Dasatinib resensitizes BRAF/MEK inhibitor efficacy in patient-derived xenografts from patients with progression on BRAF/MEK inhibitor treatment

Summary

Although BRAF/MEK inhibitor (BRAFi/MEKi) therapy initially shows high efficacy in patients with BRAF^{V600 E/K} cutaneous melanoma, resistance develops in over 75% of cases. We tested robustness of the umbrella trial strategy in this population by analyzing relationships between genomic status of a gene and associated downstream consequences at the protein level. The results revealed weak relationships among mutations, copy-number amplification, and protein expression and activation. An in vivo compound repurposing screen using 11 clinically relevant agents from an NCI-portfolio with pan-RTK, non-RTK, and/or PI3K-mTOR specificity identified dasatinib as most capable of restoring sensitivity to BRAFi/MEKi in patient-derived xenograft (PDX) models originating from tumors that had progressed on BRAFi ± MEKi. High baseline expression of BRAFi/MEKi resistance-associated proteins (e.g., AXL, YAP, HSP70, and phospho-AKT) was predictive of the response to BRAFi/MEKi + dasatinib combination therapy. These findings suggest that adding dasatinib may help overcome resistance and restore anti-tumor activity in patients with BRAFi/MEKi-refractory cutaneous melanoma.

Graphical abstract

Highlights

• •Dasatinib restores BRAFi/MEKi response in resistant melanoma PDX models
• •Protein activity, not genotype, predicts therapeutic response in melanoma
• •BRAFi/MEKi + dasatinib suppresses resistant YAP⁺ tumor subpopulations
• •PDX-guided screening identifies dasatinib as a rational salvage therapy

Cancer; Therapeutics

Introduction

There has been immense progress in the first-line treatment for patients with metastatic cutaneous melanoma, with >14 therapies having been approved by the FDA since 2011. In addition to immune checkpoint inhibitor (ICI) therapy (e.g., ani-PD-1 +/− anti-CTLA4), BRAFi/MEKi treatment is an important strategy for 40%–50% of metastatic cutaneous melanoma patients whose tumors harbor activating BRAF^{V600 E/K} mutations. Vertical targeting of the mitogen-activated protein kinase (MAPK) pathway with the use of dual BRAFi/MEKi therapy has improved the overall survival^1,2,3; however, long-term responses beyond 5 years are limited by diverse mechanisms of resistance that encompass hyperactivation of compensatory signaling through the PI3K/AKT/mTORC1,⁴ SRC,⁵ and STAT3⁶ pathways via upregulation of receptor tyrosine kinases (RTKs) including insulin growth factor receptor 1 (IGFR-1),⁷ AXL,⁸ PDGFRβ,⁹ EGFR,¹⁰ c-MET,¹¹ and HER3 (ERBB3).¹² Tumor-infiltrating lymphocyte (TIL) therapy has been recently approved by the FDA as a standard of care option for patients with progressive disease following ICI and/or BRAFi/MEKi therapy.¹³ TIL therapy is highly involved and complex, and additional therapeutic strategies are sorely needed.

Due to the immense tumoral heterogeneity among cutaneous melanoma patients and the need for more personalized therapeutic strategies, the National Cancer Institute (NCI) is currently testing a master protocol in the NCI-molecular analysis for therapy choice (MATCH) trial, whereby patients with cancer are provided therapy regimens based on the genetic changes found in their tumors.^14,15 Although highly innovative, it remains to be seen how effective this strategy will be for patients with metastatic cutaneous melanoma. Trials such as the NCI-MATCH are designed to efficiently test multiple targeted therapies for a single disease process by stratifying patients based on predictive biomarkers often determined in other diseases (e.g., other cancer types) or other risk factors. The rationale is to leverage shared resources, streamline data collection, and potentially accelerate drug development. A set of scientific assumptions are ingrained that include (1) a correlation between an activating mutation in a pathway node and elevated activation of that respective pathway at the protein level (e.g., PI3K mutations correlate with activation of PI3K/AKT/mTOR signaling), (2) copy-number increases of a gene correlate with the protein expression of that gene (e.g., tumors with EGFR amplification would exhibit increased EGFR protein expression and activation), and (3) tumors that harbor a given mutation are more sensitive to inhibitors that target the related pathway (e.g., tumors that harbor AKT mutations should be more sensitive to an mTOR inhibitor). Indeed, there are genomic alterations that predict sensitivity to certain compounds in certain tumors, including BRAF^{V600 E/K} mutations for V600 E/K selective BRAF inhibitors (e.g., vemurafenib) in cutaneous melanoma¹⁶ and EGFR mutations for EGFR inhibitors (e.g., gefitinib) in patients with advanced lung adenocarcinoma.¹⁷ However, the robustness and applicability of the relationship between genomic status and anti-tumor efficacy of a cognate inhibitor across various cancer types remain to be fully understood. The variable predictive sensitivity of certain mutations across tumor types is increasingly being recognized, as in the case of BRAF^{V600 E/K} mutant colorectal cancers that were found to be significantly less sensitive to BRAF inhibitors relative to BRAF^{V600 E/K} mutant cutaneous melanoma.¹⁸

Using a genomically, transcriptionally, and proteomically characterized set of patient-derived xenograft (PDX) models of metastatic cutaneous melanoma and cutaneous melanoma patient samples in The Cancer Genome Atlas (TCGA), we, here, analyzed the relationship between mutation, copy-number status, and total-/phospho-protein expression, revealing poor correlations. To investigate an effective salvage strategy for patients who progressed on BRAFi/MEKi, we developed an umbrella trial-styled PDX therapy repurposing screen to investigate potential triple combination strategies including BRAFi/MEKi alongside a third inhibitor intervening on the PI3K/AKT/mTOR pathway or a pan-RTK inhibitor. Although the list of reported resistance mechanisms to BRAFi/MEKi continues to grow,^{19,20,21,22,23,24,25,26} they predominately center around reactivation of the MAPK pathway and the hyperactivation of parallel survival signaling, with the latter often mediated through the PI3K/AKT/mTOR pathway via elevated receptor tyrosine kinase (RTKs) and non-receptor kinase activity. Of the 11 triple combinations screened, BRAFi/MEKi + dasatinib combination was identified as the most effective. Dasatinib has been previously identified to have synergy in combination with BRAFi against BRAF^{V600 E/K} mutant melanoma cell lines preclinically,^27,28,29 but the efficacy of this combination has not been extensively tested in vivo nor has not been tested against cell line or PDX models directly established from patients who have relapsed to BRAFi ± MEKi in the clinical setting.³⁰ These in vivo findings independently validate recent preclinical in vitro findings that dasatinib could suppress aggressive melanoma subpopulations (e.g., mesenchymal cells and neural crest stem cell-like cells)³⁰ and provide strong scientific rationale to investigate the utility of BRAFi/MEKi + dasatinib as a salvage strategy for patients with BRAF^{V600 E/K} mutated metastatic cutaneous melanoma who have relapsed to BRAFi ± MEKi.

