Yang, J.; Chereda, H.; Dünitz, J.; Bleckmann, A.; Beissbarth, T.

The concept of BRCAness describes tumors that, although lacking BRCA mutations, exhibit similar molecular vulnerabilities, particularly in the context of homologous recombination repair deficiencies. Patients which display BRCAness are often sensitive to treatments like PARP inhibitors. Here we aim to define BRCAness based on gene expression data of the tumors. Therefore, our study leverages a cutting-edge machine learning (ML) approach combining Graph Convolutional Neural Networks (GCNNs) with the interpretability approach Graph Layer-wise Relevance Propagation (GLRP) to deeply investigate the genomic intricacies of BRCAness across various cancer types. Traditional Differential Gene Expression (DEG) analysis and machine learning techniques such as Random Forest (RF) were also employed for comparative purposes. Our analysis reveals that DEG struggles to capture the nuances of BRCAness, emphasizing the superiority of GLRP in identifying genes crucial for BRCAness through their functional roles and interactions within the cancer genome. GLRP identified significant genes involved in transcription regulation and directly associated with cancer processes, underscoring the complexity of BRCAness. Gene Set Enrichment Analysis (GSEA) across different cancer types highlighted key pathways such as Signaling by Nuclear Receptors, Cellular Senescence, and ESR-mediated signaling, elucidating their roles in BRCAness and potential as therapeutic targets. This study not only advances our understanding of the genetic underpinnings of BRCAness but also opens avenues for more targeted and personalized cancer treatments, spotlighting the crucial interplay between gene expression, functional interactions, and critical biological pathways in the manifestation of BRCAness.