Abstract
Machine learning algorithms have led to the development of numerous vector embeddings for biological entities such as metabolites, proteins, genes, and enzymes. However, these embeddings often lack contextual information due to their specialized focus on individual omics. Disease progression and biosynthesis pathways are increasingly understood through complex, multi-layered networks that integrate diverse omics data and intricate signaling and reaction sequences. Capturing these relationships in a meaningful way requires embeddings that account for both functional and multi-modal dependencies. We propose an embedding approach that unifies these different biological modalities by treating them as directions in a shared space rather than as isolated data types. Similar to how word embeddings in natural language processing reveal meaningful relationships (e.g., Tokyo – Japan + UK = London, indicating a directional representation of capitals), we can model genes and proteins in a way that captures their inherent connections. A gene implies information about the protein it encodes, and vice versa, forming a structured and interpretable representation of biological pathways. Our model, inspired by NLP techniques, breaks down pathway sequences into contextual pairs spanning different omics types. By aligning pathway steps in proximity, the embeddings reflect biologically relevant relationships, enhancing their interpretability and utility. Because these embeddings are generated from pathway sequences, they can be applied to optimize reaction pathways, aiding retrosynthesis in microbiomes, drug development, and even human health interventions.