Results

BRAFi/MEKi-resistant tumors display elevated RTK/MAPK/PI3K/mTOR activity

We here tested aspects of the umbrella trial concept in a repurposing drug screen to determine whether the (1) gene mutational status correlates with activation at the protein level of the associated pathway, and/or 2) gene mutational status correlates with tumor sensitivity toward a cognate inhibitor (Figure 1A). To assess the resistance driver landscape in metastatic cutaneous melanoma with BRAF^{V600 E/K} mutations, we performed reverse-phase protein array (RPPA) analysis of 94 treatment-naïve, 22 BRAFi-resistant, and 16 BRAFi/MEKi-resistant PDX models, revealing a heterogeneous enrichment in various total and phospho-proteins involved in RTK, PI3K/AKT/mTOR, and MAPK signaling (Figure S1A). IGFBP2, phospho-MEK (S217/S221), and phospho-ERK1/2 (T202/Y204) levels were significantly elevated among BRAFi- and/or BRAFi/MEKi-resistant PDX models relative to therapy naive models, in line with previous reports.^31,32,33 AXL, a dogmatic marker of the therapy-resistant melanoma cell state,³⁴ was not significantly increased in BRAFi- and BRAFi/MEKi-resistant models. However, reduced MITF expression was observed, suggesting melanoma models with acquired resistance may occupy a de-differentiated cell state, as previously reported.^35,36 Multiple nodes of the PI3K/AKT/mTOR pathway (e.g., phospho-AKT T308/S473, phospho-mTOR S2448, phospho-4E-BP1 T37/T46/S65, and phospho-S6 S235/S236/S240/S244) were elevated in therapy-resistant PDX models relative to therapy-naïve models, albeit not to statistical significance likely due to the immense inter-tumoral signaling heterogeneity (Figure S1A).

Figure 1.

BRAFi/MEKi-resistant tumors display elevated RTK/MAPK/PI3K/mTOR activity

(A) Schematic for the role of the central dogma theory in clinical practice. Dotted lines denote assumptions with poor validation in the melanoma literature.

(B) Graphical summary of the experimental strategy of this manuscript.

(C) Targets of the 11 FDA-approved compounds used in our repurposing screen.

(D) Therapy history of the 5 patients whose tumors were used to establish R-PDX models that were subsequently screened against 11 BRAFi/MEKi-based triple combinations.

To test the umbrella trial design and identify a potentially effective salvage strategy for patients who have progressed on SOC BRAFi/MEKi therapy, we designed a custom drug screen using 11 widely used small molecule inhibitor tool compounds from an NCI drug repository, Cancer Therapy Evaluation Program and Developmental Therapy Program; the compounds possess potent in vivo pan-RTK, non-RTK, and/or PI3K-mTOR specificity (Figures 1B, 1C, and Table S1). Each of the compounds used in this study have been previously clinically tested in patients with advanced cutaneous melanoma, but neither in combination with BRAFi/MEKi nor in the context of patients who have previously progressed on BRAFi or BRAFi/MEKi in a stage II or more advanced trial. With the goal of identifying an effective three-drug combination consisting of BRAFi/MEKi in combination with a third compound, we initially selected a discovery set of 5 therapy-resistant (R)-PDX, established from patients with previous tumor progression while on treatment with BRAFi in the presence or absence of MEKi. Three of these patients showed tumor progression while receiving immune checkpoint blockade (anti-CTLA-4 or anti-PD-1), and 1 patient showed progression even after receiving radiation therapy. This discovery set was used to identify an effective triple combination we would then validate in an expanded cohort of R-PDX (Figure 1D).

Protein expression and genetic status of the PI3K/-mTOR pathway do not robustly predict efficacy of PI3K/mTOR pathway inhibitors

We tested the robustness of the umbrella trial strategy in predicting sensitivity to four PI3K-mTOR pathway inhibitors (copanlisib, AZD8186, temsirolimus, and TAK228) among the 5 discovery R-PDX models included in our repurposing screen. We performed targeted sequencing and copy-number variation (CNV) analyses of genes in the PI3K/AKT axis (Figure 2A), as well as reverse-phase protein arrays (RPPAs) to determine the baseline activation status of the PI3K/mTOR pathway in our discovery set of R-PDX models (Figure 2B). We observed no correlation between (1) AKT3 copy-number amplification and AKT activity at the protein level, and (2) NRAS^G13R mutation and hyperactivation of PI3K/AKT and MAPK signaling at the protein level. Of note, NRAS^G13R mutations occur in only 1%–2% of melanomas and may not associate with downstream PI3K and MAPK activation as other more common NRAS mutations occur at codons 12 and 61. In agreement with previous findings,^37,38,39 PTEN copy-number loss was associated with reduced PTEN protein expression and the greatest hyperactivation of PI3K/AKT signaling, as seen in the WM4008 model (Figure 2B).

Figure 2.

Protein expression and genetic status of the PI3K/-mTOR pathway do not robustly predict the efficacy of PI3K/mTOR pathway inhibitors

(A) Mutational and CNV analyses of PI3K/AKT pathway nodes across the R-PDX models in the discovery screen.

(B) RPPA analysis of PI3K/AKT pathway nodes across the R-PDX models in the discovery screen.

(C) Analysis of the relationship between AKT/mTOR protein activation status and PI3K mutations.

(D) AKT mutations and (E) PTEN copy-number loss in cutaneous melanoma patients in the TCGA.

