Blocking Genomic Instability Prevents Acquired Resistance to MAPK Inhibitor Therapy in Melanoma

Acquired resistance to targeting of oncogenic pathways is the rule rather than the exception. Genetic mechanisms of acquired resistance across malignancies often occur as focal amplifications of resistance-driver genes. Scientific efforts to reverse the consequences of resistance evolution have focused largely on targeting acquired vulnerabilities. In contrast, relatively little is known regarding the mechanisms—for example, genomic instability pathways—that enable rapid resistance evolution of cancer in patients treated with pathway-targeted therapies. Variable levels of focal gene amplifications can confer phenotypic plasticity, fine-tuning tumor cell fitness in response to the selective pressure of therapies.

BRAF inhibitor (BRAFi) therapy of BRAF^V600MUT metastatic cutaneous melanoma targets the MAPK pathway and leads quickly to acquired resistance (1). Genetic mechanisms of acquired BRAFi resistance most often result in reactivation of the MAPK pathway, which can be suppressed by adding a MEK inhibitor (MEKi; refs. 2–5). However, with combined BRAFi and MEKi therapy of patients with BRAF^V600MUT melanoma, less than 20% of patients survive past 5 years (6). Acquired resistance to BRAFi + MEKi therapy in patients occurs through high-amplitude gene amplifications that, individually or in combination, can reactivate the MAPK pathway (7–9). Hence, there is a need to understand the origins of such genomic instability. For patients with melanoma driven by NRAS mutations, MAPK inhibitor (MAPKi) therapy is currently not available, as MEKi monotherapy is of limited clinical activity (10). However, agents added to MEKi to suppress acquired MEKi resistance may deliver clinically meaningful efficacy (11). Recent analysis of patient-derived xenografts (PDX) of NRAS^MUT melanoma with acquired MEKi resistance points to similar genomic events (i.e., focal amplifications of BRAF^WT, CRAF^WT, and NRAS^MUT genes) as drivers of acquired resistance (11).

Here, we test the hypothesis that chromothripsis and derivative amplicons of resistance-driver genes confer melanoma with genetic variants necessary to resist MAPKi therapy. Cutaneous melanoma is a cancer in which the chromothripsis burden is already high without prior targeted therapy, and chromothripsis appears to be a key evolutionary mechanism by which cancer rapidly generates and accumulates highly dynamic structural variants (SV; ref. 12). SV-related amplicons can be identified as intrachromosomal complex genomic rearrangements (CGR) and extrachromosomal DNAs (ecDNA, aka double minutes), which may be temporally related structures that confer a range of genomic–signaling plasticity (13). Earlier studies have established the prognostic importance of ecDNAs in clinical tumor tissues and the potential relevance of ecDNAs to therapeutic resistance in cancer cell lines (14, 15). The non-Mendelian inheritance of ecDNAs, their accessible chromatin, as well as enhancer hijacking and transcriptional hub congregation by ecDNAs are all mechanisms that facilitate rapid adaptations to extreme stress such as oncogene-targeted therapy (16–19). Finally, iterative cycles of CGR and ecDNA biogenesis drive the amplification of a model resistance gene (DHFR) to the chemotherapy methotrexate (20).

To test the hypothesis that chromothripsis, CGRs, and ecDNAs play key roles in the evolution of MAPKi resistance in clinical melanoma, we analyzed by whole-genome sequencing (WGS) three tumor cohorts (along with patient-matched normal tissues): (i) patient-matched pre-MAPKi and acquired MAPKi-resistant tumors from patients with BRAF^V600MUT melanoma, (ii) acquired MAPKi-resistant melanoma metastatic to multiple organ sites (including the brain) from rapid autopsies of deceased patients with BRAF^V600MUT melanoma, and (iii) acquired MEKi-resistant melanoma and untreated melanoma from BRAF^V600MUT and, importantly, NRAS^MUT melanoma PDXs. We evaluated (i) resistance-specific recurrence of amplicons harboring known and putative resistance-driver genes in CGRs and ecDNAs, (ii) overlaps in the genomic coordinates between CGR–ecDNAs and chromothripsis, and (iii) inferred double-stranded DNA break (DSB) repair pathway(s) based on analysis of breakpoint–junctional sequences of CGR and ecDNA amplicons. Using human melanoma cell lines and PDXs, we tested the efficacy and dissected the mechanisms of blocking genomic instability via targeting of NHEJ-mediated DSB repair to prevent the evolution of acquired MAPKi resistance in both BRAF^V600MUT and NRAS^MUT melanoma.

To date, mutational profiles of acquired MAPKi resistance (in patients with BRAF^V600MUT melanoma and treated with BRAFi or BRAFi + MEKi therapy) have been limited to whole-exome analysis (5, 7–9, 21). Thus, we assembled three cohorts of tissues (Supplementary Table S1) for WGS-based analysis of SVs (Supplementary Table S2). The first cohort consisted of patient-matched normal tissues as well as BRAF^V600MUT melanoma tumors before MAPKi therapy and, after initial responses, at disease progression (n = 10 normal tissues; n = 10 pretreatment tumors; n = 17 acquired-resistant tumors; n = 10 patients). The second cohort consisted of rapid autopsy melanoma (RAM) tissues (n = 3 normal tissues; n = 1 sensitive BRAF^V600MUT tumor; n = 12 acquired resistant tumors; n = 3 deceased subjects who were treated with MAPKi; n = 6 metastatic organ sites). The third cohort consisted of BRAF^V600MUT or NRAS^MUT PDX tumors. We subjected PDXs (n = 6 models; 1 BRAF^MUT and 5 NRAS^MUT models) to MAPKi therapy in NOD-scid IL2R gamma null (NSG) mice at doses sufficient to elicit tumor regression and then generated acquired MAPKi-resistant tumors (n = 6 vehicle-treated tumors; n = 12 acquired-resistant tumors; n = 6 normal tissues; Supplementary Fig. S1A; refs. 11, 22). Sequencing depths across MAPKi-sensitive/naive tumors (mean, 38×; median, 32×; range, 18–94×), acquired-resistant tumors (mean, 33×; median, 27×; range, 11–85×), and patient-matched normal tissues (mean, 38×; median, 33×; range, 12–98×) were comparable. The median numbers of genetic alterations [somatic single-nucleotide variants (SNV), indels (ID), and copy-number variations (CNV)] for sensitive and resistant tumors were, respectively: (i) 81 SNVs/megabase (Mb), 5 IDs/Mb, and 261 CNVs/genome and (ii) 97 SNVs/Mb, 7 IDs/Mb, and 253 CNVs/genome.

Using WGS data, we reconstructed ecDNA- and CGR-derived, high copy-number (CN) amplicons in 58 BRAF^V600MUT and NRAS^MUT MAPKi-sensitive/naive and acquired-resistant tumors (n = 17 sensitive tumors; n = 41 acquired-resistant tumors; n = 19 patients). We detected ecDNAs and CGRs at a highly recurrent rate: ∼70% of sensitive and ∼77% of acquired-resistant genomes (Supplementary Table S3). With all three cohorts combined, we observed a greater number of ecDNAs and CGRs in acquired-resistant (versus sensitive) genomes (acquired-resistant: n = 941 total, n = 31 average per tumor; sensitive: n = 241 total, n = 16 average per tumor; unpaired Student t test, P = 0.09; Fig. 1). In a patient-matched analysis of the clinical cohort, 12 of 15 pairwise comparisons showed a higher number of ecDNA and CGR amplicons in acquired-resistant genomes (Wilcoxon test, P = 0.033; Supplementary Fig. S1B). Overall, chr7p (26 of 53 genomes) and chr7q (23 of 53 genomes) emerged as hotspots (fragile sites) for CGR and ecDNA amplicon formation (Fig. 1). In addition, we observed recurrent resistance-associated CGRs in chr9p (8 of 41 resistant tumors across all three clinical–autopsy–PDX cohorts) and lower-frequency resistance-associated CGRs (in 2q, 4q, 6p, 8q, 9q, 10q, 18p, 22p, Xp) and ecDNAs (in 6q, 10p). In the total cohort, regardless of whether the tumor was sensitive or resistant to MAPKi, the number of rearranged segments was higher within ecDNAs (versus CGRs; P = 3.5e-08, unpaired Student t test; Supplementary Fig. S1C), suggesting ecDNAs as highly dynamic foci of genomic instability. Furthermore, the mean number of rearranged segments (ecDNAs + CGRs) was highest in the RAM (∼10 segments per tumor) followed by the clinical (∼5 segments per tumor) and PDX (∼4 segments per tumor) cohorts (Supplementary Fig. S1D; Kruskal–Wallis rank test, P = 0.01). This may reflect the highly advanced evolutionary state of terminal metastatic melanoma in the RAM cohort. Polyploidy has been associated with the onset of chromothripsis (12). Accordingly, we observed higher numbers of CGRs and ecDNAs in genomes with higher ploidy (Supplementary Fig. S1E).

