MAT-Cell: A Multi-Agent Tree-Structured Reasoning Framework for Batch-Level Single-Cell Annotation

Yehui Yang; Zelin Zang; Xienan Zheng; Yuzhe Jia; Changxi Chi; Jingbo Zhou; Chang Yu; Jinlin Wu

q-bio.QMApr 7, 2026

MAT-Cell: A Multi-Agent Tree-Structured Reasoning Framework for Batch-Level Single-Cell Annotation

Authors: Yehui Yang, Zelin Zang, Xienan Zheng, Yuzhe Jia, Changxi Chi, Jingbo Zhou, Chang Yu, Jinlin Wu,

Organizations: Westlake University, Hangzhou, China · Shenzhen University of Advanced Technology, Shenzhen, China · Center for Artificial Intelligence and Robotics, Hong Kong Institute of Science and Innovation, Chinese Academy of Sciences, Hong Kong, China · University Key Laboratory of Information and Communication Security Backup and Recovery, Yunnan Minzu University, Kunming, China

Abstract

Automated single-cell annotation is difficult when the most abundant genes are not the most discriminative ones, or when a target state is poorly covered by a fixed reference atlas. GPTCelltype-style one-shot prompting allows large language models (LLMs) to produce plausible labels from generic expression signals, while reference-based annotators can force unfamiliar states into the nearest known category. We propose MAT-Cell, a prompt-driven framework for batch-level single-cell annotation that separates evidence grounding from label decision. MAT-Cell first uses Reverse Verification Query (RVQ) to combine tissue context, observed differentially expressed genes, and LLM-elicited biological priors into structured candidate-specific premises. Verifier agents then convert these premises into explicit premise-to-claim reasoning trees, and bounded multi-round debate compares,challenges, and revises the resulting claims before consensus or final adjudication.The returned Syllogistic Derivation Tree (SDT) provides an auditable debate trace rather than a formal proof of the annotation. In open-candidate benchmarks across five datasets, a locally deployed Qwen3-30B model with MAT-Cell achieves 75.5% average accuracy, compared with 64.2% for the strongest evaluated CoT baseline and 51.9% for the strongest evaluated scPilot variant. In oracle-candidate bench-marks across three species,MAT-Cell remains competitive across backbones, and local inference substantially reduces monetary cost for batch annotation. Code is available at: https://anonymous.4open.science/r/MATCell-4067

Figures & tables

Figure 1: Cellular Reasoning Challenges and Solutions. (A) Challenges: GPTCelltype-style annotation can be distracted by abundant but weakly discriminative genes, while reference-based systems can map unfamiliar states to the nearest known label. (B) MAT-Cell: RVQ grounds the input in observed DEGs and tissue context. Each verifier agent then builds a premise-to-claim reasoning tree; multi-round debate compares these trees to reach consensus or final adjudication.

Figure 2: The MAT-Cell Framework. For each analytical unit, MAT-Cell takes observed gene evidence, tissue context, and optional candidate labels as input. RVQ queries an LLM for candidate priors and combines them with context and observed evidence into a structured premise. Verifier Agents build premise-to-claim reasoning trees, then compare, rebut, and revise them through a shared SDT. The final stage returns either a normalized exact-match consensus annotation or routes unresolved focus-set cases to a Decision Agent fallback, together with the SDT trace.

Method	Model	PBMC3K	Liver	Retina	Brain	Heart	Avg Acc.
CellTypist (Sci’22)	Non-LLM	46.4	56.3	38.8	24.2	69.0	46.9
CellMarker2.0 (NAR’23)	Non-LLM	30.4	25.0	63.2	62.5	26.7	41.6
CoT (NeurIPS’22)	Gemini 2.5 Pro	62.5 $\pm$ 4.4	78.2 $\pm$ 0.8	49.5 $\pm$ 1.2	66.9 $\pm$ 1.8	64.0 $\pm$ 2.8	64.2
	GPT-o1	67.5 $\pm$ 5.2	70.4 $\pm$ 1.3	49.0 $\pm$ 1.4	66.9 $\pm$ 4.2	61.3 $\pm$ 1.8	63.0
	Qwen3-14B	57.8 $\pm$ 3.1	64.8 $\pm$ 8.5	66.3 $\pm$ 7.3	64.0 $\pm$ 6.0	51.1 $\pm$ 10.5	60.8
	GPT-4o	68.8 $\pm$ 0.0	80.4 $\pm$ 2.5	44.2 $\pm$ 2.2	45.5 $\pm$ 2.0	58.0 $\pm$ 3.8	59.4

