AI-Guided Ferroptosis Biomarker Discovery and Routine Laboratory–Based Machine Learning for Predicting Nodal Metastasis in Colon Cancer


BEŞLİ N., Vartanoglu Aktokmakyan T., Koyuncu A., Sarikamis Johnson B., Celik U.

Metabolites, cilt.16, sa.8, 2026 (SCI-Expanded, Scopus)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 16 Sayı: 8
  • Basım Tarihi: 2026
  • Doi Numarası: 10.3390/metabo16080557
  • Dergi Adı: Metabolites
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, BIOSIS, Chemical Abstracts Core, EMBASE, Directory of Open Access Journals, Natural Science Collection (ProQuest), Biological Science Database (ProQuest)
  • Anahtar Kelimeler: colon cancer, ferroptosis, lymph node metastasis, machine learning, biomarkers, iron metabolism
  • Sağlık Bilimleri Üniversitesi Adresli: Evet

Özet

Background: Accurate preoperative prediction of lymph node metastasis (LNM) remains challenging in colon cancer. Ferroptosis, an iron-dependent regulated cell-death pathway, is implicated in tumor invasion and metastasis. Yet, the clinical value of routine iron-metabolism indices as indirect systemic indicators of ferroptosis-relevant biology remains underexplored. Methods: A multi-stage workflow was implemented. First, transformer-based literature mining (PubMedBERT, BioLinkBERT, BioBERT) prioritized ferroptosis-associated genes and metabolites. Second, transcriptomic validation was performed in TCGA-COAD and GSE39582 cohorts using differential expression analysis. Third, machine-learning models (logistic regression, random forest, XGBoost) were developed to predict pathologically confirmed LNM (n = 421) under three feature configurations: preoperative baseline, preoperative plus iron-metabolism markers, and a pathology-augmented model. LVI prediction was separately evaluated (n = 416) using ferroptosis-only, clinical-only, and combined sets. Results: Literature mining identified a core ferroptosis axis dominated by redox and iron-handling regulators. Transcriptomic analyses demonstrated robust tumor–normal separation and consistent perturbation of key ferroptosis genes (SLC7A11, GPX4, ACSL4, PTGS2). Preoperative models achieved moderate LNM discrimination (best AUC = 0.704), whereas a postoperative pathology-augmented explanatory benchmark achieved an AUC of 0.818. Out-of-fold risk stratification identified a low-risk subgroup with 26.7% nodal positivity, compared with 44.7% in the overall cohort. For LVI prediction, models based on iron-handling indices showed modest discrimination, and their addition to clinical variables produced model-dependent numerical changes in performance (maximum ΔAUC = +0.114). These routinely measured indices provided exploratory predictive information but should not be interpreted as direct measures of tumor ferroptosis. Conclusions: This integrative framework links ferroptosis-related biology with clinical machine learning and suggests that routine iron-handling indices may provide exploratory complementary information for LVI prediction and preoperative nodal-risk stratification, whereas postoperative pathology variables provide additional explanatory value but are not part of a deployable preoperative model.