The addition of irinotecan to standard neoadjuvant chemoradiotherapy did not improve pathological complete response or survival outcomes across the overall population of the phase 3 ARISTOTLE trial. However, a post-hoc analysis suggests that artificial intelligence-derived tumour cell density may help identify patients with locally advanced rectal cancer who are more likely to benefit from irinotecan intensification.
Patients with high tumour cell density in their pretreatment biopsies experienced significantly longer disease-free and overall survival with irinotecan-based chemoradiotherapy than with standard chemoradiotherapy. No corresponding benefit was observed among patients with low tumour cell density.
The article, titled “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,” was published in eBioMedicine on July 27, 2026.
Authors: Zhuoyan Shen, Douglas Brand, Mikaël Simard, Nicholas P. West, Andre Lopes, Rubina Begum, Ying Zhang, Gary Royle, David Sebag-Montefiore, Charles-Antoine Collins Fekete, and Maria A. Hawkins.
The Need for Biomarker-Guided Treatment Selection
Locally advanced rectal cancer is generally defined as clinical stage II or III disease and is associated with an increased risk of recurrence. Treatment has increasingly moved towards intensive neoadjuvant strategies, including total neoadjuvant therapy, in which chemotherapy and radiotherapy are administered before surgery.
Despite these developments, current approaches have largely focused on altering treatment sequencing rather than tailoring treatment selection according to the biological features of an individual tumour.
The tumour microenvironment contains tumour cells, immune cells, fibroblasts, blood vessels, extracellular matrix, and other stromal components that can influence tumour progression and resistance to therapy. A greater stromal component may create a protective environment around the tumour and reduce its sensitivity to treatment.
Tumour cell density provides a cell-level measurement of tumour content. In this study, it was defined as the proportion of tumour cells among all classifiable tumour and non-tumour cells within neoplastic epithelial and tumour-associated stromal regions.
Visual assessment of tumour–stroma composition may be affected by interobserver variability, while manual tumour cell density assessment is time-consuming and difficult to implement at scale. The investigators therefore developed an artificial intelligence framework to quantify tumour cell density automatically from routine haematoxylin and eosin-stained whole-slide images.
Post-Hoc Analysis of the Phase 3 ARISTOTLE Trial
The phase 3 ARISTOTLE trial evaluated whether adding irinotecan to long-course capecitabine-based chemoradiotherapy could improve outcomes in patients with MRI-defined locally advanced rectal cancer.
The present hypothesis-generating analysis included 414 trial participants with available digitised pretreatment biopsy samples. Investigators developed an AI framework that combined tissue classification with cell detection to quantify tumour cell density from routine haematoxylin and eosin-stained whole-slide images.
Tumour cell density was defined as the proportion of tumour cells among all classifiable tumour and non-tumour cells within the neoplastic epithelial and tumour-associated stromal regions detected by the AI model, with morphologically normal tissue excluded. Patients were classified as TCD-high when tumour cell density was at least 0.5 and TCD-low when it was below 0.5. The cutoff was not optimised according to clinical outcomes.
Disease-free survival and overall survival were the primary outcomes of the analysis, while pathological complete response was assessed as a secondary outcome among patients with available surgical pathology. Of the 414 patients, 188, or 45%, were classified as TCD-high and 226, or 55%, as TCD-low.
Treatment Effect Differed According to Tumour Cell Density
A statistically significant interaction was observed between tumour cell density and treatment assignment for both disease-free survival and overall survival. For DFS, the treatment-by-TCD interaction had a hazard ratio of 0.44, with a likelihood-ratio chi-square value of 6.88 and a p value of 0.009. For OS, the interaction hazard ratio was 0.31, with a chi-square value of 10.61 and a p value of 0.001. The interaction remained significant in a sensitivity analysis restricted to patients with available mutation data after adjustment for KRAS and TP53 alterations.
Across the overall study cohort, irinotecan-based chemoradiotherapy did not significantly improve DFS compared with standard chemoradiotherapy. The HR was 0.89, with a 95% confidence interval of 0.66–1.21 and a p value of 0.46. There was also no significant improvement in overall survival, with a hazard ratio of 0.92, a 95% confidence interval of 0.65–1.29, and a p value of 0.62.
However, treatment outcomes differed when patients were analysed according to baseline tumour cell density. Among patients with TCD-high tumours, irinotecan-based chemoradiotherapy was associated with significantly longer DFS. The hazard ratio was 0.57, with a 95% confidence interval of 0.36–0.90 and a p value of 0.014. OS was also significantly longer with the irinotecan-containing regimen, with a hazard ratio of 0.50, a 95% confidence interval of 0.30–0.84, and a p value of 0.008.
