Identication of Mechanisms of Resistance to ALK Inhibitors. Next-generation sequencing-based liquid biopsy proling: A step towards personalized treatment.

Background: Despite impressive and durable responses, patients treated with ALK inhibitors (ALK-Is) ultimately progress. We investigated potential resistance mechanisms in a series of ALK-positive non-small cell lung cancer (NSCLC) patients progressing on different types of ALK-Is. Methods: 26 plasma and 2 cerebrospinal uid samples collected upon disease progression to an ALK-I, from 24 advanced ALK-positive NSCLC patients, were analyzed by next-generation sequencing (NGS). A tool to retrieve variants at the ALK locus was developed. Results: 61 somatic mutations were detected in 14 genes: TP53, ALK, PIK3CA, SMAD4, MAP2K1 (MEK1) FGFR2, FGFR3, BRAF, EGFR, IDH2, MYC, MET, CCND3 and CCND1. Overall, We identied at least one mutation in ALK locus in 10 (38.5%) plasma samples, being the G1269A and G1202R mutations the most prevalent among patients progressing to rst- and second-generation ALK-I treatment, respectively. An exon 19 deletion in EGFR was identied in a patient showing primary resistance to ALK-I. Likewise, the G466V mutation in BRAF and the F129L mutation in MAP2K1 (MEK1) were identied as the underlying mechanism of resistance in three patients who gained no or little benet from second-line treatment with an ALK-I. Putative ALK-I resistance mutations were also found in PIK3CA and IDH2. Finally, a c-MYC gain, along with a loss of CCND1 and a FGFR3, were detected in a patient progressing on a rst-line treatment with crizotinib. Conclusions: NGS analysis of liquid biopsies upon disease progression identied putative ALK-I resistance mutations in most cases, being a valuable approach to devise therapeutic strategies upon ALK-I failure. determine the clinical utility of liquid biopsies for ALK-Is sequencing. In we provide a pipeline to detect somatic mutations in domain. Finally, we evaluate the clinical of alterations in other than ALK.

frequently mutated in cancer, in which hotspots (single nucleotide variants; SNVs), fusions, copy-number variations (CNVs) and exon skipping can be studied. AMPureXP magnetic beads (Beckman Coulter, Inc., Brea, CA, USA) were used to purify all libraries. Subsequently, the individual libraries were quanti ed using the Ion Library TaqMan® Quantitation Kit (Thermo Fisher, Palo Alto, CA, USA) in a StepOnePlus™ qPCR machine (Thermo Fisher, Palo Alto, CA, USA) and adjusted to a nal concentration of 50 pM. Eight samples were pooled. Templating and Ion 550™ Chip loading were carried out with an Ion Chef™ System (Thermo Fisher, Palo Alto, CA, USA). Finally, an Ion GeneStudio™ S5 Sequencer (Thermo Fisher, Palo Alto, CA, USA) was used to sequence loaded Ion 550™ chips.
The mutations identi ed by NGS were con rmed by digital PCR (dPCR) on cfDNA samples using a QuantStudio® 3D Digital PCR System (Applied Biosystems, South San Francisco, CA, USA). In accordance with the manufacturer's speci cations, dPCR reactions were performed in an 18-µL volume comprising 9 µL of 20X QuantStudio 3D Master Mix, 0.45 µL of 40X commercially available predesigned or custom TaqMan® assays and 8.55 µL of cfDNA. Subsequently, 14.5 µL of the PCR reaction were loaded onto a QuantStudio 3D Digital PCR 20K chip using QuantStudio™ 3D Digital PCR Chip Loader. Each dPCR run included a negative control DNA, as a wild-type (wt) control, a blank (with no cfDNA) and a positive control. PCR reactions were performed in a thermal cycler (Applied Biosystems) at 96 °C for 10 min, then 40 cycles at 56 °C for 2 min and 98 °C for 30 s, and a nal elongation step at 72 °C for 10 min. Finally, samples were maintained at 22 °C for at least 30 min. After dPCR was completed, chips were read using two independent QuantStudio™ 3D Digital PCR Instruments and the uorescence was read twice. Results were visualized and analyzed using QuantStudio® 3D Analysis Suite™ Cloud Software and the automatic call assignments for each data cluster were manually adjusted when needed. The mutant allele frequency (MAF) was calculated as the ratio of mutant DNA molecules to the sum of mutant and wild-type (wt) DNA molecules.