(F) Therapy sensitivity prediction analysis based on the unique genomic and/or proteomic signature of each R-PDX and the protein targets of a given agent.

(G) Tumor volumes for NSG mice implanted with 1 of the 5 R-PDX models organized by row and treated with vehicle (black line), BRAFi/MEKi (blue line), or BRAFi/MEKi plus a third agent (red line) as labeled. n = 3 for each individual treatment arm per R-PDX model. Tumor volumes are shown with imputed values after mouse death to avoid artificial decreases in group averages; imputation was applied solely for graphical continuity and not for statistical analysis.

(H) Differences in growth rate with log₂ scale of tumor volume between treatment groups (Dab/Tra + other treatments versus Dab/Tra).

(I) Rank of tumor sensitivity across the 5 R-PDX models for the four drugs. ∗ signifies p < 0.05.

Given the small sample size of our R-PDX models, we analyzed the relationship between genomic status and pathway activation in 90 distinct patients with cutaneous melanoma in the TCGA for whom RPPA and genomic information were available. Notably, patients whose tumors harbor PIK3C mutations did not display elevated AKT (pAKT S473) or mTORC1 (pS6 S240/244) activity at the protein level relative to patients with wild-type PIK3C tumors (Figure 2C). This is on par with previous reports showing variable levels of pAKT in cancer cell lines depending on whether helical mutations or kinase-domain mutations are present in PIK3CA.³⁹ Although sample size was limited in the TCGA, patients whose tumor harbor AKT mutations also did not display elevated AKT or mTORC1 activity at the protein level relative to patients with wild-type AKT tumors (Figure 2D). The patients’ PTEN status again inversely correlated with AKT activity; however, the downstream mTORC1 activity did not correlate with PTEN status (Figure 2E).

On the basis of our multi-omic characterization, we classified R-PDX models in our discovery set as potentially sensitive or insensitive to PI3K or mTOR inhibition, whereby WM4008 was predicted to be the most sensitive (based on PTEN copy-number loss and elevated phospho-AKT and phosho-S6 levels) and WM4262 was predicted to be the most insensitive (based on low phospho-AKT and phospho-S6 levels) (Figure 2F and Table S1). To begin our therapy trials, we again confirmed that all 5 of the discovery set R-PDX models were resistant to BRAFi/MEKi, as seen in the near identical tumor growth patterns in NSG mice treated with vehicle control versus a combination of BRAFi (dabrafenib, 150 mg/kg) and MEKi (trametinib, 1.5 mg/kg) (Figure 2G). Treatment with the pan-PI3K inhibitor copanlisib in combination with BRAFi/MEKi conferred a significant delay in tumor growth in 2 of the 5 R-PDX models relative to BRAFi/MEKi treatment. A weak relationship between PI3K/AKT pathway status and copanlisib + BRAFi/MEKi efficacy was observed, whereby the R-PDX WM4008, predicted to be the most sensitive, did not display sensitivity. Treatment with the PI3K β/δ isoform-specific inhibitor AZD8186 elicited a significant reduction in tumor growth in 1 of the 5 R-PDX models relative to BRAFi/MEKi treatment. We next tested the efficacy of the allosteric and catalytic mTOR inhibitors temsirolimus and TAK228, respectively. Temsirolimus + BRAFi/MEKi modestly increased the anti-tumor activity of BRAFi/MEKi in 2 of the 5 R-PDX models relative to BRAFi/MEKi treatment; however, the R-PDX predicted to be sensitive (highest mTOR activity as evidenced by high phospho-S6 and phospho-4E-BP1 levels) did not respond. Treatment with TAK228 + BRAFi/MEKi demonstrated mild anti-tumor activity in 1 of the 5 R-PDX models, and again did not have efficacy in the R-PDX predicted to be the most sensitive (Figures 2G, 2H, 2I, and Table S2). Immunohistochemical staining of WM4011-2 tumors treated with the various PI3K/mTOR inhibitor regimens revealed TAK228 as being capable of most potently inhibiting pAKT S473 and pS6 S240/244, despite not having the greatest anti-tumor activity (Figure S2C). Altogether, these data from our small discovery panel of R-PDX models suggest that treatment decisions for patients enrolled in umbrella trials with PI3K/mTOR inhibitors may not be optimal if solely based on genetic, genomic, or proteomic information of the PI3K/mTOR pathway. Further, the genomic status of the PI3K/AKT pathway is not predictive of AKT and mTORC1 activity at the protein level, suggesting the need for better genetic biomarkers of PI3K/AKT pathway activity.

Protein expression, not genetic status, of RTKs partially predicts efficacy of pan-RTK inhibitors

We next tested the robustness of the umbrella trial approach in predicting R-PDX sensitivity to the 6 RTK inhibitors in our repurposing screen (Figure 1C). Mutational and CNV analyses of a panel of RTK genes in our targeted sequencing panel revealed 2.5- to 4-fold copy-number amplifications in EGFR and MET in 1 and 2 R-PDX models, respectively (Figure 3A). In addition, at least 1 mutation in EGFR, ERBB2, ERBB4, FGF4, and/or FGFR3 was detectable in 4 of the 5 R-PDX models. The R-PDX model WM4008 harbors (1) an EGFR^G721C mutation, which is associated with non-small cell lung cancer oncogenesis and is a predictive biomarker for EGFR inhibitors including afatinib, (2) an ERBB2^E2X mutation that is predicted to be deleterious, though no literature is available characterizing this mutation, and (3) an FGF4^E42Q mutation that is a variant of unknown significance (Figure 3A). The R-PDX model WM4011 harbors a ERBB4^G500E mutation that is a variant of unknown significance. The R-PDX model WM4258 harbors a FGFR3^S764L mutation that is predicted to be likely deleterious. The R-PDX model also possesses FGFR3^A558V and FGFR3K612N mutations that are variants of unknown significance (Figure 3A). The R-PDX model WM4262 did not harbor any mutations in the RTKs monitored. As could be expected, RPPA analyses revealed heterogeneous RTK protein expression and activation status across the discovery set of R-PDX models (Figure 3B). There was a weak relationship between the CNV status and protein expression of a given RTK. For example, the R-PDX WM4262 exhibited EGFR copy-number amplification from 2.5 to 4 but did not display elevated EGFR protein expression relative to R-PDX without EGFR copy-number gains. Given the small sample size of our R-PDX models, we again analyzed the relationship between the genomic status and protein expression of RTKs in 90 melanoma patients from the TCGA where RPPA and genomic information are available (Figures 3C–3H). There was no significant correlation detected between the copy-number status and protein expression of EGFR (Figure 3C), PDGFRβ (Figure 3D), HER2 (Figure 3E), VEGFR2 (Figure 3F), HER3 (Figure 3G), and AXL (Figure 3H) between patients with gene amplifications and patients with normal copy numbers.

