Magnetic resonance imaging-based machine learning radiomics predicts CCND1 expression level and survival in low-grade gliomas

doi:10.21203/rs.3.rs-3243686/v1

Download PDF

Research Article

Magnetic resonance imaging-based machine learning radiomics predicts CCND1 expression level and survival in low-grade gliomas

https://doi.org/10.21203/rs.3.rs-3243686/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Low-grade glioma (LGG) is associated with increased mortality owing to the recrudescence and tendency for malignant transformation. Therefore, novel prognostic biomarkers must be identified as the current traditional prognostic biomarkers of glioma, including clinicopathological features and imaging examinations, are unable to meet the clinical demand for precision medicine. Accordingly, we aimed to evaluate the prognostic value of cyclin D1 (CCND1) expression levels and construct radiomic models to predict these levels in patients with LGG. A total of 412 LGG cases from The Cancer Genome Atlas (TCGA) were used for gene-based prognostic analysis. Using magnetic resonance imaging (MRI) images stored in The Cancer Imaging Archive with genomic data from TCGA, 149 cases were selected for radiomics feature extraction and model construction. After feature extraction, the radiomic signature was constructed using logistic regression (LR) and support vector machine (SVM) analyses. Involved in the regulation of the cell cycle and immune response, CCND1 was identified as a differentially expressed prognosis-related gene in tumor and normal samples. Landmark analysis revealed that high expression levels of CCND1 were beneficial for survival (P < 0.05) in advanced LGG. Four optimal radiomics features were selected to construct radiomics models. The performance of LR and SVM achieved areas under the curve of 0.703, 0.705, as well as 0.724 and 0.726 in the training and validation sets, respectively. CCND1 expression levels could affect the prognosis of patients with LGG. MRI-based radiomics can serve as a novel tool for predicting the prognosis.

lower-grade glioma

CCND1

radiomics

Magnetic resonance imaging

prognosis

bioinformatics analysis.

Gliomas, characterized by poor prognosis, are highly aggressive and the most prevalent central nervous system tumors [1]. According to the classification of gliomas (grade Ⅰ–Ⅳ) by the latest 2021 World Health Organization (WHO) Classification of Tumors of the Central Nervous System [2], adult low-grade gliomas (LGGs, grades II and III), which harbor an isocitrate dehydrogenase (IDH) mutation and a 1p/19q-codeletion, grow slowly because of their low potential for malignant transformation [3]. Current treatments administered to patients include surgery, chemotherapy, radiotherapy, and immunotherapy. However, disability and mortality are relatively high owing to the recrudescence and tendency for malignant transformation [4; 5]. The classical prognostic biomarkers of gliomas are clinicopathological features, Ki67, IDH, and imaging examinations, such as computed tomography (CT) and magnetic resonance imaging (MRI), which cannot satisfy the clinical demand for precision medicine. Therefore, new prognostic biomarkers must be identified, and patients must be stratified to identify novel indicators for individualized precision treatment.

Cyclin D1 (CCND1), the core gene of the cell cycle, is a frequently deregulated and relevant biomarker in some tumors [6; 7]. In the early G1 phase, CCND1 is either a promoter of the cell cycle during the G1-S phase[8] or a growth factor sensor. Cyclin-dependent kinase 4/6 (CDK4/6) activity is required for the cell cycle G1/S transition. CCND1 binds to CDK4/6 to form a complex, which can phosphorylate the tyrosine residue 172 of CDK4/6 and activate the kinase activity of CDKs [9]. CDK4/6 inhibitors (CDK4/6i) act against tumors, restrain Rb phosphorylation, and prevent cell cycle progression from G1 to S phases [10]. Consequently, the cells became senile and apoptotic. In clinical trials (https://clinicaltrials.gov/ct2/home), CDK4/6i was identified as a suitable treatment for gliomas. Hence, CCND1 may be a prognostic or therapeutic biomarker for LGG.

Radiomics data are a high-throughput “image sequence” used to obtain a large number of imaging parameters, as well as a non-invasive, dynamic, and quantitative method to reflect tumor characteristics [11]. CT, MRI, or digital radiography combined with other –omics data and clinical characteristics into radiomics could benefit clinical decision-making, disease testing, assessment of therapeutic response, or determine a suitable conjecture for the possibility of disease progression and recurrence, death, and therapeutic schedule[12; 13]. Notably, radiogenomics is an emerging field that explores the relationship between genomics and radiomics [14; 15]. Using a deep learning method to analyze MRI radiomics data, scholars have confirmed that immunophenotype-associated mRNA signatures (IMriskScore) can offer personalized immunotherapy protocols and estimate the 1p/19q and IDH status [16] and patient prognosis [17–19]. However, to date, we lack studies demonstrating the use of radiomic features to predict CCND1 expression levels.

Therefore, we aimed to explore the relationship between CCND1 expression levels and the prognosis of patients with LGG using The Cancer Genome Atlas (TCGA) and The Cancer Imaging Archive (TCIA). Furthermore, based on magnetic resonance imaging, we constructed radiomics models to predict CCND1 expression levels, which could be used as a reference for clinical decisions and further studies.

Study design

This study was designed as a multi-factor exploratory study to probe and verify the heterogeneity of MRI radiomics features in LGG and reveal the underlying prognostic values of the biological mechanisms and heterogeneity (Fig. 1).

Data retrieval and processing for this study

Data for 199 LGG patients with enhanced MRI results (T1W1) were downloaded from TCIA database (www.cancerimagingarchive.net). The genomic data of 515 LGG patients were downloaded from the TCGA database (portal.gdc.cancer.gov). Genomic data from 149 patients in TCGA were used to extract radiomic features and construct the model, and data from 412 patients were used to analyze the gene-based prognosis. The R package “survminer” [20] was used to divide patients into high- and low-expression groups using a cutoff value of 4.55 for CCND1 expression. We prepared a radiomics model of CCND1 and explored its prognostic value.

The exclusion criteria were samples without survival data, samples without clinical data, samples with survival time < one month, samples with poor image quality, and samples without enhanced T1W1 images. The inclusion criteria were as follows: samples of WHO grades II and III and samples with clinical and genomic data from TCGA. The detailed inclusion and exclusion criteria are shown in Table 1.

