#14 · ML / AI methods

LLM agents orchestrate omics tools and models through multi-step, validated workflows

The dominant thread is LLM-centered agents that coordinate biological tools, ML models, and literature retrieval through multi-step workflows. Medea runs verified multi-step analyses for therapeutic reasoning across biological contexts, and BiOmics pairs explicit reasoning with latent embeddings to interpret multi-omics data spanning variant and cell annotation. PersonaAI mines 560,000 aging papers via RAG to propose hypotheses, then validates them with autonomous single-cell RNA-seq. Deep Research adds interactive multi-agent coordination over a persistent world state for scientific discovery. MARBLE differs in target, iteratively revising bioinformatics model architectures through debate-driven changes. What is new is the object of optimization: the workflow itself, with validation embedded in the loop rather than a single predictive output. All five entries are preprints, so performance claims remain unreviewed.