Potential predictive value of serum targeted metabolites and concurrently mutated genes for EGFR-TKI therapeutic efficacy in lung adenocarcinoma patients with EGFR sensitizing mutations

Abstract

There is a discrepancy in the efficacy of epidermal growth factor receptor-tyrosine kinase inhibitor (EGFR-TKI) treatment for advanced lung adenocarcinoma (LUAD) patients with EGFR sensitizing mutations (mEGFR). Molecular markers other than mEGFR remain to be investigated to better predict EGFR-TKI efficacy. Here, 49 LUAD patients with mEGFR (19 deletions or 21 L858R mutations) who received the first-generation EGFR-TKI icotinib therapy were included and stratified into 25 good-responders with a progression-free survival (PFS) longer than 11 months and 24 poor-responders with a PFS shorter than 11 months. We conducted targeted metabolomic detection and next-generation sequencing on serum and tissue samples, respectively. Subsequently, two metabolomic profiling-based discriminant models were constructed for icotinib efficacy prediction, 10 metabolites overlapped in both models ensured high credibility for distinguishing good- and poor-responders. Seven of the 10 metabolites displayed significant differences between the two groups, which belong to lipids including ceramides (Cers), lysophosphatidylcholines (LPCs), lysophosphatidylethanolamines (LPEs), sphingomyelins (SMs), and free fatty acids (FAs). Briefly, LPC 16:1, LPC 22:5-1, and LPE 18:2 decreased in poor-responders, while Cer 36:1-3, Cer 38:1-3, SM 36:1-2 and SM 42:2 increased in poor-responders. In parallel, we identified 6 co-mutated genes (ARID1A, ARID1B, BCR, FANCD2, PTCH1, and RBM10) which were significantly correlated with a shorter PFS. Additionally, 4 efficacy-related metabolites (Cer 36:1-3, Cer 38:1-3, SM 36:1-2, and LPC 16:1) showed significant differences between the mutant and wild-type of 4 efficacy-related genes (ARID1A, ARID1B, BCR, and RBM10). SM 36:1-2 elevated while LPC 16:1 decreased in ARID1A, BCR, and RBM10 mutant groups compared to the wild-type groups. Cer 36:1-3 increased in the ARID1A and BCR mutant groups, and Cer 38:1-3 only rose in the ARID1A mutant group. Furthermore, we observed a causal-mediator-network-based interrelation between the 4 concurrently mutated genes and the 4 metabolites related metabolic genes in glycerophospholipid metabolism and sphingolipid metabolism pathways. This study demonstrated that lipids metabolism and concurrently mutated genes with mEGFR were associated with the icotinib efficacy, which provides novel perspectives in classifying clinical responses of mEGFR LUAD patients and reveals the potential of non-invasive pretreatment serum metabolites in predicting EGFR-TKI efficacy.

Introduction

Lung cancer remains the leading cause of cancer-related death worldwide and non-small cell lung cancer (NSCLC) is the most common histological subtype which accounts for about 85% of lung cancer cases [1]. For advanced NSCLC patients with EGFR sensitizing mutations (mEGFR), epidermal growth factor receptor tyrosine kinase inhibitor (EGFR-TKI) is the paradigm of target treatment and significantly prolongs patients survival [2,3]. However, the efficacy of EGFR-TKI demonstrates diversity in mEGFR NSCLC patients with the progression-free survival (PFS) ranging from 1 month to 2.5 years [4-6]. Thus, factors beyond genetic aberrance that influence the therapeutic effect of EGFR-TKI is warranted to be further investigated [7-9].

Metabolic reprogramming has recently been recognized as a significant feature of tumor cells resistance to therapy. Therefore, metabolomics profiling, complementary to genomics and proteomics, can directly delineate cellular biochemical changes and provide useful information on disease status and therapeutic efficacy [10-12]. Nevertheless, only a few studies characterized metabolites signatures relevant to EGFR-TKI treatment response, either demonstrating single metabolites like 25-hydroxyvitamin D3 with prognostic value or revealing preclinical metabolic markers on mice models. More efforts on serum metabolites profiling and metabolites modeling predictive of EGFR-TKI efficacy are needed.

Additionally, contemporary studies have depicted concurrent genetic alterations with EGFR mutations that may affect patients’ survival [13-15]. The commonly reported coexistent gene alterations include TP53, PIK3CA, BRAF, and de-novo c-MET mutations, ALK fusion, PTEN deletion, and BIM deletion polymorphism [16-21], among which TP53 and PIK3CA mutations were the most detected ones [18,22]. Patients with these concurrent genetic alterations generally have poor clinical outcomes including shorter PFS and overall survival (OS).

In general, the aforementioned studies mainly focused on uncovering driver mutations or secondary mutations of mEGFR to investigate EGFR-TKI efficacy. There are very limited studies that integrated the serum metabolites and somatic concomitant gene mutations to determine the relevance of these factors to EGFR-TKI treatment in mEGFR NSCLC patients. In this study, utilizing samples obtained from a rigorous clinical trial, we revealed the targeted metabolites, the concurrently mutated genes that may be associated with the efficacy of first-generation EGFR-TKI icotinib in advanced mEGFR lung adenocarcinoma (LUAD) patients. We expect to further the understandings of molecular signatures predictive of EGFR-TKI efficacy and provide additional evidence for more precise identification of LUAD patients who are most likely to benefit from EGFR-TKI therapy.

