Development and validation of a nomogram model based on blood-based genomic mutation signature for predicting the risk of brain metastases in non-small cell lung cancer

Abstract

Purpose

Available research indicates that the mammalian target of rapamycin complex 1 (mTORC1) signaling pathway is significantly correlated with lung cancer brain metastasis (BM). This study established a clinical predictive model for assessing the risk of BM based on the mTORC1-related single nucleotide polymorphisms (SNPs).

Methods

In this single-center retrospective study, 395 patients with non-small cell lung cancer were included. Clinical, pathological, imaging, and mTORC1-related single nucleotide polymorphism data were collected. Lasso regression was used to identify variables related to the risk of BM in lung cancer, and a nomogram was constructed. Internal validation was performed using 1,000 bootstrap samples. We plotted the receiver operating characteristic (ROC) curve and calculated the area under the curve (AUC). The calibration of the model was assessed using calibration curves and the Hosmer–Lemeshow goodness-of-fit test, and decision curve analysis (DCA) was plotted to evaluate the net clinical benefit.

Results

The nomogram's predictive factors included lung cancer histology, clinical N stage, CEA, neutrophil to lymphocyte ratio (NLR), lymphocyte to monocyte ratio (LMR), RPTOR: rs1062935, and RPTOR: rs3751934. The AUC of the model in the training set and internal validation were 0.849 and 0.801, respectively. The calibration curves and Hosmer–Lemeshow test both indicated a good fit.

Conclusion

The nomogram has practicality and efficacy in predicting the high risk of BM in lung cancer patients, confirming that single nucleotide polymorphisms in the mTORC1 pathway genes may be good predictors in clinical prediction models.

Background

Lung cancer, with the highest incidence among malignant tumors in China, is notorious for its strong propensity for brain metastasis (BM) [1]. The afflicted patients are often left with an average survival period of less than 6 months [2]. Given the high incidence and mortality rates associated with BM, endeavors targeting early detection and intervention will bring substantial benefits for patients with lung cancer BM.

Currently, existing literature has documented a strong correlation between BM of non-small cell lung cancer (NSCLC) and the mammalian target of rapamycin complex 1 (mTORC1) signaling pathway [3,4,5,6,7]. As a serine/threonine kinase, mammalian target of rapamycin (mTOR) integrates various signals from the extracellular environment and intracellular metabolic changes to regulate cell growth, proliferation, and metabolism. mTOR promotes various anabolic processes such as the biosynthesis of lipids, proteins, and organelles, and limits catabolic processes like autophagy, thus playing a key role in regulating cell growth and metabolism. mTOR also stimulate cell proliferation and growth, which is closely related to tumorigenesis by PI3K/ALK/mTOR signaling pathway. There are two forms of mTOR: mTORC1 and mTORC2, with mTORC1 consisting of mTOR, RPTOR, mLST8, and multiple non-core subunits [8, 9]. Due to its key role in sensing growth factor signals and regulating cell growth, gene variations in mTORC1 may influence tumor development [10]. Single nucleotide polymorphism (SNP) is one of the most common genetic variations, mainly referring to the diversity of DNA sequences caused by variations of a single nucleotide at the genome level. mTORC1-related single nucleotide polymorphisms can lead to differences in mTOR gene expression. A study on a Chinese population, including 1,125 gastric cancer patients and 1,196 cancer-free controls, recorded an association between gastric cancer risk and mTORC1-related SNPs [11]. Our previous research on lung cancer also reported a correlation between SNPs in the mTORC1 pathway and BM in lung cancer patients. Further studies revealed that mTOR may regulate sphingolipid metabolism by modulating the ceramide signaling pathway, thereby affecting the progression of BM in lung cancer [5, 7]. However, despite these findings establishing a close link between the mTORC1 pathway and lung cancer BM, its clinical application as a biomarker for predicting BM risk in lung cancer patients has not been well established.

To date, existing prediction models have mainly focused on BM and clinical characteristics in advanced NSCLC patients [12]. However, predictive models for BM risk in early-stage NSCLC populations remain scarce. A nomogram is a graphical prediction tool that calculates the probability of various diseases. Compared with the traditional TNM staging system, it is more accurate for personalized prediction of BM risk in lung cancer. Therefore, based on a single-center NSCLC database, this study uses this tool combined with mTORC1 signaling pathway SNPs to construct a risk prediction model for BM in lung cancer. The aim is to explore the value of mTORC1 pathway-related gene SNPs in predicting the risk of BM in lung cancer.

Materials and methods

The construction and validation of the models observed the TRIPOD guidelines [13].

