Understanding and overcoming innate and acquired MAPK inhibition resistance in anaplastic thyroid cancer Authors: Peter Y.F. Zeng, Jalna Meens, Amir Hossein Karimi, Harrison Pan, Matthew J. Cecchini, Laura Jarycki, Sarah B. Ryan, Alice E. Dawson, Mushfiq H. Shaikh, David A. Palma, Eric Winquist, Lakshman Gunaratnam, Joe S. Mymryk, John W. Barrett, Paul C. Boutros, Laurie Ailles, Anthony C. Nichols Published: 2026 ABSTRACT SummaryAnaplastic thyroid cancer (ATC) is highly lethal. Although patients with the BRAFV600E alteration respond to the type I RAF inhibitor (RAFi) dabrafenib with trametinib, most rapidly develop adaptive or acquired resistance. Here, multi-region whole-genome, high-coverage whole-exome, and single-nuclei RNA sequencing of tumors from ATC patients undergoing type I RAFi and MEKi therapy reveals that reactivation of the mitogen-activated protein kinase (MAPK) pathway, along with immunosuppressive macrophage proliferation, may underlie the development of acquired resistance. Screening of several RAFi reveals that ATC cell lines are exquisitely sensitive to the type II RAFi naporafenib, which inhibits EphA2-mediated MAPK signaling. Further, naporafenib and trametinib overcome both innate and acquired treatment resistance to dabrafenib and trametinib in ATC cell lines and patient-derived xenograft models. Finally, we describe a mechanism of acquired resistance to naporafenib through compensatory mutations in MAST1. Taken together, our work rationalizes the clinical investigation of type II RAFi in the setting of thyroid cancer. Introduction Somatic alterations in the mitogen-activated protein kinase (MAPK)-RAS pathway are found in over 70% of human cancers.1 For example, activating mutations in the BRAF gene exemplified by the prototypical BRAFV600E alteration are observed in more than 70% of melanomas, 30%–70% of thyroid cancers, and 10% of colorectal cancers. BRAFV600E is a clinically actionable alteration as the RAF inhibitor (RAFi) and MEK inhibitor (MEKi) combination, dabrafenib and trametinib, is FDA approved tissue agnostically.2 Nevertheless, response rates differ greatly between individual cancer types, suggestive of tissue-specific mechanisms of resistance. Thyroid cancers serve as an intriguing example of such differential response to MAPK inhibition.3 The thyroid is the most common tissue of origin in endocrine malignancy.4 Differentiated thyroid cancers (DTCs) such as papillary thyroid cancer (PTC) and follicular thyroid cancer make up over 90% of all thyroid cancers and are indolent. In contrast, anaplastic thyroid cancer (ATC) accounts for just 1.3% of thyroid cancers but is one of the most lethal human malignancies,5 with a median survival of 3–5 months. Indeed, some patients succumb to the disease within days or weeks of diagnosis. Although TP53 is the most frequently mutated gene in ATC, 40%–80% of ATC tumors have alterations in at least one of the members of the RAS-MAPK or PI3K-AKT pathways, such as BRAF, NRAS, and PIK3CA.6,7,8,9 The phase 2 single-arm, open-label Rare Oncology Agnostic Research (ROAR) trial that treated BRAFV600E patients with dabrafenib (type I RAFi) and trametinib (MEKi) reported an impressive response rate of 56% and a median overall survival of 86 weeks—a dramatic improvement from historical survival times.10 These results underscore the importance of inhibition of the MAPK pathway as a therapeutic strategy in ATC. Intriguingly, these response rates are also in stark contrast to the ∼30% response rate seen in the less aggressive DTCs.11 Unfortunately, almost all ATC patients rapidly developed resistance to type I RAFi/MEKi and have limited further approved treatment options.5 Although compensatory alterations in oncogenic RAS,12 HER3 activation,13 and MET amplifications14 have been linked with resistance to type I RAFi, our understanding of the mechanisms of treatment resistance to RAFi and MEKi in ATC remains incomplete. Furthermore, patients with BRAF wild-type tumors have very few treatment options, underscoring the urgent need to mechanistically understand innate and acquired resistance mechanisms to develop more efficacious treatment alternatives. Recently, preclinical and clinical studies of type II RAFi have found that they inhibit RAF dimers more effectively and avoid paradoxical activation,2,15 making them suitable for a broader range of RAF- and RAS-driven cancers. Nevertheless, the activity of these compounds in ATC has not been explored. To address this gap, we sought to understand the mechanisms of innate and acquired resistance to RAFi in ATC. We began with deep molecular phenotyping of genomic, transcriptomic, and cellular evolution of an ATC patient (ATCWGS42) who developed acquired resistance to type I RAFi and MEKi using multi-region whole-genome sequencing (WGS), high-coverage whole-exome sequencing (WES), and single-nucleus RNA sequencing (snRNA-seq). Subclonal reconstruction identified MAPK and PI3K-AKT reactivation from an early-branching low-frequency subclone present in the baseline tumor. We further showed that a type II RAFi, naporafenib, in combination with trametinib is more efficacious than other RAFi in preclinical models of ATC. In addition, we demonstrate that naporafenib and trametinib can overcome innate and acquired resistance to MAPK inhibition in patient-derived xenograft (PDX) models of ATC, including one from patient ATCWGS42. Mechanistically, naporafenib inhibits EphA2 S897 hyperphosphorylation enriched in thyroid and colorectal cancer and thus co-represses both MAPK and PI3K-AKT signaling. Finally, we demonstrate that compensatory MAST1 alteration serves as a mechanism of acquired resistance to naporafenib. Taken together, our results reveal the mechanisms of innate and acquired resistance to MAPK inhibition in lethal MAPK-driven cancer and rationalize the clinical investigation of naporafenib in MAPK cancers that are unresponsive to dabrafenib and trametinib. Results Genomic, transcriptomic, and microenvironmental evolution during RAFi and MEKi treatment in ATC patient ATCWGS42 We studied an ATC patient with BRAFV600E-mutant tumor (ATCWGS42, Figure 1A, detailed clinical description can be found in STAR Methods) who was treated with the RAFi and MEKi combination, vemurafenib and cobimetinib. The patient demonstrated complete response to this treatment for 5 months before disease progression, underscoring the rapid occurrence of resistance to type I RAFi and MEKi combination typically seen in patients with ATC. To understand genomic and nongenomic factors that underlie rapid resistance to RAFi and MEKi in ATC, we molecularly dissected the ATCWGS42 tumors pre- and post-RAFi and MEKi treatment (Figure 1B).Figure 1Genomic, transcriptomic, and microenvironmental evolution of ATCWGS42 following RAFi and MEKi treatment(A) ATCWGS42 was a 44-year-old male patient diagnosed with BRAFV600E anaplastic thyroid cancer (ATC). He demonstrated complete response to the RAFi and MEKi combination vemurafenib and cobimetinib. After 6 months, he progressed. Detailed clinical history can be found in STAR Methods. Five distinct tumor regions of the treatment-naive primary tumor and one region of the tumor upon progression were biopsied.(B) Details of molecular data available from patient. Black means present, while white denotes absent.(C) Copy-number alteration landscape across each tumor region.(D) Subclonal reconstruction of the tumor genomic evolution. The length of the line is proportional to the number of SNV changes, except for between normal and clone 1. Cl., clone; N, normal.(E) Cancer cell clonal fraction of each subclone, identified in subclonal reconstruction, in each tumor region.(F) Uniform manifold approximation and projection (UMAP) visualization of cell states identified using single-nucleus RNA sequencing (snRNA-seq) and annotated with finely annotated (level 4) cell type labels. F, fibroblast population; M, myeloid population; L, lymphocyte population; T, tumor; D, endothelial population.(G) Proportion of cell states within each tumor region. F, fibroblast population; M, myeloid population; L, lymphocyte population; D, endothelial population; T, tumor; EC, endothelial cell; CAF, cancer-associated fibroblast.(H) Principal-component analysis (PCA) visualization of ATC tumor cell states, sample, and PROGENy-inferred MAPK and JAK-STAT pathway output activity scores.