, EGFR-mutant PC9 cells were passed through a persister bottleneck (38) by applying strong drug selection pressure to generate drug-tolerant erlotinib persister cells. We created four erlotinib-resistant clones from one parental population; S1–34, S2–10, S2–17, and S2–30. We found AXL to be overexpressed in each of these clones (Fig. 5B). To investigate the role of AXL in acquired resistance in persister-derived erlotinib-resistant cells, we treated PC9 and each of the clones either with erlotinib, RXDX106, or bemcentinib alone or in combination and assessed for changes in cell viability and signaling events (Fig. 5A and B). Erlotinib inhibited PC9 cell viability and induced complete shutdown of EGFR, AKT, and ERK signaling, as expected. AXL inhibitors, RXDX106 and bemcentinib, on the other hand had minimal effect on PC9 cell viability by themselves. Interestingly, while RXDX106 had no effect on EGFR and its downstream signaling, bemcentinib showed inhibition of ERK phosphorylation without altering EGFR or AKT phosphorylation (Fig. 5A and B). The unique effect of bemcentinib on ERK phosphorylation could be due to the MAPK targets of this drug that we identified via ABPP. These persister-derived cells were all uniformly resistant to erlotinib as well as AXL inhibitors as single agents, but had diverse responses in corresponding EGFR, AKT, and ERK activity downregulation. Combination of EGFR and AXL TKIs were in general ineffective in reversing erlotinib resistance in these cells, except for the unique effect of bemcentinib and erlotinib cotreatment in S2–10. S2–10 is resistant to erlotinib and bemcentinib as single agents, but when exposed to the combination, this clone showed inhibition of EGFR and ERK activity and a reduction in cell viability that the other persister-derived resistant cells did not show. The other AXL inhibitor, RXDX106, did not exhibit this combination effect in any of the resistant cells (Fig. 5A and B).

Whole-exome and RNA-seq analyses were performed to probe the differences in erlotinib-resistance mechanisms present in these persister-derived erlotinib-resistant cells. Overall, our whole-exome and RNA-seq data reveals the heterogeneity of the clones that emerged from the persister bottle neck (Supplementary Figs. S2 and S3). From the WES data that identified genetic changes between parental PC9 and each of the resistant cells, we first looked for mechanisms of erlotinib resistance commonly reported in clinical studies. EGFR T790M mutation allele frequency for each of the clones and their corresponding sensitivity to osimertinib are shown in Fig. 5E. Mutations in the other commonly implicated pathways, such as MET, MAPK, PI3K, were not detected in our resistant clones. We filtered DNA mutations based on their PolyPhen-2 score, a tool used to characterize mutations as “possibly damaging” and “probably damaging.” Supplementary Fig. S2 and Supplementary Table S6 show the whole spectrum of probably damaging variants which are present in any of the resistant clones compared with the parental PC9 cells. Our data reveals 57 variants that were uniquely present or absent in one clone (for instance, IBTK mutation was only present in S2–10 while OTUD7A and ARAP2 were mutated only in S2–30). We also found 10 variants that were either gained or lost across all resistant cells compared with PC9 (like MAP7D1, RGS7BP, C7orf26, etc). mRNA analysis focused on the “hallmarks” of cancer were plotted as a heatmap to show changes in each of the clones compared with parental PC9 cells (Supplementary Fig. S3). Some hallmarks, such as WNT_BETA_CATENIN, REACTIVE_OXYGEN_SPECIES, and TGF_BETA_SIGNALING, were found to be commonly upregulated in all resistant cells while others, such as OXIDATIVE_PHOSPHORYLATION, E2F_TARGETS, and NOTCH_SIGNALING, were downregulated in all of the clones. The expression of different EMT-related genes derived from the RNA-seq data is shown in Fig. 5E and shows diverse profiles of these markers amongst the different resistant clones. Consistent with our previous data (Fig. 5B), the only gene found to be upregulated in all four persister-derived resistant cells was AXL.