Figure 3.

Protein expression, not genetic status, of RTKs partially predicts efficacy of pan-RTK inhibitors

(A) Mutational and CNV analyses of RTK nodes across the R-PDX models in the discovery screen.

(B) RPPA analysis of RTK nodes across the R-PDX models in the discovery screen.

(C) Analysis of the relationship between RTK protein expression and gene copy-number gains in EGFR, (D) PDGFRβ, (E) HER2, (F) VEGFR2, (G) HER3, and (H) AXL.

(I) Therapy sensitivity prediction analysis based on the unique genomic and/or proteomic signature of each R-PDX model and the protein targets of a given agent.

(J) Tumor volumes for NSG mice implanted with 1 of the 5 R-PDX organized by row and treated with vehicle (black line), BRAFi/MEKi (blue line), or BRAFi/MEKi + a third agent (red line) as labeled; n = 3 for each individual treatment arm per R-PDX model. Tumor volumes are shown with imputed values after mouse death to avoid artificial decreases in group averages; imputation was applied solely for graphical continuity and not for statistical analysis. Of note, the same data points for control and BRAF/MEK inhibitor treatment are represented multiple times for each PDX from Figure 2G.

(K) Waterfall plots depicting median overall survival for the R-PDX discovery set treated with the 11 BRAFi/MEKi-based triple combinations. N = 15 per treatment arm, representing data from 5 R-PDX models (WM4011-2, WM4262-3, WM4258-2, WM4008-2, and WM4380-2). A “death” is defined as a tumor whose volume is ≥1,000 mm³

(L) Differences in growth rate with log₂ scale of tumor volume between the treatment groups (Dab/Tra + other treatment versus Dab/Tra).

(M) Rank of tumor sensitivity across the 5 R-PDX models for the seven drugs. ∗ signifies p < 0.05.

Based on this genomic and proteomic characterizations, we classified the R-PDX models in our discovery set as potentially sensitive or insensitive to pan-RTK inhibition (Figure 3I and Table S1). WM4258 was predicted to be the most sensitive due to greatest pan-RTK expression; WM4008 and WM4262 were predicted to be the least sensitive (due to lowest pan-RTK expression) (Figure 3I). Of note, phospho-cMET levels did not vary significantly across the R-PDX models; therefore, differences in the sensitivity to cMET inhibition could not be predicted. Treatment with the selective cMET inhibitor savolitinib + BRAFi/MEKi did not confer anti-tumor benefit in any of the discovery R-PDX panel (Figures 3J, 3K, and S2A). Treatment with the pan-RTK inhibitor cabozantinib (XL-184) + BRAFi/MEKi conferred anti-tumor activity in 3 of the 5 R-PDX models predicted to have sensitivity based on their RTK protein expression (e.g., PDGFRβ). The pan-RTK inhibitor cediranib revealed a similar trend, conferring anti-tumor activity in 2 of the 5 R-PDX models, both of which were predicted to be sensitive to pan-RTK inhibition. To a lesser extent, pazopanib and sunitinib also demonstrated modest efficacy in R-PDX with elevated RTK protein expression, except for pazopanib + BRAFi/MEKi having anti-tumor activity in 1 R-PDX (WM4008) not predicted to respond due to low RTK expression. Treatment with the SRC family kinase and pan-RTK inhibitor dasatinib + BRAFi/MEKi conferred the greatest anti-tumor activity among all 11 triple combinations consisting of BRAFi/MEKi plus a third agent tested (Figures 3J, 3K, 3M, and 3L; Table S2). Treating our R-PDX models with 7 RTK inhibitors revealed/uncovered that receptor mutations and copy number alterations were poorly predictive of the response to their corresponding inhibitors. In contrast, an encouraging relationship between RTK protein expression and pan-RTK inhibitor sensitivity emerged.

Dasatinib resensitizes R-PDX models with high expression of canonical melanoma BRAFi/MEKi resistance signatures