In the TCD-low subgroup, no significant disease-free survival benefit was observed with irinotecan-based chemoradiotherapy. The hazard ratio was 1.29, with a 95% confidence interval of 0.85–1.95 and a p value of 0.22.
Patients with TCD-low tumours who received irinotecan also showed a non-significant trend towards shorter overall survival. The hazard ratio was 1.55, with a 95% confidence interval of 0.96–2.51 and a p value of 0.07. The investigators proposed that irinotecan-associated toxicity may outweigh its therapeutic benefit in patients with limited tumour cellularity. However, this remains a possible explanation and cannot be established from the current post-hoc analysis.
Pathological Complete Response
Among the 355 patients with available pathological complete response data, pCR occurred in 31 patients, or 17%, receiving standard chemoradiotherapy and 32 patients, or 18%, receiving irinotecan-based chemoradiotherapy. There was no significant difference between the treatment groups, with an odds ratio of 1.11, a 95% confidence interval of 0.64–1.91, and a p value of 0.72.
Among patients with TCD-high tumours, the pathological complete response rate was 23% with irinotecan-based chemoradiotherapy and 11% with standard chemoradiotherapy. The unadjusted odds ratio was 2.46, with a 95% confidence interval of 1.01–5.98 and a p value of 0.042.
However, this association was no longer statistically significant after adjustment for multiple testing, when the adjusted p value was 0.13. In the TCD-low subgroup, pathological complete response occurred in 14% of patients receiving irinotecan-based chemoradiotherapy and 22% of those receiving standard chemoradiotherapy. The difference was not statistically significant, with an odds ratio of 0.60, a 95% confidence interval of 0.28–1.29, and a p value of 0.19.
The pathological complete response findings should therefore be considered exploratory, particularly because the apparent difference in the TCD-high subgroup did not remain statistically significant after correction for multiple comparisons.
Potential Role as a Predictive Biomarker
Tumour cell density was not significantly associated with disease-free or overall survival when treatment assignment was not considered. Its potential value instead appeared to lie in modifying the effect of treatment, supporting its further evaluation as a candidate predictive biomarker.
A prognostic biomarker provides information about expected outcomes regardless of the therapy received. A predictive biomarker identifies patients who may be more or less likely to benefit from a particular treatment.
The significant interactions between TCD status and treatment assignment for disease-free and overall survival therefore provide a rationale for further investigation of AI-derived tumour cell density as a potential method of identifying patients who may benefit from irinotecan intensification.
The AI framework processed each slide in approximately two minutes. By comparison, a previously described manual tumour cell density method required approximately 20 minutes per slide.
Because the AI-derived measurement can be obtained from routinely collected pretreatment biopsy slides, it may potentially be incorporated into digital pathology workflows. However, the current findings do not support using tumour cell density to guide treatment decisions in clinical practice.
Study Limitations
This was a post-hoc, hypothesis-generating analysis that was not conducted according to a prespecified protocol. Although the patients came from a randomised phase 3 trial, the trial was not designed to prospectively test treatment selection according to tumour cell density.
The analysis also included only the 414 ARISTOTLE participants with available pretreatment biopsy specimens, rather than the full trial population of 564 patients.
The observed predictive signal requires validation in independent cohorts of patients with locally advanced rectal cancer treated with irinotecan-containing neoadjuvant chemoradiotherapy. The investigators identified the CinClare trial as one potential cohort in which the findings could be evaluated retrospectively.
Prospective clinical validation would require a biomarker-stratified trial in which pretreatment tumour cell density is incorporated into the study design. Although the AI pipeline showed strong agreement with manually assessed tumour cell density, further methodological refinement may improve measurement precision.
Molecular profiling was also incomplete. Microsatellite instability status was available for 242 of the 414 patients, and all patients with available data had microsatellite-stable disease. Future studies will need to determine whether tumour cell density provides predictive information independently of, or in addition to, established genomic and transcriptomic biomarkers.
Conclusion
This post-hoc analysis of the phase 3 ARISTOTLE trial suggests that AI-derived tumour cell density may identify subgroups of locally advanced rectal cancer with different outcomes following standard versus irinotecan-intensified neoadjuvant chemoradiotherapy.
Patients with TCD-high tumours experienced significantly longer disease-free and overall survival when irinotecan was added to standard chemoradiotherapy. Patients with TCD-low tumours did not demonstrate a similar benefit. The findings indicate heterogeneity of treatment effect according to baseline tumour cell density. However, this predictive signal remains hypothesis-generating and requires independent confirmation.
AI-derived tumour cell density therefore remains a candidate predictive biomarker. Independent retrospective validation and prospective clinical evaluation are required before it can be used to guide treatment selection in locally advanced rectal cancer.
The full article is available in eBioMedicine.