Torrent Suite Software (v5.12) was used to analyze the raw sequencing data. The CoverageAnalysis (v. 5.12.0.0) plugin was used for sequencing coverage analysis (Thermo Fisher, Palo Alto, CA, USA). As recommended by the manufacturer, a median read coverage > 25,000 and a median molecular coverage > 2500 were required to detect a variant with a MAF of 0.1%. Raw  To detect variants in the ALK locus, we analyzed data by developing a bioinformatic pipeline, called the VALK tool, which is capable of fully automating the ltering of mutations in the ALK locus and generating a .csv le containing the output variants and a list of their properties. To increase the detection rate for variants at the ALK locus, speci c conditions for SNVs, indels, multiple-nucleotide polymorphisms (MNP), fusions and copy-number variation (CNV) calls were de ned. The pipeline uses the raw data in the non-ltered-oncomine.tsv, which contains variants that have passed the Oncomine Variants (v.5.12) lter and variants that have not. Brie y, parameters such as the LOD, mutant allele frequency, molecular coverage of the mutant allele, number of reads supporting a speci c variant, and clinical signi cance, among others, were considered in the selection of the different thresholds. Detailed information about the pipeline is available in the Supplementary Methods and in Supplementary Fig. 1. Positive and negative percentage agreement (PPA and NPA) and overall rates of agreement (ORA) of the VALK tool for detecting the ALK mutations speci ed in Supplementary Table 1 were calculated considering the imperfect reference standard the dPCR result, using the two independent data sets, the ALK cohort and the Valencia cohort, which consists of 53 cfDNA samples from NSCLC patients.

Statistical analysis
Median follow-up was estimated by the reverse Kaplan-Meier method. Overall survival (OS) was de ned as the time from the start of treatment with an ALK inhibitor to death or last follow-up. Progression-free survival (PFS) was de ned as the time between the start of an ALK-inhibitor and disease progression (as ascertained by RECIST criteria), death, or the censored date of the last assessment, whichever occurred rst. The log-rank test was used to assess statistical differences between Kaplan-Meier survival curves. Hazard ratios (HRs) were estimated from the Cox model using a multivariable approach adjusted for sex, Cooperative Oncology Group (ECOG) performance status, and lines of ALK-TKI. Association between clinicopathological variables and genomic features were assessed by Fisher's exact test. Values of P < 0.05 were considered statistically signi cant. Statistical analyses were performed using Stata 15.1.

Study cohort.
We collected and analyzed 26 plasma and 2 cerebrospinal uid (CSF) specimens from 24 metastatic patients diagnosed with an ALK-positive NSCLC who were progressing on an ALK-I. Baseline clinicopathological characteristics of the study population (N = 24) are presented in Table 1. The median age at diagnosis was 53 years (range, 36-72 years) and 58.3% were females. The majority of the patients were never-smokers (62.5%) and the most frequent histology was adenocarcinoma (95.8%). ECOG Performance Status at study entry varied from 0 to 2. As shown in Fig. 1, two samples from two patients were collected upon disease progression while on two consecutive lines of treatment with an ALK-I; three samples were obtained from one patient upon failure to three consecutive lines of ALK-I; all 21 other members of the cohort each provided a single sample. As presented in Fig. 1, 13 samples corresponded to ALK-I-naïve patients who progressed on a rst-line crizotinib (N = 11) or alectinib (N = 2) treatment. For these patients, the median PFS and OS were 11.6 months (95%CI: 6.5-20.9 months) and 24.6 months (95%CI: 11.8-NR months), respectively. In addition, 12 samples corresponded to patients who had received previously crizotinib and were treated with a second-generation ALK-I. Finally, the cohort included two samples from patients progressing on lorlatinib after failure of a prior second generation ALK-I and one patient progressing on alectinib who had previously received crizotinib plus two second generation ALK-I. Detailed information about treatment lines is presented in Supplementary table 2. The median PFS and OS for patients progressing on a second or subsequent line with an ALK-I were 5.4 months (95%CI: 2-9.1) and 11.2 months (95%CI: 3-NR) months respectively.