Materials and methods

Study design and patients

We retrospectively selected 49 stage IIIB or IV LUAD patients with mEGFR of exon 19 deletions or 21 L858R mutations who had received icotinib from CONVINCE study (NCT01719536) [4]. EGFR mutations were detected by the amplification refractory mutation system (ARMS) assay and PFS was defined as the duration from recruit time point to disease progression identified by response evaluation criteria in solid tumor (RECIST V1.1) standard. Demographic characters including age, gender, smoking history, Eastern Cooperative Oncology Group (ECOG) performance status (PS), clinical stage and EGFR sensitizing mutation type were collected. Informed consent was obtained from all patients for the use of tissue and serum samples in CONVINCE study which had been approved by the Ethics Committee of National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences & Peking Union Medical College with Protocol #14-001/791.

A total of 49 patients were stratified into good-responders with PFS longer than 11 months and poor-responders with PFS shorter than 11 months. At baseline, the serum samples were collected from 44 patients for targeted metabolites profiling analysis, and the paraffin-embedded (FFPE) tissue samples were collected from 19 patients for target sequencing. Fourteen patients of the total 49 provided both serum samples and FFPE samples as shown in Figure 1.

Figure 1.

Flowchart of the study. PFS, progression-free survival; LUAD, lung adenocarcinoma; mEGFR, epidermal growth factor receptor sensitizing mutation; NGS, next-generation sequencing.

Targeted metabolites detection with UPLC-ESI-Qtrap

Targeted metabolites in this study included amino acids (AA), fatty acids (FA), and lipid (LP). Serum samples were thawed on ice and prepared as follows: For extraction of AA, 10 µl serum sample was mixed with 30 µl acetonitrile. After vortex for 1 minute, the mixture was centrifuged at 13,200 rpm for 5 minutes at 4°C. 10 µl of the supernatant was transferred into a fresh tube and mixed with 10 µl ddH₂O. Then, 10 µl of the mixture was added with 70 µl Borate buffer (AccQ•Tag kit, Waters, Milford, MA, USA). After vortex for 1 minute, 20 µl of AccQTag reagent was added. The mixtures were heated at 55°C for 10 minutes and transferred to autosampler vials (DIKMA, Beijing, China) with a 250 µl insert tube for test. For extraction of FA and LP, 50 µl serum sample was added with 20 µl of the inter-standard solution (Avanti Polar Lipids, Alabaster, AL, USA) and 280 µl of chloroform/methanol (3:1; Mreda/Fisher Scientific, Waltham, MA, USA). Then the mixture was ultrasonicated at room temperature for 1 hour, added with 100 µl ddH₂O and thoroughly mixed. After centrifugation at 120,000 rpm for 10 minutes, the supernatant was dried using Centrifugal vacuum evaporator (Thermo Fisher Scientific, Waltham, MA, USA) for 4 hours at 4°C, added 100 µl of isopropyl alcohol/acetonitrile with 1:1 ratio (Fisher Scientific, Beijing, China) and dissolved in ultrasound (Kunshanhechuang, Shanghai, China). After centrifugation at 12,000 rpm for 10 minutes, 100 µL of supernatant was transferred to a 250 µl vial insert tube for test.

After that, ultra-high performance liquid chromatography (UPLC Waters, Milford, MA, USA) and Rapid Separation LC (RSLC, Shimazu LC-20AXR, Kyoto, Japan) was used to separate AA and FFA/LP, respectively. The mobile phases and gradient conditions of metabolites detection in liquid chromatography separation were shown in Figure S1. Then, multiple reaction monitoring (MRM) analysis was performed using Xevo TQ-S mass spectrometer for AA and 5500Qtrap mass spectrometer for FFA/LP (AB SCIEX, Boston, USA). The electrospray ionization (ESI) for AA and LP was positive mode (ESI⁺), while the negative mode (ESI^-) was used for FFA. The ion source temperature and capillary voltage were set to 150°C/2.0 kV for AA, 650°C/4.5 kV for FFA, and 650°C/5.0 kV for LP. The cone gas flow rate was 150 L/h, the desolvation temperature was 600°C and the desolvation gas flow was 1000 bar for AA detection. The system was controlled by the Masslynx Analysis software (SCIEX, Boston, MA, USA). The Skyline software (MacCoss, WA, USA) was used to analyze the raw data and analysis model was established according to their precise molecular weight and optimal retention time, while the integral of target metabolite peaks were assessed using analysis mode.

We used the quality control sample pooled from all the serum samples in the testing. 5 blank samples were used to balance the chromatographic column, and then 3 quality control samples were used to balance the column conditions. We inserted a quality control sample into injection after every 6-8 samples to monitor the stability and repeatability of the entire liquid-mass system.

