Classification and regression tree for estimating predictive markers to detect T790M mutations after acquired resistance to first line EGFR-TKI: HOPE-002

Background and objective. Osimertinib as first-line treatment for patients with non-small cell lung cancer (NSCLC) harboring epidermal growth factor (EGFR) mutations remains controversial. Sequential EGFR-tyrosine kinase inhibitor (TKI) might be superior to the first line osimertinib in patients at risk of developing acquired T790M mutations. Methods. We enrolled consecutive patients with EGFR-mutated (deletion 19 or L858R) advanced NSCLC treated with first-line drugs and evaluated predictive markers using classification and regression tree (CART) for the detection of T790M mutations based on patient backgrounds prior to initial treatment. Results. Patients without acquired T790M mutations had worse outcomes than those with T790M mutations (median OS: 798 days vs. not reached; HR: 2.70; P < 0.001). CART identified three distinct groups based on variables associated with acquired T790M mutations (age, CYF, WBC, liver metastasis, and LDH; AUROC: 0.77). Based on certain variables, CART identified three distinct groups in deletion 19 (albumin, LDH, bone metastasis, pleural effusion, and WBC; AUROC: 0.81) and two distinct groups in L858R (age, CEA, and ALP; AUROC: 0.80). The T790M detection frequencies after TKI resistance of afatinib and first-generation EGFR-TKIs were similar (35.3% vs. 37.4%, P = 0.933). Afatinib demonstrated longer PFS (398 vs. 279 days; HR: 0.67; P = 0.004) and OS (1053 vs. 956 days; HR: 0.68; P = 0.051) than first-generation EGFR-TKIs. Conclusion. Identification of patients at risk of acquiring T790M mutations after EGFR-TKI failure may aid in choice of first-line EGFR-TKI. Furthermore, afatinib may be the more effective 1st-line EGFR-TKI treatment for patients at risk of developing T790M as initial EGFR-TKI resistance.


Introduction
Epidermal growth factor receptor (EGFR)-tyrosine kinase inhibitors (TKIs) are key drugs for patients with nonsmall cell lung cancer (NSCLC) harboring EGFR mutations. First -, second, and third generation (1st-, 2nd-, and 3rd-G) EGFR-TKIs have been developed and evaluated for toxicity and efficacy in three randomized trials (LUX-Lung 7, ARCHER 1050, and FLAURA). All three trials showed more clinical benefits in the 2nd-and 3rd-G than the 1st-G EGFR-TKIs. While all studies showed that the 2nd-and 3rd-G EGFR-TKIs improved progression-free survival (PFS), only osimertinib in FLAURA exhibited a statistically significant improvement in overall survival (OS). While OS improvement was also observed with dacomitinib in ARCHER1050, statistical analysis was not performed in this study due to adoption of the gatekeeping method [1][2][3]. Therefore, osimertinib is the most strongly recommended drug for patients with common EGFR mutations. Japanese guidelines have judged the quality of this evidence as B and the strength of this recommendation as 1 [4].
Osimertinib was developed as an anti-cancer drug due to its activity against Thr790Met (T790M) mutations through covalent binding [5]. T790M is the most common resistance mechanism that impairs the activity of TKIs detected in approximately 50% of patients with 1st and 2nd-G EGFR-TKI-refractory tumors [6,7]. Thus, osimertinib is effective against T790M-mediated acquired resistance (AR) [8], making it a reasonable option for several EGFR-mutant patients initially treated with 1st-or 2nd-G EGFR-TKIs.
At present, there are two main first-line strategies for advanced EGFR-mutant NSCLC: first-line osimertinib and salvage osimertinib for T790M-positive AR initially treated with first-line 1st-or 2nd-G EGFR-TKIs. However, prevention of the development of AR must also be considered as a key therapeutic strategy. Importantly, firstline and salvage therapy have different mechanisms of the development of osimertinib AR [9][10][11][12] that are more complicated than those of 1st-or 2nd-G EGFR-TKIs.
Treatment strategies for AR to first-line osimertinib, are currently still limited. On the other hand, salvage osimertinib therapy was found to be a simple and effective treatment for patients with T790M detected after the development of AR to 1st-or 2nd-G EGFR-TKIs. However, the occurrence of T790M after developing AR to first-line 1st-or 2nd-G EGFR-TKIs was found to have a low incidence (approximately 30%) in real-world FLAURA trial data [3,13]. To more effectively utilize salvage osimertinib therapy for these cases, the detection rate of T790M must be increased.
In the present study, we aimed to increase rates of T790M detection after first-line 1st-or 2nd-G EGFR-TKI therapy by evaluating predictive markers for T790M mutation detection in patients prior to first-line EGFR-TKI treatment.