Analysis of amplicons due to intrachromosomal CGRs and ecDNAs uncovered a significant (unpaired Student t test, P = 0.0002) association between acquired-resistant tumors and CGRs/ecDNAs harboring bona fide MAPKi resistance–driver genes. BRAF (CN range, 4.5–27), NRAS (CN, 5–13), HRAS (CN, 13–16), MYC (CN, 12–15), and EGFR (CN, 4.6–5; Figs. 1 and 2A–D) when amplified are known to drive acquired MAPKi resistance and MAPK pathway reactivation (5, 7–9, 11). RAC1 (CN, 6–7) in an ecDNA amplicon was specifically observed in acquired resistance in NRAS^MUT melanoma PDXs, which suggests regulation of MLK3–CRAF by RAC1 (23). Moreover, we validated a recurrent ecDNA by direct isolation and high-depth sequencing [NRAS amplification, CN 13, Pt9-double drug-disease progression (DD-DP)1; Fig. 2A] using a new approach referred to as CRISPR-CATCH (24). This alternative technique confirmed the circularized junctions of an 890-kb driver ecDNA within this acquired-resistant clinical tumor sample. In addition, we used DNA-FISH to validate the resistance-specific presence and the CNs of BRAF and NRAS amplicons (Fig. 2E). In total, among the 17 of 19 patients included in this study whose tumors harbored ecDNAs/CGRs, MAPK-reactivation genes were amplified via ecDNAs/CGRs specifically in acquired-resistant tumors (but not in any MAPKi-sensitive/naive tumor) at a high frequency (60% or 23 of 38 resistant tumors; Supplementary Table S3). Expectedly, MAPKi resistance–driver gene amplification via ecDNAs/CGRs often co-occurred with other resistance-specific CN alterations and somatic mutations reported earlier (Supplementary Table S4; refs. 7–9). Using tumor-matched RNA-sequencing (RNA-seq) data available from a subset of tumors, CN alterations (resistant versus sensitive) of MAPK-reactivation genes were highly positively correlated with mRNA level changes (Spearman correlation r_s = 0.64, P = 0.001; Supplementary Table S5 and Supplementary Fig. S1F). In a minority of patients (Pt5, ecDNA⁺CGR⁺; Pt11, ecDNA⁻CGR⁻; Supplementary Table S6), we observed lower-level but linear amplifications of BRAF (CN, 3–4) specifically in acquired-resistant tumors. The BRAF^V600MUT melanoma tumors from both patients were treated with only BRAFi (not BRAFi + MEKi), suggesting that low-level linear amplifications suffice in driving acquired resistance to BRAFi monotherapy.

Nonhomologous end-joining (NHEJ) and homologous recombination repair (HRR) of DSBs are thought to be critical for CGR and ecDNA generation after chromothripsis (20). Intriguingly, we observed that XRCC2 (RAD51-like; CN, 49), a key HRR gene, and XRCC6 (KU70; CN, 14–15), a key NHEJ gene, were amplified specifically in acquired MAPKi-resistant melanomas as ecDNAs (Fig. 2F; Supplementary Table S3). Moreover, acquired MAPKi resistance was associated with ecDNA or CGR amplicons spanning other DSB repair genes in NHEJ (TRIM33 CN, 7–11; PAXIP1 CN, 3) and HRR (SSBP1 CN, 5; BRCA2 CN, 6; RFC3 CN, 6; TRIM33 CN, 7–11; SYCP1 CN, 7–9, TRIM24 CN, 5–11) pathways (Supplementary Table S3). Thus, acquired MAPKi resistance is specifically associated with CGR and ecDNA amplicons in MAPK-reactivation and DSB repair genes, which suggests functional interplay as codrivers of resistance.

Other resistance-specific genes amplified by ecDNAs and/or CGRs may also contribute functionally to the resistant phenotype. Therefore, we identified the genes and their CNs in ecDNA (Fig. 2G) or CGR (Fig. 2H) amplicons specifically associated with sensitivity, resistance, or both. Importantly, genes amplified by either ecDNAs or CGRs specifically in resistance highly outnumbered those specific to sensitive tumors, which indicates that gene amplification by ecDNAs and CGRs contributes to disease progression on MAPKi therapy. Moreover, genes amplified by either ecDNAs or CGRs in both sensitive plus resistant tumors constituted a very small fraction, which is also consistent with the notion that gene amplification by ecDNAs and CGRs contributes to disease progression. We rarely detected overlapping genes in patient-matched sensitivity-associated CGRs and resistance-associated ecDNAs (and vice versa; Fig. 2G and H). To explore the functional contributions of ecDNA- and CGR-amplified genes to the acquired MAPKi-resistant phenotype, we performed pathway enrichment analysis using genes amplified by ecDNAs/CGRs specifically detected in resistance from eight evaluable cases (see Methods; Fig. 2I). This analysis nominated alterations in MAPK (in Pt3, PDX27), PI3K–AKT (RAM12.01, PDX4), immune (Pt4, RAM12.01, PDX4), G protein-coupled receptor (GPCR; Pt4, PDX6), cellular differentiation/morphogenesis (Pt3, Pt4, Pt9, PDX4, PDX6), and metabolism (Pt9, RAM12.01, PDX27, PDX1, PDX6) pathways as resistance phenotypes driven by ecDNAs/CGRs (Supplementary Table S7).

Recent studies have suggested that enhancers within ecDNAs influence oncogene expression by either coamplification with target oncogenes within the ecDNAs (i.e., cis interaction) or regulation of intrachromosomal genes (i.e., trans interaction; refs. 16, 25, 26). Thus, we annotated CGR and ecDNA amplicons in both sensitive and acquired resistant genomes with enhancers listed in GeneHancer (27). Notably, MAPKi-resistance genes (e.g., BRAF, NRAS, HRAS, and EGFR) and known oncogenes (e.g., EZH2, CREB3L2, CARD11, and EP300) were coamplified with their corresponding enhancers within CGRs and ecDNAs of MAPKi-resistant genomes (Supplementary Fig. S1G and Supplementary Table S8). We also observed enhancers and associated DSB repair genes, such as XRCC2, XRCC6, and RAD21, within close proximity (−0.5 kb to +1.1 kb; Supplementary Fig. S1G and Supplementary Table S8). Furthermore, 10 of 11 and 4 of 5 acquired-resistant genomes with CGR/ecDNA-associated BRAF and NRAS amplifications, respectively, harbored coamplification of superenhancers (Supplementary Table S8), which suggests that genomic plasticity may facilitate epigenomic plasticity and alter gene expression on a large scale in MAPKi-resistant tumors. In further support of ecDNAs playing a role in epigenomic plasticity, we observed that H3K27 acetylated genomic regions (extracted from ENCODE data) were rearranged to the proximity of and coamplified with ecDNA genes (e.g., c-MYC and BRAF, respectively) in acquired resistance (Supplementary Fig. S1H). Enhancer docking sites within ecDNA amplicons may also influence the differentiation state(s) of MAPKi-resistant tumors by acting in trans on intrachromosomal genes such as HOXA9, HOXA11, HOXA13, and LIFR (RAM12.01-Brain-DD-DP10, RAM12.01-Brain-DD-DP3, and RAM12.01-Brain-DD-DP9) and acting as mobile regulatory elements for genes such as SMO (Pt3-DP1; Supplementary Table S8) and ATP1A1 (Mel_PDX2-R2; Supplementary Fig. S1G). Thus, CGR and ecDNA amplicons coamplify MAPKi resistance–specific and –driver genes with their enhancers and may harbor additional enhancer or superenhancer activities in cis or trans that concomitantly reprogram the transcriptome of acquired resistance.

Chromothripsis, defined as a mutational phenomenon leading to extensive genomic rearrangements and extensive CN oscillations, drives cancer initiation and progression (12). Prior studies have colocalized chromothriptic regions with ecDNAs (20, 28–30), and a recent study suggested chromothripsis as a pathway for ecDNA formation in general (30) and specifically in ecDNA-driven methotrexate resistance (20). To evaluate the specific contributions of chromothripsis before and after evolution of MAPKi resistance, we analyzed WGS data (Supplementary Table S2) from 58 patient-matched, BRAF^V600MUT or NRAS^MUT MAPKi-sensitive/naive and acquired-resistant tumors [along with 19 patient-matched normal genomic DNAs (gDNA)]. We observed a high prevalence of chromothripsis (high-confidence calls based on recently published criteria; ref. 12) in MAPKi-sensitive (16 of 17; mean total span per genome, 464 Mb) and acquired-resistant (40 of 41; mean total span per genome, 476 Mb) melanomas. The chromosome distributions of chromothripsis overlapped but were distinct between sensitive and acquired-resistant genomes (Supplementary Fig. S2A and S2B). Chromothriptic genomic spans in MAPKi-sensitive/naive and acquired MAPKi-resistant melanomas overlap 76% and 36%, respectively, with those reported earlier for a subset of The Cancer Genome Atlas skin cutaneous melanoma (12), indicating again that MAPKi selects for chromothripsis in distinct genomic regions. Interestingly, the single MAPKi-sensitive melanoma without chromothripsis (Mel_PDX27-V1) still gave rise to two (of three) acquired-resistant tumors with chromothripsis. The recurrence and span sizes of chromothriptic events in our melanoma cohorts are higher than those reported for cutaneous melanoma in the Pan-Cancer Analysis of Whole Genomes study (mean total span per genome, 120 Mb; ref. 12). These differences may be due to advancing disease and therapy resistance.

