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

EnsAgent is a novel tool-ensemble multi-agent system that enhances the robustness and accuracy of spatial transcriptomics annotation by decoupling structural partitioning from semantic labeling through a consensus-driven workflow and a multi-expert feedback loop, effectively overcoming the limitations of single-method approaches and hallucinations.

The Gist

Imagine you have a massive, incredibly detailed map of a city (the human body), but instead of streets and buildings, the map is made of tiny dots, each representing a cell with its own unique set of instructions (genes). Your goal is to label every neighborhood on this map: "This is the brain's thinking center," "This is a tumor," or "This is a healing wound."

The problem is that looking at just one type of clue (like gene activity) often leads to confusion. One method might say a neighborhood is a "park," while another says it's a "factory," and both might be partially right or wrong depending on the lighting (technical noise) or the map's scale.

Enter EnsAgent, a new digital tool designed to solve this labeling puzzle. Think of it not as a single detective, but as a high-tech "Council of Experts" working together to get the answer right.

Here is how EnsAgent works, broken down into three simple stages:

1. The "Tool Runner" (The Scout Team)

Imagine you need to draw the borders of a neighborhood. Instead of asking one person to draw it, EnsAgent sends out a whole team of different scouts, each using a different map-making tool.

• The Analogy: It's like asking five different cartographers to draw the same city. One uses satellite photos, one uses street surveys, and one uses historical records.
• What they do: They all generate their own version of the map. Some maps might have jagged lines; others might be too smooth. This creates a "portfolio" of different possibilities, ensuring no single mistake ruins the whole project.

2. The "Scoring Agent" (The Referee)

Now you have five different maps. Who decides which one is best? EnsAgent has a referee who looks at all the maps at once.

• The Analogy: Think of a sports referee who doesn't just look at the ball, but also checks the players' positions, the weather, and the crowd noise.
• What they do: This agent uses two eyes:
  1. The Science Eye: It checks the gene data (the "stats").
  2. The Visual Eye: It looks at the actual tissue image (the "photo") to see if the shapes make sense biologically.
• The Result: It combines the best parts of all the scouts' maps into one "Consensus Map." If one scout made a weird mistake, the referee ignores it. If three scouts agree on a border, the referee locks it in.

3. The "Proposer-Critic Loop" (The Debate Club)

This is the most unique part. Once the map is drawn, the system has to give names to the neighborhoods (e.g., "Tumor Core" or "Healthy Liver").

• The Proposer (The Optimist): This agent looks at the evidence and suggests a name. "I think this area is a Tumor because I see these specific genes."
• The Critic (The Skeptic): This agent is the tough auditor. It checks the Proposer's work against a giant library of biological facts. It asks: "Are you sure? Does the shape match? Do the genes actually support this? Or are you just guessing?"
• The Loop: If the Critic finds a flaw (e.g., "The genes say it's a tumor, but the shape looks like healthy tissue"), it sends the Proposer back to the drawing board with a specific instruction: "Go check the immune cells again." They keep debating and re-checking until they are 100% confident.

Why is this a big deal?

Previous tools were like a single person trying to solve a puzzle in the dark. If they made one mistake, the whole picture was wrong. They often got "hallucinations" (making up facts) or got confused by messy data.

EnsAgent is like a team of experts in a well-lit room, debating, checking each other's work, and using multiple tools to ensure the final map is accurate.

Real-world wins:

• In the Brain: It successfully mapped the tiny, thin layers of the human cortex (the brain's outer shell) that other tools missed, distinguishing between layers that look almost identical.
• In Cancer: It found hidden "neighborhoods" inside tumors that were previously invisible. It could tell the difference between a "dead zone" in a tumor and an "active war zone" where the immune system is fighting back, which is crucial for deciding how to treat a patient.
• In Noisy Data: Even when the data was messy (like a photo taken with a shaky camera), EnsAgent could still figure out the correct structure, whereas other tools got confused and scattered.

In short: EnsAgent takes the messy, confusing world of cellular biology and uses a "teamwork" approach to create clear, reliable, and trustworthy maps of our bodies, helping doctors and scientists understand diseases better.

Technical Summary

1. Problem Statement

Spatially Resolved Transcriptomics (SRT) enables the mapping of gene expression within tissue architecture, but automated domain annotation (assigning biological labels like cell types or tissue layers to spatial regions) remains a significant bottleneck. Current challenges include:

• Fragility of Single-Method Pipelines: Existing approaches often rely on a single upstream clustering algorithm. If this initial partition is flawed due to batch effects, resolution mismatches, or tissue heterogeneity, the downstream annotation fails.
• Hallucinations and Overfitting: Large Language Model (LLM)-based agents often generate semantic labels based on compressed or single-source evidence, leading to "hallucinations" when molecular signals are ambiguous or conflicting.
• Lack of Iterative Reasoning: Most agentic frameworks perform "one-shot" inference without explicit mechanisms to reconcile data-driven signals with biological knowledge or to quantify uncertainty.
• Absence of Auditability: Current systems often output narrative labels without structured evidence trails, confidence scores, or decision logs, hindering reproducibility.

