DAG-guided kernel learning framework to mitigate racial disparities in breast cancer

As the second leading cause of cancer death among women in the United States, breast cancer (BC) mortality is disproportionately higher among Black women than among non-Hispanic White women. Artificial intelligence (AI) and machine learning (ML) approaches have been successfully applied to mitigate racial disparities in BC. For instance, previously we developed a multi-modal transfer learning (TL) framework that can help address health disparities by transfer knowledge learned on the majority group (e.g., non-Hispanic White women) to improve performance of AI/ML models for a minority group (e.g., Black women). However, these approaches integrate different omics data in a simplistic manner, but failed to capture the complex cross-omics interaction among them, leading to minor or modest performance improvement for reducing BC health disparities. To address these concerns, we propose a directed acyclic graph (DAG)-guided kernel learning framework to mitigate racial disparities in breast cancer. Specifically, we first encoded the cross-omics interaction knowledge into a DAG graph by a method called non-combinatorial optimization via trace exponential and augmented Lagrangian for structure learning (NOTEARS). Then, we hard-masked the learned graph to retain only biology-consistent projections. From these projections, we computed projection-specific RBF kernels to capture non-linear sample similarity, yielding interpretable, direction-aware kernels used for training a TL model equipped with a data augmentation (DA) method, which were used to reduce breast cancer disparities. By using three omics modalities (i.e., mRNA, miRNA, and DNA methylation) from The Cancer Genome Atlas (TCGA) BRCA cohorts of 1,085 female BC patients, we demonstrated that our proposed framework achieves remarkably better performance (using various performance metrics like ROC-AUC, PR-AUC, Accuracy, and F1-score) compared with state-of-the-art approaches for reducing BC health disparities in terms of time-specific Progression-Free Interval (PFI) predictions. In addition, our framework could quantify and identify multiple key cross-modality interactions to mitigate BC racial disparities. In summary, our proposed framework would provide a novel approach to explore cross-omics interactions for boosting a robust TL model performance to reduce health disparities in BC. We also believe our approach can be extended to mitigate racial disparities for other types of cancer.