DNA extraction, next-generation sequencing (NGS) and data processing

Genomic DNA was extracted from tissue using the QIAamp DNA FFPE Tissue kit (Qiagen, Valencia, CA, USA) and quantified using Qubit 3.0 Fluorometer and Qubit dsDNA HS Assay kit (Invitrogen, Carlsbad, CA, USA) according to the manufacturers’ instructions. The DNA samples were prepared for DNA library using The KAPA Hyper Prep kit (Kapa Biosystems, Wilmington, MA, USA) and captured by SureSelect XT Target Enrichment System (Agilent Technologies, Santa Clara, CA, USA) using 474 cancer-related genes panels for NGS according to the manufacturers’ protocols. The library concentration and fragment size were evaluated with Qubit 3.0 Fluorometer (Invitrogen) and 2100 Bioanalyzer (Agilent Technologies), respectively. DNA library was then sequenced on Illumina HiSeq-X10 platform (Illumina, San Diego, CA, USA).

FASTQ files of raw sequencing reads were generated using bcl2fastq Conversion Software (Illumina, Version: 2.17.1.14) for NGS data analysis. Low quality reads were filtered out and short reads were aligned to human reference genome (UCSC hg38) using BWA-MEM (Version: 0.7.17-r1188). Alignments were converted to BAM format using Samtools (Version: 1.4). PCR duplicates were removed using GATK Picard (Version: 2.18.0). Variants were then called using VarScan2 and Vardict (Version: 1.6.0). Variant annotation was done by Jannovar (Version: 0.20). Synonymous and intronic variants were subsequently filtered out. To identify high-confidence SNVs and indels, we only retained variants that were called by both VarScan2 and Vardict, and with the allele frequency (AF) ≥ 5% and variant read depth > 5. For known hotspot variants (EGFR L858R, 19del), a variant was retained if its AF ≥ 2%. To distinguish common population polymorphisms from somatic mutations, we further filtered out variants with population frequency > 0.5% in the gnomAD database.

Pathway enrichment analysis and interaction network analysis of co-mutant genes

Pathways enrichment analysis and interaction network analysis for co-mutated genes were conducted in Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway database (https://www.kegg.jp/kegg/pathway.html) and STRING (https://string-db.org/) database, respectively.

Building of metabolites discriminant models predictive of icotinib efficacy

The multivariate method partial least squares discrimination analysis (PLS-DA) implemented in the R package MetaboAnalystR was utilized to establish predictive model [4]. Moreover, the variable importance in projection (VIP) score was utilized to weigh the influence of each metabolite in the classification model. The permutations were repeated 1000 times and P≤0.05 was considered statistically significant. Besides, the receiver operating characteristic (ROC) analysis was conducted using the R package ROCR (R version 4.0.2; R Foundation for Statistical Computing; www.r-project.org). The missing value imputation and data normalization procedures were reached according to the R package MetaboAnalyst [23]. For deep learning modeling, AutoGenome package was used to search for net structures and hyperparameters for skip-connection neural networks and automatically train a model to predict patient PFS using metabolites profiles [24]. AutoGenome-built-in model-explanation function was used to weight the importance of each metabolite to the prediction. t-distributed stochastic neighbor embedding (t-SNE) algorithm was used to perform dimension reduction on patients’ metabolite profiles by R package tsne.

Network-based omics data integration analysis of metabolites and co-mutated genes

eResponseNet was used to build a causal-mediator network [25]. eResponseNet algorithm is a minimum-cost flow optimization algorithm, which finds the maximal flow from source nodes to target nodes in a given network while minimizes the overall cost of the network. We first obtained the corresponding IDs of metabolites in Human Metabolome (HMDB) Database and then searched the KEGG IDs from the online systems-level MetaboAnalyst analysis web site (https://www.metaboanalyst.ca/). We defined PFS-relevant mutated genes as upstream causal genes (source nodes) and genes of PFS-relevant metabolites pathways as downstream response genes (target nodes) respectively. The eResponseNet was then used to find corresponding mediator genes that links the upstream genes and the downstream metabolite pathways in a manually curated PPI reference network. Each of the upstream, downstream, mediator genes, and each edge connecting the genes, were assigned with a weight of flow value, reflecting the importance degree in the current causal-mediator networks. The causal-mediator networks were visualized by Cytoscape software. The PPI reference network consists of 238,897 interactions collected from Reactome FI, STRING, and HPRD databases. For STRING database, only high-confidence interactions with a confidence score ≥ 900 were used.

Statistical analysis

χ² test and Fisher’s exact test were used to compare differences in clinical and genetic characteristics between patients with different PFS. The survival package in R (https://cran.r-project.org/web/packages/survival/) was used to analyze survival data of patients stratified by gene mutations, while the Wilcoxon-test was used to analyze continuous variables. Survival curves were plotted using the Kaplan-Meier method and a log-rank test was used to assess the statistical significance of survival differences. Univariate and multivariate COX regression models were used to investigate the relationship between mutated genes, clinical characteristics and PFS.

Results

Clinical characteristics of patients

Out of the 49 LUAD patients, 24 were poor-responders and 25 were good-responders. The median PFS of patients in these two groups were 3.62 months and 19.34 months (HR, 5.63, 95% CI 2.53-12.53; P < 0.0001), respectively. The median age of all the patients was 56.4 years old (range: 41.3-69.7 years old). The majority of patients were female (65.3%) and non-smokers (73.5%). Moreover, 23 (46.9%) and 26 (53.1%) of patients had EGFR 19 deletions and 21 L858R mutations detected by ARMS assay, respectively. The distributions of age, gender, smoking history status, ECOG PS, EGFR sensitizing mutations, and clinical stage did not exhibit significant differences between the two groups (Table 1).