Next-generation sequencing analysis upon disease progression.
Overall, using the Pan-Cancer Cell-Free Assay, 61 somatic variants in ctDNA from 24 samples were detected. There were no signi cant differences in cfDNA input between samples in which a somatic mutation was detected in the cfDNA and those in which it was not (N = 4). One of the patients with undetectable plasma ctDNA had progressed exclusively at the brain level. The average number of mutations per patient was 2.18 and the median MAF was 0.39%. As expected, SNPs were the most frequent mutation type (N = 48). In addition, we identi ed 10 indels and three CNVs (Fig. 2). Speci cally, a c-MYC gain in conjunction with a CCND1 and an FGFR3 loss were detected in a patient progressing on a rst line with crizotinib. This patient also harbored a mutation in TP53 ( Fig. 2).
As illustrated in Fig Table 3. Thirteen variants (12 of which were in the ALK locus and one was in the EGFR gene) were categorized as being of strong clinical signi cance.

Identi cation of acquired resistance mutations in the ALK locus upon disease progression
To increase the sensitivity for detecting somatic mutations in the ALK locus from ctDNA we developed an algorithm (VALK tool) for this purpose (available upon request). The tool has been speci cally designed for the analysis of the NGS data obtained from liquid biopsies. Among other parameters, the algorithm takes into account the molecular depth and molecular counts as well as speci c regions that are more likely for false positive calls. In order to test the analytical performance of the tool, all SNPs in the ALK locus that were present in the non-ltered-oncomine.tsv le were analyzed by dPCR. In total, 19 ALK variants from 22 samples were analyzed by dPCR (Supplemental Table 1). Considering the non-reference standard the dPCR result PPA, NPA, and overall percent agreement of ALK mutation detection for the VALK tool were 67% (95%CI: 35-90%), 93% (95%CI: 75-99%), and 85% (95%CI: 69-94%), respectively.
Overall, in the ALK-cohort we detected at least one ALK mutation in 10 (38.5%) plasma samples collected upon disease progression ( Table 2, Supplementary  Fig. 2). Notably, the Oncomine variants v5.10 lter only detected ALK mutations in three patients (Table 2). The G1202R mutation was the most prevalent among patients treated with a second-generation ALK-I (N = 15), being identi ed in four patients who had progressed on alectinib (N = 3) and ceritinib (N = 1). The S1206Y mutation was detected along with the G1202R mutation in one of the aforementioned alectinib-progressing patients. This patient had been treated with crizotinib before initiating alectinib treatment. The low MAF of the S1206Y mutation suggests that it could be responsible for the previous crizotinib failure. In addition, the G1269A mutation was detected upon crizotinib failure in two cases and the L1196M mutation was identi ed after progression to crizotinib and lorlatinib. The latter was detected together with the R1275Q mutation in a patient diagnosed with an ALK-positive neuroendocrine carcinoma. Finally, the A1200V mutation (N = 1) arose as a result of the failure of crizotinib, and the V1180L mutation (N = 1) was detected in a patient progressing on a rst-line treatment with alectinib (Table 2).
Other molecular mechanisms underlying resistance to ALK-I.
A deletion in exon 19 of the EGFR gene, a non-V600 BRAF mutation and the F129L mutation in MAP2K1 (MEK1) were identi ed in four patients who showed no objective survival bene t from ALK-Is. None of these patients had a secondary mutation in ALK locus.