Data source

Data of NSCLC patients who visited Fujian Provincial Hospital were collected from May 2015 to October 2017. NSCLC patients who were histopathologically diagnosed and consented to blood sample collection for DNA analysis were included and those with secondary lung cancer or with unclear histopathological features were excluded. Each enrolled patient signed a written informed consent form and contributed 3 ml of peripheral blood for blood sample collection and genomic DNA extraction. The study protocol was reviewed and approved by the Medical Ethics Committee of Fujian Provincial Hospital (Ethics Approval Number: K2017-11–006).

The demographic data, including age, gender, body mass index, were collected using structured questionnaires (see Supplementary Material, questionnaire). Clinical pathological data, including disease stage, depth of invasion, lymph node and metastasis status, and treatment regimens, were obtained from electronic medical records. Pre-treatment imaging results, serum tumor markers, complete blood counts, and biochemical results were also collected for analysis. NSCLC staging was based on the TNM criteria of the American Joint Committee on Cancer (AJCC; 8th edition, 2017). The diagnosis of BM was confirmed by computed tomography or magnetic resonance imaging.

Among all of 530 patients included in this database, 135 patients were excluded: 19 due to incomplete disease staging data, 20 due to a lack of BM information, and 96 due to the absence of blood test data. Therefore, 395 patients was finally enrolled for the subsequent analyses.

Polymorphism selection

In order to select common and potentially functional SNPs within three key genes (mTOR, mLST8, and RPTOR) of the mTORC1 signaling pathway in the Chinese Han population, we utilized the publicly available HapMap SNP database (http://www.ncbi.sg) and SNPinfo (http://snpinfo.niehs.nih.gov/). The selection criteria included: a minor allele frequency of ≥ 5%, a linkage disequilibrium coefficient (r²) of < 0.8, a location within the regulatory regions of the gene, and yielding impact on the binding sites of transcription factors or microRNA.

Genomic DNA isolation and genotype analysis

As previously reported [2], the genomic DNA from the peripheral blood samples was extracted with a EasyPure Blood Genomic DNA Kit (EE121, TransGen Biotech) according to the manufacturer's protocol and preserved at −80 °C. For genetic typing, primers for eight SNPs were designed (Table S1). Allele-specific extension products were detected by matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF–MS). The data were analyzed using Mass ARRAY platform (Agena Bioscience) and Sequenom TYPER software (version 4.1).

To assess data reproducibility, 5% of DNA samples were subjected to blind, random, and duplicate analyses, which reported a reproducibility rate of 99%.

Missing data and data processing

Due to random missing data, multiple imputation was performed to supplement missing data for BMI, Ki67, and CEA (Fig. S1). The convergence of the model was assessed by comparing the distribution of observed values with that of estimated values through convergence plots (Fig. S2). Among the 20 imputed datasets generated, the best set of data was selected for subsequent analysis based on the Bayesian Information Criterion (BIC). Statistical tests were performed for both the original and imputed datasets, as presented in Table S2 [14] neutrophil to lymphocyte ratio (NLR), platelet to lymphocyte rate (PLR), albumin to alkaline phosphatase ratio (AAPR), and lymphocyte to monocyte ratio (LMR) were dichotomized according to the optimal cutoff points determined from the receiver operating characteristic (ROC) curve (Table S3).

Candidate predictive variables and variable selection

Factors known to be associated with lung cancer BM were selected as potential variables, including gender, age, weight, tumor type, stage, chemotherapy status, and CEA. Additionally, we incorporated novel biomarkers, such as PLR, AAPR, NLR, and LMR. The least absolute shrinkage and selection operator (Lasso) regression was employed to screen potential predictors in the training set, and a multiple logistic regression model was constructed.

Model validation

The predictive performance of the model was evaluated by the area under the ROC curve (AUC). The goodness of fit of the model was assessed by the Hosmer–Lemeshow test and calibration curves. The internal validation of the model was performed using 1,000 bootstrap samples. The clinical utility of the model was analyzed by the decision curve analysis (DCA) to assess the practicality of the nomogram and net benefit was quantified across different probability thresholds. Additionally, the integrated discrimination improvement (IDI) were calculated to assess the predictive performance of the two models.

Statistical analysis

All statistical analyses were performed using IBM SPSS Statistics (version 26.0) and R software (version 4.1.3). Continuous variables with a normal distribution were expressed as mean ± standard deviation and compared by independent t-tests. Non-normally distributed continuous variables were expressed as medians with interquartile ranges and compared by the Mann–Whitney test. Categorical variables were presented as counts and percentages and compared by the chi-square test.

Redundant features were eliminated using the "caret" package and clinical parameters were selected by LASSO logistic regression using the "glmnet" package. The AUC was calculated using the "pROC" package. The models were evaluated by the Hosmer–Lemeshow test using the "ResourceSelection" package and decision curve analysis was conducted using the "rmda" package.