To further investigate the dasatinib + BRAFi/MEKi regimen, we screened an expanded validation set of 15 additional R-PDX models and found that combination dasatinib + BRAFi/MEKi treatment (1) caused significant reductions in tumor volumes at denoted timepoints in 11 PDX models (Figure S3A; Table S3) and (2) increased the median overall survival (mOS) of R-PDX-bearing mice by 2-fold relative to that of the mice treated with BRAFi/MEKi or dasatinib alone (Figure 4A). Notably, treatment of a therapy-naive PDX model with BRAFi/MEKi + dasatinib did not extend the time to relapse relative to BRAFi/MEKi treatment alone, suggesting this triple combination may be better suited as a second- or third-line therapy to address acquired, rather than intrinsic, resistance mechanisms (Figure S2B; Table S2). We next interrogated differential total- and phospho-protein expression levels between the R-PDX predicted to be the most sensitive and least sensitive to the dasatinib + BRAFi/MEKi cocktail. R-PDX models predicted to be most sensitive to the dasatinib + BRAFi/MEKi combination treatment exhibited elevated levels of several reported canonical drivers of therapy resistance in melanoma, including HSP70,⁴¹ AXL,³⁴ phospho-AKT S473, phospho-AKT T308,⁴ and YAP.⁴² In addition, R-PDX predicted to be the most sensitive to dasatinib + BRAFi/MEKi displayed (1) elevated protein levels of MRAP, p27, IRS1, UBAC1, LAD1, Notch, XPF, and c-ABL and (2) reduced protein levels of PTEN, ATG4B, MYT1, GAB2, and DDB1 (Figure 4B). Transcriptomic analyses of 8 SRC family members revealed a significant correlation between SRC mRNA and sensitivity to the dasatinib + BRAFi/MEK regimen (Figure S3B). However, phosphorylation of SRC at Y416 did not correlate with dasatinib + BRAFi/MEKi sensitivity (Figure S3C). Interestingly, the expression of 8 additional reported dasatinib targets (e.g., AURKA, STK25, and RIPK2) correlated with sensitivity to the dasatinib + BRAFi/MEKi regimen (Figures S3D and S3E). Markers of sensitivity to the BRAFi/MEKi + dasatinib combination (i.e., AXL, pAKT S473, pAKT T308, and c-ABL) negatively correlated with expression of the melanocyte lineage factor MITF across a panel of >300 melanoma PDX models, suggesting that the dasatinib combination therapy is most effective in de-differentiated melanomas that are considered the most difficult to address clinically (Figures 4C–4F).^19,43 Interestingly, analysis of the reported dasatinib targets across minimal residual disease (MRD) subpopulations shown to drive therapy resistance in previously published patient scRNAseq data revealed the expression of dasatinib targets in distinct subpopulations (Figure 4G).^43,44 For example, the dasatinib target PDGFRβ is highly expressed in neural crest-like stem cell melanoma subpopulations but lowly expressed in invasive, pigmented, and starved-like melanoma cell populations.⁴³ In contrast, AXL is highly expressed in the invasive subpopulation but lowly expressed in other MRD subpopulations. These data suggest that dasatinib may have potent activity against R-PDX due to its broad target profile capable of concurrently suppressing distinct MRD subpopulations. This corroborates a recent report demonstrating the ability of dasatinib to inhibit certain MRD subpopulations (e.g., mesenchymal, neural crest stem cell [NCSC]-like, and invasive cells³⁰). In agreement, the expressions of ALDH1A and KDM5B, each of which is the marker of stem cell-like melanoma cells, correlated significantly with sensitivity to the dasatinib + BRAFi/MEKi regimen (Figures S3F and S3G). In the final set of experiments, the WM4258 (R-PDX)-bearing mice were treated with BRAFi/MEKi ± dasatinib and characterized by single-cell RNA sequencing (scRNAseq), which revealed that three melanoma subpopulations depleted following the addition of dasatinib to BRAFi/MEKi and exhibited the following: (1) the YAP-associated downstream effector CENPF,⁴⁵ (2) dormancy and neural crest stem cell-like markers (e.g., NR2F1 and SOX4), (3) RTK nodes (e.g., FGFR1, MET, and HGF), and (4) antiapoptotic nodes (e.g., BID; cluster 5) (Figures 4H–4J). Further, the addition of dasatinib to BRAFi/MEKi reduced the expression of ABL1, MET, NGFR, and YAP1 (Figure 4K). We performed clonal analysis on the scRNAseq datasets of control DT and those treated in combination with dasatinib by leveraging a recently reported DNA barcoding technology, FateMap.⁴⁰ With FateMap, we found that the resistant clones from BRAFi/MEKi + dasatinib treatment were smaller than those from BRAFi/MEKi treatment alone (average clone size normalized for tumor volume: 7.81 for BRAFi/MEKi and 2.06 for BRAFi/MEKi + dasatinib; UMI cutoff = 2). Furthermore, clones from the BRAFi/MEKi + dasatinib condition existed predominantly as singlets (i.e., clones that are not dividing) as compared to those treated with BRAFi/MEKi (percentage singlets: 32.2% for BRAFi/MEKi + dasatinib and 20.1% for BRAFi/MEKi; UMI cutoff = 2) (Figures 4L and 4M). Collectively, these findings suggest that the reduced resistant tumor volume in BRAFi/MEKi + dasatinib treatment can, at least in part, be explained by an attenuated ability of resistant clones to divide in the presence of drug. We found these results to be consistent for different UMI cutoffs for the clonal assignment of single cells.

Figure 4.

Dasatinib resensitizes R-PDX models with high expression of canonical melanoma therapy resistance mechanisms

(A) Overall survival plots of the expanded validation set of 9 R-PDX-bearing mice treated as shown; n > 26 mice per arm.

(B) Differential total- and phospho-protein correlation analysis with sensitivity to BRAFi/MEKi + dasatinib treatment.

(C) Paired protein expression analysis of MITF with AXL, (D) phospho-AKT S473, (E) phospho-AKT T308, and (F) c-ABL.

(G) Correlation analysis of dasatinib targets and distinct melanoma populations dominant in minimal residual disease in scRNAseq data from melanoma patients (Tirosh et al.⁴⁴).

(H) NSG mice implanted with the R-PDX WM4258 and treated as shown once tumors were palpable.

(I) t-SNE plot of scRNAseq data from WM4258 tumor cells treated in vivo with BRAFi/MEKi ± dasatinib from (H).

(J) Top gene markers for melanoma subpopulations that are enriched (clusters 1 and 3) or depleted (clusters 2, 4, and 5) following the addition of dasatinib to BRAFi/MEKi. ∗ signifies p < 0.05.

(K) UMAPs visualizing changes in ABL1, MET, NGFR, and YAP1 between the tumors treated with BRAFi/MEKi ± dasatinib.

(L) Clonal analysis on the single-cell RNA sequencing datasets of BRAFi/MEKi and those treated in combination with dasatinib by leveraging a recently reported DNA barcoding technology, FateMap.⁴⁰

(M) Differential expression analysis on five largest clones from DT alone as compared to the singlets from BRAFi/MEKi ± dasatinib.

Finally, we asked whether the genes that are differentially expressed in BRAFi/MEKi + dasatinib as compared to BRAFi/MEKi alone treatment exhibit any of the known dasatinib targets.^46,47 At the sample level, we found several genes from the “dasatinib-ome” to be downregulated specifically in BRAFi/MEKi + dasatinib treatment, namely STK25, RPS6KA3, EGFR, CDK2, NEK9, PRKDC, PKN2, EPHB2, CSK, DTYMK, and TP53RK (Figure 4M). Next, for a similar analysis at the clonal level, we performed differential expression analysis on five largest clones from BRAFi/MEKi alone treatment as compared to the singlets from BRAFi/MEKi + dasatinib treatment. Again, the dasatinib-ome genes STK25, GSK3A, and RPS6KA3 were downregulated specifically in BRAFi/MEKi + dasatinib treatment (Figure 4M). Here again, the findings were consistent for different UMI cutoffs for the clonal assignment of single cells. These results suggested that dasatinib targets putatively act to reduce clone size and tumor volume in BRAFi/MEKi + dasatinib-treated mice and that the BRAFi/MEKi + dasatinib combination has the potential to address acquired BRAFi/MEKi resistance, possibly due to its broad target profile capable of concurrently eliminating distinct MRD subpopulations.