Notably, the patient harboring the E746_A750del mutation in the EGFR gene had a PFS time of 1.8 months under rst-line crizotinib treatment. The patient was subsequently treated with alectinib but tumor progression was assessed 3.1 months later prompting a switch of treatment to lorlatinib, but that also failed after 1.8 months, suggesting that the tumor had primary resistance to ALK-Is (Fig. 1, Table 3). Similarly, a non-V600 BRAF mutation, namely G466V, was identi ed in the CFS collected upon disease progression to ceritinib. Remarkably, while PFS with rst-line crizotinib was 21 months, disease progression was assessed within 3 months of starting second-line ceritinib treatment (Fig. 1), suggesting that the G466V mutation was acquired promoting resistance to second-line ALK-Is. Likewise, two patients harboring the F129L mutation in MAP2K1 (MEK1) obtained little bene t from second-line ALK-I (Fig. 1, Table 3).
This mutation was detected upon disease progression to alectinib and lorlatinib. Noteworthy, the median PFS and OS for second-line treatment for these patients were less than one month (0.97) and 3 months, respectively, whereas median the PFS and OS for patients progressing on a second or subsequent lines with an ALK-I but without mutations in MAP2K1 were 5.9 and 11.2 months (P log rank < 0.05 in both cases; supplementary Fig. 3). Putative ALK-I resistance mutations were also found in IDH2, PIK3CA and MYC (Table 3). Speci cally, the oncogenic mutations E545K and E545A in the PIK3CA gene were detected in the plasma sample of two patients progressing on ceritinib and brigatinib (Table 3). Likewise, the gain-of-function mutation in IDH2, R140Q, was detected upon disease upon progression to rst-line alectinib treatment. Finally, as previously mentioned, a c-MYC ampli cation was detected jointly with a loss of CCND1 and of FGFR3.

Other concomitant mutations
In the present cohort, TP53 was the most frequently mutated gene. The frequency of TP53 mutations identi ed is presented in Supplementary Fig. 4. As shown, the P92A mutation accounted for 36.36% of mutations in this gene. Curiously, the median PFS was 7.7 months, for patients with tumors harboring the P92A mutation and progressing on a rst-line ALK-I, compared with 14.7 months for patients in whom this mutation was not detected. Similarly, the median OS was 11.8 months for patients progressing on a rst-line treatment with an ALK-I with tumors harboring the P92A mutation compared with 34.9 months for patients testing negative for this mutation. However, these differences were not statistically signi cant.
Discussion ALK-Is have dramatically improved outcomes in NSCLC patients as well as in several other hematological and solid malignancies (17). However, despite the impressive responses they elicit, patients invariably relapse due to acquired resistance mutations. Solid biopsies remain the gold standard for biomarker testing. However, logistics for obtaining repeat tumor biopsies are complicated and seldom feasible since many patients are unable to endure an invasive procedure, which at the end of the day leads to an empirical prescription of sequential ALK-Is. Nevertheless, blinding sequential strategies might have a deleterious effect on patient's survival due to the incompletely overlapping ALK mutation coverage of different ALK-Is. In this exploratory analysis, we show, as proof of concept, that plasma NGS is feasible, enabling the detection of resistance mechanisms in patients with ALK-positive NSCLC upon progressive disease. We also provide an algorithm capable of retrieving somatic mutations in the ALK locus that would otherwise be discarded by the commercial bioinformatic pipeline. Remarkably, the developed algorithm performs well in terms of discarding samples with no mutations. Measuring the abundance of DNA molecules in a given sample by NGS is subjected to PCR ampli cation bias, as not all targeted amplicons are ampli ed with the same e cacy during library preparation. This limitation can be, at least partly, alleviated by ensuring that all molecules are distinguishable before ampli cation using unique molecular identi ers (UMIs) (18,19). With this approach, instead of counting reads, reads are grouped by UMIs, where each distinct UMI identi es the original molecule. In this scenario, parameters such as molecular coverage and molecular counts are pivotal, however free bioinformatic tools, based on the optimization of these parameters remains lacking. As presented in Table 2, the commercial pipeline only detected 3 out of 12 mutations. MAFs of variants detected by the commercial pipeline were 2.8%, 2.1% and 0.4%. According to the manufacturer's speci cations, the limit of detection, in terms of MAF, for mutations is 0.1%. However, in our hands, mutations with a MAF below 0.5% are seldom detected by the commercial pipeline. By using the VALK pipeline, some mutations that would otherwise have been missed can be rescued. Yet, con rmation using an alternative technique such as dPCR would be required to rule out false-positive calls.
