The bio-pharmaceutical industry's growing interest in computational approaches for drug development is driven by the extremely high costs associated with bringing new drugs to market. In this context, the in-silico identification of new therapeutic uses for already-approved drugs has gained significant attention in recent years. This approach, known as drug repurposing, offers substantial advantages in terms of cost reduction, risk mitigation, and faster development timelines.

In this work, we introduce a novel framework for drug repurposing, combining network theory with state-of-the-art computational techniques that utilize Large Language Models (LLMs) originally developed for natural language processing, now adapted for the analysis of protein sequences. Using this approach, we have mapped and characterized the binding-site landscape of over 3,000 pharmacologically relevant ligands and more than 180,000 putative pockets from protein complexes, drawing on 6,619 crystallographic records from the Protein Data Bank. Our results demonstrate how the embedding space generated for this 'pocketome' can expand predictive capabilities, offering new avenues for drug repurposing tasks.