#31 · ML / AI methods

LLMs as reasoning and agentic components in drug discovery workflows

The common thread is LLMs embedded as components within agentic or reasoning workflows rather than as standalone predictors. DrugR adds step-by-step pharmacological reasoning to molecule optimization, reporting improved ADMET properties while preserving efficacy. The JCIM modular agent integrates retrieval, RAG question answering, molecule generation, 75-property prediction, iterative refinement and Boltz-2 complex prediction. Beyond SMILES evaluates agentic systems across autonomous synthesis planning, literature mining and molecular design, while DrugPlayGround benchmarks LLMs and embeddings on drug characteristics, synergism, drug-protein interactions and perturbation responses. What is new is the parallel effort to test these systems against standardized tasks and to make explicit the reasoning steps added to optimization, rather than only broadening task coverage.