acceptodds
Under review as a conference paper at ICLR 2027

CHORUS: A Generalist Multi-Agent System for Complex Spatial Biology Analysis

Abstract

Large language model agents increasingly automate biological data analysis through workflow planning and tool use, yet existing systems are often either flexible but weakly constrained or reproducible but experience-blind. This limitation is particularly consequential in spatial biology, where datasets differ substantially in measurement platform, spatial resolution, available modalities, image registration, and preprocessing state, such that the same analytical workflow may be valid for one dataset but inappropriate for another. We introduce START, a State-Aware Tool Reasoning and Adaptive Trajectory Agent for memory-augmented spatial omics analysis. START represents prior executions as structured state–workflow–outcome experiences and retrieves relevant positive and negative evidence to rank candidate workflows. Scientific state constrains which trajectories are admissible, while experience memory guides preference among valid alternatives. Execution outcomes are evaluated and written back to memory, allowing future trajectories to adapt without updating the underlying language model. Experiments across multimodal and cross-platform spatial-omics settings demonstrate reliable and adaptive biological analysis.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI4SCI, Spatial biology, Agent

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_SQYDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents CHORUS, an LLM-based multi-agent system for spatial biology that treats spatial transcriptomics (ST) and H&E histology as independent sources of evidence. A Coordinator Agent routes each question to an ST Agent, an H&E Agent, or both. In joint tasks, each specialist first analyzes only its own modality and records its tool outputs as a provenance-grounded evidence logic tree that separates evidence from inferences and conclusions.

Cross-modal interaction is constrained in four ways: (i) a scientific-state gate compares claims only after specimen, coordinate, registration, and ROI correspondence are verified, and otherwise labels them incomparable rather than conflicting; (ii) during discussion, evidence nodes are immutable, and specialists can only keep or revise their own inferences/conclusions or request new evidence (formalized in Theorem 1 as evidence preservation and order independence); (iii) a Judgment Agent turns unresolved disagreements into targeted evidence requests instead of forcing consensus; and (iv) a state-conditioned experience memory reuses past tool trajectories only under compatible scientific states, so behavior adapts without model updates.

Experiments mainly use 12 Visium oral squamous cell carcinoma sections (GSE208253), plus several Xenium sections and case studies, with Qwen3-8B and DeepSeek-flash as backbones. Reported results:

  • Higher agreement with published pathologist region annotations than Leiden, Louvain, SpaGCN, and an spEMO-CONCH adaptation (e.g., ARI 0.332 vs. 0.223 for Leiden).
  • Exposing the correspondence matrix to the judgment step makes it visible in 16/22 runs (vs. 0/22) but leaves partition agreement essentially unchanged.
  • Removing the gate, evidence acquisition, or discussion barely affects workflow completion, whereas the backbone matters much more (27/28 vs. 8/28 ST-verified dual-branch executions with DeepSeek-flash vs. Qwen3-8B).
  • In a five-task user study, CHORUS completes tasks faster than human experts (13.4 vs. 22.0 min on average) and achieves a higher success rate than a DeepSeek baseline (86.4% vs. 30.0%).
  • In an appendix experiment, state-conditioned memory raises accuracy from 0.40 to 1.00, compared with 0.22 for state-shuffled memory.
Critical strengths

S1. A principled design for evidence-grounded cross-modal collaboration. Tool-grounded evidence is separated from inferences and conclusions in provenance-tracked logic trees; evidence stays immutable during discussion, and specialists may revise only their own inferences or conclusions. A scientific-state gate returns "incomparable" rather than "conflict" when specimen, coordinate, registration, or ROI correspondence is unverified. These choices target real failure modes of free-form multi-agent debate (persuasion overriding observations, forced consensus) and of naive multimodal pipelines (comparing unregistered data), and they transfer to other agentic-science settings. This is the main consideration in favor of acceptance.

S2. Evaluation that separates process completion from verified execution. Table 4 reports process completion, ST-verified dual-branch execution, and conclusion production separately, and shows that they can diverge sharply (e.g., Qwen3-8B completes 26/28 runs but has verified dual-branch execution in only 8/28). This is a useful methodological point for auditing agentic scientific workflows.