If chromothripsis were a precursor step for ecDNA and CGR generation during melanoma disease progression, we expected a nonrandom overlap of affected genomic spans. Consistent with expectation, we observed that the genomic spans of ecDNAs and CGRs overlapped significantly or nonrandomly with chromothriptic regions in ∼29% (5 of 17) of MAPKi-sensitive/naive genomes and in ∼54% (22 of 41) of acquired MAPKi-resistant genomes (Fig. 3A). For every genome, we computed the sizes of ecDNAs + CGRs (a − x), chromothriptic regions (b − x), their overlap (x), and genomic regions devoid of these genomic alterations (g − a − b). These four quantities were represented in 2 × 2 contingency tables to carry out the Fisher exact test. Using hypergeometric distribution, we then estimated the probability of observing random overlaps between ecDNA + CGR spans and chromothriptic spans. In each tumor genome with overlaps between ecDNA + CGR and chromothriptic regions, we observed a significant nonrandom convergence (Fisher exact test, P < 0.00001), suggesting chromothripsis as an origin of ecDNAs and CGRs (Fig. 3A).

Chromothripsis generation is hypothesized to involve micronuclei, which expose entrapped chromosomes to mutagens and predispose them to replication, transcription, as well as DNA repair defects (31). Therefore, we identified and characterized the somatic SNVs within the chromothriptic regions across the melanoma genomes in our three cohorts. The numbers of chromothripsis-associated SNVs ranged from 31 to 2,526 per Mb (mean 154 mutations/Mb, median 96 mutations/Mb; Supplementary Table S2). We next tested the hypothesis that chromothripsis may disproportionally contribute to the tumor mutational burdens (TMB), especially in acquired-resistant genomes, by calculating the numbers of SNVs within chromothriptic regions relative to the numbers of SNVs within nonchromothriptic regions. Acquired-resistant tumors harbored an average of 195 SNVs/Mb in chromothriptic regions versus an average of 93 SNVs/Mb in nonchromothriptic regions. Because the ratio of the chromothriptic region (∼476 Mb) to nonchromothriptic region (∼2,619 Mb) is 0.18 and given the TMB of 170 SNVs/Mb in acquired resistance, we expected only 31 SNVs/Mb (0.18 × 170) in chromothriptic regions of acquired-resistant tumors if SNVs were evenly distributed across the genome. Moreover, MAPKi-sensitive or pretreatment tumors harbored an average of 121 SNVs/Mb in chromothriptic regions versus an average of 84 SNVs/Mb in nonchromothriptic regions. Again, an observed average of 121 SNVs/Mb in chromothriptic regions was higher than the expected 20 SNVs/Mb in the chromothripsis regions, which was calculated based on a 0.17 ratio of the chromothriptic region (∼464 Mb) to nonchromothriptic region (∼2631 Mb) and the observed average TMB of 115 SNVs/Mb in MAPKi-sensitive tumors. Hence, in both acquired-resistant and MAPKi-sensitive melanoma genomes, chromothriptic regions were enriched for SNVs compared with nonchromothriptic regions (Fisher exact test, P < 0.00001). Furthermore, the ratio of SNVs in chromothriptic to nonchromothriptic genomes was higher in acquired-resistant versus MAPKi-naive/sensitive tumors (Wilcoxon rank sum test, P = 0.005; Supplementary Fig. S2C), suggesting that MAPKi-selected chromothriptic regions in some individuals feature a relatively stronger mutator phenotype.

We have reported that MAPKi selection of clinical melanomas alters the mutational spectra (9). To characterize how a MAPKi-associated mutator phenotype affects chromothriptic genomic spans selected by MAPKi, we culled chromothripsis-associated somatic mutations unique to MAPKi-sensitive/naive (mean, 64 SNVs/Mb) versus MAPKi-resistant genomes (mean, 58 SNVs/Mb) and analyzed signatures of single-base substitutions (SBS; Fig. 3B), double-base substitutions (DBS), and IDs (Supplementary Fig. S2D; 32, 33). In both MAPKi-sensitive/naive and acquired-resistant tumors, we frequently detected signatures of tobacco smoking (SBS4) and ultraviolet radiation (SBS7a and SBS38). Among chromothripsis-associated somatic mutations unique to either MAPKi-sensitive/naive or acquired-resistant tumors, we also detected a signature of defective HRR (SBS3), and this trended higher in resistant (12 of 31) versus sensitive (3 of 16) tumors, although this difference was not significant. Unique but infrequent to acquired MAPKi resistance was the detection of mutational signatures of polymerase eta somatic hypermutation (SBS9) and defective POLD1 proofreading activity (SBS10d, SBS20). Unique and recurrent in acquired MAPKi resistance was the detection of mutational signatures of (i) defective DNA mismatch repair (MMR; SBS6, SBS20, SBS26, and SBS44; 14 of 31 resistant tumors; 10 of 16 patients) and (ii) defective base excision repair (BER) as well as damage due to reactive oxygen species (ROS) or mutations in NTHL1 and MUTYH (SBS18, SBS30, and SBS36). As exceptions, SBS26 and SBS36 signatures were also detected at a low percentage in two sensitive tumors (one of which had been exposed to MAPKi). BER/ROS signatures were observed more frequently in the BRAF^V600MUT (10 of 22 resistant tumors, 6 of 11 patients) than in the NRAS^MUT (1 of 9 resistant tumors, 1 of 5 patients) subset. This differential frequency may be due to the longer MAPKi exposure in the clinical (versus PDX or experimental) setting. Consistently, the BER mutational signature (SBS18) was detected in Mel_PDX3-R5, which was among the acquired-resistant PDX tumors with longer durations of MAPKi treatment (Supplementary Fig. S1A). Overall, resistance-specific (versus sensitivity-specific) chromothriptic somatic SBSs enriched for signatures of defects in BER (Wilcoxon rank sum test, P = 0.04) and MMR (Wilcoxon rank sum test, P = 0.005). Moreover, resistance-specific (versus sensitivity-specific) nonchromothriptic SBSs also enriched for signatures of defects in BER (Wilcoxon rank sum test, P = 0.05), MMR (Wilcoxon rank sum test, P = 0.008), as well as HRR (Wilcoxon rank sum test, P = 0.04; Supplementary Table S9). We did not identify any resistance-specific or -enriched DBS and ID signature (Supplementary Fig. S2D).

We also identified SBS signatures within ecDNA and CGR sequences and determined whether certain SBS signatures are enriched in acquired MAPKi-resistant (versus MAPKi-sensitive/naive) tumors or ecDNA/CGR amplicons (versus noninvolved genomic regions regardless of tumor sensitivity or resistance; Fig. 3C). We extracted somatic SNVs unique to MAPKi-sensitive/naive (n = 12) and acquired-resistant (n = 28) tumors (in a patient-matched fashion) within ecDNA/CGR amplicons. CGR + ecDNA-associated TMB trended higher in acquired-resistant (mean, 130 SNVs/Mb) compared with MAPKi-sensitive/naive (mean, 80 SNVs/Mb) tumors (Wilcoxon rank sum test, P = 0.06). Enrichment of SBS3 (defective HRR) was more recurrent in the ecDNAs/CGRs of resistant tumors (11 of 28) compared with patient-matched sensitive tumors (1 of 12), although this difference did not reach statistical significance (Wilcoxon rank sum test, P = 0.08; Fig. 3C). Moreover, we derived mutational signature enrichment scores by calculating the normalized ratios of signature proportions within ecDNAs/CGRs and within regions devoid of these events (scores >1 defined as positive enrichment; see Methods). We found that 25% (7 of 28) of acquired-resistant tumors displayed positive enrichment scores for defective HRR signatures (SBS3) within ecDNAs and CGRs, whereas only 8% (1 of 12) of sensitive tumors displayed positive enrichment (Wilcoxon rank sum test, P = 0.3; Fig. 3C). In either sensitive or resistant tumors, we observed positive enrichment scores for SBS signatures reflective of APOBEC cytidine deaminase activity (SBS2; 2 of 12 in sensitivity; 3 of 28 in resistance) and polymerase epsilon exonuclease domain mutations (SBS10b; 3 of 12 in sensitivity; 3 of 28 in resistance). Positive enrichment scores were also noted with lower recurrence for SBS signatures of defective POLD1 proofreading (SBS10c), defective MMR (SBS44), chemotherapy treatment (SBS86, SBS87), and tobacco smoking (SBS82). Overall, 50% (6 of 12) of sensitive/naive and 32% (9 of 28) of acquired-resistant tumors displayed enrichment of unique SBS signatures within ecDNA + CGR sequences compared with uninvolved genomic regions. Lastly, we addressed whether the enrichment of unique SBS signatures differed within ecDNAs versus CGRs in acquired resistance (Supplementary Fig. S2E). We observed positive enrichment scores of SBS2 (APOBEC cytidine deaminase activity) in 26% (5 of 19) of ecDNA⁺ acquired-resistant tumors compared with 4% (1 of 24) of CGR⁺ acquired-resistant tumor (Wilcoxon rank sum test, P = 0.05; Supplementary Fig. S2F). This finding is consistent with the recent discovery of the co-occurrence of APOBEC3 kataegis in ecDNAs (33). We also observed positive enrichment scores of SBS3 (defective HRR) in 47% (9 of 19) of ecDNA⁺ acquired-resistant tumors compared with 17% (4 of 24) of CGR⁺ acquired-resistant tumors (Wilcoxon rank sum test, P = 0.07; Supplementary Fig. S2F). In short, ecDNA and CGR amplicons, whether in MAPKi-sensitive/naive or acquired-resistant tumors, harbor SBS patterns suggestive of specific defects or deficiencies in HRR. In acquired MAPKi resistance, APOBEC activity likely contributes to ecDNA mutagenesis and hence shapes the mutanome of ecDNA amplicons. Further analysis might lead to insights into ecDNA biogenesis and suppressive strategies.