Table 1

Clinical characteristics of the population with high and low CCND1 expression group
Variables	Total (n = 412)	Low (n = 242)	High (n = 170)	P value
Age, n (%)				1
~ 40	198 (48)	116 (48)	82 (48)
41~	214 (52)	126 (52)	88 (52)
Gender, n (%)				0.228
Female	183 (44)	101 (42)	82 (48)
Male	229 (56)	141 (58)	88 (52)
Histology, n (%)				0.684
Oligodendroglioma	158 (38)	93 (38)	65 (38)
Astrocytoma	154 (37)	87 (36)	67 (39)
Oligoastrocytoma	100 (24)	62 (26)	38 (22)
Grade, n (%)				< 0.001
WHO II	194 (47)	134 (55)	60 (35)
WHO III	218 (53)	108 (45)	110 (65)
IDH_status, n (%)				0.057
WT	80 (19)	55 (23)	25 (15)
Mutant	332 (81)	187 (77)	145 (85)
Chr_1p_19q_codeletion, n (%)				0.65
Non-Codel	273 (66)	163 (67)	110 (65)
Codel	139 (34)	79 (33)	60 (35)
MGMT_promoter_status, n (%)				0.024
Unmethylated	73 (18)	52 (21)	21 (12)
Methylated	339 (82)	190 (79)	149 (88)
Chemotherapy, n (%)				0.001
NO	173 (42)	118 (49)	55 (32)
YES	239 (58)	124 (51)	115 (68)
Radiotherapy, n (%)				0.24
NO	185 (45)	115 (48)	70 (41)
YES	227 (55)	127 (52)	100 (59)

From UCSC XENA (xenabrowser.net/datapages), the obtained GBMLGG (glioma LGG) RNAseq data in fragments per kilobase of transcript per million mapped fragments format were TCGA and GTEx (gtexportal.org/home) data processed by Using Toil flow unified [21]. To compare the samples, the data were transformed to transcripts per million reads format via log₂ and demonstrated in a boxplot using the R package “ggplot2” [22].

Survival analysis

Every variate was subjected to survival analysis using the R packages “survival” [23] and “survminer” [20]. Using Kaplan–Meier (KM) survival curve analysis and the log-rank test, we showed the changes and significance of survival rates among the different groups. A survival rate of 50% was defined as the median survival time.

Landmark analysis plotted the KM curve in time periods, comparing the 12, 18, ..., 78, and 84 months defined as “early stage” from diagnosis to time point, and “late stage” from time point to end of follow-up. In the KM curve analysis via landmarks, the abscissa is the survival time, and the ordinate is the mortality risk.

Univariate landmark analysis

Setting 12, 18, …, 78, and 84 months as time nodes after diagnosis of LGG, we used the R packages “jskm” (CRAN.R-project.org/package = jskm)and “survival” [23] to plot a KM curve via landmark analysis. In addition, “early grade” was defined from diagnosis to time nodes, and “advanced grade” was defined from time nodes to the end of follow-up.

Univariate and multivariable COX analysis

Cox proportional hazards regression can be used to explore the relationship between one factor or more factors and survival outcomes. We used univariate Cox regression to perform an association analysis to explore the factors influencing overall survival (OS). Multivariate Cox regression was used to explore whether one independent factor or various factors would influence OS. To identify significant factors, all meaningful variables identified using univariate Cox regression analysis [hazard ratio (HR) < 1] were subjected to multivariate analysis. We used the R packages “survival” [23] and “forestplot” (CRAN.R-project.org/package = forestplot) for these analyses.

Subgroup analysis and interaction test

Based on univariate Cox regression, we used exploratory subgroup analysis to analyze the impact of the main variable (high vs. low CCND1 expression) on patient prognosis in every covariate subgroup. Furthermore, we analyzed the interactions between CCND1 and other covariates using the log-likelihood ratio test. We used the “cmprsk,” “survival,” and “forestplot” R packages for these analyses.

Correlation analysis

Based on Spearman’s rank correlation coefficient, we explored the correlation between CCND1 and clinical characteristics and presented the results using a correlation heat map.

Analysis of the correlation between CCND1 expression and immune cell infiltration

The gene expression matrix of the LGG samples uploaded to the ImmuCellAI database (bioinfo.life.hust.edu.cn/ImmuCellAI#!) was used to measure immune cell infiltration in each sample [24]. Based on Spearman’s rank correlation coefficient, we explored the correlation between CCND1 expression and immune cell infiltration and presented the results using a correlation heat map.

Gene Set Variation Analysis (GSVA) between low and high CCND1 expression

GSVA is a non-parametric, unsupervised analysis method to evaluate gene set enrichment in microarrays and transcriptomes [25]. Based on the KEGG pathways and hallmark gene sets, GSVA was used to calculate the pathway enrichment score for each of the 412 LGG samples from TCGA-LGG. The R package “limma” [26; 27] was used for differential analysis of low and high CCND1 expression.

Screening of radiomics features

An open-source software (“pyradiomics” python package) [28] containing 107 radiomics factors (RFs) was used to analyze size, shape, intensity, morphology, and texture. Based on the CCND1 threshold value (4.554), the samples were divided into low- and high-expression groups. By intersecting TCGA transcriptome data and TCIA radiomics data, 149 TCGA-TCIA intersection samples were standardized via “pyradiomics.” Then, using the R packages “caret” and “CBCgrps” [29], we divided the data into the training and validation sets randomly at a ratio of 6:4 and analyzed the difference between them via z-score standardization (i.e., mean 0 and standard deviation 1).

The maximum relevance, minimum redundancy (mRMR) screening feature considers the correlations between the features and the predicted variables, as well as the correlation between the features themselves. The metric used was mutual information. For the mRMR method, the correlation between feature subsets and categories was calculated as the mean of the information gain of each feature and category, while the redundancy among features was calculated as the sum of the mutual information among features divided by the square of the number of features in the subset.

