Regorafenib response prediction in metastatic colorectal cancer by a novel genomic and transcriptomic model.

Abstract

3135 Background: The multi-kinase inhibitor regorafenib (Rego) is approved for the treatment of refractory metastatic colorectal cancer (CRC). However, its efficacy is limited, and its use is frequently associated with substantial toxicities. Identifying biomarkers predicting Rego-response could improve therapeutic outcomes and reduce unnecessary treatment-related adverse effects in non-responders. Methods: A predictive model for Rego-response was developed based on transcriptomic and genetic data from 41 CRC cell lines. Cell lines were classified into Rego-sensitive versus -resistant groups based on drug sensitivity data from the CTRP2 database. Several machine-learning algorithms were evaluated, with the Generalized Linear Model via Elastic Net (GLMNET) achieving the highest predictive performance. Model accuracy was assessed using leave-one-out cross-validation. Further validation was performed using transcriptomic (WTS) data from 24,384 real-world CRC patients assessed by Caris Life Sciences, which included 720 patients treated with Rego. Results: The predictive model identified key cell line features associated with Rego-response, including gene expression signatures (e.g., ZNF441, CCDC82, ZFP69 ) and specific mutations (e.g., RALGAPA1, MORC1 ). Transcriptome profiling showed that Rego responders exhibited enrichment in cell-cycle regulation and DNA-repair mechanisms, while non-responders showed a stroma-rich microenvironment with significant endothelial and fibroblast infiltration. External validation using WTS data from real-world Rego-treated CRC patients revealed that predicted responders had a prolonged time-on-treatment (p = 0.02, HR = 0.79) and median overall survival (p = 0.01, HR = 0.76) compared to predicted non-responders. This association was specific to Rego-response, as there was no survival difference between predicted responders and non-responders among patients not treated with Rego (p = 0.72, HR = 1.0). Conclusions: This novel predictive model successfully identified and validated molecular features associated with Rego-response in CRC. The transcriptomic and genetic signature holds significant potential for improving personalized treatment strategies by identifying patients most likely to benefit from Rego and prevents unnecessary Rego-associated toxicities in non-responders.