Results

Study population

This study included a total of 395 patients from Fujian Provincial Hospital, involving 222 males and 173 females, with an average age of 59.2 years old. The patient inclusion process was detailed in Fig. 1. Among them, 74 individuals were diagnosed with lung cancer BM, while the remaining 321 reported no BM (Table 1). Due to the differing incidence of BM among various subgroups of NSCLC, we conducted a subgroup analysis based on different pathological types (Table S4). The results showed that patients with the pathological type classified as "other" (including 6 cases of large cell carcinoma and 2 cases of adenosquamous carcinoma) had the highest proportion of BM, followed by those with lung adenocarcinoma and squamous cell carcinoma. There were no significant differences among the three pathological types in terms of BMI, tumor M stage, CEA levels, NLR, PLR, and LMR. However, differences were observed in gender, Ki-67, tumor T stage, tumor N stage, clinical stage, chemotherapy status, and AAPR.

Fig. 1

Flowchart of the process of patient enrollment, model development and validation

Table 1 Association between clinical variables and BM

Correlations among variables

Before variable selection, a correlation analysis was performed for the variables using Spearman correlation to ensure no multicollinearity among the selected variables. The results of the correlation analysis were shown in Fig. S3.

Correlation between snps and BM

The univariate logistic regression was employed to explore the correlation between 8 SNPs genes and BM in the NSCLC patients, which yielded a P value of < 0.05 for RPTOR:rs1062935, and RPTOR:rs3751934, respectively. The multifactor logistic analysis showed that RPTOR:rs1062935 and RPTOR:rs3751934 was correlated with lung cancer BM (Table 2). We further conducted a subgroup analysis based on different pathological types and found that RPTOR:rs1062935 and RPTOR:rs3751934 were still closely associated with BM in the lung adenocarcinoma population. In the lung squamous cell carcinoma population, univariate logistic analysis revealed no significant association for RPTOR:rs3751934 (P = 0.076). Due to the small sample size of BM patients in this group, multivariate logistic analysis could not be performed. Additionally, we did not conduct subgroup analyses for other pathological types because of the limited sample size (Table S5).

Table 2 Correlation between diferent genotypes of genes in MTORC1 pathway and brain metastasis

Predictor selection

We included BMI, gender, age, Ki67, tumor stage (including T, N), histology, chemotherapy status, CEA, NLR, PLR, AAPR, and LMR as variables for predictor selection. LASSO logistic regression was performed to identify significant predictive features influencing the risk of lung cancer BM. Based on one standard error of the minimum lambda value (0.0458), five non-zero coefficient features were selected: histology, clinical N stage, CEA, NLR and LMR (Fig. 2).

Fig. 2

Predictor selection using the least absolute shrinkage and selection operator logistic regression. a The binomial deviance is plotted versus log(λ). The black vertical lines are plotted at the optimal λ based on the minimum criteria and one standard error for the minimum criteria. b The LASSO coefficient profiles of the 13 clinical features. A coefficient profile plot is produced versus the log (λ)

Development and verification of nomogram

Based on the above results, we constructed a nomogram that includes seven predictive factors: RPTOR: rs1062935, RPTOR: rs3751934, lung cancer histology, clinical N stage, CEA, NLR, and LMR (Fig. 3a). The model had an AUC of 0.849 in the training set (Fig. 3b), and the calibration AUC after 1,000 bootstrap internal validation was 0.801, with an accuracy of 0.806 and a Kappa value of 0.260. The Hosmer–Lemeshow test for the model showed a Chi-square of 6.118545 and a P-value of 0.73. We also plotted the calibration curve (Fig. 3c), indicating that the model has a good fit. We plotted the DCA for the model (Fig. 4). Our results indicate that the model demonstrates good clinical utility at probability thresholds ranging from 0.03 to 0.60. IDI is commonly used to evaluate performance differences among various models. We further constructed a model without SNPs (Fig. S4) and found that the IDI for the two models was 0.0222 (95%CI 0.001–0.044), with a P-value of 0.04178, suggesting that including SNPs can improve the model's predictive performance.

Fig. 3

Establishment and validation of a nomogram risk prediction model for BM in NSCLC. a The nomogram for predicting the risk of brain metastasis in NSCLC patients. To use it, draw a vertical line upward from each variable and record the corresponding score. Finally, sum all the scores, and the vertical projection from Total Points to Risk of BM corresponds to the predicted probability of brain metastasis in lung cancer. The predicted low and high values of CEA are interpreted as < 5 ng/ml and ≥ 5 ng/ml, respectively. Low and high NLR values are interpreted as < 5.021 and ≥ 5.021, and low and high LMR values are interpreted as < 1.589 and ≥ 1.589. b The ROC curve of the model, with an area under the curve (AUC) of 0.849 (95% CI 0.802–0.896). c The calibration plot of the nomogram. The horizontal axis represents the predicted probability, and the vertical axis represents the actual probability. A perfect prediction corresponds to the 45° dashed line. The red and blue lines represent the observed (apparent) performance of the nomogram before and after bootstrapping (Hosmer–Lemeshow test: P = 0.73)