Critical weaknesses

W1. The central claims are not supported by the reported evidence. The abstract and conclusion state that preserving evidence and enforcing state compatibility improve cross-modal decisions and that state-conditioned memory improves accuracy, and the contribution list says the experiments "demonstrate the superiority of our agent." However:

  • In the component ablations (Table 4), removing the alignment gate, evidence acquisition, or discussion leaves DeepSeek-flash essentially unchanged (26–27/28 completed and ST-verified), and for Qwen3-8B every removal increases ST-verified dual-branch execution (8/28 → 10–13/28). DeepSeek-flash produces conclusion text in 0/28 runs under every setting.
  • Exposing the executed correspondence matrix to the judgment step (Table 2) leaves partition agreement unchanged.
  • The memory result (Fig. 4) appears only in the appendix, without a description of its 90 probes, and the appendix ("Cost and interpretation") itself states that the results do not establish a general accuracy benefit from memory or universal superiority.

This matters because the contribution rests on these mechanisms making cross-modal agent reasoning more reliable, and the reported evidence does not show that they do.

W2. The evaluation is incomplete, and its main favorable comparison is hard to interpret.

  • Table 1 and the appendix describe the multi-step quantitative ST tasks (T2), the flat vs. tree evidence-format comparison (T3), and the adjudication replay and constraint checks (T4), but report no results for them beyond inference cost (Table 5). These are the experiments best suited to test numerical correctness, evidence organization, and the interaction constraints.
  • Table 3 compares "one previously selected CHORUS output" per section with five runs per comparator (Appendix, "Partition agreement and comparisons"), although the table lists 60/12 runs for CHORUS, and it includes section s9, which was reserved for development. The joint-workflow runs in Table 2 reach only ARI ≈ 0.14, below Leiden's 0.223 in Table 3, and this gap is not explained.
  • There is no comparison with the spatial-biology agents discussed in the related work (e.g., STAgent, SpatialAgent, ChatSpatial, STAT). The user study (Fig. 2) compares only with a DeepSeek baseline whose configuration is not described and gives no protocol details. The main quantitative evaluation uses 12 sections from a single OSCC dataset without established patient-level independence.

This matters because the only quantitative evidence favoring CHORUS over baselines may reflect output selection, and the experiments meant to validate the core mechanisms have no reported outcomes.

Questions for authors

W1. The central claims are not supported by the reported evidence. The abstract and conclusion state that preserving evidence and enforcing state compatibility improve cross-modal decisions and that state-conditioned memory improves accuracy, and the contribution list says the experiments "demonstrate the superiority of our agent." However:

  • In the component ablations (Table 4), removing the alignment gate, evidence acquisition, or discussion leaves DeepSeek-flash essentially unchanged (26–27/28 completed and ST-verified), and for Qwen3-8B every removal increases ST-verified dual-branch execution (8/28 → 10–13/28). DeepSeek-flash produces conclusion text in 0/28 runs under every setting.
  • Exposing the executed correspondence matrix to the judgment step (Table 2) leaves partition agreement unchanged.
  • The memory result (Fig. 4) appears only in the appendix, without a description of its 90 probes, and the appendix ("Cost and interpretation") itself states that the results do not establish a general accuracy benefit from memory or universal superiority.

This matters because the contribution rests on these mechanisms making cross-modal agent reasoning more reliable, and the reported evidence does not show that they do.

W2. The evaluation is incomplete, and its main favorable comparison is hard to interpret.

  • Table 1 and the appendix describe the multi-step quantitative ST tasks (T2), the flat vs. tree evidence-format comparison (T3), and the adjudication replay and constraint checks (T4), but report no results for them beyond inference cost (Table 5). These are the experiments best suited to test numerical correctness, evidence organization, and the interaction constraints.
  • Table 3 compares "one previously selected CHORUS output" per section with five runs per comparator (Appendix, "Partition agreement and comparisons"), although the table lists 60/12 runs for CHORUS, and it includes section s9, which was reserved for development. The joint-workflow runs in Table 2 reach only ARI ≈ 0.14, below Leiden's 0.223 in Table 3, and this gap is not explained.
  • There is no comparison with the spatial-biology agents discussed in the related work (e.g., STAgent, SpatialAgent, ChatSpatial, STAT). The user study (Fig. 2) compares only with a DeepSeek baseline whose configuration is not described and gives no protocol details. The main quantitative evaluation uses 12 sections from a single OSCC dataset without established patient-level independence.

This matters because the only quantitative evidence favoring CHORUS over baselines may reflect output selection, and the experiments meant to validate the core mechanisms have no reported outcomes.