We analyzed the breakpoint–junctional sequences of CGRs and ecDNAs to infer DSB repair processes underlying resistance-associated amplicons (Fig. 3D). Alternative end-joining refers to mechanisms of DSB repair that may compensate for HRR- and NHEJ-based repairs and comprises single-strand annealing (SSA), microhomology-mediated end-joining (MMEJ), and other end-joining pathways (34). SSA is indicated in the breakpoint junctions by complementary repeat sequences >25 nucleotides and MMEJ by shorter tracks of sequence homology (2–20 nucleotides). Moreover, replicative processes, such as fork stalling and template switching as well as MMEJ, can contribute to the generation of CGRs. Breakpoint junctions derived from replicative processes (e.g., replicative fork collapsing when it encounters a nick) are expected to have microhomologies, insertions, and relatively long templated insertions (35, 36). Indeed, signatures of replication processes and templated insertion, as well as that of NHEJ, were detected pan-cancer (12). Analysis of breakpoint–junctional sequences of all resistance- and sensitivity-associated CGRs and ecDNAs inferred NHEJ as the main mechanism of double-stranded DNA fragment joining or rearrangement (Fig. 3D; Supplementary Fig. S3A). A lower number of breakpoint–junctional sequences displayed short and long homologous sequences as well as insertions (>10 bp), which suggested DSB repair by alternative NHEJ (alt-NHEJ) and HRR, respectively (Fig. 3D). Analysis of breakpoint–junctional sequences of specifically ecDNAs and CGRs harboring MAPK-reactivation or MAPKi resistance–driver genes also revealed a similar pattern, with NHEJ dominating the landscape of DSB repair mechanisms (Fig. 3E). Of unknown significance or etiology, a higher incidence of double-stranded DNA ligation by HRR was detected in NRAS^MUT melanoma, either sensitive or acquired-resistant to MAPKi, compared with BRAF^V600MUT melanoma (Supplementary Fig. S3B). In acquired-resistant genomes, we also detected a higher incidence of double-stranded DNA ligation by NHEJ for CGRs compared with ecDNAs (Supplementary Fig. S3C).

DNA-dependent protein kinase catalytic subunit (DNA-PK_CS) and PARP1/2 are involved in multiple DSB repair pathways, particularly NHEJ (DNA-PK_CS), HRR (DNA-PK_CS, PARP1/2), and MMEJ (PARP1/2; refs. 37, 38). Thus, we tested the activity of the specific DNA-PK_CS inhibitor (DNA-PKi) NU7026 and PARP1/2 inhibitor (PARPi) ABT888 (veliparib), individually and combinatorially, in preventing the clonal emergence of drug-tolerant proliferating persisters (DTPP; refs. 39, 40) from human BRAF^V600MUT (n = 3; M229, M249, M395) or NRAS^Q61MUT (n = 3; M202, M207, M245; Fig. 4) melanoma cell lines chronically treated with BRAFi + MEKi or MEKi, respectively. We hypothesized that DNA-PKi interferes with MAPKi-elicited, de novo rearrangement of specific SVs including ecDNAs and CGRs. We further hypothesized that DNA-PKi is more effective at preventing rather than reversing resistance once fully established by chronic MAPKi treatment. Hence, we tested treatments with NU7026 and/or ABT888 in combination with MAPKi in acquired MAPKi-resistant sublines (n = 9) that are isogenic to the parental human BRAF^V600MUT and NRAS^Q61MUT cell lines (Fig. 4A–O). BRAF^V600MUT sublines with acquired-resistance to BRAFi + MEKi are annotated as double-drug resistant (DDR), whereas NRAS^MUT sublines with acquired-resistance to MEKi are annotated simply by their clone (C) numbers. We began by testing the single-agent anticlonogenic growth activity of DNA-PKi and PARPi on parental melanoma (and nonmelanoma, see below) cell lines without MAPKi treatment in order to select noninhibitory concentrations to test in combination with MAPKi (Supplementary Fig. S4A–S4G). In combination with MAPKi, NU7026 synergistically and dose dependently prevented DTPP formation in all parental cell lines tested (Fig. 4A, D, G, I, J, and N), whereas ABT888 displayed activity in 3 of 6 parental cell lines (Fig. 4A, D, and J). In cases in which ABT888 individually was active in preventing DTPP formation, NU7026 plus ABT888 led to even greater suppression of acquired MAPKi resistance (Fig. 4A, D, and J). Consistently in all acquired MAPKi-resistant sublines, NU7026 and ABT888 at the lower range of the concentrations tested displayed no or reduced anticlonogenic activity compared with their activities observed in isogenic parental lines (Fig. 4A–O). More recent generations of DNA-PKi display improved selectivity and potency (37, 41). Hence, we used next-generation DNA-PKi (VX984 and AZD7648) and PARPi (olaparib) and corroborated their combinatorial efficacy with MAPKi in parental cell lines (Supplementary Fig. S4C–S4E and Supplementary Fig. S5A–S5F). In M245, we further corroborated the pharmacologic findings with genetic studies. Using independent short hairpin RNAs (shRNA) against two critical NHEJ genes (PRKDC, which encodes DNA-PK_CS, and DNA ligase IV, or LIG4), we showed that their protein knockdown, while not having a significant effect on growth without MAPKi, strongly suppressed the frequency of DTPP emergence (Fig. 4P and Q).

The acquired-resistant sublines used in the prior experiments have been adapted chronically (i.e., months to years) to MAPKi. Hence, we also tested whether a shorter delay (i.e., weeks) in the cotreatment of parental cell lines with DNA-PKi (and/or PARPi) would hinder the efficacy in preventing DTPP formation. In both cell lines (M229 BRAF^V600MUT and M245 NRAS^MUT), we found that a 3- to 3.5-week delay in cotreatment with DNA-PKi and/or PARPi reduced the suppression of resistance (Fig. 4R and S compared with Fig. 4A and J, respectively). In the same parental cell lines, we then compared cotreatment with DNA-PKi (and/or PARPi) during the first versus second half of the overall MAPKi treatment course (Fig. 4T and U). Notably, we observed that cotreatment with DNA-PKi, PARPi, or DNA-PKi + PARPi during the first half of the MAPKi treatment course was remittive and during the second half inferior (compared with continuous cotreatment). Thus, the upfront and initial phase of MAPKi therapy appears to be the key window of opportunity for cotreatment with DNA-PKi (and/or PARPi) to suppress acquired MAPKi resistance, suggesting a preventive mechanism of action.

To support further a causal link between acquired MAPKi resistance and ecDNAs or homogeneously staining regions (HSR) in human melanoma cell lines, we first sought to determine the modes of amplification of bona fide resistance-driver genes. In three parental cell lines (BRAF^V600MUT, M249 and M395; NRAS^MUT, M245) and five isogenic acquired MAPKi-resistant sublines, we performed DNA-FISH on cells in metaphase and probed against BRAF (amplification of which drives acquired MAPKi resistance in BRAF^V600MUT melanoma; refs. 5, 7–9), RAF1, and NRAS (amplification of which drives acquired MAPKi resistance in NRAS^MUT melanoma; refs. 11, 22), as well as centromeres of chr7, chr3, and chr1, respectively (Fig. 5A–C). We observed in acquired resistance that BRAF is amplified either as HSRs (M249 DDR4 and DDR5) or as ecDNAs (M395 DDR), RAF1 as a mixture of HSRs and ecDNAs (M245 C3), and NRAS as HSRs (M245 C5). Expectedly, these driver ecDNAs and HSRs were not detected in the isogenic parental cell lines (Fig. 5A–C). We then tested whether MAPKi withdrawal would select against resistant clones with ecDNAs and/or HSRs that amplify the aforementioned resistance-driver genes and, consequently, restore at least partial MAPKi sensitivity. Importantly, we observed that MAPKi withdrawal from acquired-resistant sublines significantly reduced metaphase cells with driver ecDNAs and HSRs (Fig. 5A–C) and significantly enhanced their MAPKi sensitivity (Fig. 5D–F). Using the M249 parental cell line, which is highly sensitive to resistance suppression by either DNA-PKi or PARPi (Fig. 4D), we tracked the emergence of BRAF HSR early on treatment (day 31) with BRAFi + MEKi or combinations with DNA-PKi, PARPi, or DNA-PKi + PARPi (Fig. 5G). Consistent with BRAF HSRs as a driver of acquired MAPKi resistance, we observed their induction by BRAFi + MEKi but suppression by combinatorial treatments with DNA-PKi and/or PARPi. During the earliest window on MAPKi treatment, expansion of ecDNA- and CGR-involved genomic spans may be critical for specific resistance-driving ecDNAs and CGRs to emerge later. We reasoned that DNA-PKi, which is generally more effective than PARPi in preventing acquired resistance (Fig. 4), should suppress MAPKi-elicited ecDNA + CGR genomic spans. To test this hypothesis, we performed WGS on two parental cell lines (M229 and M245) treated with vehicle, MAPKi, or MAPKi + DNA-PKi for a short duration (10–11 days); identified ecDNAs and CGRs; and calculated treatment-specific genomic spans of ecDNAs + CGRs (by filtering out ecDNA and CGR spans detected in vehicle-treated genomes; Supplementary Table S10). Consistent with our hypothesis, DNA-PKi cotreatment with MAPKi blunted the expansion of ecDNA + CGR genomic spans in both cell lines (Fig. 5H). Compared with treatment with MEKi alone, M245 cells treated with MEKi + DNA-PKi gained de novo ecDNAs (0.06 Mb in chr3) and lost CGRs (0.4 Mb in chr1 and chr2; Fig. 5H).