Assuming that the dataset contains m features, based on the mRMR formula, the importance of the NTH feature can be expressed as follows:

$${f}^{mRMR}\left({X}_{I}\right)=I(Y, {X}_{i})\frac{1}{\left|S\right|}{\sum }_{{X}_{S}\in S}I({X}_{S}, {X}_{i})$$

$I\left(Y, {X}_{i}\right)$ is the mutual information of ${X}_{i}$ and target variable Y. $\frac{1}{\left|S\right|}{\sum }_{{X}_{S}\in S}I({X}_{S}, {X}_{i})$is average of mutual information between variable ${X}_{i}$ and all the variable in subset of existing variables.

Relief (relevant features), a well-known filtration feature selection method, is a feature-weighting algorithm that ensures different weights for each feature according to its relevance and class. Features are removed when the weight is below a certain threshold. In the relief algorithm, the correlation between features and categories is based on the strength of the ability to distinguish close samples based on features. The larger the weight of the feature, the stronger its classification ability; conversely, a lower weight indicates a weaker classification ability.

Construction of the radiomics model

The logistic regression (LR) model, based on variations in linear regression, is a generalized regression algorithm widely used in classification. LR renders the output value of model distribution between 0 and 1 using the sigmoid function to transform the linear regression using the formula:

$\text{g}\left(\text{z}\right)=\text{y}\left(\text{x}\right)=1/(1+{\text{e}}^{{x}^{}}-omega^T$ )

The R package ‘stats’ was used to fit the screened radiomic features using the LR algorithm to construct a model for predicting gene expression.

Support vector machines (SVMs) with a Gaussian kernel were employed in the reduced feature space as the classifiers of choice to discriminate the two cohort groups. The rationale for adopting SVMs as classifiers is primarily because of their ability to generate classification hyperplanes, so the margins between the hyperplane and the nearest instances of the classified sample categories are maximized. The R packages “stats” and “caret” were respectively used to fit the screened radiomics features with the LR algorithm and SVM to construct a model for predicting gene expression.

The radiomics model was evaluated using training and validation sets. The evaluation indexes were accuracy, specificity, sensitivity, positive predictive value, and negative predictive value. The X-axis of the receiver operating characteristic curve was the false positive rate, whereas the Y-axis was the true positive rate. A larger area under the curve (AUC) and a more convex curve to the upper-left corner indicated a better model effect. Calibration of the radiomics prediction model was evaluated by drawing the calibration curve and performing the Hosmer–Lemeshow goodness-of-fit test. The Brier score was used to quantify the comprehensive performance of the radiomic prediction model. A Brier score indicated a more consistent model prediction. The degree of clinical benefit of the radiomic prediction model was determined by drawing a decision curve (DCA). The DeLong test was used to compare the AUC values of the training and validation sets to determine the condition of fit.

The radiomics model score was used to predict the probability of gene expression. A Wilcoxon test was used to compare the radiomics scores (RSs) between the high and low CCND1 expression groups.

Intraclass correlation coefficient (ICC)

ICC was used to evaluate the consistency of radiomic features extracted from volume of interest delineated by two radiologists [30]. First, one radiologist delineated all patient images. Another radiologist delineated 20 randomly selected samples and extracted the radiomics features. ICC ≥ 0.80 is typically considered an optimal agreement, 0.51–0.79 is considered moderate agreement, and < 0.50 was considered poor agreement [30]. The correlation between CCND1 and immune-related genes was analyzed based on SVM model prediction.

Statistical analysis

All statistical analyses were conducted using R version 4.0.0 (https://www.r-project.org). P < 0.05 was considered statistically significant.

Patient characteristics

A total of 412 patients from TCGA were included in this study; among them, 170 patients belonged to the high CCND1 expression group and 242 patients to the low CCND1 expression group. The baseline clinical data of the two groups were compared (Table 1). We observed that WHO grade (P < 0.001), 6-oxo-methylguanine DNA methyltransferase (MGMT) promoter status (P = 0.024), and chemotherapy (P = 0.001) had a significant difference, whereas age, gender, and IDH status had no significance (P ≥ 0.05).

Differences and survival analysis of CCND1 and clinical features

First, compared to normal tissues, CCDN1 expression was significantly higher (P < 0.001) in LGG tissues (Fig. 2A). The median survival time was 81.1 and 76.2 months for patients with low and high CCND1 expression, respectively. The KM results demonstrated that a high CCND1 expression level was related to OS in most curves, but this was not statistically significant (P = 0.117) (Fig. 2B). However, 12, 18, ……, 78, or 84 months after diagnosis was used as a landmark, and univariate landmark analysis of survival was used to assess the survival difference between the early- and late-period LGG. The KM curve demonstrated that when 78 or 84 months were used as landmarks, there was no significant difference between the low- and high-CCND1 group (P < 0.05) in the early period, but the difference in the later period was significant (P < 0.05) (Fig. 2C–D). Thus, CCND1 mainly affects LGG development in the later period.

We further explored whether the clinical features had an effect on patient survival and found that only WHO grade was significantly related to OS (P = 0.029) (Fig. 2E).

Univariate and multivariate Cox regression analysis

Univariate Cox regression analysis revealed that high CCND1 expression was a protective factor for OS (HR = 0.729, 95% CI: 0.491–1.084), but no significant correlation was observed (P = 0.118). The grade was a risk factor for OS (HR = 1.54, 95% CI: 1.042–2.278), and a significant correlation was observed between grade and clinical outcomes (P = 0.03). However, both CCND1 and tumor grade were independent positive prognostic factors in multivariate Cox regression analysis (P < 0.05). The other covariates were not significantly correlated with clinical outcomes (Fig. 3A).

Subgroup analysis and interactive effect

Furthermore, subgroup analysis was performed to explore the influence of the covariates. In WHO grade II, high CCND1 expression was a protective factor for OS (HR = 0.862, 95% CI: 0.402–1.847); however, we observed no significant correlation (P = 0.7). Nevertheless, in WHO grade III, high CCND1 expression was a protective factor for OS (HR = 0.555, 95% CI: 0.343–898), which was significantly correlated with clinical outcomes (P = 0.016). In WHO grade III, tumor grade had no significant interactive effect (P = 0.38) on the correction of CCND1 and OS in LGG patients. Other covariates had no significant correlation with clinical outcomes, nor did they have an interactive effect on the correction of CCND1 and OS (Fig. 3B).

