Abstract / Summary
Background: The fecal lipidomic changes underlying the colorectal adenoma-carcinoma sequence remain incompletely characterized, particularly regarding the continuous metabolic trajectory and heterogeneity of precancerous adenomas. We aimed to use publicly available multi-omics data and an integrative computational framework to identify candidate lipidomic features associated with a CRC-like metabolic state in adenomas.
Methods: We developed an extreme-phenotype machine learning strategy using fecal lipidomics from healthy controls and colorectal cancer (CRC) patients (Study ID: ST003798) to build a diagnostic model, which was then blindly applied to adenoma patients to compute a Fecal Lipidomic Malignancy Risk Score (FL-MRS), interpreted here as a CRC-like lipidomic similarity score. SHAP analysis prioritized influential lipid features, and cross-sectional pseudotime trajectory inference reconstructed the metabolic continuum. Transcriptomic data from TCGA-COAD and single-cell RNA-seq (Broad Institute) were integrated to explore potential tissue-level correlates. Targeted single-molecule trend verification of the top lipid candidate was performed in an independent cohort (ST002787), as full model replication was not feasible due to limited inter-cohort feature overlap. Multiple sensitivity analyses were conducted to assess the robustness of the computational pipeline.
Results: The Random Forest model showed robust performance on the independent test set (AUC = 0.864). When applied to adenomas, 51.7% of patients exceeded the FL-MRS threshold derived from extreme phenotypes; however, this proportion far exceeds the known clinical adenoma-carcinoma progression rate (approximately 5-10%), indicating that FL-MRS should be interpreted as a metabolic similarity metric rather than a direct cancer risk probability. SHAP analysis prioritized arachidonic acid-derived cholesterol ester CE(20:4) as the top predictive feature, with 100% bootstrap selection frequency. Integrative analysis of independent transcriptomic datasets identified upregulation of PTGS2 (COX-2) in tumor tissue and its predominant expression in stromal cells. External single-molecule targeted verification confirmed an accumulation trend of CE(20:4) and an early adenoma-phase peak of eicosanoid mediators (including oxidative stress and LOX-pathway metabolites). Notably, CE(20:4) levels did not differ significantly between adenoma and CRC (P = 0.055), consistent with its proposed role as an early event marker. Pseudotime trajectory inference was insensitive to root node assignment.
Conclusions: This computational study suggests that fecal CE(20:4) and the associated COX-2 pathway may represent candidate features of a CRC-like metabolic state in colorectal adenomas. FL-MRS is not proposed as a clinical diagnostic or risk-prediction tool; rather, it serves as a research instrument for quantifying CRC-like metabolic similarity and guiding future validation studies. All findings are derived from publicly available retrospective data without independent experimental validation and should be regarded as hypothesis-generating. The complete analysis code is publicly archived to ensure reproducibility and facilitate future validation.