Clinical and molecular feature-based nomogram model for predicting benefit from bevacizumab combined with first-generation EGFR-tyrosine kinase inhibitor (TKI) in EGFR-mutant advanced NSCLC

All the patients with EGFR sensitizing mutation-positive locally advanced/advanced NSCLC treated with either EGFR-TKI monotherapy or EGFR-TKI plus bevacizumab as first-line therapy were retrospectively enrolled in this study for efficacy evaluation and resistance mechanism investigation. The main inclusion criteria included (1) age of 18–75 years, (2) stage IIIB-IV (according to the 8th American Joint Committee on Cancer Staging System), (3) histologically confirmed lung adenocarcinoma/adenosquamous carcinoma with EGFR activating mutations (exon 19 deletions or L858R), (4) tumor biopsy sample available at baseline and at progression submitted for next-generation sequencing (NGS) detection for panel of 168 cancer-related genes (Additional file 1: Table S1), (5) received at least two cycles of regimen at standard dosing, and (6) having at least one radiological response evaluation according to Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1. The major exclusion criteria included treatment with other previous systemic therapy. The study protocol was approved by the Hunan Provincial Ethics Committee and written informed consent was obtained from each participant for the use of his or her clinical information for research analysis.

Our study included a total of 318 patients with EGFR-mutant locally advanced/advanced NSCLC who submitted samples for 168-gene panel NGS testing and were treated with either first-generation EGFR-TKI combined with bevacizumab (A+T) or EGFR-TKI monotherapy (single T). For the nomogram model construction, the cohort was further stratified into two groups as training and validation cohorts. The training cohort A was comprised of 176 patients from Hunan Cancer Hospital wherein 88 patients respectively received A+T and T regimens, while the validation cohort B was comprised of 142 patients from Xiangya Hospital wherein 71 patients respectively received A+T and T regimens. To ensure that the two groups had comparable baseline clinical characteristics, the patients included in the T group were selected using propensity score matching with the A+T group using a 1:1 ratio. Retrospective analyses were performed on clinical data, survival outcomes, and mutation profiles at baseline and at progression. Figure 1 illustrates the study design. The data cutoff was August 31, 2020, with median follow-up of 31 months in the training cohort A and 30 months in the validation cohort B. The standard initial dosing in clinical practice is 15 mg/kg, every 21 days for bevacizumab, 150 mg once daily for erlotinib, and 250 mg once daily for gefitinib. Drug dose adjustments were determined by the treating physicians.

Clinical information was collected, including but not limited to demographic characteristics, ORR, PFS, adverse events, and somatic mutation profile. Tumor tissue samples were collected at the initial diagnosis (baseline) and at disease progression by resection or needle biopsy and stored as formalin-fixed paraffin-embedded (FFPE) samples. NGS testing was submitted to Burning Rock Biotech.

Tissue DNA was isolated from FFPE tumor tissue samples using QIAamp DNA FFPE tissue kit (Qiagen, Hilden, Germany). The quality and the size of the fragments were assessed using Qubit 2.0 Fluorimeter with the dsDNA high-sensitivity assay kit (Life Technologies, Carlsbad, CA, USA). NGS library was prepared for each sample. Target capture was performed using a 168 gene panel (Lung Plasma, Burning Rock Biotech, Guangzhou, China), which included 68 lung cancer-related genes and 100 other genes related to cancer development, spanning 0.273 megabases (Mb) of the human genome (Table S1). The indexed samples were sequenced on Nextseq500 (Illumina, Inc., CA, USA) with paired-end reads and average sequencing depth of 1000X. DNA sequence data were analyzed using optimized bioinformatics pipelines based on mapping with the reference human genome (hg19) using Burrows-Wheeler Aligner version 0.7.10, local alignment optimization, duplication marking, and variant calling using Genome Analysis Tool Kit version 3.2 and VarScan version 2.4.3. Single nucleotide variations (SNV) and insertion-deletion variations (INDEL) with population frequency over 0.1% in the ExAC, 1000 Genomes, dbSNP, or ESP6500SI-V2 databases were grouped as single nucleotide polymorphisms (SNPs) and excluded from further analysis. Remaining variants were annotated using ANNOVAR (2016-02-01 release) and SnpEff version 3.6. Analysis of structural variations (SVs) was performed using Factera version 1.4.3. Copy number variations (CNV) were analyzed based on the depth of coverage data of capture intervals. The limit of detection for copy number (CN) gain is CN > 2.25 for hotspot genes (such as EGFR, TP53, ERBB2, and MET) and CN > 2.5 for others, while for CN loss is CN < 1.75 for hotspot genes and CN < 1.5 for others.

