Abstract
BACKGROUND: Recent advancements in immunotherapy, particularly pembrolizumab, have shown promising results in treating metastatic colorectal cancer (CRC) and triple-negative breast cancer (TNBC). Accurate detection of predictive biomarkers, such as microsatellite instability (MSI)/mismatch repair deficiency (MMRd) and programmed death-ligand 1 (PD-L1), is key to efficacy of these treatments. Traditional methods like immunohistochemistry (IHC) and next-generation sequencing are effective but are labor intensive and require subjective interpretation. METHODS: We developed a dual-modality transformer-based model for predicting MSI/MMRd and PD-L1 status using hematoxylin & eosin and IHC stained whole slide images. We evaluated the model using area under the receiver operating curve (AUROC). Time-on-treatment (TOT) and overall survival (OS) were derived from insurance claims and analyzed by Kaplan-Meier method. Hazard ratios (HR) were determined using the Cox proportional hazard model. RESULTS: Our AI framework achieves clinical-grade performance, with AUROC exceeding 0.97 for MSI/MMRd prediction in CRC and 0.96 for PD-L1 prediction in breast cancer. Patients with biomarker-positive model predictions demonstrated prolonged TOT and OS when treated with pembrolizumab. For breast cancer patients, the model's predictions were superior to PD-L1 IHC in stratifying patients with improved outcomes on pembrolizumab, suggesting a reevaluation of existing PD-L1 status thresholds. CONCLUSIONS: This study promotes the integration of advanced AI tools in clinical pathology, aiming to enhance the precision and efficiency of cancer biomarker evaluation and offering a customizable framework for varied clinical scenarios. Our model enhances predictive accuracy, integrating features from both staining methods, and exhibits superior prognostic precision compared to current biomarker assessments.