Study design
We conducted a multicenter, retrospective cohort study across nine medical institutes belonging to the Hanshin Oncology Clinical Problem Evaluation group (HOPE) in Japan. The clinical data of the patients were retrospectively extracted from their medical charts and added to a database. Because this was a retrospective observational study, sample size calculation based on hypothesis testing was not performed.
This study was approved by the ethical review board or institutional review board of each participating institute. Informed consent was not required owing to the retrospective nature of the study, and an opt-out method was utilized so that patients and their families could refuse to participate in the study.

Patient selection
Patients > 20 years of age were consecutively enrolled if they had pathologically confirmed stage IV non-squamous NSCLC (excluding recurrent cases, such as those who had undergone post-operation or post-chemoradiation therapy) with sensitizing EGFR mutations (deletion 19 or L858R) and had received gefitinib, erlotinib, or afatinib as first-line therapies between January 1, 2015 and March 31, 2017.
The patients were classified as either never-smokers (those reported to have never smoked), current smokers (those who had smoked within 1 year of diagnosis), or former smokers (the remaining). The clinical stages of all the patients were determined according to the eighth edition of the tumor, node, and metastasis classification of malignant tumors. Anti-tumor responses were assessed using the RECIST version 1.1. The intervals between dates of commencing EGFR-TKI therapy and disease progression or death (PFS) and overall survival (OS) of the patients were calculated. The cutoff date for data collection was set at August 31, 2018.

Statistical analysis
Data were analyzed by independent statisticians. We evaluated predictive markers for T790M mutation detection based on the patients' backgrounds before first-line EGFR-TKI treatment using a classification and regression tree (CART).
The Gini coefficient was used to determine the best split. The data collected included sex, age, type of EGFR mutation, smoking history, Eastern Cooperative Oncology Group (ECOG) performance status (PS), type of EGFR-TKI, site of metastasis (brain, bone, liver, adrenal, lung, and pleural effusion), and laboratory data.
Fisher's exact test was used for categorical comparisons of data. Kaplan-Meier curves were used to evaluate PFS and OS. Hazard ratios (HRs) and 95% confidence intervals (CIs) were estimated using the Cox proportional hazards model. All statistical analyses were conducted using R software (version 4.0.5; http://R-proje ct. org; The R Foundation for Statistical Computing, Vienna, Austria). The following R libraries were used: rpart (version 4.1.15), rpart.plot (version 3.0.9), pROC (version 1.17.0.1), survival (version 3.2.10), and survminer (version 0.4.9). Differences were considered statistically significant at P < 0.05.