Correlation between CCND1 and clinical features

Counterintuitively, CCND1 expression level correlated positively with chemotherapy (P < 0.01), grade (P < 0.01), IDH status (P < 0.05), MGMT promoter status (P < 0.05), and radiotherapy (P < 0.05) in LGG.

Correlation between CCND1 and immune infiltration

As demonstrated in Fig. 5, eleven immune cell types showed significance in the correlation analysis between CCND1 and immune infiltration (P < 0.05). Therefore, natural killer T, gamma delta T, central memory, and effector memory cells were positively correlated with CCND1 (cor > 0). Additionally, B, Tr1, monocyte, DC, CD4⁺ T cells, iTregs, and exhausted cells were negatively correlated with CCND1 expression (cor < 0).

GSVA

Based on the high and low CCND1 expression levels in LGG, we analyzed the enrichment of DEGs. In the KEGG gene set (Fig. 6A), we observed that high CCND1 expression significantly enriched pathways, the PPAR signaling pathway etc., and low expression significantly enriched those, cell cycle and DNA repair etc. In the hallmark gene set (Fig. 6B), we observed that high CCND1 expression significantly enriched the KRAS signaling pathway, and low expression significantly enriched the DNA repair and E2F target signaling pathways. Overall, in both the KEGG and hallmark gene sets, low expression of CCND1 significantly enriched 7 signaling pathways: apoptosis, DNA repair, MTORC1 signaling, Notch signaling, p53 pathway, TGF beta signaling, and WNT/beta-catenin signaling.

Clinicopathologic characteristics of radiomics participants

Radiomics feature extraction, model establishment consistency evaluation, feature selection, and construction of the radiomics model

Using PyRadiomics [28], we performed radiomics feature extraction and ICC. The interobserver ICC was ≥ 0.8, 0.5–0.79, and < 0.5 for 91, 15, and 1 of the features, respectively. Subsequently, using mRMR and the relief algorithm, we performed the top 20 features, and obtained 5 features by intersection (Fig. 7). The data were divided into training (90) and verification (59) sets. In the analysis of differences in each variable between groups, P > 0.05 indicated comparability between the training and validation sets (Table 2).

Table 2

Clinical characteristics of the population with high and low Radiomics score groups
Variables	Total (n = 149)	Train (n = 90)	Validation (n = 59)	P value
CCND1, n (%)				1
Low	90 (60)	54 (60)	36 (61)
High	59 (40)	36 (40)	23 (39)
Age, n (%)				0.54
~40	69 (46)	44 (49)	25 (42)
41~	80 (54)	46 (51)	34 (58)
Gender, n (%)				0.892
Female	76 (51)	45 (50)	31 (53)
Male	73 (49)	45 (50)	28 (47)
Histology, n (%)				0.528
Astrocytoma	50 (34)	30 (33)	20 (34)
Oligoastrocytoma	34 (23)	18 (20)	16 (27)
Oligodendroglioma	65 (44)	42 (47)	23 (39)
Grade, n (%)				0.348
WHO II	75 (50)	42 (47)	33 (56)
WHO III	74 (50)	48 (53)	26 (44)
IDH_status, n (%)				0.735
Mutant	117 (79)	72 (80)	45 (76)
WT	32 (21)	18 (20)	14 (24)
Chr_1p_19q_codeletion, n (%)				0.244
Codel	42 (28)	29 (32)	13 (22)
Non-Codel	107 (72)	61 (68)	46 (78)
MGMT_promoter_status, n (%)				1
Methylated	119 (80)	72 (80)	47 (80)
Unmethylated	30 (20)	18 (20)	12 (20)
Chemotherapy, n (%)				0.421
NO	56 (38)	31 (34)	25 (42)
YES	93 (62)	59 (66)	34 (58)
Radiotherapy, n (%)				0.712
NO	57 (38)	36 (40)	21 (36)
YES	92 (62)	54 (60)	38 (64)
OS, n (%)				1
Alive	119 (80)	72 (80)	47 (80)
Dead	30 (20)	18 (20)	12 (20)
OS.time, Median (Q1,Q3)	22.37 (13.8, 40.43)	20.83 (13.83, 37.81)	24.13 (13.23, 46.12)	0.357

In the training set, the radiomics model was constructed using the LR algorithm and used to analyze the importance of the selected features (Fig. 8A, Table 3), whose ICC values were all > 0.8, with good consistency. Subsequently, the receiver operating characteristic curve was used to evaluate the radiomics model, and the AUC values were 0.703 and 0.705 for the training and validation sets, respectively (Fig. 8B–C). We compared the AUC values between the training and validation sets and found no significant differences (P = 0.978). Furthermore, the calibration curve analysis and Hosmer–Lemeshow goodness-of-fit were consistent with the true values (P > 0.05) (Fig. 8D–G). The DCA model demonstrated clinical practicability. In addition, compared with the low CCND1 expression group, the high CCND1 expression group was significantly higher (P < 0.01) in both the training and validation sets (Fig. 8H–I).

Table 3

Radiomics features
Item	Importance
ngtdm_Busyness	1.353484245
firstorder_Kurtosis	1.031935343
glcm_Imc2	0.518207678
glszm_SizeZoneNonUniformity	0.606335873
ngtdm_Complexity	1.983987285

To determine the reliability of the radiomics model, we used the SVM algorithm to conduct radiomics and analyze the importance of the selected features (Fig. 9A, Table 3). In the training and validation sets, AUC values were 0.724 and 0.726 respectively (Fig. 9B–C), which were higher than those in the LR. We also compared the AUC values between the training and validation sets and found no significant differences (P = 0.986). Furthermore, we observed equivalent predictions of the two models by comparing the AUC values between the LR and SVM in the training and validation sets (P = 0.153, 0.346). Furthermore, the calibration curve analysis and Hosmer–Lemeshow goodness-of-fit were consistent with the true values (P > 0.05) (Fig. 9D–G). The DCA model demonstrated clinical practicability. In addition, compared with the low CCND1 expression group, the high-expression group was significantly higher in both the training and validation sets (P < 0.001, 0.01, respectively) (Fig. 9H–I). In summary, the radiomics model predicted by the SVM was superior to that predicted by LR.