Table 1: Open-candidate annotation results. Accuracy is the scPilot-style graded annotation score, reported as mean

\pm

std (%) across five benchmarks without oracle candidates. Best and second-best values are shown in bold and underline . Red values indicate the relative change in average accuracy against the corresponding strongest non-MAT-Cell baseline; for unmatched backbones, the strongest non-MAT-Cell average is used.

Method	Model	Human			Mouse			Monkey			Avg Acc.
Method	Model	Both	DEG	HEG	Both	DEG	HEG	Both	DEG	HEG	Avg Acc.
scCATCH (iScience’20)	Non-LLM	33.9 $\pm$ 1.0	30.2 $\pm$ 1.2	28.6 $\pm$ 1.4	37.4 $\pm$ 0.9	36.7 $\pm$ 0.5	30.3 $\pm$ 1.0	41.8 $\pm$ 0.7	42.2 $\pm$ 0.8	20.0 $\pm$ 0.8	33.5
Cell-ID (Nat Biotech’21)	Non-LLM	43.1 $\pm$ 0.4	43.2 $\pm$ 0.3	21.8 $\pm$ 1.1	28.8 $\pm$ 1.3	28.9 $\pm$ 1.1	16.0 $\pm$ 0.7	51.7 $\pm$ 0.5	49.8 $\pm$ 1.1	30.3 $\pm$ 0.7	34.8
Cell-o1 (arXiv’25)	Qwen2.5-7B	42.9 $\pm$ 1.2	40.9 $\pm$ 1.5	24.3 $\pm$ 1.1	39.0 $\pm$ 1.6	39.4 $\pm$ 1.7	23.2 $\pm$ 0.8	69.5 $\pm$ 1.8	68.5 $\pm$ 1.9	50.3 $\pm$ 1.2	44.2
CoT (NeurIPS’22)	Qwen3-14B	48.7 $\pm$ 0.5	61.8 $\pm$ 1.7	21.6 $\pm$ 0.4	47.9 $\pm$ 0.8	53.4 $\pm$ 1.5	27.4 $\pm$ 0.9	67.2 $\pm$ 0.8	72.0 $\pm$ 0.6	40.4 $\pm$ 1.0	48.9
	Qwen3-30B	55.5 $\pm$ 0.3	62.7 $\pm$ 0.5	21.3 $\pm$ 0.4	47.6 $\pm$ 2.2	52.1 $\pm$ 0.7	25.6 $\pm$ 0.4	64.9 $\pm$ 1.1	71.6 $\pm$ 1.1	35.7 $\pm$ 0.2	48.6

Table 2: Oracle-candidate cross-species results. Accuracy is mean

\pm

std (%) across Human, Mouse, and Monkey under Both , DEG-only , and HEG-only views with oracle candidate labels. Best and second-best values are shown in bold and underline . Red values report relative gains in Avg Acc. over the same-backbone CoT baseline; “-c” denotes self-generated RVQ priors.

ID	Config	H	M	Mk	Gap $\downarrow$	RI $\uparrow$
M0	GPTCelltype+HEG	16.5%	17.8%	33.1%	16.6%	73%
M1	GPTCelltype+DEG	45.0%	38.7%	64.4%	25.7%	78%
M2	CoT+HEG	21.3%	25.6%	35.7%	14.4%	75%
M3	CoT+DEG	62.7%	52.1%	71.6%	19.5%	85%
M4	MAT-Cell+HEG	28.2%	49.9%	39.5%	21.7%	80%
M5	MAT-Cell+DEG	76.4%	82.5%	75.9%	6.6%	97%

Table 3: System-level ablation. All variants use Qwen3-30B in the oracle-candidate setting. H/M/Mk denote Human, Mouse, and Monkey DEG-only accuracy; Gap is the max-min species gap, and RI is the relative invariance score defined in Appendix D .