(I) PCA visualization of ATC tumor cell transcript abundance of PAX8, EGFR, FOXP1, NOTCH3, MYC, and WNT5B. Genomic, transcriptomic, and microenvironmental evolution of ATCWGS42 following RAFi and MEKi treatment (A) ATCWGS42 was a 44-year-old male patient diagnosed with BRAFV600E anaplastic thyroid cancer (ATC). He demonstrated complete response to the RAFi and MEKi combination vemurafenib and cobimetinib. After 6 months, he progressed. Detailed clinical history can be found in STAR Methods. Five distinct tumor regions of the treatment-naive primary tumor and one region of the tumor upon progression were biopsied. (B) Details of molecular data available from patient. Black means present, while white denotes absent. (C) Copy-number alteration landscape across each tumor region. (D) Subclonal reconstruction of the tumor genomic evolution. The length of the line is proportional to the number of SNV changes, except for between normal and clone 1. Cl., clone; N, normal. (E) Cancer cell clonal fraction of each subclone, identified in subclonal reconstruction, in each tumor region. (F) Uniform manifold approximation and projection (UMAP) visualization of cell states identified using single-nucleus RNA sequencing (snRNA-seq) and annotated with finely annotated (level 4) cell type labels. F, fibroblast population; M, myeloid population; L, lymphocyte population; T, tumor; D, endothelial population. (G) Proportion of cell states within each tumor region. F, fibroblast population; M, myeloid population; L, lymphocyte population; D, endothelial population; T, tumor; EC, endothelial cell; CAF, cancer-associated fibroblast. (H) Principal-component analysis (PCA) visualization of ATC tumor cell states, sample, and PROGENy-inferred MAPK and JAK-STAT pathway output activity scores. (I) PCA visualization of ATC tumor cell transcript abundance of PAX8, EGFR, FOXP1, NOTCH3, MYC, and WNT5B. We performed WGS and high-coverage WES of 5 spatially distinct regions of the primary pre-treatment tumor (ATCWGS42-A to ATCWGS42-E) and one post-treatment tumor (ATCWGS42-R). Subclonal reconstruction identified a truncal alteration BRAFV600E, a TERT C228T promoter alteration, and multiple chromosomal copy-number alterations, including chromosome 6p loss and chromosomes 8, 9, 11, and 20 gain (Figures 1C and 1D). We identified a subclone (clone 2) enriched in ATCWGS42-R that makes up more than 70% of the tumor profiled at RAFi/MEKi resistance, suggestive of clonal sweep (Figure 1E). Clone 2 exhibits somatic alterations in NOTCH3 and DUOXA1 and was inferred to have branched early from clone 1 (Figure 1D). Its early divergence and presence in the primary pre-treatment tumor were further supported by an identical NOTCH3 alteration in one read within the high-coverage WES of region ATCWGS42-A. These alterations have not been linked to acquired resistance to dabrafenib and trametinib previously. We did not identify compensatory mutations in other members of the RAS-MAPK pathway, as past studies in melanoma and thyroid cancers have reported.12,16,17,18,19,20 Nevertheless, the high clonal proportion of clone 2 in the post-treatment ATCWGS42-R tumor suggested that genomic and/or nongenomic features in this clone may offer evolutionary advantages in resistance to MAPK inhibition. To characterize the transcriptional and tumor microenvironment changes following RAFi and MEKi therapy, we performed snRNA-seq of two tumor components, the pre-treatment (ATCWGS42-E) and the post-treatment tumor (ATCWGS42-R). After stringent quality control to remove low-quality cells and doublets, we clustered the remaining 6,367 high-quality cells using the Leiden algorithm21 to reveal 16 cell clusters. Manual hierarchical iterative annotation revealed cell types, including two ATC tumor populations, CD4 T cells, CD8 T cells, and tumor-associated macrophages (TAMs) (Figure 1F). The ATC tumor cell annotation is supported by inferred copy-number states (Figures S1A and S1B). In the post-treatment ATCWGS42-R tumor, the vast majority of the macrophage population was annotated as Y1_macrophages. The Y1_macrophage populations express a high abundance of CD163 and thus are anti-inflammatory CD163 TAMs (Figures 1G and S1E). This is concordant with previous studies of melanoma patients who develop resistance to dabrafenib,22,23 suggesting that strategies targeting these populations may relieve the immunosuppressive tumor microenvironment (TME) and prolong treatment response. We next characterized transcriptional changes in the tumor cell populations underlying acquired resistance. ATC cells in the pre-treatment and post-treatment tumors were transcriptionally distinct (Figure 1H). Furthermore, the post-treatment resistance tumor showed increased inferred activity of both MAPK and JAK-STAT pathways through PROGENy,24 suggesting MAPK pathway reactivation. Intriguingly, this MAPK activation is accompanied by loss of expression in NOTCH pathway genes NOTCH3, FOXP1, and WNT5B and increased EGFR, MYC, and PAX8 expression (Figure 1I). The decreased abundance of NOTCH3 in post-treatment tumor cells suggests that the NOTCH3 alteration in clone 2 identified using WGS and WES may be a loss-of-function alteration. Furthermore, feedback activation of EGFR is known to cause acquired resistance to RAFi.25 An increase in the MAPK output score in the post-treatment sample further supported this notion. High expression of the thyroid differentiation marker and transcription factor PAX8 (Figures 1I and S1F) as well as the thyroid differentiation score (Figure S1G) served as potential evidence of RAFi/MEKi-induced redifferentiation in ATC.26,27 Taken together, our detailed genomic and single-cell transcriptomic analysis highlighted the genomic and tumor microenvironmental changes underlying acquired resistance to RAFi/MEKi and strongly underscored reactivation of the MAPK pathway in the development of resistance to RAFi and MEKi. Naporafenib can overcome intrinsic treatment resistance in thyroid cancer Given that the MAPK pathway appeared reactivated following acquired resistance to RAFi and MEKi, vertical inhibition of the pathway may both lead to a more durable and efficacious response in treatment-naive and resistant tumors.25,28,29,30,31,32,33,34 Type II RAFi inhibit ARAF, BRAF, and/or CRAF, making them more effective against RAF dimers and broadly applicable in tumors driven by RAS mutations or RAF dimerization-led MAPK-activated cancers. They are also efficacious in preclinical models of RAF- and RAS-driven melanoma and a subset of patients in other cancer types.35,36,37,38,39,40,41,42,43,44 However, their efficacy in thyroid cancers remains largely unexplored. We screened a panel of 9 thyroid cancer cell lines (8 ATC and 1 PTC cell line K1) with dabrafenib or one of four type II RAFi (belvarafenib, lifirafenib, naporafenib, and MLN2480). No thyroid cell lines were sensitive (defined as IC50 < 10 μM) to dabrafenib (Figure 2A), as supported by previous work.13 The type II RAFi belvarafenib, MLN2480, and lifirafenib inhibited growth in a minority of the screened thyroid cell lines. As a positive control, the BRAFV600E-mutant melanoma cell line A375 was sensitive to all tested RAFi. Strikingly, single-agent naporafenib genotype agnostically inhibited the growth of 8/9 thyroid cancer cell lines, including BRAFV600E-mutant (8505C, BHT101, K1, and SW1736), NRAS-mutant (ASH3 and KMH2), and BRAF NRAS-wild-type cell lines (THJ29T and UHTh7).Figure 2Naporafenib overcomes thyroid cancer resistance to RAFi by repressing EphA2 phosphorylation(A) In a panel of 9 thyroid cancer cell lines, the melanoma cell line A375, and its isogenic A375-NRAS Q61K cell line, naporafenib was found to be efficacious. IC50, half-maximal inhibitory concentration; MUT, mutant; WT, wild type.(B) The addition of the MEK inhibitor trametinib at nanomolar concentration further sensitizes thyroid cancer cell lines to naporafenib, including the previously resistant cell line THJ16T.(C) Volcano plot of reverse-phase protein array of cell lines SW1736 and ASH3 treated with either vehicle or naporafenib for 24 h linear regression model. Estimate and Benjamini-Hochberg adjusted p value from linear regression model, with the following terms: treatment (vehicle or naporafenib), cell line (SW1736 or ASH3), and their interaction term. Proteins/phosphoproteins in yellow are repressed by naporafenib while proteins in dark blue are induced by naporafenib, respectively. Horizontal red lines indicate an adjusted p value cutoff of 0.05, while vertical red lines indicate an effect size cutoff of +0.5 or −0.5.(D) Dotplot of protein/phosphoproteins found to pass both the adjusted p value and effect size cutoff. Shade indicates the adjusted p values, while the color of the dot and text inside the circle indicates the estimate.(E) Abundance of EphA2 pS897, Aurora-A, and CRAF pS338 in SW1736 and ASH3.(F) EphA pS897 levels in The Cancer Genome Atlas (TCGA) reverse-phase protein array of colorectal cancer (TCGA-COAD), skin cutaneous melanoma (TCGA-SKCM), and thyroid cancer (TCGA-THCA) cohorts. p values from Benjamini-Hochberg adjusted Wilcoxon’s rank-sum test. ∗∗∗FDR < 0.001, ∗∗∗∗FDR < 0.0001.