Correlation between radiomics model and CCND1, immune-related genes

As shown in Fig. 10 and Table 4, CCND1 was positively correlated with RS (cor = 0.36, P < 0.001), and nine genes were significantly correlated with CCND1 and immune-related genes (P < 0.05). Namely, CD28, TNFRSF9, LAG3, CD276, and NRP1 were positively correlated with CCND1 expression (cor > 0). Additionally, HHLA2, TNFRSF14, CD200R1, and ICOSLG were negatively correlated with CCND1 (cor < 0). Three genes were significantly correlated with RS and immune-related genes (P < 0.05). TNFRSF8, TNFRSF9, and CCND1 were positively correlated with CCND1 (cor > 0). In addition, TIGIT was negatively correlated with CCND1 (cor < 0). We observed that TNFRSF9 was correlated with CCND1 (cor = 0.24, P < 0.01), as well as with RS (cor = 0.30, P < 0.001).

Table 4

Correlation between Radiomics score and CCND1, immune-related genes
Variable1	Variable2	Correlation	P value
CCND1	HHLA2	-0.37	2.80E-06
CCND1	TNFRSF14	-0.30	0.0003
CCND1	CD200R1	-0.18	0.0292
CCND1	ICOSLG	-0.17	0.0440
CCND1	CD28	0.18	0.0245
CCND1	TNFRSF9	0.24	0.0027
CCND1	LAG3	0.26	0.0014
CCND1	CD276	0.31	0.0001
CCND1	NRP1	0.32	5.77E-05
CCND1	RS	0.36	5.83E-06
RS	TIGIT	-0.17	0.0409
RS	TNFRSF8	0.16	0.0498
RS	TNFRSF9	0.30	0.0003
RS	CCND1	0.36	5.83E-06
RS: radiomics model, CCND1: Cyclin D1, HHLA2: HERV-H LTR-Associating 2, TNFRSF14: TNF Receptor Superfamily Member 14, CD200R1: CD200 Receptor 1, ICOSLG: Inducible T Cell Costimulator Ligand, CD28: CD28 Molecule, TNFRSF9: TNF Receptor Superfamily Member 9, LAG3: Lymphocyte Activating 3, CD276: CD276 Molecule, NRP1: Neuropilin 1

We evaluated CCND1 expression levels and their clinical prognostic value in patients with LGG by analyzing MRI images in the TCIA database. Our results demonstrated that MRI features are associated with differences in CCND1 expression levels in LGG, implying that supererogatory gene information could be explored from radiomics analysis of MRI images. The machine learning models improved accuracy by incorporating clinical and radiomic MRI parameters.

Owing to the LGG subtype, the median OS ranges from 11.7–19.9 years [31]; thus, further prognostic information must be considered. According to the WHO classification of central nervous system tumors 6th edition [2], a comprehensive, integrated diagnosis requires a combination of diagnostic features drawn from relevant histological, immunochemical, molecular, and other abnormalities in LGG, which is heterogenetic. CCND1 silencing may benefit live cancer therapy as a marker for live cancer stem cells [32; 33]. Furthermore, CCND1, which may promote tumor development and regeneration, has been actively investigated in various types of tumors, including breast, melanoma, bladder, and thyroid cancer [9; 34–37]. The CCND1 G870A polymorphism possibly increased the risk of gliomas in the Chinese population [38]. Our research focused on LGG. We demonstrated that high CCND1 expression levels are linked to poorer prognosis in advanced LGG patients (P < 0.05) through landmark analysis of 412 samples from TCGA.

MRI, a method that lacks objectivity and quantification, depends on morphological characteristics for diagnosis. Radiomics can minimize high-throughput data from medical images, recognize deep information that cannot be observed with the naked eye, and perform quantitative analyses. By distilling the MRI image features in bladder cancer using radiomics, the AUC of the predictive model for muscle invasion was 0.920 in the training cohort and 0.844 in the test cohort [39]. According to dynamic contrast enhanced-MRI and apparent diffusion coefficient maps, radiomics features can be used as imaging biomarkers to confirm Ki-67 status in patients with breast cancer [40]. Based on multi-parametric MRI, a radiomic nomogram can predict the tumor grade to guide the dilation and surgery model in endometrial cancer [41]. The multi-layer perceptron model, including clinical features, mutation status of the epidermal growth factor receptor, and radiomics, effectively predicted the risk of death, which can promote the individualized management of patients with lung cancer and brain metastases [42]. In the present study, based on low and high CCND1 expression levels, four radiomic features were selected via mRMR and relief to construct the LR and SVM models. In both models, the AUC of the radiomics features was > 0.70; moreover, both LR and SVM exhibited acceptable consistency and diagnostic efficacy. Furthermore, our study based on MRI was low-cost, easy to repeat, and non-invasive. Collectively, these findings suggest that the combination of radiomics and molecular features will be a trend for future research.

Based on radiomics features and clinical characteristics, Chen et al. [43] built a combined survival model to help patients with advanced gastric cancer in future clinical decision-making. Boot et al. [44] developed a predictive model to forecast human papillomavirus status with acceptable performance and improve survival prediction. Ming et al. [45] observed that imaging subtypes might serve as potential biomarkers. In this study, we built prognostic models based on clinical and radiomics features for survival analysis of patients with LGG. Landmark analysis revealed that the high CCND1 expression group in advanced LGG significantly differed from the low-expression group (P < 0.05), showing survival benefits. Moreover, compared with the WHO III group, the WHO Ⅱ group was significantly different (P = 0.029) in terms of survival. In addition, the high CCND1 expression group was significantly higher than the low-expression group (P < 0.05) based on the RS in the LR and SVM models. Our findings suggest that the clinical model alone has good prognostic power and that the radiomic features combined with CCND1 have clinical value.

Although radiomics models work well, our study has several limitations that may provide directions for future studies. First, all the data were obtained from public datasets (TCGA and TCIA), and the images were inevitably variable in quality, which may have influenced the analysis. Second, we only conducted research on CCND1; other biomarkers, such as CCND2 and CDK1, are worthy of further study. Third, this research was a retrospective study with 149 samples from TCIA; therefore, its generalizability must be further validated. Finally, radiomics analysis was based on manual segmentation by two radiologists. In other words, automated technology should be developed in further research to improve the reproducibility of radiomics analysis and reduce operator error.

