Multi-omics Signature Predicts Anti-EGFR Therapy Benefit in Colorectal Cancer Liver Metastases: A Multi-center Cohort Study
Yu Liu, Yuan Chen, Guofeng Zhou, Shizhao Zhou, Xueyi Zhang, Songbin Lin, Mengzu Zeng, Liansheng Wang, Jianmin Xu, Wenju ChangIntroduction:
Clinical management of RAS wild-type colorectal cancer liver metastases (CRLM) remains challenging because many patients exhibit primary resistance to anti-EGFR therapy. Our research centered on the development of a multi-omics deep learning framework, designed to bridge the gap between complex multi-omics data and the necessity for precise therapeutic response forecasting in this specific cohort.
Materials and Methods:
Patient cohorts receiving cetuximab from a prior investigation (PMID: 30305811) constituted the training and testing sets. For external validation, an independent cohort of consecutive patients with RAS wild-type CRLM patients was prospectively enrolled from other institutions between January and December 2018. Utilizing the PyTorch deep learning framework, we initially developed individual radiomic and genetic signatures. Subsequently, computed tomography (CT) images and genetic data were processed through pre-trained ResNet18 and Random Forest models, respectively. The final classification probability of the integrated model was calculated by weighted summation of the output probabilities from the two models, with weights of 3 and 7, respectively.
Results:
The developed signature demonstrated predictive capability for cetuximab sensitivity, with area under the curve (AUC) values of 0.75 for the radiomic model, 0.81 for the genetic model, and 0.86 for the combined model. In contrast, it did not predict response to chemotherapy (fusion signature AUC: 0.54). Within cohorts treated with cetuximab, the fusion signature proved superior to established biomarkers for identifying treatment-sensitive cases (hazard ratio (HR), 17.9; 95% confidence interval (CI), 3.22–154.31; P = 0.003). Furthermore, it showed a significant correlation with progression-free survival (PFS), with a median PFS of 9.0 versus 5.0 months (HR, 0.44; 95% CI, 0.20–0.99; P = 0.047).
Discussion:
The proposed multi-omics signature showed a promising ability to identify RAS wild-type CRLM patients more likely to benefit from anti-EGFR therapy using routinely available pretreatment CT and genomic data. Although the external validation cohort was relatively small and further validation is still needed, it may provide a practical tool for early treatment stratification and individualized decision-making.
Conclusions:
The developed multi-omics signature demonstrated promising performance in predicting sensitivity to anti-EGFR therapy and may help refine survival-based treatment stratification in patients with RAS wild-type CRLM.