(G) Scatterplot of EphA2 mRNA abundance and dabrafenib −log10 IC50 in bowel, skin, and thyroid cancer cell lines in the Cancer Cell Line Encyclopedia with BRAFV600E. p value and R coefficient from Pearson’s correlation. Naporafenib overcomes thyroid cancer resistance to RAFi by repressing EphA2 phosphorylation (A) In a panel of 9 thyroid cancer cell lines, the melanoma cell line A375, and its isogenic A375-NRAS Q61K cell line, naporafenib was found to be efficacious. IC50, half-maximal inhibitory concentration; MUT, mutant; WT, wild type. (B) The addition of the MEK inhibitor trametinib at nanomolar concentration further sensitizes thyroid cancer cell lines to naporafenib, including the previously resistant cell line THJ16T. (C) Volcano plot of reverse-phase protein array of cell lines SW1736 and ASH3 treated with either vehicle or naporafenib for 24 h linear regression model. Estimate and Benjamini-Hochberg adjusted p value from linear regression model, with the following terms: treatment (vehicle or naporafenib), cell line (SW1736 or ASH3), and their interaction term. Proteins/phosphoproteins in yellow are repressed by naporafenib while proteins in dark blue are induced by naporafenib, respectively. Horizontal red lines indicate an adjusted p value cutoff of 0.05, while vertical red lines indicate an effect size cutoff of +0.5 or −0.5. (D) Dotplot of protein/phosphoproteins found to pass both the adjusted p value and effect size cutoff. Shade indicates the adjusted p values, while the color of the dot and text inside the circle indicates the estimate. (E) Abundance of EphA2 pS897, Aurora-A, and CRAF pS338 in SW1736 and ASH3. (F) EphA pS897 levels in The Cancer Genome Atlas (TCGA) reverse-phase protein array of colorectal cancer (TCGA-COAD), skin cutaneous melanoma (TCGA-SKCM), and thyroid cancer (TCGA-THCA) cohorts. p values from Benjamini-Hochberg adjusted Wilcoxon’s rank-sum test. ∗∗∗FDR < 0.001, ∗∗∗∗FDR < 0.0001. (G) Scatterplot of EphA2 mRNA abundance and dabrafenib −log10 IC50 in bowel, skin, and thyroid cancer cell lines in the Cancer Cell Line Encyclopedia with BRAFV600E. p value and R coefficient from Pearson’s correlation. Naporafenib is an ARAF-sparing type II RAFi that potently inhibits monomeric or dimeric BRAF and CRAF.40,41 Notably, naporafenib in combination with MEKi trametinib demonstrated a disease control rate of 73.3% in heavily pretreated NRAS melanoma patients.45 We next tested whether the addition of trametinib further sensitized thyroid cell lines to naporafenib. Three different doses of trametinib (0, 1, or 10 nM) were tested in combination with a dose range of 32 nM to 100 μM of naporafenib. Importantly, the addition of trametinib at nanomolar concentrations further sensitized thyroid cell lines to naporafenib (Figure 2B), including the previously naporafenib-resistant cell line THJ16T, a BRAF-fusion ATC cell line known for its drug resistance. As no BRAFV600E ATC cell line is exquisitely sensitive to dabrafenib/trametinib (Figures 2A and 2B), we utilized the dabrafenib/trametinib-sensitive A375 melanoma cell line to test the ability of naporafenib in combating dabrafenib/trametinib resistance in vitro. We used both the parental A375 cell line and an isogenic CRISPR-Cas9-edited line possessing the clinically relevant NRAS Q61K point mutation identified in melanoma patients with acquired resistance to dabrafenib/trametinib, which are known to be more resistant to dabrafenib/trametinib. Naporafenib was equally effective in inhibiting the growth of both the parental and NRAS-mutant A375 lines (Figure 2A). These data established that combined naporafenib and trametinib treatment could inhibit the growth of cancer cell lines with acquired dabrafenib/trametinib resistance in vitro. Naporafenib blunts EphA2-mediated intrinsic RAFi resistance To mechanistically understand the sensitivity of thyroid cancer cell lines to naporafenib, we performed targeted phosphoproteomics using reverse-phase protein arrays (RPPAs) of BRAFV600E-mutant SW1736 and NRAS-mutant ASH3 cell lines treated with either vehicle or naporafenib for 24 h. Naporafenib induced increased phosphorylation of CRAF and decreased abundance of Aurora A. Strikingly, naporafenib treatment led to a 7-fold decrease in phosphorylation levels of EphA2 S897 (linear regression, Figures 2C and 2E, see STAR Methods). We confirmed these results using western blot in the cell lines 8505C, BHT101, and SW1736, where 1 h treatment with naporafenib, as well as dabrafenib and belvarafenib, led to reduction of EphA2 S897 phosphorylation (Figure S2A). EphA2 is a member of the receptor tyrosine kinase (RTK) family erythropoietin-producing hepatocellular (EPH) receptors and impinges on the RAS-MAPK, JAK-STAT, and RAS-PI3K-AKT pathways.46,47,48,49 Importantly, EphA2 is well-established to be a critical mediator of resistance to type I RAFi.50 In The Cancer Genome Atlas (TCGA), thyroid cancers exhibit higher phosphorylated EphA2 S897 compared to both colorectal cancer and melanoma (false discovery rate [FDR] < 0.001, Wilcoxon’s test, Figure 2F), an effect that is consistent even when restricted just to BRAFV600E mutant tumors (Figure S2B). In the Cancer Cell Line Encyclopedia (CCLE), EphA2 mRNA abundance is also higher in cell lines from bowel or thyroid, cancers more resistant to MAPK inhibition, compared to those from the skin (Figure S2C). Furthermore, EphA2 mRNA abundance is strongly correlated with dabrafenib IC50 values from cell lines in the CCLE (p < 0.001, Pearson’s correlation, Figure 2G). Collectively, these data established EphA2 as a strong candidate as an organ-site-specific mediator of RAFi resistance, with the highest abundance in thyroid cancers. Naporafenib and trametinib overcome resistance to RAFi and MEKi in an ATC PDX model of ATCWGS42 Next, we assessed whether naporafenib is efficacious in PDX models of ATC. We generated a treatment-naive PDX model (Figure 3A) from the pre-treatment tumor from the same patient with the BRAFV600E alteration that we studied in detail (ATCWGS42, Figure 1). We treated the PDX model with DMSO, dabrafenib, trametinib, naporafenib, dabrafenib + trametinib, or naporafenib + trametinib. Whereas dabrafenib alone failed to control tumor growth, naporafenib monotherapy successfully restrained tumor growth until the endpoint. Furthermore, treatment with the naporafenib and trametinib combination led to a complete response (100% maximal tumor control) in 3/5 mice (Figure 3B) and was superior at tumor control compared to the other conditions (p < 0.001 vs. all other groups, linear mixed-effects model, Figures 3C and 3D).Figure 3Naporafenib and trametinib induce tumor regression in an ATCWGS42-derived xenograft model by co-inhibiting MAPK and PI3K-AKT signaling(A) Schematic of experiment. One treatment-naive region of tumor from ATCWGS42 was used to generate a patient-derived xenograft (PDX) model, which was passaged to generate the indicated treatment cohorts. When tumors reached 100 mm3, mice were randomized into one of six groups: vehicle, dabrafenib, naporafenib, trametinib, dabrafenib and trametinib, and naporafenib and trametinib and treated daily for 5 days per week. After day 23, all mice were sacrificed. Tumors were collected for molecular correlative analysis using bulk RNA-seq and histology.(B) Maximal percentage tumor size shrinkage, as defined using RECIST version 1.1.(C and D) Tumor size plot of each arm grouped (C) and spaghetti plot (D) showing individual mice. p value from the linear mixed-effect model. Error bars indicate standard deviation.(E and F) Volcano plot of transcripts induced and repressed by dabrafenib (E) or naporafenib (F) treatment. Horizontal red lines indicate an adjusted p value cutoff of 0.05, while vertical red lines indicate a log2 fold changecut off of +1 or −1. p values from DESeq2.(G) Pathway analysis of transcripts induced and repressed by naporafenib, identified using GProfiler.(H) Scaled mean pathway output activity of MAPK, PI3K-AKT, and WNT inferred using PROGENy, normalized between groups. Naporafenib and trametinib induce tumor regression in an ATCWGS42-derived xenograft model by co-inhibiting MAPK and PI3K-AKT signaling (A) Schematic of experiment. One treatment-naive region of tumor from ATCWGS42 was used to generate a patient-derived xenograft (PDX) model, which was passaged to generate the indicated treatment cohorts. When tumors reached 100 mm3, mice were randomized into one of six groups: vehicle, dabrafenib, naporafenib, trametinib, dabrafenib and trametinib, and naporafenib and trametinib and treated daily for 5 days per week. After day 23, all mice were sacrificed. Tumors were collected for molecular correlative analysis using bulk RNA-seq and histology. (B) Maximal percentage tumor size shrinkage, as defined using RECIST version 1.1. (C and D) Tumor size plot of each arm grouped (C) and spaghetti plot (D) showing individual mice. p value from the linear mixed-effect model. Error bars indicate standard deviation. (E and F) Volcano plot of transcripts induced and repressed by dabrafenib (E) or naporafenib (F) treatment. Horizontal red lines indicate an adjusted p value cutoff of 0.05, while vertical red lines indicate a log2 fold changecut off of +1 or −1. p values from DESeq2. (G) Pathway analysis of transcripts induced and repressed by naporafenib, identified using GProfiler. (H) Scaled mean pathway output activity of MAPK, PI3K-AKT, and WNT inferred using PROGENy, normalized between groups. To untangle in vivo effects of naporafenib compared to dabrafenib and trametinib, we performed bulk RNA sequencing (RNA-seq) of three tumors from each group at the endpoint. After the removal of mouse reads (see STAR Methods), we utilized a multi-factor design to account for the perturbational effects of RAFi (vehicle, dabrafenib, or naporafenib) and MEKi (vehicle or trametinib) and their interactions. Both dabrafenib (Figure 3E) and naporafenib (Figure 3F) repressed the abundance of a similar subset of transcripts such as DUSP6, a negative regulator of MAPK signaling, suggesting that both dabrafenib and naporafenib repressed MAPK signaling.51,52 Furthermore, naporafenib repressed the expression of transcripts critical to epithelial-mesenchymal transition, such as VIM53,54 (Figure 3F). Pairwise comparison between mice treated with naporafenib and trametinib versus dabrafenib and trametinib also identified VIM to have higher abundance in the dabrafenib and trametinib group (Figure S2D). Pathway analysis using GProfiler255,56 (Figure 3G) suggested that naporafenib induced the expression of genes related to epithelial function, while downregulating ERK and PI3K-AKT