Discussion

Our findings highlight a key limitation of the central dogma in metastatic cutaneous melanomas—the identification of weak correlations between RTK copy number amplifications and protein expression undermines their utility for guiding treatment selection. Mutations in PI3K/AKT do not robustly correlate with the downstream activation of AKT and mTORC1 at the protein level, which suggests that clinical decisions regarding therapy may not be completely based on genomic characterization alone. For example, a patient whose tumor harbors an AKT mutation and a BRAF^{V600 E/K} mutation would be predicted to (1) possess elevated AKT pathway activity at the protein level (e.g., elevated phospho-AKT S473), (2) be relatively more sensitive to an AKT or downstream mTOR inhibitor relative to a patient with wild-type AKT, and (3) experience a longer lasting response to therapy. However, prior reports suggest this may not be the case, with evidence indicating that PIK3CA-mutant cancer cell lines and human breast tumors display only (1) minimal AKT activity and (2) partial sensitivity to AKT inhibitors.³⁹

Overall, our correlation studies between mutations, CNVs, and protein activity underscore the potential of using single-cell approaches and/or IHC analyses of tumor tissue complementing genetic characterization to improve therapy efficacy predictability. We propose that preclinical trials using PDX to test therapies being considered for clinical translation represent an effective mechanism to address the profound inter-tumoral heterogeneity that serves as a hurdle in therapy durability. In contrast to in vitro studies that have reported significant benefit to targeting the PI3K/AKT/mTORC1 axis or the c-MET RTK to address BRAFi/MEKi resistance, our in vivo R-PDX therapy trials recapitulate the clinical trial findings demonstrating a lack of efficacy from PI3K, mTOR, and cMET inhibitors in combination with BRAFi/MEKi in melanoma patients. Interestingly, the WM4011-2 model was the most sensitive R-PDX to the combination treatment with BRAFi/MEKi + any one of the four PI3K pathway inhibitors tested; however, the WM4011-2 model did not harbor elevated PI3K pathway activity relative to the other R-PDX models that were less sensitive to the PI3K pathway inhibitors. There may be a unique suite of mutations and/or protein expression underlying the elevated sensitivity of WM4011-2 to PI3K pathway inhibitors, warranting further investigation. We propose leveraging this melanoma R-PDX paradigm for the pre-clinical screening of potential therapeutic regimens toward optimizing candidates for phase-1 and phase-2 clinical trials.

Although the broad spectrum RTK inhibitors used in our screen can present challenges in identifying associations between drug responses and specific genetic alterations, our data reveal potential relationships between pan-RTK protein expression and sensitivity to pan-RTK inhibitors. In addition, our R-PDX repurposing screen identifies the combination BRAFi/MEKi + dasatinib as having the greatest potential as a salvage strategy for melanoma patients that have already relapsed to BRAFi ± MEKi therapy. A role for dasatinib in BRAFi-resistant melanoma has been previously suggested^27,29,30; however, the present study’s extensive in vivo anti-tumor validation of the BRAFi/MEKi + dasatinib combination therapy trials in PDX mouse models from cutaneous melanoma patients that have progressed on BRAFi ± MEKi has not been previously executed. We identify putative dasatinib targets to be informative of sensitivity, including RTK and non-RTK targets that might explain the elevated efficacy of dasatinib compared with other pan-RTK inhibitors. Future work will focus on the analysis of phospho-proteins of downstream signaling pathways targeted by dasatinib. Although the BRAFi/MEKi + dasatinib combination therapy demonstrated significant activity against a subset of R-PDX and represents a promising therapeutic strategy for patients, optimization of this approach should be explored within context of dasatinib’s broad target profile.