Patient demographics
A total of 289 consecutive stage 4 NSCLC patients (287 eligible) were enrolled at nine medical institutes belonging to the HOPE in Japan from January 1, 2015 to March 31, 2017. Among them, two patients with de novo T790M mutations were excluded. A total of 225 patients experienced disease progression (PD) after first-line EGFR-TKI treatment, of which 166 underwent re-biopsy by tissue or plasma. Among these patients, 147 patients remained for CART method analysis after excluding patients with missing background data, such as programmed cell death-ligand 1 (PD-L1), thyroid transcription factor-1 (TTF-1), sialylated carbohydrate antigen KL-6 (KL-6), neuron specific enolase (NSE), pro-gastrin releasing peptide (Pro-GRP) patients missing ≥ 10% data (Fig. 1

Development of T790M mutation after PD in first line EGFR-TKI therapy
A total of 36.7% (61/166) of the patients acquired T790M mutations after first-line EGFR-TKI therapy. The frequencies of T790M after each first-line EGFR-TKI failure were as

Evaluation of predictive markers for detecting T790M through CART
First, we evaluated predictive markers for the detection of T790M in the patients included by CART analysis (N = 147). CART identified three distinct groups of patients based on variables strongly associated with acquired T790M mutations (age, cytokeratin 19 fragmen (CYF), white blood cell count (WBC), liver metastasis, and lactate dehydrogenase (LDH)), with an area under the receiver operating characteristic (AUROC) of 0.77 (95% CI: 0.69-0.84) (Fig. 2a).
Although we classified the three groups according to the frequency of T790M, the AUROC was found to be low. Next, we analyzed predictive markers for detecting T790M mutations with each type of EGFR mutation since a previous analysis found that its frequency varies according to EGFR type [14]. The CART identified three distinct groups in deletion 19 based on certain variables (albumin (Alb), LDH, bone metastasis, pleural effusion, and WBC), with an AUROC of 0.81 (95% CI: 0.71-0.91) (Fig. 2b). Furthermore, CART identified two distinct groups in L858R based on certain variables (age, carcinoembryonic antigen (CEA), and ALP), with an AUROC of 0.80 (95% CI: 0.69-0.90) (Fig. 2c). The detection rates of the highest groups in deletions 19 and L858R were ≥ 80% and ≥ 60%, respectively.

PFS and OS of first-line EGFR-TKIs
As mentioned above, the T790M mutation affected OS, and the frequency of the development of T790M after developing EGFR-TKI resistance was similar between afatinib and 1st-G EGFR-TKIs. However, afatinib was previously reported to have better outcomes than gefitinib [1].  (Fig. 4b).

