DeepGR: a deep-learning prognostic model based on glycolytic radiomics for non-small cell lung cancer

Abstract

Background: Glycolysis proved to have a prognostic value in lung cancer; however, to identify glycolysis-related genomic markers is expensive and challenging. This study aimed at identifying glycolysis-related computed tomography (CT) radiomics features to develop a deep-learning prognostic model for non-small cell lung cancer (NSCLC).

Methods: The study included 274 NSCLC patients from cohorts of The Second Affiliated Hospital of Soochow University (SZ; n=64), the Cancer Genome Atlas (TCGA)-NSCLC dataset (n=74), and the Gene Expression Omnibus dataset (n=136). Initially, the glycolysis enrichment scores were evaluated using a single-sample gene set enrichment analysis, and the cut-off values were optimized to investigate the prognostic potential of glycolysis genes. Radiomic features were then extracted using LIFEx software. The least absolute reduction and selection operator (LASSO) algorithm was employed to determine the glycolytic CT radiomics features. A deep-learning prognostic model was constructed by integrating CT radiomics and clinical features. The biological functions of the model were analyzed by incorporating RNA sequencing data.

Results: Kaplan-Meier curves indicated that elevated glycolysis levels were associated with poorer survival outcomes. The LASSO algorithm identified 11 radiomic features that were then selected for inclusion in the deep-learning model. They have shown significant discrimination capability in assessing glycolysis status, achieving an area under the curve value of 0.8442. The glycolysis-based radiomics deep-learning model was named the DeepGR model. This model was able to effectively predict the clinical outcomes of NSCLC patients with AUCs of 0.8760 and 0.8259 in the SZ and TCGA cohorts, respectively. High-risk DeepGR scores were strongly associated with poor overall survival, resting memory CD4⁺ T cells, and a high response to programmed cell death protein 1 immunotherapy.

Conclusions: The DeepGR model effectively predicted the prognosis of NSCLC patients.

Keywords: Non-small cell lung cancer (NSCLC); deep learning; glycolysis; prognostic model; radiomics.

Figure 1

Patients’ enrollment in the study. TCGA, The Cancer Genome Atlas; NSCLC, non-small cell lung cancer; CT, computed tomography; GEO, Gene Expression Omnibus; SZ, The Second Affiliated Hospital of Soochow University; DICOM, Digital Imaging and Communications in Medicine.

Figure 2

The architecture of the DeepGR model.

Figure 3

Study design. TCGA, The Cancer Genome Atlas; NSCLC, non-small cell lung cancer; SZ, The Second Affiliated Hospital of Soochow University; ICC, intraclass correlation coefficient; NA, not available; TME, tumor microenvironment; CI, confidence interval; DL, deep learning; LASSO, least absolute shrinkage and selection operator; NK, natural killer cell.

Figure 4

Clinical characteristics of TCGA cohort (A) and SZ cohort (B). TCGA, The Cancer Genome Atlas; SZ, The Second Affiliated Hospital of Soochow University; NA, not applicable.

Figure 5

Survival analysis and cut-off calculation of the glycolysis scores. The cut-off values of the glycolysis scores of TCGA (A), GSE19188 (B), and GSE87340 (C) cohorts, respectively; blue line: patients with glycolysis score lower than cutoff, red line: patients with glycolysis score higher than cutoff. The survival curves of the samples from TCGA (D), GSE19188 (E), and GSE87340 (F) cohorts based on the best cut-off values, respectively. TCGA, The Cancer Genome Atlas.

Figure 6

Glycolysis-related radiomic features derived from the LASSO model. (A,B) LASSO method for the screening of the radiomics features; the red dotted line: the number of enrolled variables that yields the minimum bias; the blue dotted line: enrolled fewer variables with relative fewer bias. (C) Coefficients of the selected features; (D) performance of the model. LASSO, least absolute shrinkage and selection operator; AUC, area under the curve.

Figure 7

Evaluation of the DeepGR model. Survival analysis of the SZ cohort (A) and TCGA cohort (B) based on risk scores. AUC curves of the SZ cohort (C) and TCGA cohort (D) based on the DeepGR model, clinical features and radiomics features. (E) Survival analysis conducted using restricted cubic splines. TCGA, The Cancer Genome Atlas; SZ, The Second Affiliated Hospital of Soochow University; AUC, area under the curve; CI, confidence interval; DL, deep learning; Ra, radiomics; Cli, clinical features.

Figure 8

Heatmaps showing the relationship between risk grouping and clinical features. TCGA, The Cancer Genome Atlas; SZ, The Second Affiliated Hospital of Soochow University.

Figure 9

Biological function of the DeepGR model. (A) DEGs of the model; (B) network analysis of the DEGs; #: CYP2B6; (C) GO analysis of the DEGs; (D) KEGG analysis of the DEGs. DEGs, differentially expressed genes; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes.

Figure 10

Immune heterogeneity between the DeepGR-high and DeepGR-low groups. (A) The ESTIMATE score, immune score, and stromal score of the DeepGR-high and DeepGR-low groups; (B) the correlation between the infiltration levels of 22 immune cells and the DeepGR analyzed by the CIBERSORT algorithm; (C-F) the relative probabilities of responding to anti-CTLA-4 antibody and anti-PD-1 antibody in the DeepGR-high and DeepGR-low groups. TME, tumor microenvironment; CTLA-4, cytotoxic T-lymphocyte-associated protein 4; PD-1, programmed cell death protein 1; NK, natural killer cell.