Association of CT-based deep learning–derived consensus molecular subtypes with survival outcomes in locally advanced colon cancer: A secondary analysis of the phase III OPTICAL trial.

Abstract

3639 Background: The Consensus Molecular Subtypes (CMS) provide a robust framework for the biological stratification of colorectal cancer (CRC), but their clinical use is limited by the cost and complexity of transcriptomic profiling. We evaluated the prognostic and predictive utility of a non-invasive, CT-based deep learning model for CMS classification (CT-CMS) in patients with locally advanced colon cancer (LACC) enrolled in the OPTICAL trial. Methods: Patients from the OPTICAL trial with available baseline venous-phase contrast-enhanced CT images were included. A previously validated deep learning model was applied to predict CMS subtypes (CMS1-4). The association between CT-CMS and survival outcomes, including disease-free survival (DFS) and overall survival (OS), was assessed using multivariable Cox proportional hazards regression models adjusted for age, sex, TNM stage, and tumor location. Results: The study included 394 patients, with 189 patients in the neoadjuvant chemotherapy (NAC) group and 205 in the upfront surgery group. The median follow-up was 48.0 months. The distribution of CT-CMS subtypes was as follows: CMS1 (n=64, 16.2%), CMS2 (n=135, 34.5%), CMS3 (n=69, 17.5%), and CMS4 (n=126, 32.0%). In the total population, CMS3 showed the most favorable outcomes, with a 3-year OS rate of 96.1% (95% CI, 90.9%-100.0%) and a 3-year DFS rate of 87.9% (95% CI, 80.3%-96.2%). Multivariable analysis confirmed that CMS3 was independently associated with superior prognosis compared to CMS1 (adjusted hazard ratio [aHR] = 0.39, 95% CI 0.17-0.92, p = 0.030). In the NAC group, CMS was a significant independent prognostic factor for DFS (p = 0.044), with CMS3 (p = 0.027) and CMS2 (p = 0.046) showing significant improvement over CMS1. Notably, patients with CMS2 tumors exhibited significant survival benefits from neoadjuvant chemotherapy compared with upfront surgery, with improved OS (aHR = 0.08, 95% CI 0.01-0.62, p = 0.016, log-rank p = 0.002) and DFS (aHR = 0.45, 95% CI 0.20-0.99, p = 0.047). Conclusions: CT-based CMS3 was associated with favorable prognosis, while CMS2 identified patients were most likely to benefit from neoadjuvant chemotherapy. These findings support the clinical utility of CT-CMS as a non-invasive, pre-treatment tool for molecular stratification in LACC. All 3-year DFS rate HR (95% CI) P CMS1 76.4% (66.2%-88.2%) Reference - CMS2 79.8% (73.1%-87.1%) 0.67 (0.36-1.27) 0.219 CMS3 87.9% (80.3%-96.2%) 0.39 (0.17-0.92) 0.030 CMS4 78.2% (71.3%-85.8%) 0.89 (0.49-1.64) 0.715 NAC vs surgery (reference) * CMS1 70.3% vs 81.4% 1.20 (0.45-3.15) 0.717 CMS2 86.5% vs 72.8% 0.45 (0.20-0.99) 0.047 CMS3 90.2% vs 85.2% 0.36 (0.08-1.60) 0.180 CMS4 78.2% vs 78.3% 1.04 (0.50-2.17) 0.915 * For treatment comparisons, 3-year DFS rates are presented as NAC versus surgery.