EnsAgent: a tool-ensemble multiple Agent system for robust annotation in spatial transcriptomics

Qishi Dong

Abstract:

Motivation: Automated domain annotation in spatially resolved transcriptomics (SRT) remains challenging since it depends on gene expression,
morphology, and clinical conventions, which vary across cohorts and platforms. While Large Language Model (LLM)-driven agents show promise,
current approaches typically condition semantic reasoning on static, single-method partitions. This reliance makes annotation pipelines fragile
to upstream partition errors and prone to hallucinations when molecular evidence is ambiguous. A robust framework integrating ensemble
intelligence with iterative, evidence-based reasoning is required to ensure reproducibility and accuracy.
Results: We introduce EnsAgent, a tool-ensemble multi-agent system designed for robust SRT annotation. Uniquely, EnsAgent decouples
structural partitioning from semantic labeling via a Consultation–Review workflow. A Tool-Runner Agent orchestrates a diverse portfolio of
clustering algorithms via the Model Context Protocol (MCP), generating a consensus partition optimized by a multimodal Scoring Agent.
Subsequently, a Proposer–Critic feedback loop coordinates four specialized experts (Marker, Pathway, Spatiality, and Visual) to formulate
annotations with explicit evidence trails and uncertainty estimates. Benchmarking on three SRT datasets demonstrates that EnsAgent effectively
neutralizes batch effects and resolves subtle tumor microenvironment niches missed by single-paradigm baselines, delivering state-of-the-art
accuracy and interpretability.