Figure 3: Sensitivity to Council Scale (

R

). Accuracy peaks at

R=5

and then declines under excessive over-pruning.

Figure 4: Representative SDT workflow and non-consensus sources. Green labels indicate correct predictions and red labels indicate incorrect predictions. Solve Verifier Agent candidates are examined by Expert Verifier Agent templates and Verifier Agent debate in the shared SDT; unresolved units are sent to the Decision Agent for fallback adjudication. The pie chart reports the main non-consensus sources that trigger fallback decisions.

Metric	CoT	CoT	GPTCelltype	MAT-Cell
Model	GPT-4.1	Gemini 2.5 Flash	GPT-4.1	Qwen3-30B
Deployment	API	API	API	Local A100
Accuracy	68.1%	62.9%	64.9%	82.5%
Batch size	1	1	1	20
Time/query	$\sim$ 8.0s	$\sim$ 6.2s	1.93s	9.17s
Cost/query	$\sim$ 0.412 ¥	$\sim$ 0.356 ¥	$\sim$ 0.1100 ¥	$\sim$ 0.0079 ¥

Table 5: Accuracy, time, and monetary cost under API and local inference. Accuracy is the Mouse, DEG-only column in Table 2 ; time/cost is measured on the Human DEG efficiency setting. Local cost is amortized over 20-sample A100 batches.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Analytical Dimension	Metric Description	Value
Part I: Overall Performance Gains	Cell-level Avg. Partial Accuracy (GPTCelltype baseline)	0.4278
Part I: Overall Performance Gains	Cell-level Avg. Partial Accuracy (MAT-Cell)	0.7268
Part II: Debate Convergence	Pre-debate (First Round) Full Agreement Rate (Mean)	0.7632
	Post-debate (Last Round) Full Agreement Rate (Mean)	0.9052
	Pearson $r$ (Pre vs. Post Agreement, per batch)	0.7538
Part III: Decision Agent (DA) Efficacy	DA Fallback Frequency (Disputed Batches)	11.50% ( $n=23$ )

Appendix

Table 6: Empirical Metrics on the Necessity of Multi-Agent Reasoning. This table reports the internal state metrics from the Human (DEG view) dataset (total

n=200

batches) to quantitatively demonstrate the necessity and efficacy of multi-agent debate and contextual synthesis.

Cell type	Count
CD4+ T Cells	1,157
Classical Monocytes	479
B Cells/Dendritic Cells	341
Effector/Activated T Cells	297
Effector/Activated T Cells/NK Cells	158
Non-Classical Monocytes	157

Appendix

Table 7: Top cell types in the open-candidate datasets. Counts report the most frequent annotated cell types in each dataset.

Property	Human	Mouse	Monkey
Number of batches	2,400	2,400	2,400
Total cells	27,588	27,583	25,871
Number of cell types	75	123	121
Unique top genes	5,583	6,941	7,757
Unique DEG genes	1,434	2,432	2,485
Cells per batch	7–15	7–15	7–15

Appendix

Table 8: Summary statistics for the oracle-candidate benchmarks.

Cell type	Count
Oligodendrocyte	1,593
L2/3–6 intratelencephalic projecting glutamatergic neuron	1,560
Astrocyte	1,544
Oligodendrocyte precursor cell	1,508
L2/3 intratelencephalic projecting glutamatergic neuron	1,497
Pvalb GABAergic cortical interneuron	1,488

Appendix

Table 9: Top cell types in the oracle-candidate benchmarks. Counts report the most frequent annotated cell types for each species.