Discussion
This is the first analysis to identify predictive markers using CART for the detection of T790M mutations based on patient backgrounds prior to first-line EGFR-TKI treatment. This study also identified markers that distinguish between EGFR mutation types, leading to more accurate predictions of T790M detection. CART classified the groups according to the detection rate of T790M based on certain variables (age, CYF, WBC, liver metastasis and LDH). Furthermore, deletion 19 mutations were classified into three distinct groups based on certain variables (Alb, LDH, bone metastasis, pleural effusion, and WBC), and L858R mutations were classified into two distinct groups based on certain variables (age, CEA, and ALP). As first-line treatment, osimertinib has been found to be associated with longer PFS and OS than 1st-G EGFR-TKIs against advanced NSCLC harboring EGFR mutations (exon-19 deletion and L858R) [3]. However, in the Asian subset (especially in the Japanese subset) analysis of OS in the FLAURA study, osimertinib was not observed to be superior to 1st-G EGFR-TKIs [15]. Furthermore, no additional molecular targets for therapy are known due to the heterogeneity of resistance mechanisms that are not well understood [11,16]. As a result, in the clinical care of most patients following cancer progression after osimertinib treatment, chemotherapy is the only remaining option for secondline treatment. In contrast, the most common mechanism of the development of resistance to 1st-and 2nd-G EGFR-TKI treatment is the T790M mutation. Fortunately, osimertinib overcomes the T790M mutation and provides significantly longer PFS compared to standard platinum-based chemotherapy in advanced T790M positive NSCLC patients with AR to first line 1st or 2nd-G EGFR-TKIs (median PFS:  [8]. In particular, a non-interventional GioTag study demonstrated clinical benefit with sequential afatinib and osimertinib in patients with EGFR mutation-positive NSCLC with T790Macquired resistance; this trend was more pronounced among the Asian population [17].
In this study, we showed that patients with acquired T790M mutations had better outcomes than patients without T790M mutations after AR to first-line 1st-or 2nd-G EGFR-TKIs. Therefore, the use of osimertinib as first-line treatment for NSCLC patients harboring EGFR mutations remains controversial in practice. Furthermore, sequential EGFR-TKI treatment may be superior to first-line osimertinib in patients who will likely develop acquired T790M mutations. When considering sequential therapy, the benefit is likely to be diminished if T790M is not detected since osimertinib remains an important and beneficial drug that is essential for patients with advanced NSCLC harboring EGFR mutations. However, the incidence of T790M after AR to first-line first line 1st-or 2nd-G EGFR-TKIs was found to be low based on real-world and FLAURA trial data [3,13].
Previous reports have shown the value of performing a re-biopsy since patients who initially present as T790Mnegative exhibited T790M-positive conversion after repeated re-biopsy. In these studies, performing re-biopsies increased their T790M detection rates from 36 to 80% and 45% to 67% [14,18]. While this method increases the frequency of T790M detection, it also increases the number of invasive procedures done on patients. Liquid biopsies, a type of rebiopsy, are less invasive for patients but have lower T790M detection rates [19]. On the other hand, the droplet digital polymerase chain reaction (ddPCR) is another method that increases the sensitivity and rate of T790M [20,21]. These studies suggest that patients with low T790M allele frequency had longer PFS with osimertinib than those with high T790M allele frequency. Thus, and detection and measurement of T790M using ddPCR is an effective method for increasing T790M detection rates that guarantees the efficacy of osimertinib. However, the use of ddPCR is nonreimbursable and impractical for daily clinical practice. Therefore, the development of other methods to increase the rate of T790M detection after first-line 1st-or 2nd-G EGFR-TKI therapy remains important. We evaluated predictive markers for the detection of T790M mutations based on patient backgrounds prior to first-line EGFR-TKI treatment through CART. CART analysis is a prediction model constructed by recursively partitioning a dataset and fitting a simple mode with machine learning methods for constructing prediction models from data [22]. CART classified three distinct groups of patients based on variables that were strongly associated with detecting acquired T790M mutations (age, CYF, LDH, and liver metastasis); however, the AUROC was not satisfactory. In our study, the T790M detection rates between cases with deletion 19 and L858R mutations were different, in which is consistent with a previous report [14,23,24]. Accordingly, we decided to analyze predictive markers for the detection of T790M mutations for each type of EGFR mutation. We demonstrated that CART highly stratified the T790M detection markers according to EGFR type, with a more satisfactory AUROC than that of the total population. This may be due to the biological and clinical differences between deletion 19 and L858R mutations. The affinity between ATP and EGFR in cases with deletion 19 mutations is higher than that of L858R [25]. The phosphorylation of Akt and Erk in cases of deletion 19 mutations, along with downstream signals of EGFR, are inhibited in a concentration-dependent manner compared with in cases of L858R [26]. Furthermore, a difference in the mechanisms of EGFR activation has been reported [27], and differences in the efficacy of EGFR-TKI treatment have been reported in clinical practice [3,28].
The present study had some limitations. First, this was a retrospective study, in which only Japanese patients were eligible for the analysis. Thus, our findings may not be generalizable to other ethnic populations. Second, the patients were treated according to the physician's choices, and the treatment and examinations may not have been standardized. Therefore, this study was with a little missing data within eligible group that would limit the applicability of results. Third, the patients underwent various assays for primary and secondary EGFR mutations in this study. The sensitivity of each assay may have differed. However, this cohort is representative of true real-world practice.
In conclusion, identification of patients at risk of acquiring the T790M mutation after failure of EGFR-TKI may help in the selection of first-line EGFR-TKI treatment options. Furthermore, prediction of T790M mutations after initial EGFR-TKI resistance aids in recommending afatinib as the more effective first-line EGFR-TKI treatment compared with 1st-G EGFR-TKIs.