Table 1.

Clinical characteristics in LUAD patients

Characteristic	Patients		Good-responders		Poor-responders		P-value
	No.	Percent (%)	No.	Percent (%)	No.	Percent (%)
	49		25		24
Age							0.879
≥ 57 years	23	46.9	12	52.2	11	47.8
< 57 years	26	53.1	13	50	13	50
Gender							0.108
Female	32	65.3	19	59.4	13	40.6
Male	17	34.7	6	35.3	11	64.7
Smoking history							0.088
yes	13	26.5	4	30.8	9	69.2
no	36	73.5	21	58.3	15	41.7
ECOG-PS							1.000
0-1	47	95.9	24	51.1	23	48.9
≥ 2	2	4.1	1	50	1	50
Clinical stage							0.322
IIIb	10	20.0	7	0.70	3	0.30
IV	39	80.0	18	46.2	21	53.8
EGFR mutation types							0.062
19 deletions	23	46.9	15	65.2	8	34.8
21 L858R mutations	26	53.1	10	38.5	16	61.5

LUAD, lung adenocarcinoma; ECOG-PS, Eastern Cooperative Oncology Group performance status; EGFR, epidermal growth factor receptor.

Identification of targeted metabolites associated with the efficacy of icotinib using two discriminant models

44 pretreatment serum samples were subjected to targeted metabolomics profiling that covers 199 metabolites to explore the potential relevance of metabolites to the efficacy of icotinib. The 44 patients consisted of 22 poor-responders with a median PFS of 2.91 months and 22 good-responders with a median PFS of 19.34 months (HR, 5.33, 95% CI 2.34-12.17; P < 0.0001). We assigned the 44 patients into a training set of 30 patients and a testing set of 14 patients to establish and test metabolites models distinguishing good-and poor-responders.

Firstly, we built a discriminating model using partial least squares discrimination analysis (PLS-DA) and identified 27 differential metabolites with VIP score > 1.5 (Figure 2A, 2B and Table S1) in the training set. The PLS-DA scores plot showed a clear separation between good-responders and poor-responders with a model accuracy of 0.77 (the cross-validated R²=0.74 and pperm =0.016). Next, the ROC analysis revealed that the model distinguished good-responders from poor-responders in the testing set with a sensitivity of 0.8, a specificity of 0.67, and an area under the ROC curve (AUC) of 0.62 (Figure 2C). These 27 metabolites were mainly involved in linoleic acid, glycerophospholipid, and histidine metabolism pathways (Figure 2D). Among the 27 metabolites, the levels of 15 metabolites including 7 sphingomyelins (SMs), 3 ceramides (Cers), 2 phosphatidylcholines (PCs), 2 amino acids (AAs, L-β-Amino-N-butyric acid, 3-Methyl-L-Histi dine), and a fatty acid 18:0 (FA 18:0), were higher in poor-responders than those in good-responders, whereas 12 metabolites including 5 lysophosphatidylcholines (LPCs), 3 phosphatidylethanolamines (LPEs), 3 amino acids (L-tyrosine, ethanolamine, histamine) and FA 18:2 showed a contrary trend (Figure 2E and Table S1).

Figure 2.

Correlation between the 27 metabolites and the efficacy of icotinib using PLS-DA. A. PLS-DA scores plot showed the clustering of good-responders and poor-responders, S1-S30 represented 30 LUAD patients in the training set. B. 27 metabolites with a VIP score > 1.5. C. ROC curve to validate the capacity of the PLS-DA model to distinguish good- and poor-responders in the testing set including 14 LUAD patients. D. Enrichment metabolism pathways of 27 metabolites. E. Hierarchical clustering of 27 metabolites in good- and poor-responders. The top columns in red and cyan represented good-responders and poor-responders, respectively. Generally, the level of metabolites above the horizontal black line was higher in poor-responders than that in good-responders, whereas the level of metabolites below the black line was higher in good-responders. LUAD, lung adenocarcinoma; PC, phosphatidylcholine; SM, sphingomyelin; Cer, ceramides; FA, fatty acid; PLS-DA, partial least squares discrimination analysis; LPC, lysophosphatidylcholine; LPE, phosphatidylethanolamine. Amino acids included L-β-amino-N-butyric acid, 3-methyl-L-histidine, Histamine, L-tyrosine, and ethanolamine.

Secondly, we built a deep-learning neural network model using the same cohorts through an AutoGenome auto-machine learning tool. Trained by data from 30 patients, the AutoGenome model achieved an AUC of 0.85 tested by 14 patients data in the testing set (Figure 3C). Among the top 80 metabolites ranked by feature importance contribution in the AutoGenome model’s prediction, 10 metabolites overlapped with the PLS-DA-filtered 27 metabolites (Table S2). Specifically, the 10 overlapped metabolites could best distinguish the good- and poor-responders visualized by tSNE, with the highest intra- vs. inter-similarity ratio (silhouette score =0.097) superior to 199, 80 and 27 metabolites (silhouette score =0.013, 0.010, and 0.074; Figure 3D).