For our second model, two EGFR mutation–driven NSCLC cells (HCC4006 and HCC827) were exposed over time to increasing doses of erlotinib, a first-generation EGFR TKI, to generate HCC4006ER and HCC827ER cells that did not acquire EGFR T790M mutation and exhibited erlotinib resistance and EMT features (37). The HCC827ER cells generated did not have MET amplification. Both these resistant cell lines were found to overexpress AXL compared with the parental cells (Fig. 5C). These cells were then treated either with erlotinib, RXDX106, or bemcentinib alone or in combination and assessed for changes in cell viability and signaling events (Fig. 5C and D). As expected, the HCC4006 and HCC827 parental cells are sensitive to erlotinib, whereas the HCC4006ER and HCC827ER cells are resistant (Fig. 5D). Erlotinib treatment completely inhibited EGFR Tyr¹⁰⁶⁸ phosphorylation in both parental and ER cells. Downstream AKT and ERK activity was concurrently inhibited in the parental cells whereas this activity was retained in the ER cells (Fig. 5C), thereby leading to drug resistance. The AXL inhibitors, alone or in combination with erlotinib, were unable to increase inhibition of AKT and ERK or reverse erlotinib resistance in the ER cells (Fig. 5C and D).

Our data also shows that EGFR and AXL TKI combination treatment fails to reverse resistance in two different cell line models of erlotinib resistance, except for the unique effect of bemcentinib on one persister-derived erlotinib-resistant cell, where it resensitizes the resistant cell to Erlotinib in an AXL-independent manner. The heterogeneity among the various persister-derived erlotinib-resistant cells is probably an underlying mediator of the differential effects observed.

Studies surveying AXL expression in human lung cancer patient samples have reported a wide range of AXL overexpression levels, ranging from 33% to 93%, possibly due to differences in antibodies and staining methods used (32). The varying subtypes, stage, and histology of the patient samples as well as the size of the cohort studied may also contribute to the differing numbers from each report. The largest study on AXL expression in clinical samples so far consists of 134 NSCLC tumors (51). For more comprehensive view of the landscape of AXL expression in patients with lung cancer, we used AXL IHC to screen patient tumor microarrays (TMA), consisting of more than 500 tissue samples from small cell, squamous cell, and adenocarcinomas, to assess the overall expression of AXL protein in lung cancers. A score of 0 to 3 was assigned to each tissue section based on staining intensity (Fig. 6A). Our data shows that AXL expression is high in approximately 30% to 40% of patients with NSCLC, irrespective of their histology or mutational status (Fig. 6B). High AXL expression was detected in 27.8% of the squamous sample and 38.5% of the adenocarcinoma samples. There is however very low prevalence of AXL overexpression in small cell lung cancer (SCLC) patient tissues with only approximately 8% of the samples having high AXL staining (Fig. 6B). Prevalence of AXL positivity was not affected by driver mutation status with 42.5% of the EGFR mutant, 41.2% of the KRAS mutant, and 40.9% of the ALK-positive tissue samples having high AXL expression by IHC (Fig. 6B). Interestingly, significant fraction of KRAS mutant cancers having high AXL can have implication for KRAS therapeutics in development as recent data suggests that AXL can be a bypass signaling mechanism in KRAS G12C inhibitor treated cells (52). Twenty lung cancer cell lines were also screened for AXL protein expression using Western blotting and we found a wide range of AXL expression across these lines (Fig. 6C). Most (approximately 75%) of the cell lines screened had some level of AXL protein expression. Interestingly, AXL was detected to be endogenously activated by phosphorylation at Tyr⁷⁰² only in three of the 20 cell lines screened (H1299, Calu1, H2009), with H1299 having the highest levels of phosphorylated AXL (Fig. 6C). The comprehensive data from this large set of TMAs and cohort of cell lines provides a view of the landscape of AXL expression across lung cancer tumor tissues and cell line models.

In this study, we utilize multiple system-wide proteomic approaches to identify AXL-associated interacting partners and signaling networks, and drug target profiles for AXL TKIs used as probes and in clinics for lung cancers. To our knowledge, this is the first study to use multiple proteomics approaches to probe the AXL interactome, phosphoproteome, and response to pharmacologic AXL inhibition at a systems level. This study will aid in improving our understanding of AXL-dependent signaling as there is limited information available about this. In fact, a search of BioGRID, a curated interaction repository, for known AXL interactors yields a list of only 70 unique interactors, as compared with a list of 1,293 unique interactors for EGFR, clearly demonstrating the gap in knowledge about AXL-driven signaling. Most of these previously reported AXL interactors were identified from low-throughput assays and the ones identified via high-throughput screening were not from studies designed to investigate AXL biology. Our data from AXL protein complex annotation by BioID identified 118 novel AXL interacting proteins and reveal a primary role for AXL in mediating cell adhesion/invasion and identify several proteins involved in this process downstream of AXL. Several groups have reported overexpression of AXL associated with EMT, cell migration/invasion, and cancer metastasis (11, 12, 53, 54), but specific protein partners coupled to AXL that are involved in this were not identified.