Dataset	Top-25 genes
Brain	CNTNAP2, DSCAM, DPP10, ROBO2, KCNIP4, GRIP1, ZNF385D, EPIC1, CA10, FSTL4, ARL15, HTR1F, FOXP2, FSTL5, MYO16, PTCHD4, CLSTN2, CPNE4, NRG1, DTNA, CBLN2, CDH9, SLIT3, SLIT2, UNC13C
Heart	RGS6, ANK3, LINC02147, CLIC5, HIGD1B, PIK3R5, SLC38A11, MLIP, TRDN-AS1, MYBPC3, XIRP2, MYH6, ACACB, PRKAG2, ACTA1, SH3RF2, PPP1R3C, ENSG00000230490, MYL7, LINC02552, ENSG00000258231, ENSG00000271959, HECW2, FRMD5, G0S2
Liver	H3f3a, Ubb, Tmsb10, Hba-a1, Gpx1, Hba-a2, Hbb-bt, S100a8, S100a9, Apoe, Igfbp7, Sparc, S100a6, Cd24a, Pglyrp1, Gm5483, BC100530, Stfa1, Anxa1, Ifitm6, Serpina3k, Mup20, Gnmt, Cd7, Ccl4
PBMC3K	B2M, RPL13, MALAT1, RPL21, TPT1, RPL10A, ACTB, RPL8, H3F3B, RPS3A, RPS5, EEF1D, RPS27A, FTH1, MT-CO2, CD74, FTL, OAZ1, CD37, CD79A, FCGR3A, LST1, COTL1, GNLY, GZMB
Retina	FTH1, RHO, APOE, RBP3, WIF1, GLUL, CLU, PTGDS, TF, MFGE8, MPP4, FRZB, CRABP1, SPP1, ENO1, DKK3, RLBP1, CA2, GPX3, CRYAB, CADPS, NEAT1, NR2E3, YPEL2, CNGA1
Human	MALAT1, ACTB, ACTG1, B2M, TMSB4X, FTH1, GAPDH, RPL13, RPS27, RPL41, RPL21, EEF1A1, TPT1, RPS3A, RPL32, RPL3, RPS2, RPS18, RPS6, RPS12, RPL10, RPL34, RPS27A, RPL13A, RPL11

Appendix

Table 10: Top-25 highly expressed genes after excluding mitochondrial and ribosomal genes.

Rule ID	Markers (IF)	Cell Type (THEN)
R1	CD4, IL7R, TCF7	CD4+ T-Cell
R2	CD8A, CD8B, GZMA	CD8+ T-Cell
R3	CD14, LYZ, FCGR3A	Monocyte
R4	PPBP, PF4, TUBB1	Platelet

Appendix

Table 11: Biological Syllogistic Rules (Example)

Parameter	Value
Number of Agent Types	4
#ExpertAgent (EA) Instances	1
#SolveAgent (SA) Instances	1
Council Scale ( $R$ , #VerifierAgent/VA)	5
#DecisionAgent (DA) Instances	1
Top-K genes (Both view)	25

Appendix

Table 12: Hyperparameters for MAT-Cell

Figure 5: Example MAT-Cell reasoning tree. A single-cell SDT trace in which multiple verifier agents compare candidate annotations and converge on the final cell-type label through structured debate.

Metric	Value
Consensus Accuracy	$0.7859$
Partial Cell Accuracy	$0.7850$
Pairwise Disagreement Rate	$0.5694\pm 0.0414$
Prediction Correlation	$0.3542\pm 0.0688$
Cohen’s Kappa	$0.3752\pm 0.0486$
Mutual Information (nats)	$1.3691\pm 0.1517$

Appendix

Table 13: Agent diversity and consensus statistics for EA-generated Verifier Agents.

Top-K	Human (DEG)	Monkey (DEG)	Mouse (DEG)
Top-5	$0.371\pm 0.006$	$0.630\pm 0.008$	$0.402\pm 0.016$
Top-10	$0.513\pm 0.003$	$0.690\pm 0.009$	$0.461\pm 0.010$
Top-15	$0.532\pm 0.004$	$0.705\pm 0.007$	$0.463\pm 0.004$
Top-25	$0.564\pm 0.009$	$\mathbf{0.721\pm 0.007}$	$\mathbf{0.503\pm 0.009}$
Top-50	$\mathbf{0.568\pm 0.005}$	$0.708\pm 0.006$	$0.470\pm 0.005$