Limitations of the study

We cannot disregard potential genetic heterogeneity in our PDX models (e.g., co-occurrence of an NRAS mutation and BRAF mutation in the same R-PDX) further confounding any correlations between genetic status with pathway activation at the protein level. We included the WM4262 model in this study, which harbors an NRAS and BRAF mutation, to represent this genetic heterogeneity. Inclusion of just one model harboring an NRAS-mutation limits our ability to make definitive conclusions and future studies further investigating the role of genetic heterogeneity are warranted. Of note, we do not observe tumor regression, but rather significant cytostasis of R-PDX models treated with the dasatinib + BRAFi/MEKi combination. Further therapy screening in the future should continue to identify therapy strategies that can achieve tumor regression preclinically to guide the development of strategies most likely to confer the greatest anti-tumor effect for cutaneous melanoma patients that have exhausted existing BRAFi/MEKi strategies.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Vito W. Rebecca (vrebecc2@jh.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •Bulk RNAseq genomics data have been deposited. GSE314755 “Dasatinib resensitizes BRAF/MEK inhibitor efficacy in patient-derived xenografts from patients with progression on BRAF/MEK inhibitor treatment.” scRNA-seq and barcode data have been deposited in NCBI GEO under the accession number GEO: >GSE273120 (access token: qpslieegthodbc).
• •No custom code was used. All code packages used in the study are described in methods.
• •All other data supporting the findings of this study are available within the article and its supplemental files.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Validated RPPA antibody collection	Functional Proteomics RPPA Core, MD Anderson Cancer Center	https://www.mdanderson.org/research/research-resources/core-facilities/functional-proteomics-rppa-core/education-and-references.html
Anti-phospho-S6 (Ser240/244, clone D68F8)	Cell Signaling Technology	Cat# 5364
Anti–phospho-AKT (Ser473, clone D9E)	Cell Signaling Technology	Cat# 4060
Bacterial and virus strains
E. coli Stbl3 (for lentiviral plasmid amplification)	Thermo Fisher Scientific	Cat# C737303
Biological samples
Human melanoma patient tumor samples (for PDX generation)	Wistar Institute Biorepository (IRB-approved)	–
Chemicals, peptides, and recombinant proteins
Dasatinib	CTEP/DTP	–
Dabrafenib	CTEP/DTP	–
Trametinib	CTEP/DTP	–
Copanlisib	CTEP/DTP	–
AZD8186	CTEP/DTP	–
Temsirolimus	CTEP/DTP	–
TAK228	CTEP/DTP	–
Savolitinib	CTEP/DTP	–
Cabozantinib	CTEP/DTP	–
Cediranib	CTEP/DTP	–
Pazopanib	CTEP/DTP	–
Sunitinib	CTEP/DTP	–
Sorafenib	CTEP/DTP	–
Critical commercial assays
QIAamp DNA Mini Kit	Qiagen	Cat# 51306
Chromium Next GEM Single Cell 3′ Kit v3.1	10× Genomics	Cat# 1000121
Deposited data
Bulk RNAseq genomics data scRNA-seq and barcode sequencing data	This paper This paper	GSE314755 GEO: GSE273120 (token: qpslieegthodbc)
Experimental models: Cell lines
WM4258 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM4011 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM4380 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM4262 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM4008 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM3960 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM4298 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM4239 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM4351 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM3939 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM4505 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM3942 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM3901 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM4398 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM4335 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM4509 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM4336 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM4276 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM4404 (human melanoma PDX)	Wistar Institute Melanoma Core	–
WM4262 (human melanoma PDX)	Wistar Institute Melanoma Core	–
HEK293FT	Thermo Fisher Scientific	Cat# R70007
Experimental models: Organisms/strains
NOD.Cg-Prkdcscid Il2rgtm1Wjl/SzJ (NSG) mice	Jackson Laboratory	Cat# 005557
Oligonucleotides
Lentiviral barcode oligo pool (FateMap library)	Goyal et al., 2023; synthesized by IDT	–
Recombinant DNA
psPAX2 packaging plasmid	Addgene	Cat# 12260
pMD2.G (VSVG envelope plasmid)	Addgene	Cat# 12259
Lentiviral barcode plasmid (FateMap vector)	Goyal et al., 2023	–
Software and algorithms
Cell Ranger (v3.1.0)	10× Genomics	https://www.10xgenomics.com
Seurat (v4–v5)	Satija Lab	https://satijalab.org/seurat/
ALRA (v0.0.2)	Linderman et al., 2022	https://github.com/KlugerLab/ALRA
Starcode (barcode collapsing, v1.3)	Zorita et al., 2015	https://github.com/gui11aume/starcode
RPPA SPACE	MD Anderson Cancer Center	Shehwana et al., 2022
GraphPad Prism (v9)	GraphPad Software	https://www.graphpad.com
R (v4.3.2)	R Project	https://www.r-project.org
Python (v3.9)	Python Software Foundation	https://www.python.org
Other
Huron TissueScope Scanner	Huron Digital Pathology	–
Grace Bio-Labs Nitrocellulose Slides	Grace Bio-Labs	Cat# 305152

Experimental model and study participant details

Patient-derived xenograft (PDX) models

All PDXs used in this study were previously established under institutional review board (IRB)-approved protocols with informed consent and reported.⁴⁸ These PDX experiments were performed under the approval and supervision of the WistarInstitute Institutional Animal Care and Use Committee (IACUC).

PDX tumors were propagated in NOD.Cg-Prkdc<scid> Il2rg<tm1Wjl>/SzJ (NSG) male mice (Jackson Laboratory) between 7-10 weeks of age. Tumors were maintained subcutaneously and passaged when volumes reached approximately 1,000-1,500 mm³. For treatment trials, R-PDX tumors were expanded until they reached 100-200 mm³ and then randomized into treatment groups (n = 3 mice per group). Tumor dimensions were measured three times per week with calipers, and volumes were estimated as (length × width × width)/2. Mice were sacrificed after 3-6 weeks of treatment or upon reaching institutional endpoint criteria.

All animal experiments were conducted in accordance with institutional and national ethical guidelines. 6-8 weeks male mice were used in animal experiments. All animal procedures were conducted in accordance with institutional guidelines and regulations for the humane treatment of laboratory animals. The experimental protocols were reviewed and approved by the Institutional Animal Care and Use Committee (IACUC) of the Wistar Institute under protocol numbers #111954.

Method details

Massively parallel DNA sequencing

Genomic DNA from PDX tumors was extracted using the QIAamp DNA Mini Kit (Qiagen) and analyzed using a custom 108-gene targeted sequencing panel. Variants were called and annotated for mutations and copy number alterations. To minimize artifacts from mouse stromal DNA contamination, variants not previously reported and with an allelic fraction < 0.15 were filtered out.

DNA barcoding library generation and transduction

FateMap barcode libraries were generated following adapted protocols from Goyal et al. (2023). Briefly, lentivirus was produced in HEK293FT cells cultured in DMEM + 10% FBS and penicillin/streptomycin. Cells were transfected at near confluency using polyethylenimine (PEI, Polysciences #23966) and plasmids encoding VSVG (5 μg), pPAX2 (7.5 μg), and the barcoding library (7.35 μg). Virus-containing supernatants were collected every ∼11 hours for 30 hours, filtered (0.45 μm PES), and stored at -80 °C.

For transduction, WM4258 melanoma cells were plated at 1 × 10⁵ cells/well in 6-well plates with virus-containing medium and centrifuged at 1,750 rpm (517 g) for 25 min (“spinfection”). After 8 h, medium was replaced with fresh growth medium. Viral titers were adjusted to achieve a multiplicity of infection (MOI) of ∼0.1-0.25 to minimize multiple integrations per cell (Zhang et al. 2024; Pillai et al. 2023). Transduced cells were implanted subcutaneously into NSG mice for in vivo treatment studies.

Barcode-based clonal analysis

Barcode reads were processed using custom Python and R pipelines (Goyal et al. 2023). Barcodes were filtered by read quality, expected sequence length (40 bp), and GFP marker presence. Erroneous barcode variants were collapsed using Starcode (https://github.com/gui11aume/starcode) with a Levenshtein distance of 8.