Our global phosphoproteomics and supporting Western blotting data shows coupling of AXL predominantly to downstream PI3K/AKT and not MAPK signaling in H1299 cells. Other groups have observed AXL-dependent activation of MEK/ERK signaling (55–57), possibly due to the chimeric receptors and/or the model systems used in these studies. Moreover, we found that inhibition of AXL signaling does not affect viability of cells that have high AXL activation at doses sufficient to block AXL activity. Our self-consistent map of the phosphoproteomics data show upregulation of phosphorylation in nodes such as ERBB2, CDK2, several MAPKs (MAPK1, 3, 7), MYC, ETS1, and JUN, which are known to be involved in prosurvival signaling, and could explain adaptive mechanisms that help these cells evade death from AXL inhibition. These observations highlight a limitation in studying AXL's mechanism of activation and action, which is the lack of an ideal in vitro model system. AXL is rarely reported to be genetically altered and hence AXL-driven cell line models of cancers are lacking. Cell lines that we and others and have used to study AXL signaling are driven by other dominant drivers and have concomitant overexpression of AXL, such as the NRAS mutation driving H1299 cells. As such AXL functions as a co-/secondary driver in these models. Data derived from such models are complicated by the presence of these dominant background mutations making it challenging to delineate effects controlled by AXL versus the ones exerted by the background mutation.

Several AXL targeting agents are in various stages of preclinical development and more are advancing into different stages of clinical investigation (27). The majority of these AXL inhibitors are actually multi-kinase inhibitors (27) whose target profiles and mode of action have not been fully elucidated. As they are given to patients, this information will help predict and circumvent possible side-effects and future resistance mechanisms. Here, we used activity-based protein profiling to identify and compare the target binding profile of three AXL inhibitors, a TAM inhibitor RXDX106, an FDA-approved multi-kinase inhibitor cabozantinib, and an AXL inhibitor bemcentinib being evaluated in clinical trials. Our data shows diverse target profiles for each AXL inhibitor with bemcentinib having the most off-targets, several of which modulate the MAPK pathway. The promiscuity of this drug may explain the bemcentinib-induced AXL-independent phenotypic and signaling alterations that we and others have observed (58).

AXL has been linked to development of resistance to targeted cancer therapeutics. The mechanistic basis for AXL-mediated drug resistance is yet to be completely elucidated. In this study, we investigated the effect of AXL inhibitors on EGFR TKI resistance in lung cancers using dose-escalation and persister-derived cell line models of erlotinib resistance. The one common characteristic between all our cell line models of erlotinib resistance was the overexpression of AXL kinase. Combination treatment with AXL and EGFR TKI failed to resensitize the resistant cells to erlotinib in either model of acquired erlotinib resistance. Several studies have reported that upregulation of AXL in response to EGFR inhibition leads to resistance to these targeted agents and that genetic and/or pharmaceutical inhibition of AXL restored sensitivity to EGFR inhibitor in these models (7, 8, 59). However, there are also other groups that have reported that suppression of AXL could not restore the sensitivity of the resistant cells to drug treatment (29, 60). These contradictory reports could be attributed to differences in model systems used to study the role of AXL in driving resistance to targeted therapy in cancers. The functional importance of AXL in a particular resistance model can vary depending on the tissue type, driver mutation of the cell line, the method of generation of resistance, and the targeted therapy used. Our data suggests that AXL TKI is unlikely to be an effective strategy in EMT resistance to EGFR TKI, but other means of targeting AXL, such as antibody–drug conjugates, may be more promising.

