Identification of a ferroptosis-related gene signature predicting recurrence in stage II/III colorectal cancer based on machine learning algorithms

Abstract

Background: Colorectal cancer (CRC) is one of the most prevalent cancer types globally. A survival paradox exists due to the inherent heterogeneity in stage II/III CRC tumor biology. Ferroptosis is closely related to the progression of tumors, and ferroptosis-related genes can be used as a novel biomarker in predicting cancer prognosis. Methods: Ferroptosis-related genes were retrieved from the FerrDb and KEGG databases. A total of 1,397 samples were enrolled in our study from nine independent datasets, four of which were integrated as the training dataset to train and construct the model, and validated in the remaining datasets. We developed a machine learning framework with 83 combinations of 10 algorithms based on 10-fold cross-validation (CV) or bootstrap resampling algorithm to identify the most robust and stable model. C-indice and ROC analysis were performed to gauge its predictive accuracy and discrimination capabilities. Survival analysis was conducted followed by univariate and multivariate Cox regression analyses to evaluate the performance of identified signature. Results: The ferroptosis-related gene (FRG) signature was identified by the combination of Lasso and plsRcox and composed of 23 genes. The FRG signature presented better performance than common clinicopathological features (e.g., age and stage), molecular characteristics (e.g., BRAF mutation and microsatellite instability) and several published signatures in predicting the prognosis of the CRC. The signature was further stratified into a high-risk group and low-risk subgroup, where a high FRG signature indicated poor prognosis among all collected datasets. Sensitivity analysis showed the FRG signature remained a significant prognostic factor. Finally, we have developed a nomogram and a decision tree to enhance prognosis evaluation. Conclusion: The FRG signature enabled the accurate selection of high-risk stage II/III CRC population and helped optimize precision treatment to improve their clinical outcomes.

Keywords: ferroptosis-related gene; machine learning; prognosis; stage II/III colorectal cancer; tumor heterogeneity.

FIGURE 1

Workflow of the study.

FIGURE 2

Identification and construction of the best performance signature. (A) C-indices of 83 combinations of machine learning prediction models in five validation cohorts. (B) Determination of the optimal lambda was obtained when the partial likelihood deviance reached the minimum value and further generated the gene features with non-zero coefficients. (C) Lasso coefficient profiles of the candidate genes for FRG signature construction. (D) Determination of the optimal number of components when the iAUC reached the maximum value. (E) Categories and coefficients of 23 genes finally obtained in plsRcox regression.

FIGURE 3

Evaluation indicators and prognostic value of the FRG signature. (A) Time-dependent ROC analysis for predicting DFS at 1-, 3-, and 5-year across the training meta-cohort and all validation datasets. (B) C-indices of the signature across all datasets. (C) Kaplan–Meier survival curve of DFS between patients with a high-signature score and with a low-signature score in the training meta-cohort and five validation datasets. (D) Kaplan–Meier survival curve of DFS between patients in stage II vs. stage III patients and with respect to the stage and the identified gene signature of the meta-cohort.

FIGURE 4

Comparisons of clinical and molecular characteristics, and published signatures with the FRG signature. (A) C-index comparisons between clinical and molecular variables and signature in the training meta-cohort and validation datasets. (B) C-index comparisons between signature and five published signatures. * means p < 0.05, ** means p < 0.01, and *** means p < 0.001.

FIGURE 5

Interaction and combinations of the FRG signature with clinical and molecular features. Univariate Cox regression analysis (A) and multivariate Cox regression analysis (B) of prognostic factors for DFS for the training meta-cohort. (C) Subgroup analysis of the identified signature in clinical and molecular markers. (D) Prognostic nomogram predicting the probability of 1-, 3-, and 5-year DFS. (E) Calibration plot of the nomogram for 1-, 3-, and 5-year DFS prediction. Model performance is shown by the plot, relative to the 45-degree line, which represents perfect prediction. (F) DCA curve of the FRG signature and established risk factors in terms of DFS in the training cohort. The x-axis indicates the threshold probability, and the y-axis represents the net benefit. (G) A decision tree classifies patients into low-risk, intermediate-risk, and high-risk according to the probability of recurrent disease. * means p < 0.05, ** means p < 0.01, and *** means p < 0.001.