As DNA-PKi plus MAPKi may constitute an effective combination for BRAF^V600MUT and NRAS^MUT melanoma, we explored this combinatorial efficacy (± PARPi) in human pancreatic ductal adenocarcinoma (PDAC) and non–small cell lung carcinoma (NSCLC) cell lines driven by KRAS^G12C. For the MAPKi regimen, we used MEKi (trametinib), KRAS^G12C inhibitor (KRAS^G12Ci; AMG510 or MRTX849) + MEKi, or type II RAF inhibitor (RAFi; BGB-283) + MEKi. As with melanoma cell lines, we first identified concentrations of DNA-PKi (NU7026) or PARPi (ABT888) that did not affect the clonogenic growth of KRAS^G12C PDAC and NSCLC cell lines without MAPKi (Supplementary Fig. S4F and S4G). We then assayed for acquired MAPKi-resistant growth in these cell lines, without or with DNA-PKi and/or PARPi (Fig. 6). Interestingly, we observed that DNA-PKi effectively suppressed acquired MAPKi resistance in two of two KRAS^G12C PDAC cell lines (Fig. 6A–F) and one of two KRAS^G12C NSCLC cell lines (Fig. 6G–J). Moreover, PARPi was generally ineffective as a combinatorial agent with MAPKi. Future studies should explore the contributions of chromothripsis, ecDNAs, and CGRs to acquired MAPKi resistance in human KRAS^G12C PDAC and NSCLC.

Given the broader resistance-preventive activity of DNA-PKi, we tested NU7026's in vivo efficacy in five cutaneous melanoma PDXs. In Mel_PDX16 (BRAF^V600MUT) melanoma, using well-established tumors (∼500–700 mm³; Fig. 7A), BRAFi + MEKi (vemurafenib at 90 mg/kg/d, trametinib at 0.7 mg/kg/d) elicited only tumor growth inhibition, and DNA-PKi treatment (8 mg/kg/d) had no discernible effect on tumor growth compared with vehicle-treated tumors. In contrast, the triplet of BRAFi + MEKi + DNA-PKi elicited tumor regression transiently for ∼14 days until the tumors acquired resistance. There were no overt signs of toxicities or significant reductions in body weight in any experimental group. In Mel_PDX27 (BRAF^V600MUT) melanoma (Fig. 7B; Supplementary Fig. S6A), BRAFi + MEKi (vemurafenib at 90 mg/kg/day, trametinib at 0.7 mg/kg/day) elicited transient tumor regression, whereas DNA-PKi treatment (8 mg/kg/day) did not elicit tumor regression compared with vehicle-treated tumors. In contrast, the triplet of BRAFi + MEKi + DNA-PKi significantly forestalled acquired resistance, without incurring overt signs of toxicities or significant reductions in body weight. In NRAS^MUT melanoma (Mel_PDX1; Fig. 7C), DNA-PKi treatment (10 mg/kg/d) did not elicit tumor regression, and MEKi (trametinib, 3 mg/kg/d) elicited only transient tumor regression. In contrast, the combination of MEKi + DNA-PKi led to minimally palpable tumors, without incurring toxicity. In two additional NRAS^MUT melanoma PDXs (Mel_PDX2 and Mel_PDX4), we consistently observed superior efficacy of MEKi + DNA-PKi over MEKi alone (Fig. 7D and E). In Mel_PDX4, we initiated dosing with both DNA-PKi at 8 mg/kg/day and MEKi at 3 mg/kg/day but stopped dosing DNA-PKi (in combination with MEKi) on day 67 (Fig. 7E). With longer follow-up (Supplementary Fig. S6B), discontinuation of DNA-PKi dosing in this group of mice treated initially with MEKi + DNA-PKi did not lead to tumor relapse. Based on cell line findings (Fig. 4), we expected that delayed DNA-PKi cotreatment in vivo would also diminish its preventive mechanism of action. In Mel_PDX27 (BRAF^V600MUT) melanoma (Fig. 7B; Supplementary Fig. S6A), we repeated BRAFi + MEKi (vemurafenib at 90 mg/kg/day, trametinib at 0.7 mg/kg/day) treatment, and, at the time of disease progression (when the mean tumor volume returned to the mean pre-MAPKi volume of ∼500 mm³), we divided the tumors/mice into two groups. The first group continued on BRAFi + MEKi treatment at the same dosages, whereas the second group received DNA-PKi treatment (8 mg/kg/d) added on top of BRAFi + MEKi. Consistent with the cell line results (Fig. 4A–O), DNA-PKi cotreatment did not elicit discernible tumor regression on tumors that had already acquired MAPKi resistance in vivo (Fig. 7F).

We then explored early on-treatment pharmacodynamic markers that are associated with the superiority of MAPKi + DNA-PKi over MAPKi alone in forestalling or preventing acquired resistance. We posited that the combination of DNA-PKi with MAPKi (versus MAPKi alone) would suppress the total size of ecDNAs and CGRs specifically generated due to MAPKi treatment, agnostic of content genes, regulatory elements, and known or putative roles in driving resistance. Using all five PDX models, we collected early on-treatment tumors and vehicle-treated tumors (Supplementary Fig. S6C and S6D). Expectedly, p-ERK was suppressed strongly in both groups of tumors treated with MAPKi or MAPKi + DNA-PKi (Supplementary Fig. S6E). Binding of the DNA-PK holoenzyme (DNA-PK_CS + Ku70 + Ku80) to DSBs elicits autophosphorylation of DNA-PK_CS on serine 2056. We found that DNA-PKi treatment, with or without MAPKi, suppressed nuclear p-DNA-PK_CS foci (Supplementary Fig. S6F), which is consistent with NU7026 being able to suppress the kinase activity and autophosphorylation of the activated DNA-PK holoenzyme. We found that short-term MEKi treatment in Mel_PDX1 induced γH2AX nuclear foci in >50% of tumor cells (Supplementary Fig. S6G), consistent with a recent report that found MAPKi treatment resulting in γH2AX nuclear foci or DSBs in melanoma cell lines (42). Expectedly, DNA-PKi cotreatment with MEKi further induced DSBs marked by γH2AX foci. We then generated WGS data to enable an analysis of CGR and ecDNA amplicons (Supplementary Table S11). CGR and ecDNA amplicons identified in vehicle-treated PDX tumors were considered as background and were removed from those identified in MAPKi- or MAPKi + DNA-PKi–treated tumors. Consistent with our hypothesis, MAPKi + DNA-PKi (versus MAPKi) treatment was associated with reductions in the average total genomic spans of CGRs and ecDNAs in five of five PDX models analyzed (Fig. 7G). We also analyzed the breakpoint–junctional sequences of MAPKi treatment–specific ecDNAs and CGRs to infer the relative contributions of DSB repair pathways (Supplementary Fig. S6H). Consistent with the functional role of DNA-PKi, its combination with MAPKi suppressed the contribution of NHEJ, with potentially compensatory increases in alt-NHEJ or HRR processes.

Preexisting tumor heterogeneity and de novo diversification in response to targeted therapy are thought to fuel the evolution of acquired resistance. Here, we provide insights into the underlying genomic instability processes that generate preexisting and therapy-elicited clonal diversification in advanced cutaneous melanoma. We identified chromothripsis as well as ecDNAs and CGRs as highly recurrent and pervasive genomic SVs in MAPKi-naive/sensitive and acquired MAPKi-resistant melanoma in both the clinical setting (BRAF^V600MUT melanoma, in which MAPKi therapy is a standard-of-care therapy) and the experimental setting (NRAS^MUT melanoma, in which there is a lack of targeted therapy options). CGRs can derive from reintegration of ecDNAs or breakage–fusion–bridge (BFB) cycles (20, 43). We did not detect any BFB event in any of our tumors, MAPKi-sensitive/naive or acquired-resistant, favoring ecDNA reintegration as the main route of CGR generation in the setting of metastatic cutaneous melanoma. Consistently, the acute stress of MAPKi therapy favors ecDNAs, whereas stable or chronic stress favors reintegration of ecDNAs into chromosomes as HSRs (13). Moreover, following chromothripsis, chimeric circularization of DNA and reintegration of DNA circles into chromosomes constitute a major source of SVs and linear genome mutagenesis (44, 45). Importantly, the selective pressure of MAPKi therapy is evidenced by the amplification of bona fide resistance-driver genes via ecDNAs and CGRs. The report here of resistance-specific ecDNAs and CGRs amplifying a wide array of coding and noncoding sequences warrants future investigations into their resistance-causative mechanisms.