Regarding acquired mutations in the ALK locus, our results are consistent with those of previous studies. Speci cally, secondary mutations were detected in the plasma samples of 4 of the 11 (36%) patients treated with rst-line crizotinib, with the G1269A mutation being detected in two cases. In this regard, mutation detection rate after crizotinib failure might vary from 60% (20) to 24% (21), G1269A being the most prevalent mutation. In this subset of patients we also detected the L1196M and S1206Y mutations, which have been reported to occur in 7% and 2% of cases, respectively, of ALK-positive NSCLC patients treated with crizotinib (11). Finally, we detected the A1200V mutation after crizotinib failure in one patient. This mutation is also known to appear upon crizotinib progression (20). On the other hand, we found that the G1202R mutation was identi ed in 3 of the 10 patients (30%) progressing on alectinib. This mutation is known to arise mainly after treatment with second-generation ALK-Is (11). Recently, Johannes N et al reported a 53% ALK mutation detection rate in samples obtained post-progression on alectinib (22) in which G1202R was the most frequent mutation. In our cohort, more than one mutation in ALK locus was detected in two samples collected during second-and third-generation ALK-I treatment. Likewise, it has been described that ALK resistance mutations become more frequent with each successive generation of ALK-I as sequential treatment may promote the appearance of resistance mutation at the ALK locus (23).
A reduced number of studies analyzing samples collected upon progression to an ALK-I by NGS have so far been conducted (11,19,20),  (11). The E545K and E545A mutations, which are two of the most common oncogenic mutations in PIK3CA, have also been detected upon progression in advanced EGFR-positive NSCLC patients (24). On the other hand, the IDH2 R140Q detected in our cohort is known to transform cells in vitro and induces myeloid and lymphoid neoplasms in mice (25,26). The R149Q mutation in IDH2 is frequent in angioimmunoblastic T-cell lymphoma (27). In NSCLC, IDH1/2 mutations are rarely detected in primary tumors but it has been suggested that they could be branching drivers leading to subclonal evolution, based on the MAFs at which these mutations are detected (28). It is therefore not surprising that we found them upon treatment failure.
In addition, we found the E746_A750del mutation in one patient who did not bene t from treatment with ALK-Is. In this way, some researchers have found that mutations in EGFR in some NSCLC tumors coexist alongside ALK rearrangements (29) which may lead to primary resistance to ALK-I (30). Likewise, a non-V600 BRAF mutation was detected after 3 months of treatment with second-line ceritinib treatment, suggesting that resistance of the tumor to the ALK-I could be due to the acquisition of the BRAF mutation. It has been reported that ceritinib enhances the e cacy of trametinib, a MEK inhibitor, in BRAF/NRAS-wild type melanoma cell lines (31), which makes it plausible that ceritinib wouldn't have any effect in BRAF-mutated cells. Finally, two patients in whose plasma sample the F129L-activating mutation in MAP2K1 (MEK1) was detected, exhibited marked resistance to second-and third-generation ALK-Is. This mutation has been identi ed as the molecular mechanism underlying MEK/ERK pathway activation in resistant clones of human HT-29 colon cancer cells (32). Moreover, the activation of this downstream pathway is critical to the survival of ALK-positive NSCLC cells (33,34). Indeed, the combination of ALK and MEK inhibition was highly effective at suppressing tumor growth in a preclinical model of EML4-ALK NSCLC (35). Taken together, it is plausible that the F129L-activating mutation in MAP2K1 is an acquired mutation that leads to tumor resistance to ALK-Is.
