#1 · Tools & infrastructure

LLM agents and multimodal models convert literature and spectra into molecular knowledge

The shared thread is learned models that convert heterogeneous scientific data into structured molecular knowledge, increasingly organised as agentic systems. BioMiner uses a multimodal LLM to extract protein-ligand bioactivity triplets from literature text, tables and figures, evaluated on the BioVista benchmark. In mass spectrometry, MSAgent orchestrates more than 50 MS tools in an evidence-grounded workflow that improves CASMI retrieval MRR by over 10%. Drug repurposing preprints follow a similar pattern: LinkD unifies diffusion-based affinity prediction, selectivity scoring, phenotypic data and clinical evidence; RepurAgent adds human-supervised hypothesis generation; and the Open Intelligence Hub coordinates 32 containerised tools for binder design and ADC prioritization. What is new is the shift from fixed predictors toward agentic, evidence-tracked pipelines. Reported results remain largely benchmark-based or computational, so prospective validation is not established across these items.