Figure 3.

Deep-learning neural network model of AutoGenome and performance of the overlapped 7 metabolites in distinguishing good- and poor-responders. A. The net structure of a metabolite-to-PFS prediction deep learning model searched by AutoGenome. B. Silhouette scores of poor- and good-responders using 4 different sets of metabolites: all 199 metabolites, 80 AutoGenome-filtered metabolites, 27 PLS-DA-filtered metabolites and 10 metabolites overlapped in AutoGenome and PLA-DA. A high silhouette score indicates better distinguishing capacity between two groups using given features of metabolites. C. AUC of AutoGenome metabolite-PFS prediction model tested in 14 patients data. D. tSNE visualization of poor- and good-responders using 199, 80, 27 and 10 metabolites mentioned above. E. Boxplot of 7 metabolites significantly different in good- and poor-responders. Red and blue boxes represented good-responders and poor-responders. *P < 0.05, **P < 0.01, ***P < 0.001. F. ROC analysis showed AUC of 0.89 for 7 metabolites in discriminating good- and poor-responders. PFS, progression-free survival; LUAD, lung adenocarcinoma; tSNE, t-distributed stochastic neighbor embedding; ROC, receiver operating characteristic; AUC, area under the curve.

Seven of the 10 metabolites exhibited significantly difference between good- and poor-responders including 2 Cers (Cer 36:1-3, Cer 38:1-3), 2 LPCs (LPC 16:1, LPC 22:5-1), 2 SMs (SM 36:1-2, SM 42:2) and LPE 18:2. Specifically, the 2 LPCs and LPE 18:2 had elevated levels in good-responders, while the other 4 metabolites had higher levels in poor-responders (Figure 3E). The 7 metabolites in combination yielded an improved AUC of 0.89 in the ROC analysis (Figure 3F), and were used in the following omics integration analysis.

Association of concurrently mutated genes with the efficacy of icotinib therapy

Among the 49 patients, 19 patients were sequenced using a 474-gene NGS panel. The median PFS of 7 poor-responders and 12 good-responders were 6.84 months and 21.38 months, respectively (HR, 6.67, 95% CI 1.17-37.89; P < 0.0001). With regard to mEGFR, 11 samples had EGFR L858R mutations and 8 samples had EGFR 19 deletions, which was in concordance with ARMS detection. Good-responders had roughly identical numbers of the EGFR L858R mutations and 19 deletions, while the majority of the poor-responders (86%, 6/7) only had EGFR L858R mutations (Figure 4). We identified 239 concurrently mutated genes using a 5% allele frequency of mutations as cutoff under the quality control (Table S3). Kaplan-Meier survival analysis showed 28 mutated genes were related to PFS. We further performed a univariate COX regression analysis of 28 genes and the clinical characteristics (P < 0.05, log-rank test). 23 concomitantly mutated genes remained significantly correlated with PFS, while age, gender, clinical stage, EGFR sensitizing mutation types (19 deletions vs 21 L858R mutations), and smoking history displayed no significant association with PFS (Table S4). We then focused on 6 genes (ARID1A, BCR, PTCH1, RBM10, FANCD2, ARID1B) with the highest mutation frequency (16%) out of the top 10 genes relevant to PFS. Mutation of the 6 genes indicated shorter PFS in mEGFR LUAD patients (Figure 5). Given the knowledge of the different therapeutic response of patients with 19 deletions or 21 L858R mutations, we conducted COX multivariate analysis for the 6 genes by considering EGFR sensitizing mutation types. The result showed that BCR and FANCD2 remained significant (P=0.020 and P=0.016) and RBM10 was marginally significant (P=0.058), while the EGFR sensitizing mutation types remained no significant difference between the two groups (P=0.695) (Figure S2A).

Figure 4.

Heatmap of mutations identified in 19 LUAD patients with mEGFR. Heatmap of most frequent somatic mutations identified in this study. The top panel shows mutational burden for each sample, red for synonymous mutations, blue for non-synonymous mutations and green for frameshift mutations. The down panel shows groups of 12 good-responders (red) and 7 poor-responders (cyan), respectively. The left panel shows the number of mutations identified in each gene. The right panel shows the genes identified in each patients. LUAD, lung adenocarcinoma; mEGFR, epidermal growth factor receptor sensitizing mutation.

Figure 5.

Kaplan-Meir survival curves of PFS and 6 concurrently mutated genes in 19 LUAD patients with mEGFR and treated with icotinib. Cyan line indicates wild-type and red line indicates mutant type of genes. LUAD, lung adenocarcinoma; mEGFR, epidermal growth factor receptor sensitizing mutation; PFS, progression-free survival.

Furthermore, mutated BCR or RBM10 indicated shorter PFS in our cohort and also had a similar effect on overall survival (OS) in the TCGA datasets. In TCGA cohorts, NSCLC patients harboring both mEGFR and BCR/RBM10 mutations had a poorer OS (Figure S2B and S2C). Restricted to the details of EGFR-TKI therapy and OS, only 2 patients with both mEGFR and RBM10 mutations receiving EGFR-TKI treatment showed a poorer OS compared with the 37 patients with only mEGFR (Figure S2D). RBM10 and BCR mutations also tended to coexist with EGFR mutations in TCGA datasets (Figure S2E).

