Machine learning derived development and validation of extracellular matrix related signature for predicting prognosis in adolescents and young adults glioma

Abstract

The mortality rates have been increasing for glioma in adolescents and young adults (AYAs, aged 15-39 years). However, current biomarkers for clinical assessment in AYAs glioma are limited, prompting the urgent need for identifying ideal prognostic signature. Extracellular matrix is involved in the development of tumors, while their prognostic significance in AYAs glioma remains unclear. By an integrated machine learning workflow and circuit training and validation procedure, we developed a machine learning-derived prognostic signature (MLDPS) based on 1,026 extracellular matrix-related genes and 3 AYAs glioma cohorts. MLDPS exhibited robust and consistent predictive performance in overall survival and could serve as an independent prognostic factor for AYAs glioma. Simultaneously, MLDPS outperformed previous 89 published prognostic signatures and traditional clinical characteristics, confirming the robust predictive capability. Besides, MLDPS had the potential to stratify prognosis in patients with other cancer types. In addition, the tumor microenvironment between high and low MLDPS groups displayed different patterns while more tumor-infiltrating immune cells were observed in high MLDPS group. Additionally, patients in low MLDPS group had significantly prolonged survival when received immunotherapy in cancers including glioblastoma, urothelial carcinoma and melanoma. Overall, our study proposes a promising signature, which can be utilized for clinicians to evaluate prognosis and might provide individualized clinical management for AYAs glioma.

Keywords: Adolescents and young adults; Glioma; Immunotherapy; Machine learning; Prognosis.

Fig. 1

The flowchart of this study.

Fig. 2

Construction of the machine learning-derived prognostic signature (MLDPS). (A) The C-index of 65 machine learning algorithms combinations in CGGA-693 training procedure. (B) The C-index of 65 machine learning algorithms combinations in CGGA-325 training procedure. (C) The C-index of 65 machine learning algorithms combinations in TCGA training procedure. (D-F) Top five average C-index in CGGA-693, CGGA-325 and TCGA training procedure, respectively. (G-I) The performance of MLDPS was compared with common clinical and molecular characteristics in CGGA-693 (G), CGGA-325 (H) and TCGA (I). *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001.

Fig. 3

Survival analysis and predictive performance evaluation of machine learning-derived prognostic signature (MLDPS). (A-C) Kaplan-Meier survival analysis for overall survival between high and low MLDPS groups in CGGA-693 (A), CGGA-325 (B) and TCGA cohorts (C). (D-F) Univariate and multivariate Cox regression analyses regarding of MLDPS in CGGA-693 (D), CGGA-325 (E) and TCGA cohorts (F). (G-I) Time-dependent receiver-operator characteristic (ROC) analysis for predicting OS at 1-, 3- and 5-year in CGGA-693 (G), CGGA-325 (H) and TCGA cohorts (I).

Fig. 4

The correlation between machine learning-derived prognostic signature (MLDPS) and clinical characteristics. (A-C) The correlation between age, gender, grade, IDH status, 1p/19q status and MLDPS in CGGA-693 (A), CGGA-325 (B) and TCGA cohort (C), respectively. (D) Kaplan-Meier survival analysis for overall survival between high and low MLDPS groups in different age, gender, grade, IDH status and 1p/19q status subgroups in CGGA-693 cohort.

Fig. 5

Comparisons between machine learning-derived prognostic signature (MLDPS) and 89 published prognostic signatures. (A) The C-index of 89 published signatures and MLDPS in CGGA-693, CGGA-325 and TCGA cohort. (B-D) Comparisons between C-index of MLDPS and 89 published signatures in CGGA-693 (B), CGGA-325 (C) and TCGA cohorts (D). (E-G) Comparisons between the area under the curve (AUC) values of MLDPS and 89 published signatures in predicting overall survival at 1-year (E) 3-year (F) and 5-year (G) in CGGA-693 cohort, respectively. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001.

Fig. 6

Pan-cancer survival analysis and functional characteristics of the high and low machine learning-derived prognostic signature (MLDPS) groups. (A-H) Kaplan-Meier survival analysis for overall survival (OS) in TCGA-ACC (A), TCGA-BRCA (B), TCGA-COAD (C), TCGA-LAML (D), TCGA-GBMLGG (E, excluded the AYAs glioma), TCGA-MESO (F), TCGA-SARC (G) and TCGA-THCA cohort (H). (I-J) The biological processes (BP) (I) and pathways (J) enriched in high MLDPS group. (K-L) The biological processes (BP) (K) and pathways (L) enriched in low MLDPS group.

Fig. 7

Immune microenvironment analyses. (A-C) The differences in ESTIMATE score, immune score, stromal score and tumor purity between high and low MLDPS groups in CGGA-693 (A), CGGA-325 (B) and TCGA cohorts (C). (D-F) The differences in immune cells between high and low MLDPS groups according to WHO II (D), WHO III (E) and WHO IV (F) in CGGA-693 cohort estimated by ssGSEA method. (G-J) Kaplan-Meier survival analysis for evaluating prognosis in patients received immunotherapy in PRJNA482620 (G), IMvigor 210 (H), GSE91061 (I) and GSE78220 cohort (J). (K-N) The stacked histogram shows the differences in immunotherapy responsiveness between high and low MLDPS groups in PRJNA482620 (K), Imvigor 210 (L), GSE91061 (M) and GSE78220 (N). *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001.