Domain-specific embeddings uncover latent genetics knowledge

The inundating rate of scientific publishing means every researcher will miss new discoveries from overwhelming saturation. To address this limitation, we employ natural language processing to overcome human limitations in reading, curation, and knowledge synthesis, with domain-specific applications to genetics and genomics. We construct a corpus of 3.5 million normalized genetics and genomics abstracts and implement both semantic and network-based embedding models. Our methods not only capture broad biological concepts and relationships but also predict complex phenomena such as gene expression. Through a rigorous temporal validation framework, we demonstrate that our embeddings successfully predict gene-disease associations, cancer driver genes, and experimentally-verified protein interactions years before their formal documentation in literature. Additionally, our embeddings successfully predict experimentally verified gene-gene interactions absent from the literature. These findings demonstrate that substantial undiscovered knowledge exists within the collective scientific literature and that computational approaches can accelerate biological discovery by identifying hidden connections across the fragmented landscape of scientific publishing.