We also compared the KEGG pathways and STRING networks of uniquely mutated genes in good- and poor-responders. Compared with the good-responders, poor-responders were found to harbor more co-occurring mutant genes (50 vs 125) which were enriched in the MAPK signaling pathway, Ras signaling pathway, and Rap1 signaling pathway (Figure 6A and Table S5). PI3K-Akt signaling pathway and EGFR tyrosine kinase inhibitor resistance pathway were seen in both groups (Figures 6A and S3A). These signaling pathways have all been demonstrated to be closely related to EGFR-TKI resistance. Additionally, STRING network revealed EP300 and PIK3CA as the hub genes exhibiting the most interactions with other genes (Figures 6B and S3B).

Figure 6.

KEGG pathways and STING interaction network in poor-responders. A. KEGG pathways of concurrently mutant genes uniquely mutated in poor-responders. Red outlined frame highlighted important pathways relevant to EGFR-TKI resistance. B. STRING network of the uniquely mutated genes in poor-responders. EP300 in red frame as the hub gene of the STRING network.

The interrelation between the targeted metabolites and concurrently mutated genes

To acquire a better understanding of the possible mechanisms leading to the difference in efficacy, we integrated metabolites and concurrent gene mutations that were both related to icotinib efficacy from two perspectives. Firstly, we analyzed the correlation between the 6 efficacy-related mutated genes and 7 metabolites mentioned above in 14 patients with both eligible sequencing and metabolites data in this study. The median PFS of 5 poor-responders and 9 good-responders was 9.18 months and 21.38 months (HR, 6.07, 95% CI 0.74-49.91; P=0.001). We found that different levels of 4 metabolites (Cer 36:1-3, Cer 38:1-3, LPC 16:1, SM 36:1-2) associated with the mutation status of 4 genes (ARID1A, ARID1B, BCR, and RBM10) (P < 0.05, Table S6). Cer 36:1-3 levels significantly increased in the ARID1A and BCR mutant group, and Cer 38:1-3 only rose in the ARID1A mutant group compared to the wild-type group. Patients with mutated ARID1A, BCR, and RBM10 genes had higher levels of SM 36:1-2, while patients with wild-type of the 3 genes had higher levels of LPC 16:1 (Figure 7A and Table S6). These results suggested that the mutant genes may have an impact on metabolism. Herein, the 4 metabolites and 4 concurrently mutant genes were referred to as our target metabolites and target genes, respectively.

Figure 7.

The interrelation between 4 target metabolites and 4 concurrently mutated genes. A. Boxplot shows the difference in the levels of Cer 36:1-3, Cer 38:1-3, SM 36:1-2, LPC 16:1 between wild-type and mutant type of ARID1A, ARID1B, BCR and RBM10 genes. Red and blue boxes represented mutant and wild-type type of the 4 genes, respectively. *P < 0.05. B. The interrelation of upstream mutated genes (ARID1A, ARID1B, BCR and RBM10) and downstream metabolic genes in sphingolipid metabolism pathway by causal-mediator network analysis. C. The interrelation of upstream mutated genes (ARID1A, ARID1B, BCR and RBM10) and downstream metabolic genes in glycerophospholipid metabolism pathway by causal-mediator network analysis. The red outlined rectangles represented 4 upstream concurrently mutated genes, blue outlined ellipse represented mediator genes, red outlined triangles represented downstream metabolic genes and black edges denoted the connections between genes.

Secondly, we performed a systematic causal-mediator network analysis to further explore the genes-metabolites relationship based on interactome databases. We obtained the KEGG IDs of 4 metabolites (Cer 36:1-3, Cer 38:1-3, LPC 16:1, and SM 36:1-2) using the MetaboAnalyst tool, queried the corresponding metabolic pathways and abstracted all the metabolic genes involved in the pathways within KEGG database. We denoted the 4 concurrently mutated genes of interest (ARID1A, ARID1B, BCR, and RBM10) as upstream nodes and metabolic genes modulating the biological processes of the 4 metabolites as downstream nodes to track any possible inner interaction between them. The 4 concurrently mutated genes were connected to metabolic genes in glycerophospholipid metabolism pathway and sphingolipid metabolism pathway through mediator genes (Figure 7B and 7C). In general, ARID1A connects to CTNNB1 (catenin beta 1)/CREBBP (CREB binding protein) and EP300 (E1A binding protein p300), ARID1B interacts with SMAD9 (SMAD family member 9) and ZNF263 (zinc finger protein 263)/CUL2 (cullin2), BCR interacts with PLCG1 (phospholipase C gamma 1) and GRB2 (growth factor receptor bind protein 2), and RBM10 interacts with PRCC (proline-rich mitotic checkpoint control factor) mediator genes to eventually impact the following downstream metabolic genes in the 2 metabolic pathways. Overall, the testification in our cohort and in interactome databases convinced us of the correlation between mutated genes and targeted metabolites.