Each barcode-cell combination was assigned based on unique molecular identifiers (UMIs). Analyses were repeated with UMI cutoffs of 2 and 3 to test robustness. Barcoded cells (273 and 146 cells at UMI ≥ 2; 186 and 119 cells at UMI ≥ 3 in replicate 1 and 2, respectively) were clustered by shared barcode identity. Clone size normalization accounted for tumor volume differences between treatment groups:

clone size (DT normalized) = measured clone size × (V_DT / V_DT+D)

Differential expression analyses between DT vs DT+D were performed using Seurat’s FindMarkers (Wilcoxon rank-sum test; logFC > 0.1, Bonferroni-adjusted p). Overlaps with Dasatinib target gene sets (Shi et al. 2012; Montenegro et al. 2020; Copland et al. 2006; Choi et al. 2020; Wei et al. 2023) were computed.

Single-cell RNA sequencing

Single-cell suspensions were processed using 10x Genomics Chromium 3′ v3.1 kits. Reads were aligned to the GRCh38.93 human genome using STARvia CellRanger (v3.1.0). Cells expressing < 200 genes or > 10-20% mitochondrial reads were excluded.

Data integration, normalization, clustering, and visualization were performed in Seurat (v4 or v5). Integration used canonical correlation analysis; normalization was performed via SCTransform (v2) and imputation via ALRA (Linderman et al. 2022). Cells were demultiplexed by Cell Hashing and HTODemux (positive quantile = 0.8). Marker identification used Wilcoxon rank-sum tests with Bonferroni correction.

Cluster annotation used canonical melanoma and stromal marker genes visualized in density plots. Differential gene expression between DT and DT+D conditions was calculated for each cluster to identify treatment-modulated transcriptional programs.

Reverse phase protein array (RPPA)

PDX tumor lysates were sent to the MD Anderson RPPA Core Facility for analysis. Proteins were extracted in 1% SDS + 2-mercaptoethanol buffer, serially diluted (five 2-fold steps), and arrayed on nitrocellulose-coated slides (Grace Bio-Labs) using the Quanterix 2470 Arrayer.

Validated antibodies (https://www.mdanderson.org/research/research-resources/core-facilities/functional-proteomics-rppa-core/antibody-information-and-protocols.html) were used for detection, and signal amplification was performed using the Agilent GenPoint system. Slides were scanned (Huron TissueScope) and quantified (Array-Pro Analyzer).

Signal quantification and normalization were performed using RPPA SPACE (Shehwana et al., 2022). Each dilution curve was fitted with a logistic model, and protein concentrations were normalized across antibodies and samples by median-centering.

Quantification and statistical analysis

Statistical analyses were performed using R (v4.3.2) and Seurat. For in vivo PDX trials, tumor volumes are reported as mean ± SEM with n = 3 mice per group. Statistical significance between groups was assessed by unpaired two-tailed t tests unless otherwise indicated.

For single-cell and barcode analyses, differential gene expression was assessed using Wilcoxon rank-sum tests with Bonferroni-adjusted p values < 0.05 considered significant.

All statistical details, including n, replicates, and test type, are provided in the figure legends and Results.

Acknowledgments

The research was supported by NIH grants R01 CA 182890, U54 CA224070, P01 CA114046, P01 CA025874, P30 CA010815, R01 CA047159, and K01 CA245124-01; Dr. Miriam and Sheldon G. Adelson Medical Research Foundation; the Melanoma Research Foundation; the DoD Melanoma Research Program Idea Award; and the Koo/Adler Fund for Cancer Research. The support for shared resources utilized in this study was provided by the Cancer Center Support Grants (CCSGs) CA010815 and S10 OD023586 to the Wistar Institute. M.A.D. is supported by Dr. Miriam and Sheldon G. Adelson Medical Research Foundation, the AIM at Melanoma Foundation, the NIH/NCI P50CA221703, 1U54CA224070-05, the American Cancer Society and the Melanoma Research Alliance, Cancer Fighters of Houston, the Anne and John Mendelsohn Chair for Cancer Research, and philanthropic contributions to the Melanoma Moon Shots Program of MD Anderson.

Author contributions

Conceptualization, V.W.R. and M.H.; data curation, V.R., M.X., A.K., X.Y., J.D., Q.L., and M.H.; formal analysis, V.W.R., A.K., M.W., H.J., Y.C., K.N., K.A.M., X.Y., Y.D., Q.L., and Y.G.; funding, V.W.R., M.A.D., and M.H.; investigation, V.W.R., M.X., A.K., T.G., G.S.B., D.F., G.M.A, M.P., J.B., V.E., Y.C., E.T., M.E.F., K.L.C., J.V., K.M.A., G.W., Y.N.V.G., M.A.D., and M.H.; methodology, V.W.R., A.K., Q.L., and M.H.; supervision, V.W.R. and M.H.; writing – review & editing, V.W.R. and M.H.

Declaration of interests

M.A.D. has been a consultant to Roche/Genentech, Array, Pfizer, Novartis, BMS, GSK, Sanofi-Aventis, Vaccinex, Apexigen, Eisai, Iovance, Merck, and ABM Therapeutics, and he has been the PI of research grants to MD Anderson by Roche/Genentech, GSK, Sanofi-Aventis, Merck, Myriad, Oncothyreon, Pfizer, ABM Therapeutics, and LEAD Pharma.

Published: January 13, 2026

Supplemental information

Table S1. RPPA ranking for Figures 2 and 3

(A) Median score calculations for Figure 2B. (B) Median score calculations for Figure 3B. (C) Rankings of therapy efficacy based on tumor growth for Figures 2 and 3.

Table S2. Tumor volume statistics for Figures 2, 3, 4, and S2

(A) Statistical methodology used in other tables. (B). Linear mixed-effect results of tumor volume slopes for Figure 2G. (C) Linear mixed-effect results of tumor volume slopes for Figure 3J. (D) Linear mixed-effect results of tumor volume slopes for Figure 4I. (E) Linear mixed-effect results of tumor volume slopes for Figure S2B.

Table S3. Statistical methodology and results for Figure 4

(A) Statistical methodology used in other tables. (B) Treatment effect comparison by day.