Development and biological characterization of a random survival forest prognostic model for colon adenocarcinoma using eight machine learning algorithms
Year: 2026
Type: article
Abstract: Background: Prognostic models for colon adenocarcinoma (COAD) frequently rely on predefined biologic gene sets or a single analytic framework, and few studies simultaneously address algorithm selection, biologic interpretation, comparison with conventional staging, and benchmarking against published COAD signatures. This study aimed to develop and biologically characterize a random survival forest (RSF) prognostic model for COAD.Methods: RNA-sequencing (RNA-seq) and clinical data from 455 The Cancer Genome Atlas (TCGA)-COAD patients were analyzed in a training cohort (n=324) and an internal test cohort (n=131). A 12-gene signature was derived through differential-expression screening, log-rank testing, and univariate Cox analysis. Eight survival algorithms were compared, and the best-performing RSF model generated the final risk score. Additional analyses included Gene Ontology (GO)/Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment, variable-importance and SHapley Additive exPlanations (SHAP) interpretation, Kaplan-Meier and subgroup interaction analyses, Cox regression, comparison of RSF vs. tumor-node-metastasis (TNM) vs. RSF + TNM using time-dependent receiver operating characteristic (ROC), concordance index (C-index), net reclassification improvement (NRI), and integrated discrimination improvement (IDI), Immuno-Oncology Biological Research (IOBR)-based immune-signature profiling, and benchmarking against published COAD signatures.Results: All 12 model genes mapped successfully for enrichment analysis, yielding 115 significant GO biological-process terms, 33 cellular-component terms, 25 molecular-function terms, and 4 KEGG pathways centered on telomere regulation, chromosome organization, and transcription-related programs. RSF showed near-perfect internal discrimination, with 1-, 3-, and 5-year areas under the ROC curves (AUCs) of 0.999, 0.999, and 0.998 in the training set, but external performance was more modest in the test set (0.631, 0.608, and 0.552). PPFIA4, GTF2H4, H4C5, ARL6IP4, LRP2, and TERT were consistently prioritized by variable importance (VIMP) and SHAP. Only pathologic M in training and lymphatic invasion in testing showed significant interaction with RSF score. In the test cohort, TNM and RSF + TNM outperformed RSF alone, and adding TNM to RSF improved NRI at 1 and 3 years. IOBR analyses identified 72 significant correlations in training and 57 in testing, dominated by exosome-, glycan-, metabolism-, and T-cell-related signatures.Conclusions: The 12-gene RSF score captures biologically meaningful transcriptomic and immune-related signals in COAD, but its most defensible translational role is as a complement to TNM rather than a replacement. Rigorous external RNA-seq validation remains necessary.
Source: Translational Cancer Research
Institutions: Hubei University for Nationalities, Minda Hospital
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: diamond
OpenAlex ID: W7224086139