In conclusion, CCND1 expression levels could significantly influence the prediction and prognosis of patients with LGG. We constructed two radiomics-based models to effectively predict CCND1 expression levels. Therefore, our model can be widely used as an effective method for non-invasive tumor characterization.

Acknowledgments

We would like to thank Editage (www.editage.cn) for their assistance with English language editing.

Funding

This work was supported by the research foundation of the Affiliated People’s Hospital of Jiangsu University (Y2021003S) and Project 333 of Jiangsu Province (BRA2019231).

Availability of data and materials

The datasets analyzed during the current study are available in the TCGA , TCIA, UCSC XENA and GTEx repository, https://portal.gdc.cancer.gov/, https://www.cancerimagingarchive.net/, https://xenabrowser.net/datapages/, https://www.gtexportal.org/home/.

Authors’ contributions

Zhao Kun and Zhang Hui: Conceptualization, Methodology, Software, Data Curation, Writing - Original Draft; Lin Jianyang, Liu Jianzhi, and Xu Shoucheng: Resources, Investigation; Gu Yongbing and Ren Guoqiang: Formal analysis; Lu Xinyu and Chen Baomin: Software, Validation; Chen Deng and Yan Jun: Visualization; Ma Jichun and Wei Wenxiang: Investigation; Wang Yuanwei: Writing - Review & Editing.

