Rectal cancer accounts for approximately one-third of all colorectal cancer cases and continues to pose major therapeutic challenges despite improvements in screening and treatment. Neoadjuvant chemoradiotherapy (nCRT) followed by surgery is the established standard of care for locally advanced rectal cancer, and more recently, total neoadjuvant therapy (TNT) has gained traction as it may further increase rates of complete tumor elimination before surgery.
However, patient responses to radiotherapy vary dramatically: while 20 to 30 percent of patients achieve a pathological complete response and enjoy favorable long-term outcomes, a substantial proportion demonstrate minimal tumor regression or develop radioresistant disease. This unpredictability leads to overtreatment of some patients and undertreatment of others, resulting in unnecessary side effects for those who would not benefit and missed opportunities for more aggressive treatment in those who need it.
Conventional clinicopathologic variables such as tumor stage, differentiation grade, and imaging findings lack the sensitivity and specificity needed for individualized treatment prediction. Transcriptomic and genomic analyses offer a vast resource for identifying molecular determinants of treatment response, but inconsistent biomarker reproducibility across patient cohorts, batch effects in gene expression data, and tumor heterogeneity have historically hindered translation of research findings into clinical practice.
From a genomic interpretation perspective, biomarker candidates supported only by expression associations may be difficult to translate clinically unless their dysregulation can be linked to interpretable genetic variation mechanisms such as copy number alterations or recurrent somatic mutations. This variant-to-biomarker framework provides a more mechanistically grounded and clinically actionable approach to biomarker development.
Researchers integrated three independent publicly available GEO datasets from rectal cancer patients treated with neoadjuvant radiotherapy. Two cohorts (GSE46862, n=69; and GSE35452, n=46) included pre-treatment tumor samples with documented responder or non-responder labels and were used for model training. A third cohort (GSE94104) provided paired pre- and post-treatment samples for differential expression analysis. All three cohorts together covered 16,124 shared genes, and cross-platform batch effects were corrected using cohort-internal gene-wise z-score standardization.
To identify robust predictive features, five complementary machine-learning algorithms were applied in parallel: LASSO regression, Elastic Net, Random Forest, XGBoost, and information gain analysis. Each algorithm generated its own ranked list of candidate genes, and a Borda-type consensus ranking was applied to integrate these lists and identify the 40 most consistently nominated candidates. Forward stepwise feature selection using repeated 10-fold cross-validation was then applied to narrow the list to a final five-gene model, adding each gene only when it improved cross-validated model performance above a conservative threshold.
The final model was evaluated using out-of-fold predictions to minimize overfitting, with discrimination assessed via ROC and precision-recall curves, calibration analysis, and decision curve analysis to evaluate clinical utility. Model interpretability was assessed using SHAP (SHapley Additive exPlanations) beeswarm plots, which ranked each gene's contribution to individual predictions.
For the top-ranked gene CES1, the researchers performed additional genomic variation profiling using TCGA-READ (The Cancer Genome Atlas rectal adenocarcinoma dataset). Copy number alteration data from the GISTIC2 algorithm were used to assign copy number states for CES1, and somatic mutation analysis was performed using maftools. In vitro functional experiments were then conducted in two human rectal cancer cell lines (HT-29 and SW480) to confirm CES1's mechanistic role in radiation response.
After cross-platform standardization, principal component analysis confirmed that batch effects between the two efficacy cohorts were substantially reduced, allowing meaningful integration of the datasets. Differential expression analysis in the paired pre/post-treatment cohort identified 1,122 significantly changed genes following chemoradiotherapy, with pathway enrichment suggesting systematic remodeling of immune and inflammatory processes as well as DNA damage and cell-cycle stress response pathways.
The multi-algorithm consensus feature selection framework identified 40 top candidate genes from the integrated transcriptomic data. Among these, SHAP analysis of the final XGBoost model consistently identified CES1 as the single most dominant contributor to treatment response prediction. The full five-gene model (comprising CES1, HAL, SLC9A3, SMAD9, and PRIMA1) demonstrated high discriminative accuracy, reliable calibration, and favorable clinical utility in decision curve analysis compared to treat-all and treat-none strategies.
CES1 expression increased significantly after clinical chemoradiotherapy in the GSE94104 cohort, suggesting it is transcriptionally induced by radiation treatment. This was confirmed in vitro: in both HT-29 and SW480 rectal cancer cell lines, CES1 mRNA expression showed dose-dependent induction following X-ray irradiation at 0, 2, and 4 Gy doses, indicating that CES1 is a direct transcriptional response to ionizing radiation in rectal cancer cells.
Genomic analysis in TCGA-READ showed that CES1 expression was positively and significantly associated with its copy number status. Tumors with higher CES1 copy numbers expressed more CES1 mRNA. In contrast, coding-sequence mutations in CES1 were infrequent, suggesting that the primary mechanism of CES1 dysregulation in rectal cancer is copy-number-driven transcriptional variation rather than functional mutation. This copy-number linkage fulfills the variant-to-biomarker requirement for mechanistic interpretability.
To establish whether CES1 plays an active mechanistic role rather than simply serving as a passive biomarker, the researchers performed CES1 knockdown experiments using siRNA in HT-29 and SW480 cells. Cells with reduced CES1 expression showed significantly decreased radiation-induced apoptosis as measured by Annexin V/PI flow cytometry, demonstrating that CES1 normally promotes cell death in response to irradiation.