Discussion

EGFR-TKI is an important treatment for advanced NSCLC patients with mEGFR, but the molecular heterogeneity of NSCLC leads to different clinical responses [5,6,26]. There are currently no validated biomarkers addressing the issues related to the difference of EGFR-TKI efficacy for patients with mEGFR. Recent studies are principally focused on a single perspective of gene variations or protein alterations, to explore mechanisms for this therapeutic responses divergence [14,16]. In this study, we revealed that targeted metabolites and concomitantly mutated genes with mEGFR were both relevant to the efficacy of icotinib treatment, which provides a promising strategy in distinguishing patients subgroups and may facilitate mEGFR to improve the personalized therapy.

Altered metabolites in tumor cells could be potential biomarkers for cancer diagnosis and prognosis because of their close relationship to phenotypes and direct reaction to stimuli from the microenvironment [12]. The fact that the sensitivity of tumor cells to anti-cancer drugs is closely related to metabolic perturbation is well-acknowledged. Meanwhile, metabolites are stable and can be quantitatively detected via non-invasive serum samples, thus making them more convenient potential prognosis biomarkers. Previous studies have demonstrated that a plurality of metabolites could assist in the diagnose of NSCLC or multiple metabolites models that can predict the efficacy of chemotherapy, yet there is a lack of relevant studies on metabolites profiling derived directly from patients pretreatment serum samples to explore their relationship to EGFR-TKI treatment [27,28]. In this study, we profiled targeted metabolites in mEGFR LUAD patients and established discriminative models by both traditional PLS-DA analysis and deep learning-based AutoGenome analysis. Our results revealed that metabolites based models possessed the potency to distinguish between different responders to icotinib treatment with moderate to good capacity (AUC of PLS-DA:0.62, AUC of AutoGenome: 0.85). It’s worth noting that the overlapped 10 metabolites contributed to both models in enhancing the confidence of their underlying relevance to the efficacy of icotinib.

More importantly, 7 out of the 10 overlapped metabolites exhibited a significant difference between good-responders and poor-responders, which were mainly involved in the lipids metabolism including Cers, LPCs, LPEs, SMs and FAs. Notably, 7 metabolites in combination further improved the distinguishment of different responders. Among them, Cers as a bioactive sphingolipid can be hydrolyzed from SMs and plays dichotomous roles in inducing apoptosis or implicating cancer cell proliferation [29]. In this study, increased Cer 36:1-3 and Cer 38:1-3 associated with poor-responders, which may preferably be associated with the tumor-promoting function of Cers. Besides, higher Cers levels produce more sphingosine-1-phosphate (S1P) which excites FAK, SRC oncoproteins and arousing MAPK/ERK, PI3K/AKT pathways activation to promote tumor cell proliferation, invasion and migration. SMs do not only serve as an essential constituent of the cell membrane but also as the sources of secondary messengers like Cers, sphingosine(Sph) and S1P, which are involved in many critical signal transduction pathways [30]. Correctly arraying SMs across the cellular membrane at certain concentrations is essential for Ras protein localization and normal functioning [31]. Elevated SM contents may be an indication of adaption to the overactivation of EGFR-Ras-ERK signaling pathway to facilitate tumor proliferation, which could explain the increased SM 36:1-2 and SM 42:2 levels in poor-responders. LPCs are a lysophospholipid intermediate in the biosynthesis and degradation of phosphatidylcholine (PC). Various studies observed that the plasma levels of LPCs were decreased in advanced metastatic cancer patients. Likewise, lower LPC 16:1 concentrations were detected in poor-responders perhaps as a result of its well-known involvement in cell proliferation, migration, and angiogenesis [32]. LPEs is hydrolyzed from phosphatidylethanolamine and functions in stimulating the calcium signaling pathway [33]. The levels of LPCs in lung cancer patients have been reported to be controversial. In this study, poor-responders had reduced levels of LPC 16:1 which indicated the deregulation of LPE relevant signaling transduction turbulence linked to cancer development. Accumulated evidence has been elucidated the important role of fatty acids in cancer progression. A higher level of FA 18:0 in patients with shorter PFS complies with the biological effect of FA affecting the proper localization and function of EGFR and STAT3/NF-κB-cyclin D1/survivin axis [34].

