Computational image features of immune architecture reveal molecular patterns associated with outcome and risk of recurrence in microsatellite-stable or MSI-low colon cancers.
Abstract
Background
Tumor-infiltrating lymphocyte (TIL) density provides prognostic information in colon adenocarcinoma (COAD), but the additional value of TIL spatial organization remains unclear.
Methods
We developed CoATIL, a computational pathology framework quantifying TIL spatial architecture from routine H&E-stained whole-slide images. The study included 461 patients with microsatellite-stable or MSI-low COAD. The model was trained using 140 patients from University Hospitals Cleveland Medical Center, validated internally in 52 patients, and externally in The Cancer Genome Atlas (n = 130), and Guangdong Provincial People's Hospital (n = 139). Prognostic performance was assessed using Kaplan-Meier analysis and multivariable Cox regression. CoATIL was compared with a clinical model and eight foundation-model-based multiple-instance learning approaches.
Results
CoATIL was associated with progression-free survival (HR per standard-deviation increase, 1.32; 95% CI 1.13-1.55; p < 0.001; C-index=0.66) and remained prognostic after adjustment for clinicopathologic factors in each validation cohort (adjusted HRs, 2.00-4.40; all p < 0.05). High-versus-low-risk stratification was retained within stage II (HR 3.07; 95% CI 1.83-5.15; p < 0.001) and stage III disease (HR 3.24; 95% CI 1.28-8.19; p = 0.013). The combined clinical-CoATIL model achieved a higher C-index than the clinical model alone (0.69 versus 0.61; paired bootstrap difference, 0.08; 95% CI 0.03-0.14). Exploratory genomic analyses associated AMER1 and FBXW7 mutations with lower and higher CoATIL risk scores, respectively.
Conclusions
Spatial immune architecture extracted from routine histology provides complementary prognostic information in COAD. Prospective validation is warranted.