Untargeted baseline metabolomic profiling to identify candidate markers of long-term severe cancer-related fatigue (CRF) in breast cancer (BC) survivors.

12118 Background: CRF is common and debilitating among BC survivors; however its predictors and underlying biology remain partially defined. We sought to identify baseline plasma metabolomic markers of subsequent CRF to inform risk prediction and symptom mechanisms. Methods: Baseline (pre-treatment [tx]) plasma samples from women with stage I-III ER+/HER2− BC in the CANTO cohort (NCT01993498) underwent untargeted metabolomic profiling (liquid chromatography and high-resolution mass spectrometry with electrospray ionization), yielding metabolic features for analysis. CRF was measured using EORTC QLQ-C30 (severe CRF: score≥40/100). Three penalized regression approaches (lasso, adaptive lasso, elastic net) were used to select baseline metabolites associated with severe CRF, adjusting for clinical covariates, including pre-tx CRF, age, body mass index, comorbidities, socioeconomic status, psychological distress, health behaviors, and tx. The cohort was split 2/3:1/3 (train:test); penalization factor λ was tuned in the train set by cross-validation. Clinical-only and clinical+metabolite features models were compared using test-set AUC, with bootstrap confidence intervals from resampled predictions (B = 1000). Covariate missingness was addressed by multiple imputation on 15 datasets. Metabolite annotation was conducted computationally using W4M, with manual curation and interpretation of spectra to identify key metabolites relevant to CRF. Results: In the train set (N = 645), mean age was 58 years (SD 10), 44% received chemo- and 90% endocrine-tx, 20% reported severe pre-tx CRF (similar characteristics in the test set; N = 325). 1935 metabolic features were analyzed. For severe global CRF at year-4 (prevalence 31%) the adaptive lasso clinical+metabolite features model included a 51-feature signature and had AUC = 0.70, specificity = 0.89, and sensitivity = 0.39 (sensitivity difference vs. clinical-only +0.17 [95% CI +0.06 to +0.29; p < .001]). Features with the largest penalized coefficients were concordant with univariate associations (all in the same direction) and showed clear differences in standardized levels between severe vs non-severe CRF. Nineteen features belonged to the annotated reference dataset, enabling putative identification of metabolites spanning endogenous pathways (including glycine [inverse association] and chenodeoxycholic acid [positive association]) and a likely exogenous/xenobiotic-related feature (mandelonitrile [positive association]). Conclusions: Baseline untargeted metabolomics identified a 51-feature signature associated with severe global CRF at year-4 after ER+/HER2− BC. Adding features selected via adaptive lasso to clinical covariates increased model sensitivity, and putative pathway mapping highlighted candidates for targeted validation and mechanistic studies.