Integrative machine learning for identifying transcriptomic signatures of neoadjuvant chemoradiotherapy response in locally advanced rectal cancer

Background: Predicting tumour response to neoadjuvant chemoradiotherapy (nCRT) in locally advanced rectal cancer remains an unresolved clinical problem. The substantial inter-patient variability in pathological response means that molecular biomarkers capable of prospective patient stratification are urgently needed. However, transcriptomic biomarker studies in this setting are prone to methodological limitations — particularly data leakage — that produce irreproducible performance estimates and hinder translation. Objective: This study aimed to identify a biologically interpretable transcriptomic signature of nCRT response in rectal cancer through an integrative machine learning framework designed with strict attention to leakage prevention and methodological transparency. Methods: Gene expression microarray profiles from 46 rectal cancer patients (24 responders, 22 non-responders; GEO: GSE35452; 22,189 gene features) were analysed. Phenotype was defined by the Mandard tumour regression grade (TRG) [1] : responders = TRG 1–2 (complete or near-complete pathological response); non-responders = TRG 3–5. All patients received fluoropyrimidine-based concurrent chemoradiotherapy followed by surgery. A preliminary variance filter — applied to the complete dataset because variance is a class-label-free operation — retained the 1000 most variable genes; sensitivity analyses at 500, 2000, and 5000 genes are reported in Supplementary Note S1. A label-leakage-controlled sklearn.pipeline.Pipeline was constructed in which all preprocessing — including standardisation and feature selection — was performed independently within each training fold. Four feature selection algorithms were applied per fold (ANOVA F-statistic, mutual information, L1-penalised logistic regression, and Random Forest importance), and a fold-specific consensus gene set was constructed at each outer iteration (Algorithm ??). Eight classifiers were evaluated as an exploratory multi-algorithm comparison. Results: The pooled out-of-fold AUC across the five outer folds was 0.987 for Gaussian Naïve Bayes (95% percentile bootstrap CI: 0.949–1.000), with a mean fold-wise AUC of 0.990 ± 0.020. Based on the pooled out-of-fold predictions, specificity was 87.0%, and balanced accuracy was 91.5%. A 109-gene consensus candidate panel was identified across the fold-specific feature-selection analyses. Within this panel, six genes showed statistically significant differential expression between responders and non-responders: WBP1L , CEBPB , ATE1 , and BAG2 showed lower expression in non-responders, whereas BSG and ZFAND2B showed higher expression. Functional enrichment analysis of the 109-gene candidate panel, using the 1000 variance-filtered genes as the background universe, identified biological processes related to DNA damage response, signal transduction, apoptotic signalling, protein regulation, and immune-related processes. The results are exploratory and require validation in independent cohorts. FAM175A was selected by feature importance but showed non-significant DE ( 𝑝 = 0 . 3 9 3 ) within the discovery cohort. Functional enrichment analysis of the consensus candidate set (clusterProfiler; Benjamini–Hochberg FDR < 0.05; background = 1000 variance-filtered genes) identified DNA damage response, apoptotic signalling, and MAPK cascade as top enriched biological processes. Conclusion: This discovery study demonstrates that an integrative, label-leakage-controlled machine learning framework can identify biologically coherent transcriptomic candidates for nCRT response prediction in rectal cancer. Given the single retrospective cohort of 46 patients and absence of independent external validation, all findings should be regarded as hypothesis-generating. External validation in independent prospective cohorts is required before any clinical or translational inference can be drawn.