Median survival was calculated by the Kaplan–Meier method. The Cox proportional hazards model was used for multivariable survival analysis. The variables with a p < 0.1 from the univariate analysis were included in the multivariate analysis. Schoenfeld residuals were used to check the proportional hazards assumption. Based on the multivariable analysis, the nomograms were constructed using R version 2.12.1 (The R Foundation for Statistical Computing, Vienna, Austria). The performance of the nomogram was evaluated by calculating the concordance index (C-index). The nomograms were validated using internal and external validation methods. Calibration of the nomogram was performed by comparing the predicted probability and the actual status after bias correction. All p values were two-sided, and p < 0.05 was considered statistically significant, unless stated otherwise. All procedures were performed using either SPSS software (version 24) or R Studio (version 2.12.1).

A total of 159 patients, with 88 patients from Hunan Cancer Hospital and 71 patients from Xiangya Hospital with EGFR-mutant locally advanced/advanced NSCLC who were treated with A+T were matched to 159 patients who were treated with single-agent T using propensity score matching with 1:1 ratio (Fig. 1). The two groups and cohorts were comparable for baseline clinical characteristics including age, sex, smoking status, Eastern Cooperative Oncology Group Performance Status (ECOG PS) score, tumor node metastasis (TNM) stage, and tumor histology (Table 1). Cohort A also had comparable baseline somatic mutational characteristics (Additional file 2: Figure S1). In cohort A, baseline brain metastasis was detected in 28 (31.4%) patients in the A+T group and 30 (34.1%) patients in the T group (p = 0.451). No significant difference was observed for the distribution of other metastatic sites in both cohorts (Table 1).

The A+T group had significantly longer PFS (median PFS: 14.0 vs. 10.5 months; hazard ratio (HR) = 0.55, 95% confidence intervals (CI) = 0.40–0.76; p < 0.001) (Fig. 2A) and improved OS as compared with the T group (median OS 37.0 vs. 26.0 months; HR = 0.65, 95% CI = 0.43–0.98; p = 0.042) (Fig. 2B). Mutational status on certain genes, including TP53 mutation, was associated with significantly different PFS and OS in the T group but not in the A+T group (Additional file 2: Figure S2).

We evaluated the impact of the brain, liver, or pleural metastasis on the therapeutic efficacy of single T and A+T regimens in the training cohort. In the single T group, the presence of brain, liver, or pleural metastasis was associated with significantly shorter PFS and OS (Additional file 2: Figure S3A). However, in the A+T group, the PFS was comparable for patients with brain, liver, or pleural metastasis, but had poor OS outcome for those with brain/liver metastasis (Additional file 2: Figure S3B).

We evaluated the impact of the concomitant mutations in tumor suppressor or oncogenic driver genes on the therapeutic efficacy of single T and A+T. Patients with only EGFR activating mutations showed comparable PFS with the two treatment regimens (p = 0.568) (Fig. 3A). Patients who concurrently harbor mutations in any tumor suppressor gene achieved longer PFS in the A+T group than in the T group (median PFS 14.5 vs. 8.0 months, p < 0.001) (Fig. 3A). The patients with concurrent oncogenic driver gene mutations had statistically comparable PFS with A+T and T treatments (16.3 vs. 9.0 months, p = 0.665) (Fig. 3A). Similar trends were observed for OS (Fig. 3B).

As the most common genetic alteration observed in our cohort, the presence of TP53 mutation had a negative impact on the efficacy of single T treatment. Patients who harbored TP53 mutations had inferior PFS in comparison to those who had wild-type TP53; however, it did not impact the efficacy of A+T treatment (Additional file 2: Figure S4A). Expectedly, as compared with single T, A+T significantly prolonged the PFS of TP53-mutant patients (median PFS 15.0 vs. 8.0 months, p < 0.001), while no difference was observed among TP53 wild-type patients (Additional file 2: Figure S4A). These observations were also consistent for OS (Figure S4B). As compared to single T, significantly longer PFS and OS were observed with A+T treatment of patients with TP53 mutations found in exons 5-8 (Additional file 2: Figure S5A-B), but not TP53 mutations resulting in loss-of-function (Additional file 2: Figure S5C-D) or those with concomitant mutations in TP53 and RB1 (Additional file 2: Figure S5E-H).