signaling. Finally, pathway activity inference (Figure 3H) using PROGENy24 demonstrated that combined treatment of naporafenib and trametinib uniquely repressed pathway activity of the MAPK, PI3K-AKT, and WNT pathways. Immunofluorescence of phospho-MEK and phospho-ERK supports the inhibition of the MAPK pathway by naporafenib and trametinib (Figure S2E). These experiments, in sum, suggest that naporafenib can overcome innate resistance to MAPK by inhibiting both MAPK and PI3K-AKT signaling, leading to more efficacious tumor control. Naporafenib and trametinib overcome acquired resistance to dabrafenib and trametinib We then assessed the efficacy of naporafenib and trametinib combination therapy in the setting of acquired resistance to dabrafenib and trametinib. We expanded the PDX tumors that were treated with dabrafenib and trametinib (as in Figure 3, dabrafenib and trametinib group). We implanted ten mice with these tumors and, upon reaching 100 mm3, began treatment with dabrafenib and trametinib until resistance (Figure 4A). Tumor sizes did not shrink but remained steady until day 21 (Figures 4B and 4C). On day 33, when the tumors had acquired resistance to dabrafenib and trametinib (mean tumor size reached 250 mm3), we randomized the mice to either continue treatment with dabrafenib and trametinib or switch to naporafenib and trametinib. Strikingly, the naporafenib and trametinib combination treatment induced significant tumor regression in dabrafenib- and trametinib-resistant tumors (Figures 4B and 4C). When all dabrafenib- and trametinib-treated mice reached the endpoint by day 64, three mice treated with naporafenib and trametinib were also collected for correlative biomarker analysis. The two mice that remained on naporafenib and trametinib maintained tumor control. From days 95 to 106, the mice went on a forced drug holiday for 11 days (due to institutional policy; see Figure 4A). Post-drug holiday and upon resumption of naporafenib and trametinib treatment, tumor sizes shrunk again before finally reaching the humane endpoint.Figure 4Naporafenib and trametinib overcome acquired resistance to dabrafenib and trametinib(A) Schematic of experiment. The ATCWGS42 PDX model was used to generate 10 mouse tumors. When tumors reached 100 mm3, mice were treated with dabrafenib and trametinib until tumor size reached 250 mm3. Mice were then randomized to either continue dabrafenib and trametinib or switch to naporafenib and trametinib treatment. On day 64 when most mice in the dabrafenib and trametinib group were close to endpoint, all five mice in the dabrafenib and trametinib group and three of five from the naporafenib and trametinib group were sacrificed for correlative studies. Two out of five in the naporafenib and trametinib group continued treatment until the development of resistance or humane endpoint. There was a drug holiday from day 95 to 106 due to lack of access to the animal facility under institutional mandate.(B) Tumor size plot of each arm grouped and spaghetti plot showing each individual mouse. p value from the linear mixed-effect model comparing naporafenib and trametinib to dabrafenib and trametinib. Error bars indicate standard deviation.(C) Schematic of experiment. The ATCWGS42 PDX model was used to generate 30 mouse tumors. When tumors reached 100 mm3, mice were treated with vehicle (n = 5) or dabrafenib and trametinib (n = 25). When the mean tumor size of the group treated with dabrafenib and trametinib reached 250 mm3, mice were randomized to either continue dabrafenib and trametinib (10 out of 25) or switch to naporafenib and trametinib treatment (15 out of 25). All mice in the dabrafenib and trametinib group were treated until endpoint, and 5 out of 15 mice treated with naporafenib and trametinib were sacrificed at that time. The remaining 10 out of 15 mice in the naporafenib and trametinib group continued treatment. On day 133, all the dabrafenib and trametinib group were sacrificed, and 2 of the remaining 6 mice treated with naporafenib and trametinib were randomly selected for vehicle treatment to assess whether removal of treatment results in tumor growth rebound.(D) Tumor size plot of each arm grouped and spaghetti plot showing each individual mouse for ATCWGS42. p value from the linear mixed-effect model comparing naporafenib and trametinib to dabrafenib and trametinib. Error bars indicate standard deviation.(E) Schematic of experiment. A treatment-naive region of tumor from patient ATC125 was used to establish a PDX model, which was passaged to generate 30 tumors. These were treated in the same way as the ATCWGS42 PDX model above.(F) Tumor size plot of each arm grouped and spaghetti plot showing each individual mouse for ATC125. p value from the linear mixed-effect model comparing naporafenib and trametinib to dabrafenib and trametinib. Error bars indicate standard deviation. Naporafenib and trametinib overcome acquired resistance to dabrafenib and trametinib (A) Schematic of experiment. The ATCWGS42 PDX model was used to generate 10 mouse tumors. When tumors reached 100 mm3, mice were treated with dabrafenib and trametinib until tumor size reached 250 mm3. Mice were then randomized to either continue dabrafenib and trametinib or switch to naporafenib and trametinib treatment. On day 64 when most mice in the dabrafenib and trametinib group were close to endpoint, all five mice in the dabrafenib and trametinib group and three of five from the naporafenib and trametinib group were sacrificed for correlative studies. Two out of five in the naporafenib and trametinib group continued treatment until the development of resistance or humane endpoint. There was a drug holiday from day 95 to 106 due to lack of access to the animal facility under institutional mandate. (B) Tumor size plot of each arm grouped and spaghetti plot showing each individual mouse. p value from the linear mixed-effect model comparing naporafenib and trametinib to dabrafenib and trametinib. Error bars indicate standard deviation. (C) Schematic of experiment. The ATCWGS42 PDX model was used to generate 30 mouse tumors. When tumors reached 100 mm3, mice were treated with vehicle (n = 5) or dabrafenib and trametinib (n = 25). When the mean tumor size of the group treated with dabrafenib and trametinib reached 250 mm3, mice were randomized to either continue dabrafenib and trametinib (10 out of 25) or switch to naporafenib and trametinib treatment (15 out of 25). All mice in the dabrafenib and trametinib group were treated until endpoint, and 5 out of 15 mice treated with naporafenib and trametinib were sacrificed at that time. The remaining 10 out of 15 mice in the naporafenib and trametinib group continued treatment. On day 133, all the dabrafenib and trametinib group were sacrificed, and 2 of the remaining 6 mice treated with naporafenib and trametinib were randomly selected for vehicle treatment to assess whether removal of treatment results in tumor growth rebound. (D) Tumor size plot of each arm grouped and spaghetti plot showing each individual mouse for ATCWGS42. p value from the linear mixed-effect model comparing naporafenib and trametinib to dabrafenib and trametinib. Error bars indicate standard deviation. (E) Schematic of experiment. A treatment-naive region of tumor from patient ATC125 was used to establish a PDX model, which was passaged to generate 30 tumors. These were treated in the same way as the ATCWGS42 PDX model above. (F) Tumor size plot of each arm grouped and spaghetti plot showing each individual mouse for ATC125. p value from the linear mixed-effect model comparing naporafenib and trametinib to dabrafenib and trametinib. Error bars indicate standard deviation. We further confirmed these findings by repeating the acquired resistance experiment starting from the treatment-naive tumor (Figures 4C and 4D). Mice were treated with either vehicle or dabrafenib and trametinib. On day 75, when the mean tumor size reached 250 mm3, mice treated with dabrafenib and trametinib were randomized to either continue dabrafenib and trametinib or switched to naporafenib and trametinib. Strikingly, combined treatment using naporafenib and trametinib led to tumor size regression (p < 0.001, linear mixed-effects model, Figure 4D). By day 133, the average tumor size of mice in the dabrafenib and trametinib group was 1,120 mm3, and all mice were sacrificed, while the average tumor size in the naporafenib and trametinib group was 162 mm3 (Figures 4C and 4D). We tested whether tumor growth relied on continuous naporafenib and trametinib administration by switching two of the remaining six mice treated with naporafenib and trametinib to vehicle treatment (gray group, Figures 4C and 4D). Indeed, withdrawal of naporafenib and trametinib led to rapid tumor growth in two mice that both reached endpoint by day 162. Meanwhile, tumor inhibition for three mice out of four mice continuously treated with naporafenib and trametinib continued until day 186, demonstrating the durable and robust response of ATC to naporafenib and trametinib treatment. Next, using a PDX model from the BRAFV600E ATC tumor from a separate ATC patient (ATC125, Figure 4E), we found that the naporafenib and trametinib combination was also able to control tumor growth more effectively than dabrafenib and trametinib (ATC125, p < 0.001, linear mixed-effects model, Figure 4F). Finally, using an immunocompetent syngeneic cell