Integrated transcriptomic, machine-learning and structural analyses prioritize NR3C1 for experimental follow-up in colorectal cancer

Objectives To construct a resveratrol-associated colorectal cancer (CRC) candidate space, distinguish tissue-discrimination features from candidates for mechanistic follow-up, and audit the stability of supervised feature selection. Methods Predicted resveratrol targets were intersected with a 2,025-gene CRC disease-associated union derived from differential expression and WGCNA. A globally prescreened nonlinear benchmark evaluated 113 workflows in a 46-gene feature space and selected Stepglm(backward)+XGBoost; its out-of-fold performance was conditional on that globally prescreened feature set. A separate fold-specific DEG-screened ridge-logistic audit used repeated nested five-fold cross-validation with fold-specific supervised screening, inner selection of feature number and ridge penalty, and outer held-out prediction. Expression direction and tissue-level discrimination were assessed in GEO external evaluation cohorts and TCGA-COAD/READ. Exploratory marker-inferred cellular, perturbational and structural context used GSE200997, L1000FWD, molecular docking and one 100 ns NR3C1–RES trajectory. A revised exploratory five-candidate evidence synthesis incorporated nested-CV final-selection frequency and prespecified sensitivity analyses. Results Forty-nine RES–CRC candidate genes were identified. The globally prescreened Stepglm(backward)+XGBoost benchmark showed conditional out-of-fold discrimination (AUC = 0.986; 95% CI 0.969–1.000) and external-evaluation AUCs of 0.990 and 0.866. In the separate ridge-logistic audit, repeat-level median held-out AUC was 0.987 (IQR 0.985–0.988); pooled and sample-mean AUCs were treated as descriptive because repeated predictions were correlated within samples. Final-selection frequencies were 100% for CA1, 93% for NR3C1 and 78% for EDNRA, compared with 0% for TACR2 and 1% for PPARG. TCGA reproduced the expression directions of all five initially prioritized genes; these resected tissue contrasts were not interpreted as clinical screening performance. Marker-inferred single-cell analysis localized NR3C1 mainly to immune and stromal compartments, and the epithelial tumor–normal comparison was not significant after FDR correction (FDR = 0.236). Glucocorticoid-related opposing perturbagens provided exploratory pharmacological context. NR3C1 yielded the most favorable Vina score among the examined structures, without implying a quantitative cross-protein affinity ranking, and RES remained associated with the predicted NR3C1 pocket during the single analyzed trajectory. In the revised five-candidate synthesis, NR3C1 ranked first after removal of each evidence domain and in 91.5% of 20,000 random-weight settings; the asymmetric evidence depth precludes interpretation as an unbiased comparison. Conclusions Leakage-aware resampling supported CA1, NR3C1 and EDNRA as comparatively stable CRC-associated candidates, whereas TACR2 and PPARG showed limited feature stability. NR3C1 was prioritized as a candidate for experimental mechanistic follow-up. These findings are hypothesis-generating and do not establish direct binding or NR3C1-mediated resveratrol activity.