#1 · ML / AI methods

LLM agents automate bioinformatics workflows with verifiable, grounded outputs

The common thread is LLM-based agents for computational biology in which outputs are constrained by external evidence rather than parametric recall. MAT-Cell grounds batch single-cell annotation premises in differentially expressed genes and debates claims in a tree-structured multi-agent setup. STAnalyzer adds self-refinement with literature and database validation to spatial transcriptomics analysis. Pipette constrains natural-language workflow construction with a Skill Graph extracted from scientific literature. For plant science, GraphRAG over provenance-linked knowledge graphs targets complete, source-attributed list queries where parametric guessing fails. MolClaw orchestrates more than 30 drug discovery tools through hierarchical skills and introduces MolBench. CompBioBench proposes 100 verifiable tasks with metadata scrubbing to benchmark such agents. The new element is the emphasis on verifiability and grounding as evaluation criteria, though the abstracts provided report few quantitative results.