Even though chromothripsis is regarded as a potential precursor of ecDNAs, it creates oscillating CNs of genomic segments but does not cause high-level amplifications. We found that, within each tumor, ecDNAs (mean total size per genome, 7 Mb; mean size per ecDNA, 343 kb) and their reintegrated CGR counterparts (mean total size per genome, 6 Mb; mean size per CGR, 598 kb) almost always span genomic regions bounded by larger chromothriptic regions (mean total size per genome, 474 Mb; mean size per chromothriptic region, 120 Mb), which is consistent with a chromothriptic origin of ecDNA and CGR amplicons. Chromothripsis can occur as a result of micronuclei formation around lagging chromosomes or chromosome bridge formation due to telomere crisis (46–48). Both aberrant processes are associated with a loss of primary or micronuclear membrane integrity and subsequent mutagenesis. This is consistent with our finding of enhanced mutational density within chromothriptic genomic regions, especially within acquired-resistant genomes, as well as a resistance-specific mutator phenotype enriched for signatures of excessive single-stranded DNA damage and/or deficient repair (BER and MMR). Intriguingly, a recent study proposed that a specific BER defect may predispose micronuclei-associated or cytoplasmic chromosomes to breakage, a key step toward chromothripsis (49).

Translationally, we produced in vivo evidence supportive of the well-studied role of DNA-PK_CS in NHEJ as critical in promoting the total segment sizes of ecDNAs and CGRs generated early on MAPKi therapy. This finding is consistent with prior literature supporting NHEJ as key to ecDNA formation (28, 50). Selection for numeric expansion of specific ecDNAs and CGRs is considered a direct mechanism promoting tumor fitness in response to a given stressor. Therefore, reduction of the total ecDNA/CGR genomic spans early on DNA-PKi cotreatment should serve to reduce the reservoir of ecDNAs/CGRs available for natural selection by MAPKi therapy.

DNA-PK_CS, the target of DNA-PKi, subserves other less-characterized cancer survival pathways (37). It is possible that another beneficial mechanism of action of DNA-PKi coexists. DNA-PKi has been proposed in combination with agents that directly induce DNA damage, such as radiotherapy or chemotherapy, with the intent of radio- or chemosensitization (37). The rationale is based on catastrophic DSBs that would result in excessive DNA damage repair stress and hence the synergistic induction of death in cancer cells, especially those overexpressing DNA-PK_CS. Here, we rationalize the combination of DNA-PKi and MAPKi based on dual concepts. First, the rapid induction of genomic instability mechanisms, in particular the generation of ecDNA and CGR amplicons, is critical for genomic diversification and perhaps epigenomic reprogramming necessary for melanomas to adapt quickly to MAPKi therapy. In this rationale, DNA-PKi suppresses NHEJ, which is necessary for the efficient formation of ecDNAs and CGRs. Second, MAPKi is potentially an inducer of DNA damage and/or DNA damage repair deficiency. In this context, MAPKi has been shown to induce DNA damage in early drug-tolerant persister subpopulations (42). A therapy-induced oxidative metabolic adaptation (51) has been proposed to cause ROS-induced mutagenesis, which can be repaired by DNA single-stranded break (SSB) repair processes such as BER and MMR. DNA-PK_CS, in addition to DSB repair, can also bind to and is activated by DNA SSBs (52). Excessive DNA SSBs can be converted into DNA DSBs, engendering DNA damage repair stress and/or chromosome breakage. The latter may be a pathway to chromothripsis and ecDNA/CGR generation (49).

Our demonstration of ecDNAs and CGRs driving acquired MAPKi resistance advances the concept that multiple resistance mechanisms, genetic and epigenetic, as well as direct (drug–target or MAPK pathway reactivation) and indirect (nondrug–target pathway activation), are simultaneously causal of clinically acquired resistance. Future work needs to dissect this hybrid genomic–epigenomic model with ecDNAs and CGRs at the center of therapeutic targeting efforts. Findings here also advance the concepts that preventing, instead of reversing, acquired resistant phenotypes may be more impactful clinically and that targeting DNA-PK_CS and NHEJ, and potentially MMEJ and HRR, lies at the center of this approach in stabilizing cancer genomes during oncogene-targeted therapies. Finally, prevention of chromothripsis, such as by reducing chromosome fusion and missegregation events and minimizing primary and micronuclei membrane ruptures, may expand our molecular armamentarium against targeted therapy resistance.

Patient characteristics related to clinical tissues are presented in Supplementary Table S1. Patient-derived tissues were obtained with written informed consent and approval by local institutional review boards.

NSG (NOD scid gamma) mice were obtained from the Radiation Oncology breeding colony at University of California, Los Angeles (UCLA). Male or female mice were used at 4 to 6 weeks of age. All animal experiments were conducted according to the guidelines approved by the UCLA Animal Research Committee.

To develop PDX models, tumor fragments derived from metastatic melanoma, with approval by the local institutional review boards, were transplanted subcutaneously in sex-matched NSG mice (4–6 weeks old). One tumor fragment was implanted in each mouse. Tumors were measured with a caliper every 2 days, and tumor volumes were calculated using the formula (length × width²)/2. Tumors with volumes around 500 mm³ were randomly assigned into experimental groups. Special mouse diets were generated to reduce stress to animals by incorporating trametinib (LC Laboratories) into chow to achieve daily trametinib dosing at 3 or 5 mg/kg/day or combined vemurafenib (LC Laboratories) and trametinib dosing at 90 mg/kg/day and 0.7 mg/kg/day, respectively (Test Diet). DNA-PKi (NU7026, Selleckchem) was dissolved in saline (UCLA Division of Laboratory Animal Medicine pharmacy) and administered intraperitoneally at 6, 8, or 10 mg/kg/day. We derived model- or patient-matched vehicle-treated tumors and acquired trametinib-resistant tumors as plotted in Supplementary Fig. S1A. We also derived model-matched vehicle-treated tumors and early on-treatment tumors as plotted in Supplementary Fig. S6C and S6D.

gDNA and total RNA were extracted from frozen tumor tissue preserved in RNALater or snap-frozen tumor tissue using the Qiagen AllPrep DNA/RNA Mini Kit and the Ambion mirVana miRNA Isolation Kit. Normal gDNA from peripheral blood mononuclear cells (PBMC) were extracted from fresh or frozen PBMCs using the Qiagen FlexiGene DNA Kit. All gDNA were quantified using a NanoDrop (Thermo Fisher Scientific) and Qubit fluorometer using the dsDNA BR Assay (Life Technologies), and then gDNA size and quality were tested using TapeStation (Agilent) to ensure gDNA libraries were prepared using equal gDNA input and presence of a high-molecular-weight band. Whole-genome libraries were prepared using the Roche KAPA HyperPrep Kit. Briefly, after enzymatic fragmentation of gDNA, the libraries were constructed by end repairing and A-tailing the fragmented DNAs, ligation of adapters, and PCR amplification. After library construction, indexed libraries were quantified for equal molar pooling and paired-end sequenced with a read length of 2 × 150 bp on the Illumina NovaSeq 6000 S4 platform.

Whole-genome libraries were prepared for patient-matched normal tissues and tumors as follows: (i) 10 BRAF^V600MUT clinical patients (baseline tumors, n = 10; resistant tumors, n = 17), (ii) three BRAF^MUT RAM patients (sensitive tumor, n = 1; resistant tumors, n = 12), and (iii) six PDX models. PDX models included one BRAF^MUT PDX (vehicle-treated tumor, n = 1; resistant tumors n = 3) and five NRAS^MUT PDXs (vehicle-treated tumors, n = 5; resistant tumors, n = 9). For early on-treatment samples, we prepared whole-genome libraries from (i) one human BRAF^V600MUT (M229; DMSO, n = 1; BRAFi + MEKi, n = 1; BRAFi + MEKi + DNA-PKi, n = 1) and one NRAS^MUT (M245; DMSO, n = 1; MEKi, n = 1; MEKi + DNA-PKi, n = 1) cutaneous melanoma cell line and (ii) two BRAF^V600MUT (vehicle, n = 6; BRAFi + MEKi, n = 6; BRAFi + MEKi + DNA-PKi, n = 9) and three NRAS^MUT (vehicle, n = 6; MEKi, n = 6; MEKi + DNA-PKi, n = 7) cutaneous melanoma PDX models. Also, whole-genome libraries were prepared for vehicle-treated tumors and early on-treatment tumors from two NRAS^MUT PDX models. In total, paired-end sequencing was performed on 123 (104 tumors or cell lines, 19 patient-matched normal tissues) genomes using Illumina NovaSeq S4 with a read length of 2 × 150 bp and at a sequencing depth of 11 to 98× (median, 26×).

WGS reads were mapped to the GRCh38/hg38 human reference genome using BWA-MEM (53). Alignments were sorted, and PCR duplicates were removed using Samtools (54). CN variations were called using two depth-of-coverage–based methods CNVkit (55) and ReadDepth (56). Default parameters were used for CNVkit. We used an FDR of 0.05 and overdispersion of 1 for ReadDepth analysis. Structural variations reported by at least two SV detection methods: SvABA (57), TIDDIT (58), and DELLY (59) were considered. SVs in both DELLY and TIDDIT are determined by combining discordant read pairs and split-reads, whereas TIDDIT additionally uses depth-of-coverage signatures. SvABA utilizes discordant reads and genome-wide local assembly strategies for predicting SVs from the genome. For high-coverage data (>15×), default parameters were considered for SvABA, TIDDIT, and DELLY. Parameter minimum number of points (-l) was set to 5 in TIDDIT for low-coverage data.