CES1 knockdown also enhanced clonogenic survival following irradiation, meaning that cells with lower CES1 expression retained a greater ability to form colonies and survive lethal radiation doses. This is a hallmark feature of radioresistant cells in laboratory models. In parallel, CES1-silenced cells displayed markedly increased migratory capacity in wound healing assays, indicating a more aggressive, invasive phenotype consistent with radioresistant tumors.
To confirm specificity of these effects, rescue experiments were performed by re-expressing CES1 in CES1-silenced cells. Restoration of CES1 expression reversed the radioresistant phenotype: radiation-induced apoptosis was re-established, clonogenic survival was reduced back toward control levels, and migratory capacity decreased. This rescue validation strongly supports CES1 as a direct and specific regulator of radiosensitivity rather than an off-target effect of gene knockdown.
Together, the functional data establish a mechanistic model in which adequate CES1 expression is required for cells to respond normally to ionizing radiation. When CES1 is reduced, rectal cancer cells evade radiation-induced apoptosis, survive with higher efficiency, and acquire an aggressive migratory phenotype, collectively defining a radioresistant and clinically dangerous cellular state.
A key innovation of this study is its use of a variant-to-biomarker framework that links expression-level biomarker findings to underlying genomic variation. Traditional biomarker studies often identify genes that are differentially expressed between responders and non-responders, but without understanding why those genes are dysregulated, translating them into clinical utility is difficult.
In this study, the researchers demonstrated that CES1 expression in rectal cancer is not simply altered at the transcriptional level by tumor biology, but is positively regulated by structural genomic events, specifically copy number amplifications. Tumors with more CES1 gene copies produce more CES1 mRNA. This means that standard genomic tumor profiling techniques, which are increasingly routine in clinical oncology, could theoretically capture CES1 copy number status and predict which patients are likely to respond favorably to radiotherapy.
The five-algorithm machine-learning consensus approach used in this study addresses a known weakness of single-algorithm biomarker studies: any single method may identify spurious associations driven by statistical noise or the specific characteristics of one dataset. By requiring candidate genes to rank highly across LASSO, Elastic Net, Random Forest, XGBoost, and information gain analyses, the final biomarker panel has substantially greater generalizability and reproducibility across independent cohorts.
Carboxylesterase 1 (CES1) is an enzyme involved in lipid metabolism, xenobiotic detoxification, and prodrug activation, making its connection to radiation response biologically intriguing. Radiation induces DNA damage and reactive oxygen species that trigger metabolic and stress-response programs, and metabolic enzymes like CES1 may play unexpected roles in coordinating cellular responses to genotoxic stress. The precise downstream mechanisms by which CES1 regulates apoptosis in the context of radiation remain to be fully characterized in future studies.
The identification of CES1 as a copy-number-linked predictor of radiotherapy response carries significant clinical implications for locally advanced rectal cancer management. The five-gene prediction model incorporating CES1 could potentially be used to stratify patients before treatment, identifying those likely to achieve a complete pathological response versus those likely to be radioresistant.
Patients predicted to be non-responders could be considered for alternative treatment intensification strategies, dose-escalation regimens, or novel radiosensitizing agents targeting the CES1 pathway. Conversely, patients predicted as responders could be spared unnecessarily aggressive treatment approaches. This kind of precision treatment selection has the potential to improve both therapeutic outcomes and quality of life for rectal cancer patients.
The clinical workflow for implementing such a biomarker would involve tumor biopsy at diagnosis, genomic profiling including copy number analysis, and integration of the five-gene expression model into treatment decision algorithms. Given that TCGA-READ data already supports the copy number association, and that the model performed well in independent external cohorts, the biomarker has a stronger foundation for clinical translation than many single-cohort biomarker studies.
The authors note that this study provides a mechanistically interpretable biomarker framework consistent with the emerging concept that integrating genetic variation profiling with functional validation can accelerate translation of research findings into clinical practice. Further prospective validation in larger clinical cohorts will be necessary before implementing CES1-based stratification in routine clinical decision-making for rectal cancer treatment.
This integrative computational and experimental study establishes CES1 as a predictive biomarker and copy-number-linked functional regulator of radiosensitivity in rectal cancer. The research demonstrates that CES1 satisfies multiple criteria for a clinically meaningful biomarker: it is consistently identified across five independent machine-learning algorithms, shows a mechanistic link to DNA copy number alterations, is functionally validated in two independent cell lines, and its knockout and rescue experiments confirm direct causal involvement in radiation response.
The five-gene predictive model combining CES1 with HAL, SLC9A3, SMAD9, and PRIMA1 demonstrated high discriminative accuracy with favorable calibration and clinical utility. The Sankey diagram visualization of the feature selection process showed strong convergence across independent algorithms toward the same final gene panel, providing confidence that the model reflects genuine biological signals rather than statistical artifacts.
The study's dual approach of computational discovery combined with genomic variation profiling and experimental functional validation represents a methodological advance over purely correlative biomarker studies. This framework can serve as a model for future biomarker development efforts in other cancer types where treatment response prediction remains a clinical challenge.
Future research should focus on prospective validation of the five-gene model in large clinical cohorts, investigation of the precise molecular mechanisms connecting CES1 to radiation-induced apoptosis signaling, and exploration of whether CES1 modulation could be leveraged therapeutically to enhance radiosensitivity in resistant rectal cancer tumors.