Mutations in the FGFR2 and FGFR3 genes were detected in two patients progressing on ALK-Is, suggesting sensitivity to broblast growth factor receptor inhibitors. It has been reported that alectinib, despite being a potent ALK-I, has limited inhibitory activity against other protein kinases such as FGFR2 (36). It may therefore be worth con rming whether the appearance of mutations in FGFR genes is a recurrent event after treatment failure with an ALK-I. If this proved to be the case, clinical trials evaluating the e cacy of combinations of ALK-Is with FGFR inhibitors would be of particular interest. Nevertheless, with respect to the FGFR2 mutation encountered, it is important to point out that although G305R has been identi ed in tumor samples (37,38), it has not been biochemically characterized so, in this study, it was classi ed as being of unknown clinical signi cance.
Three CNVs in c-MYC, CCND1 and FGFR3 were detected upon disease progression in one patient, who was being treated with crizotinib. Remarkably, c-MYC ampli cation determines many oncogenic effects, including cell growth and proliferation (39) and it has been identi ed as a potential mechanism of primary resistance to crizotinib in ALK-rearranged NSCLC patients (40). It has been previously suggested by Alidousty et al that co-occurrence of early TP53 mutations in ALK + NSCLC can lead to chromosomal instability. Speci cally, authors reported that, in a subset of 53 ALK + tumors, up to a quarter of TP53-mutated tumors showed ampli cations of known cancer genes such as MYC or CCND1 (41). Consistent with this, we detected the P92A and V157F mutations in the TP53 gene in the same plasma sample of this patient.
Our results suggest that the P92A mutation in the TP53 gene could be of prognostic signi cance, although this observation should be interpreted with caution.
Our data are consistent with those of previous studies. First of all, the median PFS was 11.6 months (95%CI: 6.5-20.9 months) for ALK-positive NSCLC patients treated with rst-line crizotinib (N = 11) or alectinib (N = 2), which was very similar to that of the 10.9 months reported in the PROFILE 1014 trial, which included 172 patients randomized to crizotinib (42

Conclusions
In conclusion, we present a thorough molecular-level description of a series of patients with ALK-positive NSCLC progressing on an ALK-I. Molecular mechanisms underlying treatment failure seem to involve different pathways. NGS analysis of liquid biopsies collected upon disease progression is a valuable approach towards personalized that will lead to better care for ALK-positive NSCLC patients.

Consent for publication
All named authors approved the content and submission to this journal.

Availability of data and materials
The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.
MP reports personal fees from Roche, BMS, MSD P zer, Lilly, Novartis and Takeda grants and personal fees from AstraZeneca, and Boehringer during the conduct of the study. VC reports personal fees from Roche BMS, MSD, P zer, Lilly, AstraZeneca, Boehringer, Novartis, Takeda, during the conduct of the study.
MD reports personal fees from Astra-Zeneca, BMS, Boehringer Ingelheim, MSD, P zer, Roche. The rest of the authors have declared no con ict of interest. Funding: This study has been funded by Instituto de Salud Carlos III through the project "PI17/01977" (Co-funded by European Regional Development Fund/European Social Fund "A way to make Europe"/"Investing in your future"). The work presented in this paper also received funding from the European Union's Horizon 2020  Figure 1 Swimmer chart showing the individual treatment responses of the study cohort. Blue, red, purple, green and orange bars correspond to patients who were treated with crizotinib, alectinib, ceritinib, brigatinib and lorlatinib, respectively. Tumor progression is denoted by triangle and patient's death is denoted by a squared. Mutations in BRAF, EGFR and MAP2K are indicated by a lightning.