Reviewer_zTZpOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors present CHORUS, a multi-agent system for analyzing spatial transcriptomics and H&E images together. Separate agents analyze gene expression (spatial transcriptomics) and tissue morphology (H&E), then compare their findings after checking that the samples and spatial regions match. The agents preserve the original analysis results while revising their interpretations and can request additional analyses when needed. The system also reuses previous workflows when their requirements match the current data and analysis state. The authors test CHORUS on Visium and Xenium datasets, compare its tissue partitions with published annotations, and examine the effects of its components and memory.

Critical strengths

The paper is well motivated and brings together two important research areas: spatial biology and multi-agent systems.

Critical weaknesses
  1. Key details of the memory experiment are missing. Figure 4 reports accuracy increasing from 40% to 100%, but the authors do not explain what the 90 evaluation cases are, where the stored experiences come from, which model is tested, or how answers are scored. It is also unclear whether the evaluation cases are kept separate from memory. The appendix describes a different test with two trajectories and 16 queries, which checks whether retrieved workflows meet task requirements rather than whether the model answers correctly (p. 14, lines 710-716). These may be separate experiments, but the paper does not provide enough information to assess the gains reported in Figure 4.

  2. The experiments do not establish whether the proposed components improve final answers. Table 4 reports whether workflows finish and produce written conclusions, but not whether their conclusions are correct. Every DeepSeek setting reports conclusion text in 0/28 runs, which needs explanation. The authors also describe T2/T3/T4 evaluations of numerical answers, evidence presentation, final judgments, and implementation constraints, but do not report the corresponding results; Table 5 provides only computational costs.

The performance comparisons have additional ambiguities. Table 3 reports 60 CHORUS runs, while the appendix describes one selected output per section without explaining how it was chosen (p. 15, lines 788–791). Comparisons with clustering algorithms also cannot establish whether gains come from the proposed agent design or from different tools and analysis settings. Figure 2’s success rates and human completion times lack clear task counts, scoring rules, and baseline settings.

Questions for authors
  1. How was the Figure 4 experiment conducted? Please describe the 90 evaluation cases, stored experiences, model, scoring rules, and separation between evaluation cases and memory. How does this relate to the two-trajectory/16-query test? Please also clarify whether the added compatibility filter was used, how confidence intervals were calculated, and what success criterion was preregistered.

  2. Can you report the results for the T2/T3/T4 evaluations described in the appendix? For Table 4, how many runs produced a correct answer supported by the evidence, or correctly concluded that the evidence was insufficient? Why does every DeepSeek setting report 0/28 runs with conclusion text?

  3. Table 3 lists 60 CHORUS runs, but the appendix describes one selected output per section. How do these counts relate, how were outputs selected, and were reference annotations used during selection? Please also distinguish results on development section s9 from those on the remaining sections.

  4. For Figure 2, how many tasks and datasets were evaluated, and what counted as success? Did CHORUS and the baseline have the same tool access and comparable computation budgets? How were the ten experts’ completion times measured, and were the timed outputs checked for correctness?

Reviewer_KcPDOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

CHORUS is an agent framework for analyzing spatial transcriptomics and H&E images. Two specialists first analyze the data separately, then compare their findings after checking that the samples, coordinates, and tissue regions match. Agents can revise tool interpretations of tool outputs or request further analysis. The system also reuses past workflows when their requirements are compatible with the current data and analysis state.

The framework addresses a useful problem: making joint molecular and image analysis more reliable and easier to check. However, the experiments mainly show whether workflows run, with limited evidence that the proposed mechanisms improve the final answers. The memory experiment is also insufficiently defined, and some comparison results have inconsistent reporting.

Critical strengths
  1. The design addresses a meaningful scientific failure mode.

Treating unmatched or unregistered observations as incomparable rather than contradictory is appropriate. Likewise, preserving independently acquired molecular and structural evidence is a useful design principle for scientific agents.

  1. The evaluation design contains useful ideas for scientific-agent assessment.

Independent numerical references, inclusion of failed assignments, and separate checks for execution, input identity, and answer support are valuable. Table 2 also reports that increasing correspondence-matrix visibility does not improve partition agreement. These choices help distinguish evidence availability from downstream utility, although several of the designed evaluations lack reported outcomes.

Critical weaknesses
  1. The claimed reasoning improvement is not clearly demonstrated.

Table 4 mainly measures whether workflows and analysis branches run, not whether their conclusions are correct. It therefore does not show whether the alignment gate, discussion, or additional evidence improves scientific decisions. The appendix describes quantitative tasks and adjudication checks (T2–T4), but omits their correctness results. These results are needed to establish what the framework improves beyond tool execution.