Several studies have demonstrated that co-existing somatic gene alterations can affect EGFR-TKI efficacy [21]. In this study, we focused on 6 mutated genes accompanied with mEGFR, which were correlated with a shorter PFS. Among them, BCR and RBM10 were also shown to affect OS in TCGA cohort. BCR gene, originally identified in the BCR/ABL complex in chronic myeloblastic leukemia, is a serine/threonine kinase that may activate RAS/RAF/MARK signaling pathway and promote DNA synthesis [35]. RBM10 is a LUAD tumor suppressor that mutated in 16% patients within our cohort which is similar to its prevalence of 5%-20% in other studies indicating its mutation co-occurred with EGFR, PIK3CA, and KRAS mutations [36-39]. It has been demonstrated that RBM10 mutations disrupted alternative splicing of NUMB (a Notch pathway regulator) and are closely related to EGFR, MAPK, PI3K-AKT, and RAS signaling pathways, hence promoting the growth of lung cancer cells [40-42]. RBM10 mutations tended to be associated with low RBM10 expression levels and a poorer prognosis in LUAD patients [37], which is consistent with the results of this study. ARID1A and ARID1B both encode key proteins in the SWI/SNF chromatin remodeling complex and their loss-of-function because of mutations promotes carcinogenesis [43]. PTCH1 encodes a protein that participates in the Hedgehog signaling pathway, which has been implicated in the regulation of cell proliferation, differentiation, and associated with the sensitivity to chemotherapy [44,45]. Similarly, we hereby observed that PTCH1 mutations are associated with a shorter PFS in icotinib therapy. FANCD2 encoding protein plays an important role in DNA cross-linking damage repair and the pathogenesis of Fanconi anemia (FA). Mutant EGFR is epistatic with FANCD2 and EGFR signaling may be linked to altered FA/BRCA functions leading to increased responses to chemotherapy [46]. However, the exact mechanism through which the coexistence of mEGFR and mutated FANCD2 influences patients’ response to EGFR-TKI therapy remains to be investigated. Simultaneously, concurrently mutated genes in poor-responders were enriched in PI3K-AKT, MAPK, Ras, Rap1 signaling pathways which are involved in the drug resistance of EGFR-TKI therapy [47-49]. Rap1 can be activated downstream by different receptor tyrosine kinases (RTKs) including EGFR, FGFR, VEGFR, IGFR, and MET [50,51], suggesting that RTKs bypass signaling induces resistance to EGFR-TKIs [52]. Additionally, we identified EP300 as the hub of the gene interaction network, which is consistent with its important role in transcription of target genes and is closely involved in the resistance of EGFR-TKI [53].

Importantly, the interrelation between concurrently mutated genes and metabolites from two aspects to better illustrate their correlation to the prognosis of LUAD patients. Firstly, the statistical associations of metabolites (Cers, LPCs, and SMs) and target genes (ARID1A, ARID1B, BCR, and RBM10) were depicted in this study. The following results were in concordance with each other and strengthened the confidence of their relationship to icotinib efficacy: (1) Elevated Cers/SMs and reduced LPCs were related to poor response; (2) Mutant type of the 4 genes were associated with poor-response; (3) Increased Cers/SMs and decreased LPCs were also observed in patients with the mutant type of the 4 mutated genes; Secondly, we illustrated possible interaction mechanisms between the target genes and the target metabolites in glycerophospholipid metabolism and sphingolipid metabolism pathways by causal-mediator network analysis. The causal-mediator network is a manually curated PPI reference network, which enables us to explore the inner causal relationship with rich data supported by 238,897 interactions collected from Reactome FI, STRING, and HPRD databases. Mediator genes connecting upstream and downstream genes are well-known cancer-related genes involved in transcriptional regulation and signal transduction. Specifically, ARID1A links to EP300 and CTNNB1/CREBBP in the 2 metabolism pathways. EP300 and CREBBP encoding transcriptional coactivators and CTNNB1 encoding β-catenin, all mediate resistance to EGFR-TKI [53,54]. ARID1B interacts with SMAD9 and SMARCs(SWI/SNF family) which both could affect the transcription of their target genes by mediating TGF-β signaling transduction and chromosome remodeling, respectively [55]. BCR interacts with signal-transducing adaptor proteins GRB2 and phosphatidylinositol signaling/AKT1 survival pathway components PLCG1 to potentially regulate metabolic genes [56,57]. RBM10 connects to PRCC which is reported as fusion partners with the TFE3 gene in renal cell carcinoma [58]. Nevertheless, further investigation is required to elucidate the mechanism through which ARID1B, BCR, and RBM10 regulate the mediator genes and downstream metabolically active genes. Overall, these results suggested that the concurrently mutated genes interacted with genes participating in metabolism and contribute to the different clinical responses to icotinib therapy. Given the easy accessibility of blood samples, the results convinced us the potential of the 4 target serum metabolites (Cer 36:1-3, Cer 38:1-3, SM 36:1-2M, LPC 16:1) in predicting the efficacy before icotinib treatment.

In this study, despite the 3-4 years storage of FFPE and serum samples before testing, the possible decrease in sample quality may not be major concern in affecting the accuracy of these tests results because sequencing and metabolites testing underwent rigorous quality control. Besides, studies have demonstrated that long-term stored FFPE for 8-32 years still showed robustness in the NGS data of mutation-analysis [59,60]. Moreover, a study reported that the human plasma metabolome is adequately stable for long-term storage at -80°C for up to seven years [61].

The limitations of this study include the small sample size, serial specimen deficiency and lack of functional validation of experiments in vitro.

Conclusion

To the best of our knowledge, this is the first study to profile the targeted metabolites in non-invasive pretreatment serum samples from mEGFR LUAD patients who received icotinib treatment. Double-modeling-filtered lipid metabolites and concurrently mutated genes demonstrated the potential to predict the efficacy of icotinib, which may also provide insights into identifying which patients could benefit from icotinib therapy. The interrelation between these dual-level molecules also mutually confirmed their respective association with icotinib efficacy. Further studies are warranted to validate these findings.

Disclosure of conflict of interest

None.