Development of a computational histology artificial intelligence (CHAI)–powered predictive biomarker for first-line chemotherapy intensification in metastatic colorectal cancer (mCRC) and validation in prospective randomized phase III trials (RCT).

Abstract

3513 Background: In unresectable mCRC, meta-analysis of RCTs suggests triplet therapy consisting of fluorouracil, oxaliplatin, and irinotecan (FOLFOXIRI) + bevacizumab (bev) may improve survival outcomes over doublet regimens of fluorouracil with irinotecan (FOLFIRI) or oxaliplatin (FOLFOX) + bev, at the expense of increased toxicity. Adoption of triplet therapy remains low. A histology-based biomarker to predict outcomes in mCRC treated with triplet vs doublet regimens could optimize treatment selection. We aimed to use the previously described CHAI platform to develop and validate a histology-based predictive biomarker for doublet vs triplet therapy in mCRC with participant-level data from 3 RCTs. Methods: Development used diagnostic H&E-stained whole slide images (WSI) from participants with mCRC treated with first-line (1L) chemotherapy enrolled in the MRC FOCUS RCT. WSIs were analyzed using the CHAI assay and trained towards clinical outcomes as a function of chemotherapy escalation. Weighting of histologic features was optimized for association with progression-free (PFS) & overall survival (OS) to construct a signature that was combined with tumor sidedness to create the mCRCpred biomarker. A cutpoint was selected to optimize treatment effect stratification, assigning each case a binary label. The locked biomarker was evaluated on the independent validation cohort of WSIs from the TRIBE + TRIBE2 RCTs. The association of the biomarker-treatment interaction with PFS & OS was evaluated with multivariable (MVA) Cox models & Likelihood Ratio tests. Results: The development cohort included 294 cases from MRC FOCUS. The independent validation cohort (n = 386) comprised cases from TRIBE (n = 159) and TRIBE2 (n = 227). 193 (50%) had 1L triplet+bev and 193 (50%) doublet+bev. Biomarker (+) (indicative of triplet benefit) disease treated with triplet+bev had superior 3-year PFS (17% vs 7%, HR = 0.51 [0.37, 0.70], p < 0.001) and OS (43% vs 22%, HR = 0.51 [0.35, 0.73], p < 0.001) versus doublet+bev. Biomarker (-) cases treated with triplet+bev vs doublet+bev had no significant differences in PFS (5 vs 15%, HR = 1.30 [0.96, 1.74], p = 0.09) or OS (31 vs 40%, HR = 1.35 [0.94, 1.92], p = 0.1). In MVA including cohort, age, ECOG, number of metastatic sites, and timing of metastasis (metachronous/synchronous), the biomarker-treatment interaction was significant (p < 0.001) indicating a predictive association between the biomarker and treatment regimen. Conclusions: The CHAI mCRCpred histology-based biomarker predicted differential benefit of chemotherapy escalation (doublet-bev vs triplet-bev) as 1L therapy in mCRC, with independent validation post-hoc in two RCTs (Simons Level of Evidence IB). This biomarker may guide optimal treatment selection for 1L mCRC. Clinical trial information: ISRCTN79877428 (FOCUS); NCT00719797 (TRIBE); NCT02339116 (TRIBE2).