Stratifying By Baseline Tumor Cell Density Indicates Treatment Effect Heterogeneity in Rectal Cancer

Patients with rectal cancer who have higher tumor cell density (TCD) experienced differential outcomes following irinotecan-intensified neoadjuvant chemoradiotherapy (CRT), compared with patients who received standard radiotherapy.¹ These results came from a post-hoc analysis of the phase 3 ARISTOTLE trial (ISRCTN09351447) published in eBioMedicine.

In Cox proportional hazards interaction models, the treatment-TCD interaction was significant for disease-free survival (DFS; HR, 0.44; 59% CI, 0.24-0.82; P = .009) and overall survival (OS; HR, 0.31; 95% CI, 0.16-0.64; P = .001). Per DFS, including of the interaction term reduced the Akaike Information Criterion (AIC) from 1936.62 to 1931.71 and increased the log-likelihood from –966.31 to –962.87, which highlights the improved model fit. The interaction was deemed statistically significant via a likelihood ratio test (χ² = 6.88; P = 0.009). Regarding OS, the interaction model also demonstrated improved fit, with an AIC of 1457.47 vs 1448.86 and a log-likelihood of –726.73 vs –721.43. This was significant as well (χ² = 10.61; P = 0.001).

Sensitivity analyses were restricted to patients with available DNA mutation data revealed that the treatment-TCD interaction was significant for DFS (HR, 0.39; 95% CI, 0.17-0.91; P = .029) and OS (HR, 0.30; 95% CI, 0.11-0.91; P = .017) after adjustment for KRAS and TP53. Improved fit was also observed for both DFS and OS.

The Cox proportional hazards regression showed that, between the TCD-low and TCD-high subgroups, no significant difference was observed for DFS (HR, 1.04; 95% CI, 0.77-1.40; P = .78) or OS (HR, 1.15; 95% CI, 0.81-1.62; P = .44).

In the group of patients who received irinotecan-intensified CRT, those who were TCD-high showed a nonsignificant trend towards longer DFS (HR, 0.68; 95% CI, 0.44-1.07; P = .10) and OS (HR, 0.63; 95% CI, 0.37-1.06; P = .08) compared with those in the TCD-low subgroup. In the regular CRT group, patients with TCD-high status had significantly shorter DFS (HR, 1.54; 95% CI, 1.02-2.33; P = .04) and OS (HR, 2.04; 95% CI, 1.26-3.29; P = .003) compared with the TCD-low subgroup.

Across the entire cohort, irinotecan-intensified CRT did not significantly affect DFS (HR, 0.89; 95% CI, 0.66-1.21; P = .46) and OS (HR, 0.92; 95% CI, 0.65-1.29; P = .62). Within the TCD-high subgroup, irinotecan-intensified CRT did confer significantly improved DFS and OS vs regular CRT; in the TCD-low subgroup, no evidence of differential treatment effect was observed for DFS, as well as a trend towards shorter OS.

“The principal finding of this study is a predictive signal: stratification by baseline TCD indicated heterogeneity of treatment effect, supported by a significant treatment-TCD interaction for both DFS and OS. This interaction remained significant after adjustment for KRAS and TP53 in the mutation-available subset,” wrote lead study author Zhuoyan Shen, of the Department of Medical Physics and Biomedical Engineering at University College London, and coauthors. “Patients with high TCD who received [irinotecan-intensified CRT] experienced clinically meaningful improvements in DFS and OS and were more likely to achieve pathological complete response [pCR], suggesting that tumours with a lesser stromal component may respond more favourably to irinotecan.”

This analysis of ARISTOTLE included 414 patients with available pre-treatment biopsy who were recruited between October 2011 and July 2018. Patients were either enrolled to receive standard CRT or irinotecan-intensified CRT, which also gave patients 60 mg/m² of intravenous irinotecan once weekly.

TCD was defined as the proportion of tumor cells relative to all classifiable cells in the neoplastic epithelial and tumor-associated stromal region that were detected by the AI model. Patients with a TCD higher than or equal to 0.5 were classified as TCD-high, and those with TCD lower than 0.5 were TCD-low.

The AI tissue classifier demonstrated a precision of 0.873 for tumor region identification and a recall of 0.984 for tumor-associated tissues. In 116 collected WSIs, the AI-derived TCD was strongly concordant with manually assessed TCD using Lin’s concordance correlation analysis (CCC, 0.831; 95% CI, 0.760-0.866).

No significant differences in baseline clinical characteristics or survival outcomes were observed between the sub-cohort analyses and the full ARISTOTLE population. Across 414 total patients, 45% (n = 188) were stratified as TCD-high and 55% (n = 226) were TCD-low.

The study was designed to evaluate the prognostic and predictive relevance of the AI-derived TCD. The primary outcomes measures in the study were DFS and OS, and pathological complete response was used as a secondary outcome measure.

The AI framework used consisted of a tissue classifier, which was an EfficientNet trained on a specific dataset, and a cell detector, which was SoftCTM.

“These findings suggest that AI-identified TCD is associated with differential outcomes after standard versus irinotecan-intensified nCRT, while warranting independent validation before future use in treatment selection,” wrote the study authors.

Reference

Shen Z, Brand D, Simard M, et al. Tumour cell density quantified by artificial intelligence is associated with differential benefit from irinotecan-based chemo-radiotherapy in locally advanced rectal cancer: a post-hoc study of the phase 3 ARISTOTLE trial. eBioMedicine. 2026;130:106397. Published online July 27, 2026. doi:10.1016/j.ebiom.2026.106397