Ethics approval and consent to participate

Not applicable

Patient consent for publication

Not applicable

Competing interests

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Ogino, H., Taylor, J. W., Nejo, T., Gibson, D., Watchmaker, P. B., Okada, K., Saijo, A., Tedesco, M. R., Shai, A., Wong, C. M., Rabbitt, J. E., Olin, M. R., Moertel, C. L., Nishioka, Y., Salazar, A. M., Molinaro, A. M., Phillips, J. J., Butowski, N. A., Clarke, J. L., . . . Okada, H. (2022). Randomized trial of neoadjuvant vaccination with tumor-cell lysate induces T cell response in low-grade gliomas. J Clin Invest, 132(3). https://doi.org/10.1172/jci151239
Louis, D. N., Perry, A., Wesseling, P., Brat, D. J., Cree, I. A., Figarella-Branger, D., Hawkins, C., Ng, H. K., Pfister, S. M., Reifenberger, G., Soffietti, R., von Deimling, A., & Ellison, D. W. (2021). The 2021 WHO Classification of Tumors of the Central Nervous System: a summary. Neuro Oncol, 23(8), 1231-1251. https://doi.org/10.1093/neuonc/noab106
Haddad, A. F., Young, J. S., Oh, J. Y., Okada, H., & Aghi, M. K. (2022). The immunology of low-grade gliomas. Neurosurg Focus, 52(2), E2. https://doi.org/10.3171/2021.11.Focus21587
Salam, R., Saliou, A., Bielle, F., Bertrand, M., Antoniewski, C., Carpentier, C., Alentorn, A., Capelle, L., Sanson, M., Huillard, E., Bellenger, L., Guégan, J., & Le Roux, I. (2023). Cellular senescence in malignant cells promotes tumor progression in mouse and patient Glioblastoma. Nat Commun, 14(1), 441. https://doi.org/10.1038/s41467-023-36124-9
Sanai, N., Chang, S., & Berger, M. S. (2011). Low-grade gliomas in adults. J Neurosurg, 115(5), 948-965. https://doi.org/10.3171/2011.7.Jns101238
Musgrove, E. A., Caldon, C. E., Barraclough, J., Stone, A., & Sutherland, R. L. (2011). Cyclin D as a therapeutic target in cancer. Nat Rev Cancer, 11(8), 558-572. https://doi.org/10.1038/nrc3090
Pestell, R. G. (2013). New roles of cyclin D1. Am J Pathol, 183(1), 3-9. https://doi.org/10.1016/j.ajpath.2013.03.001
Sherr, C. J., & Roberts, J. M. (2004). Living with or without cyclins and cyclin-dependent kinases. Genes Dev, 18(22), 2699-2711. https://doi.org/10.1101/gad.1256504
Liu, J., Lin, J., Wang, X., Zheng, X., Gao, X., Huang, Y., Chen, G., Xiong, J., Lan, B., Chen, C., Si, L., & Chen, Y. (2022). CCND1 Amplification Profiling Identifies a Subtype of Melanoma Associated With Poor Survival and an Immunosuppressive Tumor Microenvironment. Front Immunol, 13, 725679. https://doi.org/10.3389/fimmu.2022.725679
Du, Q., Guo, X., Wang, M., Li, Y., & Li, Q. (2020). The application and prospect of CDK4/6 inhibitors in malignant solid tumors. Journal of Hematology & Oncology, 13(1).
Hosny, A., Parmar, C., Quackenbush, J., Schwartz, L. H., & Aerts, H. (2018). Artificial intelligence in radiology. Nat Rev Cancer, 18(8), 500-510. https://doi.org/10.1038/s41568-018-0016-5
Huang, E. P., O'Connor, J. P. B., McShane, L. M., Giger, M. L., Lambin, P., Kinahan, P. E., Siegel, E. L., & Shankar, L. K. (2023). Criteria for the translation of radiomics into clinically useful tests. Nat Rev Clin Oncol, 20(2), 69-82. https://doi.org/10.1038/s41571-022-00707-0
Chitalia, R. D., Rowland, J., McDonald, E. S., Pantalone, L., Cohen, E. A., Gastounioti, A., Feldman, M., Schnall, M., Conant, E., & Kontos, D. (2020). Imaging Phenotypes of Breast Cancer Heterogeneity in Preoperative Breast Dynamic Contrast Enhanced Magnetic Resonance Imaging (DCE-MRI) Scans Predict 10-Year Recurrence. Clin Cancer Res, 26(4), 862-869. https://doi.org/10.1158/1078-0432.Ccr-18-4067
Chen, M., Lu, H., Copley, S. J., Han, Y., Logan, A., Viola, P., Cortellini, A., Pinato, D. J., Power, D., & Aboagye, E. O. (2023). A Novel Radiogenomics Biomarker for Predicting Treatment Response and Pneumotoxicity from Programmed Cell Death-1 Pathway Inhibition in Non-Small Cell Lung Cancer. J Thorac Oncol. https://doi.org/10.1016/j.jtho.2023.01.089
Cui, H., Sun, Y., Zhao, D., Zhang, X., Kong, H., Hu, N., Wang, P., Zuo, X., Fan, W., Yao, Y., Fu, B., Tian, J., Wu, M., Gao, Y., Ning, S., & Zhang, L. (2023). Radiogenomic analysis of prediction HER2 status in breast cancer by linking ultrasound radiomic feature module with biological functions. J Transl Med, 21(1), 44. https://doi.org/10.1186/s12967-022-03840-7
Casale, R., Lavrova, E., Sanduleanu, S., Woodruff, H. C., & Lambin, P. (2021). Development and external validation of a non-invasive molecular status predictor of chromosome 1p/19q co-deletion based on MRI radiomics analysis of Low Grade Glioma patients. Eur J Radiol, 139, 109678. https://doi.org/10.1016/j.ejrad.2021.109678
Li, Z. Z., Liu, P. F., An, T. T., Yang, H. C., Zhang, W., & Wang, J. X. (2021). Construction of a prognostic immune signature for lower grade glioma that can be recognized by MRI radiomics features to predict survival in LGG patients. Transl Oncol, 14(6), 101065. https://doi.org/10.1016/j.tranon.2021.101065
Xiao, Z., Yao, S., Wang, Z. M., Zhu, D. M., Bie, Y. N., Zhang, S. Z., & Chen, W. L. (2021). Multiparametric MRI Features Predict the SYP Gene Expression in Low-Grade Glioma Patients: A Machine Learning-Based Radiomics Analysis. Front Oncol, 11, 663451. https://doi.org/10.3389/fonc.2021.663451
Xu, C., Peng, Y., Zhu, W., Chen, Z., Li, J., Tan, W., Zhang, Z., & Chen, X. (2022). An automated approach for predicting glioma grade and survival of LGG patients using CNN and radiomics. Front Oncol, 12, 969907. https://doi.org/10.3389/fonc.2022.969907
Kassambara, A. (2017). Drawing Survival Curves using 'ggplot2' [R package survminer version 0.2.0].
Vivian, J., Rao, A. A., Nothaft, F. A., Ketchum, C., Armstrong, J., Novak, A., Pfeil, J., Narkizian, J., Deran, A. D., Musselman-Brown, A., Schmidt, H., Amstutz, P., Craft, B., Goldman, M., Rosenbloom, K., Cline, M., O'Connor, B., Hanna, M., Birger, C., . . . Paten, B. (2017). Toil enables reproducible, open source, big biomedical data analyses. Nat Biotechnol, 35(4), 314-316. https://doi.org/10.1038/nbt.3772
Wickham, H. (2009). Ggplot2: Elegant Graphics for Data Analysis. ggplot2: Elegant Graphics for Data Analysis.
Therneau, T. M., & Grambsch, P. M. (2013). Modeling Survival Data: Extending the Cox Model. Modeling Survival Data: Extending the Cox Model.
Miao, Y. R., Zhang, Q., Lei, Q., Luo, M., Xie, G. Y., Wang, H., & Guo, A. Y. (2020). ImmuCellAI: A Unique Method for Comprehensive T-Cell Subsets Abundance Prediction and its Application in Cancer Immunotherapy. Adv Sci (Weinh), 7(7), 1902880. https://doi.org/10.1002/advs.201902880
Hanzelmann, S., Castelo, R., & Guinney, J. (2013). GSVA: gene set variation analysis for microarray and RNA-seq data. BMC Bioinformatics, 14, 7. https://doi.org/10.1186/1471-2105-14-7
Phipson, B., Lee, S., Majewski, I. J., Alexander, W. S., & Smyth, G. K. (2016). Robust Hyperparameter Estimation Protects against Hypervariable Genes and Improves Power to Detect Differential Expression. Ann Appl Stat, 10(2), 946-963. https://doi.org/10.1214/16-AOAS920