Fig. 4

Decision curve analysis of the nomogram. In this analysis, the decision curve provided a larger net benefit across the range of 0.03 to 0.60

Discussion

BM significantly worsen the prognosis of lung cancer patients; therefore, early identification of populations at high risk for potential BM in lung cancer can help improve patient outcomes. Currently, prediction models for BM in NSCLC are primarily based on routine clinical data (such as peripheral blood indicators and tumor markers), which can be easily influenced by various factors including patient treatment and infections. In this study, we constructed a clinical prediction model that selects predictive factors from multiple dimensions, including genetic mutation information, treatment, systemic inflammation status, and pathological conditions. This ensures that the model has greater accuracy in clinical applications.

The predictive factors in our model include lung cancer histology, clinical N stage, CEA, NLR, and LMR. Among lung cancer histologies, other types (mainly large cell lung cancer) and adenocarcinoma serve as strong predictors of BM in lung cancer. This aligns with previous reports; a study based on the SEER database indicated that the risks of BM in NSCLC patients were 11% for adenocarcinoma, 6% for squamous cell carcinoma, and 12% for large cell carcinoma. Given that the majority of cases classified as "other" in this study were large cell carcinoma, this histological type contributed most significantly to the model [15]. In our study, the N stage of lung cancer is a strong predictor of BM, consistent with Wang's research, which found statistically significant differences in the incidence of BM based on the number of mediastinal lymph node metastases (< 4, 4–6, > 6) [16]. A higher N stage often indicates a greater likelihood of distant tumor metastasis. This study also found that high NLR and low LMR are associated with lung cancer BM. NLR and LMR are systemic inflammation markers, and chronic inflammation is considered one of the characteristics of tumors. Notably, a study by Young Wha Koh. on 260 cases of stage IV NSCLC found that a high NLR (≥ 4.95) was associated with BM, which contradicts our conclusions and may be influenced by the presence of some extracranial metastases in our non-BM group [17, 18]. Given that some clinical markers may be confounded by other factors, this study attempts to incorporate mTOR SNP into the model based on our previous research indicating that RPTOR promotes BM and that mTOR gene polymorphisms are closely related to BM, aiming to elucidate the clinical value of mTOR SNP as a predictive factor for NSCLC BM.

The two SNPs, RPTOR: rs1062935 and RPTOR: rs3751934, are located in RPTOR. RPTOR, located on chromosome 17q25.3, acts as a scaffold recruiting substrate to mTOR kinase, thereby regulating mTOR activity [19]. Several studies have indicated that mTOR and RPTOR polymorphisms are associated with cancer risk stratification. In a study of 803 bladder cancer patients, seven polymorphisms in the RPTOR gene, including rs1062935, are associated with an increased cancer risk [20]. Our previous research has discovered a correlation between SNPs in the mTORC1 signaling pathway and lung cancer BM [2]. Due to the well-known regulation of energy metabolism by the mTOR signaling pathway, the component RPTOR within mTORC1 can promote BM by regulating sphingolipid metabolism to generate S1P, aligning with a Harvard gene profiling study of over 500 tumor cell lines, which reveals a potential link between abnormal intracranial sphingomyelin metabolism and BM. These findings highlight the feasibility of incorporating more SNPs to enhance the efficacy of clinical prediction models for lung cancer BM.

In our study, subgroup analysis based on pathological type showed that univariate logistic analysis did not find a significant relationship between rs3751934 and BM in the lung squamous cell carcinoma population. Given that there were 72 cases of lung squamous cell carcinoma, with only 5 cases of BM, it is likely that the small sample size of lung squamous cell carcinoma caused bias. Additionally, the biological characteristics distinguishing lung squamous carcinoma from lung adenocarcinoma may account for these differences, but further testing is needed for confirmation. Finally, we found through the IDI test that incorporating RPTOR: rs1062935 and RPTOR: rs3751934 as predictive factors further optimized the predictive performance of the model compared to the model without SNPs. This suggests that detecting mTOR signaling pathway gene polymorphisms in peripheral blood has value in enhancing the prediction of BM risk.

However, there were several limitations remain in our study. Firstly, as a single-center retrospective study, there might be a selection bias as patients who did not undergo cranial imaging were excluded. Secondly, the models were only subjected to internal validation, without an external validation using data from other institutions. Thirdly, the sample size in our study was relatively small. Therefore, a large sample size is needed to validate the accuracy of the model.

In conclusion, to our best knowledge, this study first incorporates mTORC1-related SNPs from the peripheral blood to construct a clinical prediction model for lung cancer BM, which has yielded a superior predictive efficacy. The findings suggest that mTORC1-related SNPs can potentially serve as important predictive factors for early detection and intervention of lung cancer BM.