line model of ATC57 of acquired resistance to dabrafenib and trametinib, we again found that naporafenib and trametinib were able to better inhibit tumor growth compared to dabrafenib and trametinib (Figure S3A). Taken together, the naporafenib and trametinib combination is efficacious in PDX and syngeneic models of acquired resistance to dabrafenib and trametinib in ATC. Acquired resistance to naporafenib and trametinib through compensatory alterations in MAST1 Whereas the genomic mechanism of resistance to type I RAFi of BRAFV600E is more well characterized, the mechanisms of resistance toward type II RAFi remain largely unexplored. Compensatory alterations along MAPK pathway members, for example, in ARAF, confer resistance to the type II RAFi such as belvarafenib and naporafenib.58 Immunofluorescence staining of PDX tissues for p-S6 and p-EphA2 levels found that, as expected, phosphorylation of both proteins increases as tumors develop resistance toward dabrafenib and trametinib (Figure S3B), demonstrating reactivation of MAPK and mTORC1 proliferative signaling. Although resistance to naporafenib and trametinib is also associated with increased p-S6 signaling, the increase in p-EphA2 is more moderate, suggesting that additional alternative mechanisms of resistance may exist for naporafenib. To understand the differential mechanism of resistance, we performed high-coverage WES in ATCWGS42 PDX tumors at baseline (untreated), at the time of resistance to dabrafenib and trametinib, while responding to naporafenib and trametinib, and finally after acquired resistance to naporafenib following 128 days of naporafenib treatment. The only protein-coding alteration that was significantly enriched in the naporafenib- and trametinib-resistant PDX tumor was a MAST1 Ala733Val substitution. This alteration was found in only region E of the original ATCWGS42 tumor at very low frequency (1/493 reads) and slightly enriched in PDX tumors resistant to dabrafenib and trametinib or the tumor responding to naporafenib and trametinib. Strikingly, the variant allele frequency (VAF) for this alteration expanded to 47% of all reads in the tumor that is resistant to naporafenib and trametinib (Figure 5A). Immunofluorescence also found that MAST1 abundance appears to be increased in samples resistant to naporafenib and trametinib but not other treatment conditions (Figure S3B).Figure 5Compensatory alterations in MAST1 may lead to re-activated MAPK signaling(A) Variant allele frequency of MAST1 Ala733Val alteration in primary tumor (ATCWGS42-A to ATCWGS42-E), post vemurafenib and cobimetinib resistant tumor (ATCWGS42-R), patient-derived xenograft (PDX) treated with vehicle, PDX model that developed acquired resistance to dabrafenib and trametinib, PDX model that still respond to naporafenib and trametinib, and PDX model that demonstrate acquired resistance to naporafenib and trametinib. Percentage is the total number of reads, while the number in the bracket indicates the number of reference and alternative reads.(B) IC50 of ATC cell line BHT101 transfected with either wild-type or mutant MAST1 Ala733Val to naporafenib. Each curve represents three biological replicates.(C) Proposed model of innate and acquired resistance to type I and type II RAF inhibition in thyroid cancer. Compensatory alterations in MAST1 may lead to re-activated MAPK signaling (A) Variant allele frequency of MAST1 Ala733Val alteration in primary tumor (ATCWGS42-A to ATCWGS42-E), post vemurafenib and cobimetinib resistant tumor (ATCWGS42-R), patient-derived xenograft (PDX) treated with vehicle, PDX model that developed acquired resistance to dabrafenib and trametinib, PDX model that still respond to naporafenib and trametinib, and PDX model that demonstrate acquired resistance to naporafenib and trametinib. Percentage is the total number of reads, while the number in the bracket indicates the number of reference and alternative reads. (B) IC50 of ATC cell line BHT101 transfected with either wild-type or mutant MAST1 Ala733Val to naporafenib. Each curve represents three biological replicates. (C) Proposed model of innate and acquired resistance to type I and type II RAF inhibition in thyroid cancer. MAST1 is a member of the CRAF-MEK1 complex and can phosphorylate MEK1 independently of CRAF, leading to CRAF-independent activation of the MAPK pathway.59 As naporafenib is hypothesized to bind BRAF and CRAF with similar affinity, compensatory alterations in MAST1 may thus drive reactivation of MAPK signaling and tumor proliferation. Indeed, expression of MAST1 Ala733Val in the ATC cell line BHT101 induced resistance to naporafenib (Figure 5B) compared to wild-type MAST1, regardless of the presence or absence of trametinib. Similarly, expression of MAST1 Ala733Val also led to resistance to naporafenib and trametinib in the ATC cell line 8505C (Figure 5B). Taken together, our results suggest that we have identified a mechanism of resistance to type II RAFi that reactivates MAPK signaling2,15 (Figure 5C). Discussion In the current study, we utilize translational and functional multi-omics, paired with in vitro and in vivo functional studies, to understand and resolve the mechanism of innate and acquired resistance to MAPK inhibition in lethal thyroid cancer. Further, we demonstrate that the type II RAFi naporafenib can lead to repression of both MAPK and PI3K-AKT signaling models of ATC. In addition, we identify compensatory alterations in the MAST1 kinase as a mechanism of acquired resistance to type II RAFi. Biologically, we identified EphA2 as an important mediator of RAFi resistance that is enriched in thyroid carcinomas and may explain the lower response rate of thyroid cancers toward type I RAFi, compared to melanomas. EphA2 is a receptor tyrosine kinase that activates MAPK signaling by facilitating RAS activation and downstream ERK phosphorylation, as well as PI3K-AKT signaling.60,61 Thus, our data further underscore the importance of the PI3K-AKT pathway in the aggressivity of ATC. Our recent large-scale genomic and evolutionary analysis of the progression of ATC from the indolent PTC identified an enrichment of PIK3CA hotspot alterations,6 resulting in further dedifferentiation and increased tumor aggressivity.62 Indeed, inhibition of both MAPK and PI3K-AKT led to prolonged and sustained tumor control in PDX models of ATC.62 Thus, we add to the mounting evidence that the PI3K-AKT pathway is a critical mediator of response to RAFi. Methodologically, we demonstrate the advantages of deep molecular phenotyping at multiple time points across the treatment regime of a rare cancer type, followed by bedside-to-preclinical modeling to discover and understand mechanisms of treatment resistance in ATC. Starting from single-patient data, we delineated the genomic, transcriptomic, and microenvironmental evolution during therapy. These insights led us to identify that MAPK inhibition using type II RAFi may be efficacious in in vitro cell line models of ATC. We further generated PDX models from the same patient and demonstrated that naporafenib and trametinib can more effectively control tumor growth. Finally, we identified additional mechanisms of acquired resistance to naporafenib and trametinib treatment that could lead to additional therapeutic targets. Compared to other MAPK-altered cancer types such as colorectal cancer and melanoma, the tissue-specific mechanism of resistance to RAFi in thyroid cancer remains understudied. The extreme aggressivity of ATC makes it an exceptional case of MAPK-addicted cancer, and as we have demonstrated here, its detailed characterization can illuminate substantial insights to MAPK-directed therapies that may extend to other molecularly similar cancer types. Together, this framework could prove useful for the study of other rare cancers and diseases. Clinically, our study has important implications for both thyroid cancer and other MAPK-addicted cancers. While ATC is rare, patients suffering from this disease have almost 100% disease-specific mortality, with no standard therapeutic approach.5 A portion of the 20%–50% of patients with tumors harboring BRAFV600E alteration respond to dabrafenib/trametinib,63 although almost all patients eventually develop resistance.10 Moreover, although recent studies have identified potential roles of immunotherapeutic agents,64,65 there are no approved therapies for the ∼50%–80% of ATC patients with BRAF-wild-type tumors or those with acquired resistance to RAFi/MEKi. As most patients have rapid disease progression following diagnosis, the wait time for molecular testing (usually 2–4 weeks) precludes some from further treatment. Thus, the development of a genotype-agnostic therapeutic approach allowing for immediate treatment may be especially attractive for ATC. Furthermore, our results may extend to other MAPK-addicted cancer types, such as a subset of colorectal, pancreatic, and lung cancers. Currently, many tissue-agnostic basket trials are enrolling MAPK-altered patients, and our results suggest that it will be important to assess MAST1 alterations as a putative mechanism of resistance. As clearly demonstrated in both the clinical course and follow-up in vivo work utilizing ATCWGS42, treatment with targeted therapies can lead to clonal sweep of subclonal mutations, which leads to resistance. Strategies to overcome intratumor heterogeneity are urgently needed