Ritchie, M. E., Phipson, B., Wu, D., Hu, Y., Law, C. W., Shi, W., & Smyth, G. K. (2015). limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res, 43(7), e47. https://doi.org/10.1093/nar/gkv007
van Griethuysen, J. J. M., Fedorov, A., Parmar, C., Hosny, A., Aucoin, N., Narayan, V., Beets-Tan, R. G. H., Fillion-Robin, J. C., Pieper, S., & Aerts, H. (2017). Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res, 77(21), e104-e107. https://doi.org/10.1158/0008-5472.Can-17-0339
Zhang, Z., Gayle, A. A., Wang, J., Zhang, H., & Cardinal-Fernández, P. (2017). Comparing baseline characteristics between groups: An introduction to the CBCgrps package. Annals of Translational Medicine, 5(24), 484-484.
Leijenaar, R. T., Carvalho, S., Velazquez, E. R., van Elmpt, W. J., Parmar, C., Hoekstra, O. S., Hoekstra, C. J., Boellaard, R., Dekker, A. L., Gillies, R. J., Aerts, H. J., & Lambin, P. (2013). Stability of FDG-PET Radiomics features: an integrated analysis of test-retest and inter-observer variability. Acta Oncol, 52(7), 1391-1397. https://doi.org/10.3109/0284186x.2013.812798
Hervey-Jumper, S. L., Zhang, Y., Phillips, J. J., Morshed, R. A., Young, J. S., McCoy, L., Lafontaine, M., Luks, T., Ammanuel, S., Kakaizada, S., Egladyous, A., Gogos, A., Villanueva-Meyer, J., Shai, A., Warrier, G., Rice, T., Crane, J., Wrensch, M., Wiencke, J. K., . . . Molinaro, A. M. (2023). Interactive Effects of Molecular, Therapeutic, and Patient Factors on Outcome of Diffuse Low-Grade Glioma. J Clin Oncol, Jco2102929. https://doi.org/10.1200/jco.21.02929
Zhang, H. (2020). CCND1 silencing suppresses liver cancer stem cell differentiation through inhibiting autophagy. Hum Cell, 33(1), 140-147. https://doi.org/10.1007/s13577-019-00295-9
Ding, H., Wang, Y., & Zhang, H. (2020). CCND1 silencing suppresses liver cancer stem cell differentiation and overcomes 5-Fluorouracil resistance in hepatocellular carcinoma. J Pharmacol Sci, 143(3), 219-225. https://doi.org/10.1016/j.jphs.2020.04.006
Valla, M., Klæstad, E., Ytterhus, B., & Bofin, A. M. (2022). CCND1 Amplification in Breast Cancer -associations With Proliferation, Histopathological Grade, Molecular Subtype and Prognosis. J Mammary Gland Biol Neoplasia, 27(1), 67-77. https://doi.org/10.1007/s10911-022-09516-8
Sushma, P. S., Jamil, K., Udaykumar, P., Aldakheel, F. M., Alduraywish, S. A., Alali, B. H., Mateen, A., & Syed, R. (2021). Analysis of CCND1 protein and circulatory antioxidant enzyme activity association in oral squamous cell carcinoma. Saudi J Biol Sci, 28(12), 6987-6991. https://doi.org/10.1016/j.sjbs.2021.07.085
Ying, Y., Li, J., Xie, H., Yan, H., Jin, K., He, L., Ma, X., Wu, J., Xu, X., Fang, J., Wang, X., Zheng, X., Liu, B., & Xie, L. (2020). CCND1, NOP14 and DNMT3B are involved in miR-502-5p-mediated inhibition of cell migration and proliferation in bladder cancer. Cell Prolif, 53(2), e12751. https://doi.org/10.1111/cpr.12751
Jeon, S., Kim, Y., Jeong, Y. M., Bae, J. S., & Jung, C. K. (2018). CCND1 Splice Variant as A Novel Diagnostic and Predictive Biomarker for Thyroid Cancer. Cancers (Basel), 10(11). https://doi.org/10.3390/cancers10110437
Chen, X., Zhao, T., Li, L., Xu, C., Zhang, X., Tse, V., Zhang, T., Liu, X., & Lu, F. (2012). CCND1 G870A polymorphism with altered cyclin D1 transcripts expression is associated with the risk of glioma in a Chinese population. DNA Cell Biol, 31(6), 1107-1113. https://doi.org/10.1089/dna.2011.1521
Li, J., Qiu, Z., Cao, K., Deng, L., Zhang, W., Xie, C., Yang, S., Yue, P., Zhong, J., Lyu, J., Huang, X., Zhang, K., Zou, Y., & Huang, B. (2023). Predicting muscle invasion in bladder cancer based on MRI: A comparison of radiomics, and single-task and multi-task deep learning. Comput Methods Programs Biomed, 233, 107466. https://doi.org/10.1016/j.cmpb.2023.107466
Feng, S., & Yin, J. (2022). Radiomics of dynamic contrast-enhanced magnetic resonance imaging parametric maps and apparent diffusion coefficient maps to predict Ki-67 status in breast cancer. Front Oncol, 12, 847880. https://doi.org/10.3389/fonc.2022.847880
Yue, X., He, X., He, S., Wu, J., Fan, W., Zhang, H., & Wang, C. (2023). Multiparametric magnetic resonance imaging-based radiomics nomogram for predicting tumor grade in endometrial cancer. Front Oncol, 13, 1081134. https://doi.org/10.3389/fonc.2023.1081134
Liao, C. Y., Lee, C. C., Yang, H. C., Chen, C. J., Chung, W. Y., Wu, H. M., Guo, W. Y., Liu, R. S., & Lu, C. F. (2023). Predicting survival after radiosurgery in patients with lung cancer brain metastases using deep learning of radiomics and EGFR status. Phys Eng Sci Med. https://doi.org/10.1007/s13246-023-01234-7
Chen, W., Gao, C., Hu, C., Zheng, Y., Wang, L., Chen, H., & Jiang, H. (2023). Risk Stratification and Overall Survival Prediction in Advanced Gastric Cancer Patients Based on Whole-Volume MRI Radiomics. J Magn Reson Imaging. https://doi.org/10.1002/jmri.28621
Boot, P. A., Mes, S. W., de Bloeme, C. M., Martens, R. M., Leemans, C. R., Boellaard, R., van de Wiel, M. A., & de Graaf, P. (2023). Magnetic resonance imaging based radiomics prediction of Human Papillomavirus infection status and overall survival in oropharyngeal squamous cell carcinoma. Oral Oncol, 137, 106307. https://doi.org/10.1016/j.oraloncology.2023.106307
Ming, W., Li, F., Zhu, Y., Bai, Y., Gu, W., Liu, Y., Liu, X., Sun, X., & Liu, H. (2022). Unsupervised Analysis Based on DCE-MRI Radiomics Features Revealed Three Novel Breast Cancer Subtypes with Distinct Clinical Outcomes and Biological Characteristics. Cancers (Basel), 14(22). https://doi.org/10.3390/cancers14225507

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Magnetic resonance imaging-based machine learning radiomics predicts CCND1 expression level and survival in low-grade gliomas

Status:

Version 1

Abstract

Figures

Introduction

Methods

Study design

Data retrieval and processing for this study

Survival analysis

Univariate landmark analysis

Univariate and multivariable COX analysis

Subgroup analysis and interaction test

Correlation analysis

Analysis of the correlation between CCND1 expression and immune cell infiltration

Gene Set Variation Analysis (GSVA) between low and high CCND1 expression

Screening of radiomics features

Construction of the radiomics model

Intraclass correlation coefficient (ICC)

Statistical analysis

Results

Patient characteristics

Univariate and multivariate Cox regression analysis

Subgroup analysis and interactive effect

GSVA

Clinicopathologic characteristics of radiomics participants

Radiomics feature extraction, model establishment consistency evaluation, feature selection, and construction of the radiomics model

Correlation between radiomics model and CCND1, immune-related genes

Discussion

Conclusion

Declarations

References

Additional Declarations

Status:

Version 1