We carried out reconstruction of focal gene amplifications and elucidation of ecDNA and CGRs using AmpliconArchitect (60). Briefly, AmpliconArchitect determines the list of potential intervals for each amplicon to be reconstructed, and within each amplicon, CNs and SVs are estimated using read depth and discordant read signatures. It then constructs breakpoint graphs detailing sequence edges and breakpoint edges, and predicts copy counts of all edges. Simple cycles are then decomposed from the breakpoint graphs. Finally, AmpliconClassifier classifies the amplicons into ecDNAs, CGRs, and linear amplicons. Amplicons are classified as ecDNAs if the segment(s) forms a head-to-tail structure, with size >10 kb, and CN >4.5; as complex genomic rearrangements for noncircular amplicons containing DNA segments from different chromosomes or regions that are far apart (>1 Mb) on chromosomes; or as linear amplicons for linear amplifications. In our study, the initial set of CNV seed regions was inferred by ReadDepth and CNVKit. SV view of amplicons and circle plots of simple cycles were generated using functions available in AmpliconArchitect.

We carried out pathway enrichment analysis of genes within ecDNAs and/or CGRs using the Molecular Signatures Database with pathways listed in Gene Ontology, Kyoto Encyclopedia of Genes and Genomes, Reactome, and Pathway Interaction Databases. For each tumor sample, we identified somatic SNVs using Strelka2 (61) with default parameters. Next, we estimated SBS signatures in regions within and without ecDNAs and CGRs using the nonnegative matrix factorization–based tool MutationalPatterns (62) with COSMIC SBS signatures V3.3 as reference. We carried out mutational signature analysis using SNVs in ecDNAs and/or CGR regions of (i) acquired MAPKi-resistant (mutations in patient-matched sensitive tumors subtracted from mutations in resistant tumors) and (ii) MAPKi-sensitive genomes. To characterize whether SBS mutational signatures are preferentially detected within (versus outside of) ecDNAs and/or CGRs amplicons, we calculated the ratios of the proportions of signatures within ecDNA and/or CGR sequences and the proportion of signatures outside ecDNAs and/or CGR sequences and defined these ratios as SBS mutational signature enrichment score. We considered an enrichment score of 1 as the cutoff threshold. Scores >1 indicate SBS signatures enriched within ecDNAs and/or CGRs, whereas scores <1 indicate SBS signatures enriched in the background (non-ecDNA– and/or non-CGR–involved genomic regions). For clarity, only scores >1 and <1 were plotted.

We extracted enhancer elements and their connected genes from GeneHancer (version4-4, GeneCards). Genehancer is an integrated database of human enhancers and their inferred target genes, mined from four different genome-wide databases: ENCODE, FANTOM, the VISTA enhancer browser, and the Ensembl regulatory build. The enhancer–target genes associations were obtained from eQTLs, CHi-C, eRNA coexpression, transcription factor coexpression, and gene–enhancer distance methods. Annotations of superenhancers were obtained from dbSUPER (63). Enhancers were first assessed for their presence on CGR and ecDNA amplicons, and, if so, their connected oncogenes were searched for their presence within CGR/ecDNA amplicons (cis interaction) or elsewhere in the chromosome (trans interaction). Furthermore, we obtained H3K27 acetylation peaks from seven cell lines (GM12878, H1-hESC, HSMM, HUVEC, K562, NHEK, and NHLF) listed in ENCODE. The peaks aligned to ecDNA regions in Pt9-DD-DP2 (MYC) and Mel_PDX27-R2 (BRAF) were obtained from the UCSC Genome Browser.

Cells in metaphase were prepared by colcemid (Sigma) treatment at 10 μg/mL for several hours (depending on the growth rate). The single-cell suspensions were collected, washed with PBS, and then treated with 0.075 M KCl (Gibco) for 15 minutes. We then fixed and washed cells with 3:1 methanol:acetic acid. The final cell pellet was resuspended with the fixative and dropped onto a humidified slide.

Formalin-fixed, paraffin-embedded (FFPE) tissue samples were baked at 90°C for 25 minutes in an oven and immersed in 100% xylene and then 100% ethanol, each for 10 minutes to deparaffinize tissues. Air-dried tumor tissues were pretreated in 90°C to 95°C 10 mmol/L in citric acid buffer (pH 6.8, Thermo Fisher Scientific, 327162500) for 30 minutes and washed in 2× SSC buffer (Invitrogen, 15557044) for 5 minutes. Then, FFPE slides were digested in 37°C pepsin solution (Thermo Fisher Scientific, J6167906) for 20 to 30 minutes, washed in 2× SSC buffer for 5 minutes, and dehydrated in ascending ethanol series (70%, 85%, and 100%), each for 2 minutes. For metaphase DNA-FISH, fixed cells in interphase or metaphase on slides were dehydrated in ascending ethanol series (70%, 85%, and 100%), each for 2 minutes. We used the following DNA-FISH probes: NRAS/CEN1 amplification probe (NRAS-CHR01-20-ORGR) targeting NRAS or centromeric region of chr1; BRAF/CEN7 amplification probe (BRAF-CHR07-20-ORGR) targeting BRAF or centromeric region of chr7; and RAF1/CEN3 amplification probe (RAF1-CHR03-20-ORGR) targeting RAF1 or centromeric region of chr3, all from Empire Genomics. The probes were mixed with the provided hybridization buffer in a 1:4 ratio and applied onto the tissues or cells. FFPE samples were then denatured at 75°C in a slide moat for 7 minutes, whereas metaphase samples were denatured at 73°C for 2 minutes. We then performed hybridization overnight at 37°C in a humidified chamber. The samples were then washed at 73°C with 0.3% Igepal/0.4× SSC for 2 minutes, followed by another 2-minute wash with 0.1% Igepal/2× SSC at room temperature. Finally, the tissue samples were stained with ProLong Diamond Antifade Mountant with DAPI (Invitrogen, P36966) and covered by coverslips. Images were acquired on a Leica Confocal SP8-STED/FLIM/FCS microscope. For ecDNA quantification, we counted the number of specific gene foci in each nucleus by ImageJ (version, 1.52a). For HSR quantification, we counted the area of gene foci larger than 2 μm² to exclude ecDNA and non-HSR gene foci in each nucleus by ImageJ (version, 1.52a).

Total RNAs were extracted from frozen tumor tissue preserved in RNALater of snap-frozen tumor tissues using the Qiagen AllPrep DNA/RNA Mini Kit and the Ambion mirVana miRNA Isolation Kit. Total RNAs were quantified by the Qubit RNA High Sensitivity kit (Thermo Fisher Scientific) and/or using a NanoDrop (Thermo Fisher Scientific). RNA size and quality were measured using an Agilent 2100 Bioanalyzer (Agilent Technologies). RNA libraries were constructed using the NuGen Universal Plus mRNA-Seq Kit. Briefly, after fragmentation of total RNA and double-stranded cDNA generation using a mixture of random and oligo(dT) primers, the RNA libraries were constructed by end repairing, adapter ligation, strand selection, and PCR amplification. Libraries were quantified for equal molar pooling and paired-end sequenced with a read length of 2 × 50 bp on the Illumina NovaSeq 6000 S4 platform.

We mapped paired-end reads of patient-matched tumor samples from 10 patients and 6 PDX models to the GRChr38 human reference genome using STAR aligner (64) with default parameters. Log₂ fold change values were calculated for each acquired resistant tumor compared with patient- or model-matched pretreatment tumor or vehicle-treated tumor, respectively. Genes with an absolute fold change >2 were considered significant. Gene counts were estimated using FeatureCounts (65) with GENCODE (version 38) annotations. Feature counts were then normalized using trimmed mean of M values (TMM) approach and log₂ transformed (66). Fold changes for each gene were estimated by computing the difference between the normalized values of acquired resistant tumors and MAPKi-sensitive/naive transcriptomes.

To infer chromothriptic events in the genomes, we applied ShatterSeek (12), an in silico scoring algorithm based on the clusters of interleaved SVs, CN oscillations, and interchromosomal SVs. CN and SVs described above were considered as input for ShatterSeek. High-confidence chromothriptic events were selected based on the statistical criteria recommended by the authors. Briefly, a high-confidence chromothriptic event is characterized with (i) a cluster of SVs (>6 DUP/DEL/h2hINV/t2tINV), (ii) oscillating CN between two states (>7 CN events), (iii) chromosomal enrichment and distribution of DNA breakpoints (P < 0.05), (iv) randomness of fragment joins (P > 0.05), and/or (v) interchromosomal rearrangements between multiple chromosomes. Somatic SNVs spanning chromothripsis regions were identified using Strelka2 (61) with default parameters and defined as chromothripsis-associated SNVs. Regions with and without chromothripsis were extracted and TMBs were computed within these regions for each tumor sample. Ratios between TMBs within and outside of chromothriptic regions were calculated. Mutational signature analysis was carried out using SNVs in chromothriptic regions of (i) acquired MAPKi-resistant (mutations in patient-matched sensitive tumors subtracted from mutations in resistant tumors) and (ii) MAPKI-sensitive genomes. Mutational (SBS, DBS, and ID) signatures for each sample were predicted using nonnegative matrix factorization–based tool MutationalPatterns (62), and the extracted signatures were compared with the COSMIC SBS, DBS, and ID signature database.

Sequences of CGRs and ecDNAs identified in all the genomes were analyzed. Breakpoints of amplicons harboring key MAPKi-resistance genes, BRAF, NRAS, HRAS, and EGFR, were extracted, and sequences spanning these junctions were further analyzed. Homologous sequences and insertions for the breakpoint junctions were obtained from SvABA, and a possible mechanism for DNA DSB repair was inferred. First, HRR was identified with large homologous sequences and insertions (≥8 bp). Second, MMEJ or alt-NHEJ was identified with breakpoints comprising homologous sequences (2–8 bp). Third, NHEJ was identified when neither of the sequence types described above was identified in the vicinity of the breakpoints.