2. Methodology: The EnsAgent Framework

EnsAgent is a tool-ensemble multi-agent system designed to decouple structural partitioning from semantic labeling via a Consultation–Review workflow. It operates in three distinct stages:

Stage 1: Tool-Runner Agent (MCP Orchestration)

• Function: Acts as the execution engine, converting raw SRT data (AnnData objects) into a diverse set of candidate domain partitions.
• Mechanism: Utilizes the Model Context Protocol (MCP) to orchestrate a portfolio of spatial domain detection tools (e.g., BayesSpace, BASS, DR-SC, STAGATE, IRIS, SEDR).
• Standardization: Since different tools produce arbitrary cluster indices, the agent aligns all candidate partitions into a shared label space using a maximum-weight assignment problem (solved via the Hungarian algorithm based on Intersection-over-Union).
• Output: Generates a manifest of aligned partitions and associated evidence artifacts (visualizations, differentially expressed genes, pathway enrichments).

Stage 2: Scoring Agent (Dual-Stream Evaluation & Consensus)

• Function: Integrates the diverse candidate partitions into a single, robust consensus domain map.
• Dual-Stream Engine:
  1. Evaluation Module: Uses an MLLM to score partitions based on quantitative metrics: spatial coherence, cluster separability, marker specificity, and pathway consistency.
  2. Visual Module: Uses a Vision-Language Model (VLM) to assess morphological fidelity (geometric continuity, biological plausibility, technical artifacts).
• Consensus Construction: Calculates a reliability-weighted voting score for each spot. The final consensus map ($y^*$) is derived by aggregating labels where methods with higher reliability scores contribute more strongly.
• Output: A refined consensus partition and updated evidence bundles for downstream annotation.

Stage 3: Proposer–Critic Iterative Refinement Loop

• Function: Performs semantic annotation with explicit evidence trails and uncertainty estimation.
• Proposer Agent: Coordinates four specialized expert agents (Marker, Pathway, Spatiality, Visual) to generate top-$K$ label hypotheses. It synthesizes their outputs into a linear opinion pool to determine the primary hypothesis and confidence score.
• Critic Agent: Audits the Proposer's hypothesis against a curated Knowledge Base (KB) and checks for cross-modality consistency.
  • Worst-Case Support: A label is rejected if any single expert assigns it a low probability.
  • Consistency Check: Uses Jensen-Shannon divergence to measure agreement between experts.
• Iterative Loop: If the audit score falls below a threshold, the Critic triggers targeted re-analysis (e.g., re-running DEGs with stricter thresholds, inspecting histology overlays) by specific experts. This loop continues until convergence or a maximum iteration limit.
• Output: A structured JSON manifest containing the final label, audit scores, modality-specific posteriors, and a full decision trace.

3. Key Contributions

• Decoupling of Partitioning and Labeling: EnsAgent separates the structural clustering task from semantic interpretation, allowing for ensemble-based robustness in the former and iterative reasoning in the latter.
• Tool-Ensemble via MCP: Introduces a standardized interface (MCP) to orchestrate heterogeneous clustering tools, mitigating the bias of relying on a single algorithm.
• Active Arbitration Layer: Unlike passive agents, EnsAgent employs a Scoring Agent (to filter noisy partitions) and a Critic Agent (to audit and trigger re-analysis), effectively simulating an "expert review board."
• Evidence-Linked Uncertainty: The system outputs calibrated probabilities and structured decision traces, enabling users to audit why a label was assigned and identify low-fidelity regions.

4. Results & Benchmarks

EnsAgent was evaluated on three diverse SRT datasets:

• Human Dorsolateral Prefrontal Cortex (DLPFC):
  • Performance: Achieved the highest Adjusted Rand Index (ARI = 0.67) and Normalized Mutual Information (NMI), outperforming single-method baselines (e.g., BayesSpace, DR-SC) by 4–17%.
  • Outcome: Successfully recovered fine-grained cortical laminar architecture (Layers 1–6) and functional logic, resolving boundaries that other agents fragmented or mixed.
• Human Breast Cancer (Tumor Microenvironment):
  • Performance: Identified 20 distinct spatial domains, refining coarse pathologist annotations.
  • Outcome: Discovered subtle subgroups missed by single-paradigm baselines, such as distinguishing an "immune-suppressed" deep tumor core from an "immune-active" tumor-stroma interface. It revealed specific ligand-receptor signaling cascades (e.g., GPC3–UNC5C) driving these niches.
• Mouse Olfactory Bulb (Batch Effect Robustness):
  • Performance: Demonstrated superior resilience to synthetic batch effects, maintaining an NMI of 0.6260 and ARI of 0.5580.
  • Outcome: The Critic-guided rerun protocol corrected confidence mismatches, recovering coherent laminar organization (e.g., EPL, GCL) that single-method pipelines failed to reconstruct under noise.

5. Significance

• Reproducibility and Auditability: By providing structured JSON outputs with decision traces and evidence links, EnsAgent addresses the "black box" nature of current AI-driven bioinformatics tools.
• Robustness to Heterogeneity: The ensemble approach neutralizes platform-specific biases and batch effects, making it suitable for multi-cohort studies.
• Automated Expert Intuition: The Proposer–Critic loop effectively automates the iterative hypothesis testing and evidence verification process typically performed by human pathologists and biologists.
• Future Directions: The authors propose distilling EnsAgent's ensemble reasoning into a specialized Bio-Spatial Small Language Model (SLM) to reduce computational overhead and extending the architecture to support 3D volumetric mapping across serial tissue sections.

In summary, EnsAgent represents a paradigm shift from static, single-method annotation to a dynamic, multi-agent system that prioritizes consensus, evidence verification, and iterative refinement, significantly improving the accuracy and interpretability of spatial transcriptomics analysis.