to offer more durable responses, for example, through combination treatments with immunotherapy. Indeed, a limitation of the current study is the incomplete understanding of the role of the tumor immune microenvironment in resistance to targeted therapy. Future studies will be required to better delineate the specific role of cell subsets in mediating therapy response to type I and type II RAFi in thyroid cancer. In conclusion, our multi-faceted approach, paired with functional studies, has identified and validated therapeutic targets to potentially improve patient outcomes for one of the most lethal cancer types. Future work should examine whether our findings apply to BRAFV600E or RAS-mutant tumors from other cancer types such as melanoma and colorectal cancer to evaluate whether the identified effects are thyroid cancer specific or may extend to other MAPK-driven tumor types. Limitations of the study Our study has several limitations. First, given the rarity of ATC in humans, we were only able to test naporafenib in two ATC PDX models. Second, both PDX models that were utilized in this study originated from patient tumors harboring the BRAFV600E alteration. As such, future studies should examine whether our findings apply to both non-BRAFV600E MAPK-driven (such as RAS-mutant) thyroid cancers and other cancer types alone or in combination with immunotherapy. Resource availability Lead contact Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Dr. Anthony C. Nichols (anthony.nichols@lhsc.on.ca). Materials availability This study did not generate new unique reagents. Data and code availability •All generated bulk RNA-seq, whole-genome/exome sequencing, and single-nuclei datasets are available upon request at EGA with accession codes EGAD50000002145, EGAD50000002146, and EGAD50000002147. TCGA data were retrieved from the National Cancer Institute Genome Data Commons (https://portal.gdc.cancer.gov/). The Cancer Cell Line Encyclopedia was retrieved using CellMinerDB (https://discover.nci.nih.gov/rsconnect/cellminercdb/).•The current manuscript utilizes software and packages that have been published previously. Detailed information regarding all software packages, algorithms, and computational tools employed in this study, including version numbers and specific parameters, are provided in relevant method subsections. Code can be accessed at https://doi.org/10.5281/zenodo.19228501.•Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request. •All generated bulk RNA-seq, whole-genome/exome sequencing, and single-nuclei datasets are available upon request at EGA with accession codes EGAD50000002145, EGAD50000002146, and EGAD50000002147. TCGA data were retrieved from the National Cancer Institute Genome Data Commons (https://portal.gdc.cancer.gov/). The Cancer Cell Line Encyclopedia was retrieved using CellMinerDB (https://discover.nci.nih.gov/rsconnect/cellminercdb/). All generated bulk RNA-seq, whole-genome/exome sequencing, and single-nuclei datasets are available upon request at EGA with accession codes EGAD50000002145, EGAD50000002146, and EGAD50000002147. TCGA data were retrieved from the National Cancer Institute Genome Data Commons (https://portal.gdc.cancer.gov/). The Cancer Cell Line Encyclopedia was retrieved using CellMinerDB (https://discover.nci.nih.gov/rsconnect/cellminercdb/). •The current manuscript utilizes software and packages that have been published previously. Detailed information regarding all software packages, algorithms, and computational tools employed in this study, including version numbers and specific parameters, are provided in relevant method subsections. Code can be accessed at https://doi.org/10.5281/zenodo.19228501. The current manuscript utilizes software and packages that have been published previously. Detailed information regarding all software packages, algorithms, and computational tools employed in this study, including version numbers and specific parameters, are provided in relevant method subsections. Code can be accessed at https://doi.org/10.5281/zenodo.19228501. •Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request. Acknowledgments We would like to thank all patients and their families who participated in this study. We wish to acknowledge the contribution of Dr. Vivian Bradaschia and Dr. Karuna Kapoor at The Centre for Phenogenomics Pathology Core for assistance with histology and immunohistochemistry. We would like to thank Dr. Sareh Parangi for kindly sharing with us the TBP3743 cell line. This work was supported by Canadian Institutes of Health Research grants MOP 487005 to A.C.N., L.G., L.A., and P.C.B. and MOP 145586 to P.C.B. and by the Government of Canada through Genome Canada and the Ontario Genomics Institute (OGI-125). This work was also supported through generous donations by the 10.13039/100019711London Health Sciences Foundation (Robert and Sheila Wilkes, Betty and Harry Ostrander, and Mervyn and Joyce Dietz) and the Woodstock Foundation (Laila, Arnold, Mario, Andrea, Tony and Tomassina Spina, Susan and Bill George, Piero and Maria Manzini, Henry and Rina Deroo, Harry and Shani Loewith, Franco and Carol Castellucci, Tony and Bill Van Haeren, Billy Stevanovich, Jennifer Peace-Hall, Tom and Pat Baird, Christine and Jeff Nichols, Cliff and Linda Zaluski, Thomas Vandertuin, Laurie and Paul Green, and Kelli and Mike Koopman). A.C.N. was supported by the Wolfe Surgical Research Professorship in the Biology of Head and Neck Cancers Fund. L.A. is supported by an Ontario Institute for Cancer Research Investigator Award (OICR IA-016). P.Y.F.Z. is supported by a Vanier Canada Graduate Scholarship and PSI foundation fellowship. Author contributions Conceptualization and design, P.Y.F.Z., J.W.B., L.A., and A.C.N.; methodology and investigation, P.Y.F.Z., J.M., A.K., H.P., L.J., S.B.R., and A.E.D.; data analysis and bioinformatics, P.Y.F.Z., J.M., A.K., P.C.B., and A.C.N.; clinical sample acquisition and analysis, A.C.N., M.J.C., D.A.P., E.W., L.G., J.S.M., and J.W.B.; funding acquisition, A.C.N., L.A., P.C.B., and P.Y.F.Z.; supervision, L.A., A.C.N., and P.C.B.; writing – original draft, P.Y.F.Z., J.W.B., L.A., J.S.M., and A.C.N., writing – review and editing, all authors. Declaration of interests A.C.N. has research funding from Novartis Canada, Merck Canada, LabCorp, and Droplet Biosciences for studies that are unrelated to the submitted work. He has equity from and is a consultant for NEED Inc. M.J.C. has research funding from AstraZeneca, Merck, and Pfizer. He has received payment for speaker honorarium and/or served on advisory boards for Eli Lilly Merck, AstraZeneca, and Amgen. M.J.C. has equity from and is a consultant for NEED Inc. D.A.P. is a consultant and has equity from NEED Inc. E.W. has consulted or advised for Merck, Bayer, Eisai, Roche, Ipsen, and EMD Serono. E.W. reports research funding from Merck, Novartis, and Lilly through Western University. P.Y.F.Z., J.W.B., J.S.M., and A.C.N. hold patents for transcriptional biomarkers in head and neck cancer, unrelated to this work. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERAntibodiesrabbit anti-EphA2Cell Signaling TechnologyCat# 6997; RRID:AB_10827743rabbit anti-pMEKCell Signaling TechnologyCat# 2338; RRID:AB_490903rabbit anti-pERKCell Signaling TechnologyCat# 4370; RRID:AB_2315112rabbit anti–phospho-p70 S6 kinaseCell Signaling TechnologyCat# 9234; RRID:AB_2269803rabbit anti–phospho-EphA2Cell Signaling TechnologyCat# 6347; RRID:AB_11220420rabbit anti-MAST1Thermo FisherCat# PA5-84915; RRID:AB_2792064Opal 520Akoya BiosciencesRef. OP-001001Opal 570Akoya BiosciencesRef. OP-001003Opal 690Akoya BiosciencesRef. OP-001006Biological samplesATC patient tumor samplesThis studyN/ACritical commercial assaysReverse-phase protein arrayMD AndersonN/ADeposited dataWGS, WES, and snRNA-seqThis studyEGAD50000002145, EGAD50000002146, EGAD50000002147TCGA Genomics DataTCGA GDC portalhttps://portal.gdc.cancer.gov/Cancer cell line encyclopediaCellMinerDB66https://discover.nci.nih.gov/rsconnect/cellminercdb/Experimental models: Cell linesTBP3743 cell lineVanden Borre et al.57N/AA375 and isogenic CRISPR-Cas9 knock-in NRAS-Q61KATCCCRL-1619IG-2ASH3JCRBN/ABHT101DSMZN/AKMH2JCRBN/ASW1736University of UppsalaN/ATHJ-16TMayo ClinicN/ATHJ-29TMayo ClinicN/AUHTh7University of UppsalaN/A8505CJCRBN/AExperimental models: Organisms/strainsB6129SF1/J miceJackson LabsStrain #:101043NOD/SCID/IL2Rγ−/−(NSG) miceUniversity Health Network Animal Resources CenterN/ARecombinant DNAPlasmids expressing wildtype or A733V MAST1This studyN/ASoftware and algorithmsBWA-mem (v0.7.17)Li and Durbin67https://github.com/lh3/bwasamtools (v1.16.1)https://www.htslib.org/GATK4 (v4.3.0)https://github.com/broadinstitute/gatk/releasesManta (v 1.6.0)https://github.com/Illumina/mantatitanCNA (v 1.17.1)Ha et al.68https://github.com/gavinha/TitanCNACONIPHERGrigoriadis et al.69https://github.com/McGranahanLab/CONIPHER-wrapperxengsort (v2.0.1)Zentgraf and Rahmann70https://gitlab.com/genomeinformatics/xengsortCellRanger (v6.0.2)https://www.10xgenomics.com/support/software/cell-ranger/latestharmonypyKorsunsky et al.71https://github.com/slowkow/harmonypyscanpy (v1.9.3)Wolf et al.72https://github.com/scverse/scanpynplr (v.0.1–7)https://cran.r-project.org/web/packages/nplr/index.htmlgprofiler2 (v0.2.2)https://cran.r-project.org/web/packages/gprofiler2/index.htmlSTAR (v2.6.1day)Dobin et al.73https://github.com/alexdobin/STARRSEM (v1.3.1)Li and Dewey74https://deweylab.github.io/RSEM/DESeq2 (v 1.40.2)Love et al.75https://bioconductor.org/packages/release/bioc/html/DESeq2.htmldecoupleR (v2.6.0)Badia et al.76https://saezlab.github.io/decoupleR/R (v4.3.1)The R Projecthttps://www.r-project.org/Python (v3.8.17)Python Software Foundationhttps://www.python.org/rstatix (v0.7.2)https://cran.r-project.org/web/packages/rstatix/index.htmllme4 (v1.1-34)https://cran.r-project.org/web/packages/lme4/index.html REAGENT or RESOURCESOURCEIDENTIFIERAntibodiesrabbit anti-EphA2Cell Signaling TechnologyCat# 6997; RRID:AB_10827743rabbit anti-pMEKCell Signaling TechnologyCat# 2338; RRID:AB_490903rabbit anti-pERKCell Signaling TechnologyCat# 4370; RRID:AB_2315112rabbit anti–phospho-p70 S6 kinaseCell Signaling TechnologyCat# 9234; RRID:AB_2269803rabbit anti–phospho-EphA2Cell Signaling TechnologyCat# 6347; RRID:AB_11220420rabbit anti-MAST1Thermo FisherCat# PA5-84915; RRID:AB_2792064Opal 520Akoya BiosciencesRef. OP-001001Opal 570Akoya BiosciencesRef. OP-001003Opal 690Akoya BiosciencesRef. OP-001006Biological samplesATC patient tumor samplesThis studyN/ACritical commercial assaysReverse-phase protein arrayMD AndersonN/ADeposited dataWGS, WES, and snRNA-seqThis studyEGAD50000002145, EGAD50000002146, EGAD50000002147TCGA Genomics DataTCGA GDC portalhttps://portal.gdc.cancer.gov/Cancer cell line encyclopediaCellMinerDB66https://discover.nci.nih.gov/rsconnect/cellminercdb/Experimental models: Cell linesTBP3743 cell lineVanden Borre et al.57N/AA375 and isogenic CRISPR-Cas9 knock-in NRAS-Q61KATCCCRL-1619IG-2ASH3JCRBN/ABHT101DSMZN/AKMH2JCRBN/ASW1736University of UppsalaN/ATHJ-16TMayo ClinicN/ATHJ-29TMayo ClinicN/AUHTh7University of UppsalaN/A8505CJCRBN/AExperimental models: Organisms/strainsB6129SF1/J miceJackson LabsStrain #:101043NOD/SCID/IL2Rγ−/−(NSG) miceUniversity Health Network Animal Resources CenterN/ARecombinant DNAPlasmids expressing wildtype or A733V MAST1This studyN/ASoftware and algorithmsBWA-mem (v0.7.17)Li and Durbin67https://github.com/lh3/bwasamtools (v1.16.1)https://www.htslib.org/GATK4 (v4.3.0)https://github.com/broadinstitute/gatk/releasesManta (v 1.6.0)https://github.com/Illumina/mantatitanCNA (v 1.17.1)Ha et al.68https://github.com/gavinha/TitanCNACONIPHERGrigoriadis et al.69https://github.com/McGranahanLab/CONIPHER-wrapperxengsort (v2.0.1)Zentgraf and Rahmann70https://gitlab.com/genomeinformatics/xengsortCellRanger (v6.0.2)https://www.10xgenomics.com/support/software/cell-ranger/latestharmonypyKorsunsky et al.71https://github.com/slowkow/harmonypyscanpy (v1.9.3)Wolf et al.72https://github.com/scverse/scanpynplr (v.0.1–7)https://cran.r-project.org/web/packages/nplr/index.htmlgprofiler2 (v0.2.2)https://cran.r-project.org/web/packages/gprofiler2/index.htmlSTAR (v2.6.1day)Dobin et al.73https://github.com/alexdobin/STARRSEM (v1.3.1)Li and Dewey74https://deweylab.github.io/RSEM/DESeq2 (v 1.40.2)Love et al.75https://bioconductor.org/packages/release/bioc/html/DESeq2.htmldecoupleR (v2.6.0)Badia et al.76https://saezlab.github.io/decoupleR/R (v4.3.1)The R Projecthttps://www.r-project.org/Python (v3.8.17)Python Software Foundationhttps://www.python.org/rstatix (v0.7.2)https://cran.r-project.org/web/packages/rstatix/index.htmllme4 (v1.1-34)https://cran.r-project.org/web/packages/lme4/index.html Experimental model and study participant details Clinical timeline of the patient This study has been approved by the Western University institutional ethics board (REB#108920) and received written consent from the two ATC patients. ATCWGS42 was a 44-year-old male who presented with a thyroid mass in December 2018. Ultrasound guided fine needle aspiration (FNA) was highly suspicious of malignancy. In Mar 2019, an attempt at surgical resection was aborted due to extensive laryngotracheal involvement. A total of five spatially distinct tumor sites were biopsied (regions A, B, C, D, E). BRAFV600E mutation status was identified via IonTorrent deep-coverage targeted sequencing, and the patient was started on vemurafenib and cobimetinib in Apr 2019. In May 2019, the patient was confirmed to completely respond to treatment, with complete resolution of neck swelling. Continuous response was seen in Jul 2019. Disease progression was confirmed in Sep 2019. Tracheostomy was performed, and biopsy of the tumor was performed (region R). The patient was started on lenvatinib in Sep 2019, but experienced pain crisis. Lenvatinib was stopped and the patient started pembrolizumab. The patient showed stable disease on Jan 2020. After 8 cycles of pembrolizumab, the patient progressed once again in Mar 2020, and after discussion was transferred to palliative care. The patient unfortunately passed away in June 2020. Establishment of patient-derived xenograft (PDX) and immunocompetent syngeneic models and treatment Fresh ATC tumor specimens were received and kept at 4°C in PBS until engraftment no later than 24-h post resection, as we have previously done. Briefly, viable tumor pieces were implanted subcutaneously into the flank regions of NOD/SCID/IL2Rγ−/− (NSG) mice. The TBP3743 syngeneic ATC cell line was similarly implanted subcutaneously into the flank regions of B6129SF1/J mice, as previously described.57 Mice were sacrificed using isoflurane anesthesia followed by CO2 asphyxiation when tumors reached a diameter of 1–1.5 cm. Tumors were dissociated in culture medium (containing DNASE 1 and collagenase/hyaluronidase) and passaged subcutaneously with five mice per treatment in 1:1 Matrigel/PBS to generate the PDX models. Mice with tumor volumes 80–120 mm3 were randomly assigned to groups of 5 or more, and treated daily with vehicle, dabrafenib, trametinib, dabrafenib/trametinib, naporafenib, or naporafenib/trametinib. Dabrafenib (30mg/kg), naporafenib (15mg/kg), and trametinib (0.6mg/kg) were administered via oral gavage 5 days per week for up to 3 weeks. Tumor volume was calculated using the formula (π/6)×length×width×height. Tumor volume comparison was made using linear mixed-effects model, including tumor treatment group and day as fixed effects, and individual mice number as random effects. Method details DNA whole genome sequencing (WGS) and whole exome sequencing (WES) Whole-genome sequencing (WGS) was performed on five spatially distinct regions of the pre-vemurafenib and cobimetinib treatment tumor (regions A, B, C, D, E), one region of the post-treatment tumor (R), and matched normal tissue (blood). Genomic DNA was isolated from snap-frozen tissue and subjected to whole genome sequencing at The Center for Applied Genomics (Hospital for Sick Children, Toronto, ON) or at Beijing Genomic Institute (BGI). High-coverage whole exome-sequencing was performed at GeneSeeq Inc. (Toronto, Ontario, Canada) using IDT xGen exome panel.6,77,78 Data were aligned to the GRCh38 human reference genome, including available ALT and decoy alleles using BWA-mem (v0.7.17).67 Lane-level BAM files were merged using samtools (v1.16.1). GATK4 (v4.3.0) was used to mark duplicate reads and perform base quality recalibration. SNVs were identified using Mutect2 (v4.3.0). Indels and structural variations were called using Manta (v 1.6.0). Allele-specific subclonal copy number alterations were identified using titanCNA (v 1.17.1). Subclonal reconstruction was performed using CONIPHER using copy number calls generated from WGS data and variants generated from high-coverage WES, as previously described.69 For WES of PDX models, we used a similar approach except that human and mouse reads were first separated using xengsort70 (v2.0.1) using default settings before proceeding to alignment of human reads to the human genome and the rest of the workflow. Single nucleus-RNA sequencing (snRNA-seq) We performed single nucleus RNA-sequencing (snRNA-seq) on two regions from ATCWGS42 (ATCWGS42_E and ATCWGS42_R) as done previously.79,80 Fresh tumor tissue was minced into 1–2 mm3 pieces on dry ice using a chilled razor blade. The tissue was covered with lysis buffer and transferred to a tube to lyse on ice. The supernatant and pellet were dounced separately and transferred to chilled tubes. Nuclei viability was confirmed by SYBR staining using a Countess Cell Counter (Invitrogen) and collected by centrifugation (800xg for 10 min). This process was repeated. Nuclei were then resuspended in 1 mL suspension buffer and passed through a 40 μm Flowmi cell strainer. Nuclei were then loaded onto the Chromium single cell controller (10X Genomics) and used to construct single nucleus 3′ libraries. Sequencing was performed using NovaSeq 6000 sequencer. Alignment to GRCh38 reference genome, barcode, and unique molecular identifier counting were performed using CellRanger (v6.0.2). Nuclei were quality-controlled and filtered using the following criteria: 1) number of detected genes was above 200) the number of transcript counts detected was above 500. We further removed nuclei whose genes detected, total transcript count, total mitochondrial transcript count, and mitochondrial transcript percentage are 5 median absolute deviation (MAD) above the median. We then logged and normalized transcripts counts per nucleus. We integrated the dataset across using harmonypy.71 In total, we observe a total of 6367 high-quality cells. To identify cell states, we used the python scanpy (v1.9.3) package to identify highly variable genes (HVG) with a max_mean of 3, a minimum mean of 0.0125, and a minimum dispersion