Leveraging transcriptomic profiles and deep learning to detect homologous recombination deficiency in breast cancer with softHRD

The homologous recombination (HR) pathway is the canonical repair mechanism by which cells repair DNA double-strand breaks. Defects in this pathway, known as homologous recombination deficiency (HRD), can lead to genomic instability and are observed in approximately 13% of breast cancers. HRD is often driven by somatic or germline mutations in BRCA1/2 , which render tumors sensitive to PARP inhibitors and platinum-based therapies through synthetic lethality. However, many HRD-positive cancers lack BRCA1/2 mutations, underscoring the need for more reliable approaches to identify tumors likely to benefit from these treatments. To address this, we developed softHRD, a transcriptomics-based framework for detecting HRD in breast cancer. softHRD was trained on RNA-seq profiles from 857 breast cancer patients in The Cancer Genome Atlas (TCGA), filtered for protein-coding genes. A variational autoencoder was first used to reconstruct these transcriptomic profiles, generating latent representations that capture the underlying structure of gene expression patterns. A sparse autoencoder was then applied to these latent features to derive mechanistically interpretable components and identify an HRD-associated gene set. These genes were subsequently leveraged to train a downstream Elastic Net regression model, yielding a robust 111-gene transcriptional signature indicative of HRD. We validated softHRD in 80 breast cancer patients from the I-SPY 2 clinical trial treated with neoadjuvant chemotherapy and olaparib. The model identified a statistically significant difference in pathologic complete response between HRD-predicted and HR-proficient tumors (p = 0.00676). Unlike whole-genome sequencing, which provides a static view of mutational alterations, transcriptomic profiling captures the dynamic state of gene expression, revealing biological changes that genomic methods may overlook. softHRD demonstrated robust performance across all PAM50 breast cancer subtypes, highlighting its generalizability. With the growing integration of transcriptomics into clinical research and diagnostics, softHRD represents a scalable and adaptable framework for accurate, efficient HRD characterization, with potential applications across multiple cancer types.