Multimodal AI Pathology + ctDNA Integration for Stage III Colon Cancer Risk Stratification

Title

Multimodal AI Pathology + ctDNA Integration for Stage III Colon Cancer Risk Stratification

Content

Mayo Clinic investigators (Sinicrope et al.) presented a multimodal AI deployment at the American Society of Clinical Oncology (ASCO) 2026 Annual Meeting (May 2026, Abstract 3525) that achieves dramatically sharper risk stratification in stage III colon cancer by combining deep-learning-derived tumor microenvironment (TME) features from digitized H&E surgical slides with postoperative circulating tumor DNA (ctDNA) status. The lead result: ctDNA-positive patients had a 5-year disease-free survival of 27.7% versus 77.1% for ctDNA-negative patients — a near-50 percentage-point gap driven by hazard ratios of approximately 5.0 for time-to-recurrence and 5-6 for disease-free survival, with 20.4% of the cohort testing ctDNA-positive. The analysis drew on the phase III NCCTG N0147 trial of 3,084 stage III colon cancer patients randomized to FOLFOX ± cetuximab, with 2,260 patients evaluable for the ctDNA component. On the tissue side, a deep-learning algorithm — previously validated as QuantCRC on 6,468 colorectal cancer slides — was applied to surgical specimens to quantify six TME parameters: tumor-stroma ratio, tumor-infiltrating lymphocyte (TIL) density, tumor budding, mucin pools, poorly-differentiated clusters, and desmoplastic stroma. On the liquid-biopsy side, postoperative ctDNA was quantified using the tissue-free Guardant Reveal methylation assay, which achieved >98% specificity. The 2026 multimodal analysis extends Mayo's earlier ASCO 2023 work, which had already shown that AI-quantified tumor budding and poorly-differentiated clusters were the strongest prognostic feature in mismatch-proficient (pMMR) colon cancers (adjusted HR ~0.20 for lowest vs highest quartile). The new analysis demonstrates that AI-derived TME features correlate with ctDNA positivity and recurrence risk, supporting a combined tissue+ctDNA biomarker framework for guiding adjuvant chemotherapy escalation or de-escalation decisions in stage III colon cancer. The work signals a path to broader multimodal AI biomarker integration in oncology, analogous to other multimodal foundation-model efforts in pharmaceutical R&D. (Rochester, Minnesota, USA — academic research deployment presented at ASCO 2026.)