Interestingly, despite AXL overexpression in all our erlotinib-resistant cell line models, we found a unique vulnerability of one persister-derived erlotinib resistant clone to the combination of erlotinib and bemcentinib. This clone, S2–10, is sensitive to the combination of erlotinib and bemcentinib, but resistant to both erlotinib and bemcentinib as single agents as well as to RXDX106 as a single agent or in combination. This unique AXL-independent phenotypic effect can probably be attributed to the numerous off-targets of bemcentinib as seen in our ABPP data. The mechanism by which bemcentinib exerts its AXL-independent phenotypic effects is not fully understood and follow-up studies with functional and phenotypic screens will be needed to answer that question. In phase I/II clinical trial combining bemcentinib with erlotinib in patients with EGFR-mutant NSCLC, it was reported that 1 of 8 (12.5%) enrolled patients, who had previously progressed on erlotinib monotherapy, showed a partial response when given erlotinib and bemcentinib combination (61). Overall, our data suggests that the promiscuity of bemcentinib may be beneficial for its successful transition to clinic, but caution should be exercised if using this drug as tool compound to investigate AXL-driven signaling.

As promising AXL-targeted therapies advance into various stages of clinical trials, a major limitation that needs to be addressed is the development of assays/biomarkers to stratify patients that would most likely benefit from AXL TKI therapy. Since AXL is rarely genetically altered in patients, AXL IHC is the most feasible strategy to identify patients. However, IHC assays are highly dependent on the quality of the antibody used and have the limitation of providing a measure of protein expression and not protein activity in a sample. For instance, existing studies that have reported AXL expression levels in lung cancers have been highly inconsistent, with expression rates varying from 33% to 93.2%, probably due to differences in antibodies and evaluation methods used (32). PLA, that annotate active protein complexes, may be a better and more effective assay for selecting AXL-driven tumors (35). To this end, we have established PLA to detect active AXL:pY signaling complexes in cell line models expressing phosphorylated AXL, which can be developed as an assay to measure drug-targetable active AXL complexes in clinical samples. Interacting proteins and putative substrates identified by BioID and phosphoproteomics respectively, can be used to guide further development of specific PLA to detect active AXL signaling complexes as diagnostic tools.

In summary, our study reports a systems level analysis of AXL biology, combining AXL protein complex annotation by BioID, signaling analysis using phosphoproteomics, and characterization of AXL TKI using chemical proteomics to elucidate mechanisms and phenotypes of AXL action.

M. Stroud reports grants from NIH; and grants from Florida Department of Health Bankhead-Coley Program during the conduct of the study. F. Kinose reports grants from NIH; and grants from Florida Department of Health Bankhead-Coley Program during the conduct of the study. J.M. Koomen reports grants from BMS outside the submitted work. E.B. Haura reports grants from Ignyta during the conduct of the study; personal fees and nonfinancial support from Janssen Pharmaceuticals; and grants and personal fees from Revolution Medicines outside the submitted work; in addition, E.B. Haura has a patent for Protein Protein Interactions as Biomarkers issued to Moffitt Cancer Center. No disclosures were reported by the other authors.

A. Majumder: Formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. S. Hosseinian: Validation, investigation, visualization, writing–original draft. M. Stroud: Validation, investigation, visualization, writing–original draft. E. Adhikari: Investigation, writing–original draft. J.J. Saller: Investigation, visualization, methodology. M.A. Smith: Formal analysis, validation, investigation, visualization. G. Zhang: Investigation, methodology. S. Agarwal: Validation, investigation, visualization. M. Creixell: Formal analysis, visualization, methodology, writing–original draft. B.S. Meyer: Investigation. F. Kinose: Formal analysis, investigation, visualization. K. Bowers: Software, investigation, visualization, writing–original draft. B. Fang: Data curation, software, supervision. P.A. Stewart: Software, formal analysis, supervision, visualization, writing–original draft. E.A. Welsh: Software, formal analysis, visualization, writing–original draft. T.A. Boyle: Supervision, writing–review and editing. A.S. Meyer: Supervision, writing–review and editing. J.M. Koomen: Resources, data curation, supervision, writing–review and editing. E.B. Haura: Conceptualization, supervision, funding acquisition, writing–original draft, writing–review and editing.

We would like to thank the Gingras Lab for providing and helping with the BioID vectors. This work has been supported in part by the Proteomics & Metabolomics, Biostatistics and Bioinformatics, Molecular Genomics, Analytic Microscopy, and Tissue Core Facilities at the H. Lee Moffitt Cancer Center & Research Institute, an NCI-Designated Comprehensive Cancer Center (grant nos. P30-CA076292). Funding for this work was supported by NIH (grant no. 5U01CA215709-03) to E.B. Haura and A.S. Meyer, and Florida Department of Health Bankhead-Coley Program (grant no. 5BC07) to E.B. Haura.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.