#26 · ML / AI methods

Language and graph foundation models target drug repurposing and molecule generation

The dominant thread is machine learning that couples drug-relevant biological knowledge with generative or predictive models for drug repurposing and molecular design. NetMedGPT pretrains a transformer on a biomedical knowledge graph using masked token prediction and reports zero-shot drug repurposing. GEMGen instead conditions small-molecule generation on text representations of up- and down-regulated gene sets describing transcriptomic cell states. Causal Knowledge Graphs extend knowledge graphs with causal semantics and are applied to a Drug-Disease graph using UK Biobank and MIMIC-IV data. What is new is largely the surrounding infrastructure: DRPMKB 1.0 compiles 193 AI drug-repositioning models and 693 data entries from PubMed through March 2024, with a dual model-evaluation component.