Can Pathology Slides and Liquid Biopsy Complement Each Other? AI Biomarker Enters Validation in Early Colorectal Cancer

After colorectal cancer surgery, the most difficult question is often not how much tumor was removed, but whether tiny residual lesions that cannot be seen on imaging remain hidden in the body. Circulating tumor DNA (ctDNA) in the blood can provide a window into this question, but a negative test does not mean the risk has fallen to zero. A new study is now preparing to combine blood signals with artificial intelligence-based interpretation of pathology slides to test whether the two can produce a more complete picture of recurrence risk.

DoMore Diagnostics and Italy’s IFOM—AIRC Institute of Molecular Oncology announced a collaboration on August 10 to validate the prognostic value of the deep-learning pathology biomarker Histotype Px in early colorectal cancer. The central question is whether, beyond distinguishing risk on its own, the model can provide meaningful independent information in addition to postoperative ctDNA results.

Histotype Px analyzes routine H&E-stained digital pathology slides and assigns patients to different risk categories based on image features. This differs from the information measured by liquid biopsy: the former estimates disease behavior from the tissue morphology of the resected tumor, while the latter searches the blood for tumor DNA that may represent molecular residual disease. The study will include patients with stage II T4N0 or stage III disease who underwent postoperative ctDNA testing in the PEGASUS trial. It will also analyze a stage I to low-risk stage II population in the AlfaOmega platform who did not receive adjuvant treatment after surgery; model performance is additionally planned to be evaluated in a larger retrospective cohort spanning stages I to III.

PEGASUS is a prospective, multicenter phase II trial with a target enrollment of 140 patients with microsatellite-stable stage III or stage II T4N0 colon cancer. After surgery, the study assigns CAPOX or capecitabine based on ctDNA results, with subsequent escalation, de-escalation, or switching to FOLFIRI according to changes in later test results. The trial also incorporates the AlfaOmega observational platform, allowing clinical data, imaging, and biological specimens to be linked longitudinally and providing an important foundation for this integrated biomarker study.

A PEGASUS analysis published in 2023 showed that, among 135 patients included in the per-protocol analysis, 35 were ctDNA-positive after surgery. Of these, 11 became negative after three months of CAPOX, but 8 later experienced recurrence or tested positive for ctDNA again. Another 24 remained positive after CAPOX; among them, 11 became negative following FOLFIRI treatment and had not experienced recurrence during the follow-up period at that time. The results indicate that ctDNA carries a prognostic signal, while also highlighting that a single conversion to negative does not necessarily mean residual disease has been eradicated.

This is precisely where adding pathology AI may have value. However, the current announcement provides no new performance data and does not fully specify the sample size, primary endpoint, prespecified statistical thresholds, or approach to external validation. The study must also clarify whether the AI model can improve risk identification among ctDNA-negative patients while avoiding the misclassification of genuinely low-risk patients as high risk, which could lead to unnecessary chemotherapy.

Even if combining the two biomarkers demonstrates better prognostic discrimination, that would not mean it has been proven that adjusting treatment on this basis can prolong survival. Before entering routine care, the model’s reproducibility across different hospitals, scanning equipment, and patient populations must still be confirmed, and prospective studies with control designs must demonstrate that it can improve clinical decision-making and patient outcomes. The importance of this collaboration lies in beginning to answer these questions using clinical cohorts with long-term follow-up data, rather than declaring that the answers have already been established.