of 0.5 that resulted in 3909 highly-variables genes. Principal component analysis (PCA) was performed on the HVGs. The scanpy neighbors function was used to generate nearest neighborhood graphs. Clustering was performed using the leiden algorithm. UMAP was applied for dimensionality reduction. We performed hierarchical annotations of the identified clusters. In the first level of cell annotations, we used well-characterized cellular markers to annotate cells as epithelial (FXYD3, EPCAM, ELF3), immune (PTPRC, CD53, CORO1A), stroma (COL1A3, DCN, MFAP4), and endothelial (CLDN5, ECSCR, CLEC14A), which makes up cell type level 1. We then further annotated cell clusters using marker genes identified using the scanpy rank_genes_groups function hierarchically with increasing precision until the finest level (cell type level 4). In total, we identified 1 endothelial population, 2 ATC tumor populations, 4 fibroblast populations, and 9 immune populations (4 lymphoid and 5 myeloid derived). Thyroid differentiation score was calculated as described.6,81 In vitro drug screening All thyroid cancer cell lines were acquired from Japanese Collection of Research Bioresources Cell Bank (JCRB), Deutsche Sammlung von Mikroorganismen und Zellkulturen (DSMZ), University of Uppsala, or the Mayo Clinic and were subjected to regular STR profiling. A375 and isogenic CRISPR-Cas9 knock-in NRAS-Q61K A375 cell lines were acquired from American Type Culture Collection (ATCC). The A375 NRAS-Q61K cell line is known to be resistant to RAF inhibition (https://www.atcc.org/products/crl-1619ig-2). Cells were plated in 96 well plates. After 24 h, for screening using only RAFi, cells were treated with the RAFi in 6 replicates to generate 10-point dose-response curves ranging from 1.5 nM to 100μM. for experiments testing combined treatment of dabrafenib and trametinib, 10nM of trametinib was added to each well that contained variable amounts of dabrafenib. For experiments assessing the effect of trametinib in addition to naporafenib, three different doses of trametinib (0, 1, or 10 nM) were tested in combination with a dose range of 32 nM to 100μM of naporafenib. After 72 h, cells were incubated with PrestoBlue, fluorescence readings were taken and mean inhibitory concentrations (IC50) were calculated using the R package nplr (v.0.1–7). Functional phosphoproteomics using reverse-phase protein array Targeted functional phosphoproteomics using reverse-phase protein array (RPPA) was performed at the MD Anderson Cancer Center RPPA core. Briefly, protein extracts were diluted 2-fold in five serial dilutions (undiluted to 1:16) and spotted onto nitrocellulose-coated slides. These extracts are arranged in a 4 × 12-pin-11 × 11 subarray that accommodates 1056 serially diluted samples plus replicate vertical controls on each slide. Each protein of interest was detected by probing with a specific antibody, amplifying the signal via a tyramide amplification system, and visualized via DAB colorimetric reaction. Each slide was probed with one antibody. The detection system used was a GenPoint-based staining kit (Agilent). Digital images of slides were obtained by scanning on a Huron TissueScope scanner producing 16-bit TIFF files. Spot intensities from the TIFF files were determined via Array-Pro Analyzer software. Mean net intensities of spots from all 1056 samples were used for curve-fitting for each slide (RPPA Super Position and Concentration Evaluation). Relative protein level for each sample was determined by converging its sample dilution series to one point (EC50) and interpolating in RPPASPACE (MD Anderson). Bulk RNA-sequencing of tumors from PDX models Tissue was first homogenized with a VWR microhomogenizer, and the homogenate passed through a QiaShredder as previously described.82,83,84 Qiagen AllPrep DNA/RNA kits were used to isolate total RNA. RNA was shipped to Geneseeq Technologies (Toronto, ON) for quality control, library preparation, and sequencing.85 RNA quality was confirmed with Bioanalyzer, and libraries were prepared using a NEB Ultra II Directional mRNA library kit. Samples were then processed using random primers and paired end sequencing using an Illumina HiSeq 2500 yielded 100 million reads/sample. Human and mouse reads were first separated using xengsort70 (v2.0.1) using default settings. The human reads were then aligned to the reference genome GRCh38 using STAR (v2.6.1day) two-pass mode, and transcripts abundance were estimated using RSEM (v1.3.1). DESeq2 (v 1.40.2) was used to perform multifactorial-analysis, by including RAFi (vehicle, dabrafenib, or naporafenib), MEKi (vehicle or trametinib), and their interactions terms in the design formula. GProfiler2 (R package gprofiler2 v0.2.2) was used to perform pathways analysis, excluding gene sets with more than 400 or less than 20 genes. PROGENy24 as part of the decoupleR (v2.6.0) was used to infer the activity of MAPK, PI3K-AKT, and WNT pathways. Immunofluorescence on FFPE PDX samples Samples were submitted to The Center for Phenogenomics (TCP) Pathology Core in Toronto. Tissues were FFPE-processed, sectioned at 5 μm, and mounted on charged slides. Paraffin sections were baked and dewaxed on a Leica Bond RX system and subjected to sequential Opal multiplex immunohistochemistry. Slides were stained with rabbit anti-EphA2 (Cell Signaling Technology, Cat# 6997, 1:100) and visualized with Opal 520 (Akoya Biosciences, Ref. OP-001001), followed by antibody stripping using ER1 solution. The staining cycle was then repeated with rabbit anti-pERK (Cell Signaling Technology, Cat# 4370, 1:100) conjugated with Opal 570 (Akoya Biosciences, Ref. OP-001003), and rabbit anti-pMEK (Cell Signaling Technology, Cat# 2338, 1:100) conjugated with Opal 690 (Akoya Biosciences, Ref. OP-001006). All staining steps were performed on the Leica Bond RX using reagents from the Opal 3-Plex Automation IHC Kit (Akoya Biosciences, Ref. NEL820001KT). Similarly, a second set of slides were stained with rabbit anti–phospho-p70 S6 kinase (Cell Signaling, Cat# 9234, 1:100) and visualized with Opal 520 (Akoya Biosciences, Ref. OP-001001), followed by antibody stripping with ER1 solution. The staining cycle was then repeated using rabbit anti–phospho-EphA2 (Cell Signaling, Cat# 6347, 1:100) conjugated with Opal 570 (Akoya Biosciences, Ref. OP-001003). These staining steps were performed sequentially on the Leica Bond RX using reagents from the Opal 3-Plex Automation IHC Kit (Akoya Biosciences, Ref. NEL820001KT). Following Opal staining, sections were stained for rabbit anti-MAST1 (Cell Signaling, Cat# PA5-84915, 1:100) and detected with an Alexa Fluor 647–conjugated donkey anti-rabbit secondary antibody (Invitrogen, Cat# A32795). Following staining, nuclei were counterstained with DAPI. Tissue autofluorescence was minimized using Sudan Black B treatment, and sections were mounted with Vectashield Vibrance Antifade Mounting Medium. Image analysis was performed with ZenLite (version 3.13), registered, and fluorescence intensity was normalized across each sample. Transfection of MAST1 expression plasmids BHT-101 or 8505C cells were seeded at a density of 2 × 10ˆ6 cells per well in a 6-well dish and were incubated overnight at 37°C with a supplement of 5% CO2. The next day the cells (60–80% confluency) were transfected with the plasmids expressing either WT MAST1 or MAST1 A733V mutation respectively with Lipofectamine 3000 Transfection Reagent (Invitrogen) per the manufacturer’s protocol. The cells were then incubated for four hours with the transfection mixture followed by the replacement of the transfection media with complete media. The transfected cells were then placed in the incubator for 72 h. Blasticidin was added at 5 μg/mL to select successfully transfected cells. Non-transfected parental controls were dead within 48 h. Drug selection was continued for 1 week to ensure the maximum availability of the transfected cells. Drug treatment and testing of resistance The BHT-101 and 8505C cell lines transfected with either the plasmid carrying MAST1(wildtype) or MAST1 mutant (A733V) were counted and seeded at 5 × 103 per well in 96 well dishes. The plates were incubated at 37°C for 5 h. The spent media was removed and fresh media containing either 2 nM trametinib or 0 nM trametinib and various doses of naporafenib (0.15–10uM) were added to the cells. The plates were incubated for up to 72 h at 37°C. Cell viability was then determined using Presto Blue (Thermo Fisher Scientific, Waltham, MA, USA) viability reagent. Prestoblue was added and the plate was incubated for 60 min at 37°C and the fluorescence levels were measured using a Synergy H4 Hybrid Reader (BioTek, Winooski, VT, USA). The cells were excited at 560 nm and the emission measured at 590 nm. Response curves were generated using the R-script nplr package (v.0.1–7). Quantification and statistical analysis Statistics All statistical analyses were performed using either R (v4.3.1) or Python (v3.8.17). Wilcoxon test, linear mixed effect models, and Pearson’s correlation were performed using the packages rstatix (v0.7.2), lme4 (v1.1-34), or stats. Statistical details of experiments can be found within the figure and/or figure legends. Unless otherwise indicated in the figure legend or text, correction of multiple comparisons was performed utilizing the Benjamini-Hochberg procedure as previously described. All tests were two-sided.