Furthermore, two patients with stage IV lung adenocarcinoma detected with EGFR exon 19 deletion and either concurrent KRAS K117N (n = 1) or concurrent de novo EGFR T790M (n = 1) achieved partial response with A+T regimen, with PFS of > 16.0 months and 10.3 months, respectively. Another patient was detected with concurrent EGFR exon 21 L858R concurrent with de novo EGFR T790M and had best response of progression disease and PFS of 1 month with single-agent gefitinib.

In training cohort A, through statistical analysis and in consideration of clinical significance, the presence/absence of concurrent genetic mutation (i.e., exclusive EGFR activating mutations, concurrent oncogenic driver gene mutation, or concurrent tumor suppressor gene mutation) and the location of metastasis (i.e., locally pleural, brain, or liver metastasis) were selected as the molecular and clinical factors to generate the nomograms for estimating the 12-month and 18-month PFS rates of T treatment (Fig. 4A). As shown in the nomogram, a lower total score/point was associated with having only EGFR mutations and/or having only intrathoracic metastasis and indicated a lower risk of disease progression with T treatment, whereas a higher score indicated a higher risk of disease progression due to the presence of concurrent gene mutations and/or presence of brain or liver metastasis. To validate the prognostic utility of the nomogram model, we used the nomogram to score the patients in either the T or the A+T group in the training cohort A or validation cohort B and used the median score as the cutoff to divide the patients as low-risk (< median, low score) and high-risk (≥ median, high score). As expected, patients with lower scores presented with significantly longer PFS than those with higher scores (15.0 vs. 9.0 months, p = 0.002) (Fig. 4B). In contrast, the scores from the nomogram were not predictive of PFS with A+T treatment as shown by the comparable PFS between the subgroups (15.9 vs. 13.9 months, p = 0.2556), suggesting that the factors included in the nomogram did not impact the efficacy of A+T (Fig. 4C). Consistently, patients with higher scores in the T group of the validation cohort B had significantly shorter PFS than those with lower scores (14.5 vs. 8.0 months, p < 0.001) (Fig. 4D), while the patients in the A+T group also had comparable PFS for those with high and low scores (16.8 vs. 12.3 months, p = 0.9042), suggesting that the factors included in the nomogram did not impact the efficacy of A+T (Fig. 4E).

Likewise, we constructed a model for estimating the 12-month and 24-month OS of T treatment, using the same molecular and clinical factors included in the nomogram model for PFS (Additional file 2: Figure S6A). Similarly, lower scores predicted significantly prolonged OS in the T group (36.0 vs. 21.0 months, p = 0.0051) (Additional file 2: Figure S6B); however, the scores were not predictive of the OS for the patients in the A+T group (38.7 vs. 28.0 months, p = 0.1221) (Additional file 2: Figure S6C).

The fact that this model was not applicable for the A+T-treated patients suggested that these molecular and clinical factors negatively impacted the efficacy of T treatment but did not affect A+T treatment. Hence, we further evaluated each variable included in the nomograms on their impact on A+T efficacy using Cox multivariable analysis. We found that none of these factors impacted the PFS for A+T treatment (Additional file 2: Figure S7A), while the presence of brain metastasis (HR = 2.96, 95% CI = 1.51–5.83, p = 0.002) and liver metastasis (HR = 2.31, 95% CI = 1.02–5.22, p = 0.044) significantly deteriorated the OS of A+T-treated patients (Additional file 2: Figure S7B).

Resistance mechanisms after disease progression from single T and A+T regimens were only evaluable for 79 and 63 patients included in the training cohort A, respectively.

At progression, EGFR T790M was the major resistance mechanism in both groups and was detected from 41% of the T group and 34% of the A+T group. The distribution of acquired mutations in the T group included exclusively EGFR T790M (30%), EGFR T790M plus cell cycle gene mutations (6%), EGFR T790M plus ERBB2 amplification (2%), EGFR T790M plus MET amplification plus cell cycle gene mutation (2%), and EGFR T790M plus TP53 mutation plus RB1 mutation (1%) (Fig. 5A). The distribution of acquired mutations in the A+T group included exclusively EGFR T790M (29%), EGFR T790M plus cell cycle gene mutation (3%), and EGFR T790M plus ERBB2 amplification (2%) (Fig. 5B). As compared to the A+T group, the T group frequently presented with multiple resistant mutations.

Among the 63 patients who progressed from A+T regimen, three patients were detected with RB1 mutation wherein two patients (67%) were confirmed with histological transformation to small-cell lung cancer (SCLC). Among the 79 patients who progressed from single T regimen, six patients were detected with RB1 mutations, of which four patients (67%) were confirmed to have SCLC transformation.