  1. The memory accuracy is difficult to interpret.

Figure 4 reports an increase from 40% to 100%, but does not adequately define the 90 probes, reference answers, model settings, or separation between memory and evaluation cases. Its relationship to T4's small retrieval-compatibility test is also unclear. Without this information, it is hard to judge whether memory improves scientific answers or mainly helps satisfy predefined compatibility rules.

  1. The model differences need a failure analysis.

DeepSeek achieves 27/28 verified dual-branch executions, while Qwen achieves only 8/28 despite completing 26/28 workflows. Removing discussion raises Qwen's verified executions to 13/28. What fails in these runs, and why does discussion affect the models differently? A breakdown of the existing traces would help explain the framework's limits. The absence of conclusion text in all DeepSeek configurations also needs clarification.

Questions for authors
  1. Which concrete decision-quality endpoint supports the claimed improvement in cross-modal reasoning? Please report the existing T2/T3 correctness results and T4 adjudication outcomes, including appropriate abstention where relevant.

  2. What does “Final Accuracy” in Figure 4 measure? Please provide representative probes, scoring definitions, memory/evaluation separation, execution settings, and its relationship to the T4 retrieval diagnostic.

  3. What failure categories account for Qwen's 20 assignments that do not meet the verified dual-branch endpoint? What changes when discussion is removed, and why is conclusion text absent for all DeepSeek configurations?

  4. How were Table 3's CHORUS outputs selected, how many independent runs were performed, and how do these outputs relate to Table 2?

Reviewer_EahcOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper presents a new multi-agent system to analyze spatial transcriptomics data when paired H&E images are available, which is a very common and useful case in real-world spatial transcriptomics analysis pipelines, such as those for 10x Visium, Visium HD, and Xenium. Based on this situation, it proposes a per-modality agent framework to handle each modality specifically and then collaborate to reach the final conclusion, which is also a very interesting way to simulate real-world biologists’ research workflows. The motivation and application scenario here are useful and significant.

Nevertheless, the evaluation and benchmarking here lack solid comparisons and do not show the potential of the model, which decreases the significance of this manuscript and makes it difficult for the community to assess the actual contribution of the multi-agent system.

Critical strengths
  1. It targets a very useful and meaningful niche domain for biological agents, which is key to developing a domain-specific agent application.
  2. The per-modality agent framework is interesting and aligns with real-world biological research workflows, and the collaboration framework provides a technical contribution.
Critical weaknesses
  1. It lacks comprehensive benchmarking against current general-purpose agent harnesses and specialized biomedical agent systems.
  2. The downstream tasks are too simple to demonstrate the distinctive advantages of the agentic system over traditional bioinformatics analysis pipelines.
Questions for authors
  1. The current manuscript lacks comparisons with other agent harnesses and biomedical agent systems, such as the open-source versions of Pi/Codex, Stellar, SpatialAgent, and Biomni. The comparison is limited to the base model without the multi-agent harness and does not include other biomedical agent systems, such as Stellar, SpatialAgent, or Biomni.

  2. The evaluation tasks are too simple. The authors only evaluate spatial domain clustering, and even for this simple task, they lack comprehensive benchmarking against state-of-the-art spatial clustering methods using unimodal or bimodal inputs, such as CellCharter, Novae, STAGATE, and MISO.

  3. In Figure 2a, the authors already show results for cell annotation, deconvolution, SVG detection, and differential expression analysis. Why are these tasks not also evaluated against traditional methods using appropriate metrics, as is done for spatial domain identification? Was spatial domain identification cherry-picked for evaluation?

  4. Furthermore, for an agent system, we expect to see something that traditional bioinformatics analysis cannot accomplish. For example, could the authors use this agent system to generate hypotheses from the current data that traditional bioinformatics pipelines cannot generate, thereby demonstrating the agent’s reasoning capabilities? Could the agent identify new spatial biomarkers that previous traditional analysis pipelines could not find?

  5. It would be helpful for the authors to show how human biologists interact with this agent system to improve the final results.

Comments on the paper

1 comment

Dear reviewers,

Thank you very much for your time and effort in reviewing this submission. The authors have now submitted their rebuttal, and the discussion period is nearing its end. We would greatly appreciate it if you could take some time to carefully review the authors’ responses and share your assessment of whether your concerns have been adequately addressed. Thank you again for your valuable time and contributions to the review process.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.