All cell lines were routinely tested for Mycoplasma and profiled and identified by RNA-seq and the GenePrint 10 system (Promega) at periodic intervals during the course of this study. All M series cell lines were established from patient-derived tumors at the UCLA with institutional review board approval. All M cell lines with acquired MAPKi resistance were derived in the Lo Laboratory and published previously (3, 5, 8, 9, 11, 22, 40, 67). We maintained H358 (ATCC) in RPMI 1640 (Gibco) with 10% heat-inactivated FBS (Omega Scientific, FB-02) and 2 mmol/L glutamine in a humidified, 5% CO₂ incubator at 37°C; all other cell lines were maintained in high-glucose DMEM (Omega Scientific, DM-22) with 10% heat-inactivated FBS (Omega Scientific, FB-02) and 2 mmol/L glutamine in a humidified, 5% CO₂ incubator at 37°C. We obtained inhibitors from the following sources: PLX4032 (Plexxikon), AZD6244 (Selleck Chemicals), trametinib (LC Laboratories), NU7026 (Abcam, ab120970), ABT888 (Enzo, ALX-270-444-M005), VX984 (MCE, HY-19939S), AZD7648 (TargetMol, T7122), olaparib (LC Laboratories, 763113-22–0), MRTX849 (Selleckchem, S8884), AMG510 (Selleckchem, S8830), and BGB-283 (BeiGene, via a Material Transfer Agreement with UCLA). All inhibitors were dissolved in DMSO and stored at −20°C.

shRNAs for PRKDC and vector control (pGIPZ) were obtained from Robert Damoiseaux, PhD (Molecular Screening Shared Resource, UCLA); shRNAs for LIG4 were from Sigma. All shRNAs were packaged into lentiviral particles for infection. The lentiviral viruses were generated by transfection of the constructs together with pMD2.G, pRSV-Rev, and pMDLg/pRRE into HEK-293T cells using calcium phosphate. Fourteen hours after transfection, media were replaced with preheated fresh media. Virus particles were harvested 24 and 48 hours later and filtered by a 0.45 μm filter unit (Millipore). Cells were transduced with recombinant lentivirus with 10 μg/mL polybrene (Millipore, TR-1003-G) for 48 hours and then selected by puromycin (Sigma, P8833) for 1 week. shRNA targeting sequences were as follows:

• Human PRKDC sh1: CAGTGAAAGTCTGAATCAT

• Human PRKDC sh2: GTCATGGATTCAAGAAATA

• Human PRKDC sh3: GAGCTTACATGCTAATGTA

• Human PRKDC sh4: GGTCATGGATTCAAGAAAT

• Human LIG4 sh1: TTCGACGCCACACCGTTTATT

• Human LIG4 sh2: TATGTCAGTGGACTAATGGAT

For clonogenic assay, cells were plated at indicated cell densities in 6-well plates and treated with inhibitor(s) the next day. Inhibitors and media were replenished every 2 days for the number of days indicated. Colonies were fixed in 4% paraformaldehyde, followed by staining with 0.1% crystal violet. For the MTT assay, cells were plated at 2,000 cells per well in 96-well plates (Fig. 4P and Q); parental and acquired resistant sublines were seeded at 4,000 cells per well and treated with graded concentrations of MAPKi the next day (Fig. 5D–F); and media were replenished every 2 to 3 days. Methylthiazolyldiphenyl-tetrazolium bromide (MTT; 100 μL) solution (0.5 mg/mL, Sigma, M5655) was added to each well and incubated at 37°C for 2 hours for MTT formazan formation. Formazan was then dissolved in 100 μL DMSO, and the absorbance values were measured using an ELISA reader (SpectraMax Plus 384) at the wavelength of 570 nm, blanked with DMSO solution. Experiments were performed with 4 replicates.

Cells were lysed in RIPA buffer (Thermo Fisher Scientific) with protease inhibitor cocktail (Thermo Fisher Scientific) and phosphatase inhibitor cocktails (Thermo Fisher Scientific) for Western blotting. BCA protein assay (Thermo Fisher Scientific, PI23227) was used to determine the protein concentration. Antibodies used in Western blot were as follows: TUBULIN (Cell Signaling Technology, 2144S), PRKDC (Cell Signaling Technology, 38168S), and LIG4 (Cell Signaling Technology, 14649S).

Tumor tissues were fixed in formalin followed by paraffin embedding. After deparaffinization and rehydration, tissue sections were antigen retrieved by heat. Permeabilization and blocking were followed by overnight incubation with primary antibodies [p-ERK1/2 (Cell Signaling Technology, 4370), p-DNA-PKcs (Abcam, ab124918), p-H2AX (Cell Signaling Technology, 9718)]. Immunofluorescence was performed with Alexa Fluor–conjugated secondary antibodies (Life Technologies, A-21429). Nuclei were counterstained by DAPI. Signals were captured with a Zeiss microscope (AXIO Imager A1) mounted with a charge-coupled device camera (Retiga EXi QImaging), and images were captured by Image-Pro plus 6.0.

Statistical analysis for genomic data was conducted in R.4.02, Python 3.8.0, and Python 2.7.17. Statistical analyses, described for specific experiments, for cell line– or PDX tumor–based assays were performed using GraphPad Prism.

The BAM files of WGS data are deposited in the European Genome-phenome Archive (https://www.ebi.ac.uk/ega/) with the accession number EGAS00001006874. This study did not generate custom codes. Any additional information regarding data reported in this article is available from the corresponding author upon request. There are no restrictions on data availability.

A. Ribas reports personal fees (honoraria) from Amgen, Bristol Myers Squib, Chugai, Genentech, Merck, Novartis, Roche, Sanofi, and Vedanta, personal fees (scientific advisory board member, honoraria, and stock) from 4C Biomed, Appia, Apricity, Arcus, Highlight, Compugen, ImaginAb, Kalthera/ImmPACT Bio, MapKure, Merus, Rgenix, Lutris, PACT Pharma, Synthekine, Tango, Advaxis, CytomX, Five Prime, RAPT, Isoplexis, and Kite/Gilead, and grants from Agilent and Bristol Myers Squibb outside the submitted work. S.J. Moschos reports grants from Merck during the conduct of the study, as well as grants from Merck, Amgen, and Syndax Pharma and other support from Castle Biosciences outside the submitted work. R.S. Lo reports grants from Merck, OncoSec, Pfizer, and Bristol Myers Squibb outside the submitted work, as well as a patent for U.S. Patent Application Serial No.: 63/268,028 issued. No disclosures were reported by the other authors.

P. Dharanipragada: Data curation, formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing. X. Zhang: Formal analysis, investigation, visualization, methodology, writing–review and editing. S. Liu: Data curation, formal analysis, methodology, writing–review and editing. S.H. Lomeli: Resources, methodology, writing–review and editing. A. Hong: Formal analysis, visualization, methodology, writing–review and editing. Y. Wang: Methodology, writing–review and editing. Z. Yang: Resources, methodology, writing–review and editing. K.Z. Lo: Resources, methodology. A. Vega-Crespo: Resources. A. Ribas: Resources. S.J. Moschos: Resources, methodology, writing–review and editing. G. Moriceau: Resources, formal analysis, visualization, methodology, writing–review and editing. R.S. Lo: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing.

We thank all members of the Lo Laboratory for their critical comments. This research was supported by grants (to R.S. Lo) from the NIH (1R01CA176111A1, 1R21CA215910-01, R21CA255837-01, and 1P01CA168585), the Melanoma Research Alliance (MRA; Team Science Awards 2020 and 2022), and the V Foundation for Cancer Research (Translational Award). Additional funding was provided by NIH 1P01CA168585 (to G. Moriceau), an MRA Dermatology Fellows Award (P. Dharanipragada, S. Liu, and Z. Yang), an MRA Young Investigator Award (G. Moriceau), a Jonsson Comprehensive Cancer Center (JCCC) Postdoctoral Seed Grant (to Z. Yang), JCCC Postdoctoral Fellowships (to P. Dharanipragada, X. Zhang, S. Liu, and Z. Yang), and the NIH T32CA009120 Tumor Immunology Postdoctoral Fellowship (to A. Hong). S. Liu is the recipient of a Career Development Award from the Melanoma Research Foundation. Additional support came from Mary Tanner and Maurizio Grimaldi (R.S. Lo), the Ressler Family Foundation (R.S. Lo), and the University Cancer Research Fund [Lineberger Comprehensive Cancer Center, The University of North Carolina at Chapel Hill (UNC-CH)]. We thank Mr. Vincent J. Moylan Jr, M.S., P.A. (Assistant Professor of Pathology and Laboratory Medicine, UNC-CH), Medical Examiner for Orange County, NC, and Dr. Leigh B. Thorne (Professor of Pathology and Laboratory Medicine, UNC-CH) for performing warm autopsies and quality-control analysis of tissues used. We thank Xinmin Li, PhD (Director), and the Technology Center for Genomics and Bioinformatics at UCLA for excellent technical support. We also thank Robert Damoiseaux, PhD (Director), and the Molecular Screening Shared Resource for sharing resources.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.