Appendix

Table 14: Accuracy by Top-K Gene Count using Qwen3-30B without RVQ

Seed	Human (DEG)	Monkey (DEG)	Mouse (DEG)	Deviation from Mean
Seed-1	$0.736\pm 0.017$	$0.758\pm 0.006$	$0.827\pm 0.003$	H:-0.21%, Mk:-0.37%, M:-1.52%
Seed-2	$0.751\pm 0.005$	$0.759\pm 0.008$	$0.837\pm 0.004$	H:+1.77%, Mk:-0.19%, M:-0.34%
Seed-3	$0.744\pm 0.011$	$0.763\pm 0.005$	$0.840\pm 0.002$	H:+0.87%, Mk:+0.34%, M:+0.01%
Seed-4	$0.726\pm 0.010$	$0.755\pm 0.002$	$0.852\pm 0.010$	H:-2.43%, Mk:-0.84%, M:+1.85%
Mean $\pm$ Std	$\mathbf{0.739\pm 0.010}$	$\mathbf{0.759\pm 0.004}$	$\mathbf{0.839\pm 0.010}$	Stable

Appendix

Table 15: Robustness to Random Seed under DEG Gene Input

Method	Human (both)	Monkey (both)	Mouse (both)
RVQ (5 markers)	$0.629\pm 0.006$	$0.779\pm 0.010$	$0.674\pm 0.013$
RVQ (10 markers)	$0.643\pm 0.003$	$0.804\pm 0.005$	$0.697\pm 0.007$
RVQ (15 markers)	$\mathbf{0.664\pm 0.005}$	$\mathbf{0.808\pm 0.003}$	$\mathbf{0.700\pm 0.007}$
RVQ (20 markers)	$0.640\pm 0.005$	$0.800\pm 0.004$	$0.690\pm 0.006$

Appendix

Table 16: Ablation on RVQ Gene Budget (Qwen3-30B)

Method	Human (DEG)		Mouse (DEG)		Monkey (DEG)
Method	Acc.	95% CI	Acc.	95% CI	Acc.	95% CI
Cell-o1	$0.409$	[0.403, 0.414]	$0.394$	[0.388, 0.401]	$0.685$	[0.678, 0.692]
GPTCelltype Qwen-3-30B	$0.450$	[0.445, 0.455]	$0.387$	[0.384, 0.390]	$0.644$	[0.641, 0.647]
GPTCelltype DeepSeek-V3	$0.632$	[0.627, 0.636]	$0.567$	[0.563, 0.571]	$0.471$	[0.468, 0.474]
GPTCelltype Gemini-2.5	$0.709$	[0.705, 0.713]	$0.659$	[0.653, 0.665]	$0.859$	[0.856, 0.862]
GPTCelltype GPT-4.1	$0.733$	[0.731, 0.736]	$0.649$	[0.644, 0.654]	$0.864$	[0.863, 0.866]

Appendix

Table 17: Main Results with 95% Confidence Intervals

Comparison	Mean Diff	t-statistic	df	p-value
MAT-Cell vs. GPT-4.1	$+0.031$	$+3.12$	29	$0.0042$ **
MAT-Cell vs. Gemini-2.5	$+0.055$	$+6.28$	29	$<0.0001$ ***
MAT-Cell vs. DeepSeek-V3	$+0.132$	$+9.24$	29	$<0.0001$ ***
MAT-Cell vs. GPTCelltype-Qwen-3-30B	$+0.314$	$+15.82$	29	$<0.0001$ ***

Appendix

Table 18: Human Dataset: Paired T-test Results (DEG View)

Comparison	Mean Diff	t-statistic	df	p-value
MAT-Cell vs. GPT-4.1	$+0.110$	$+2.89$	29	$0.0071$ **
MAT-Cell vs. Gemini-2.5	$+0.100$	$+5.83$	29	$<0.0001$ ***
MAT-Cell vs. DeepSeek-V3	$+0.192$	$+8.67$	29	$<0.0001$ ***
MAT-Cell vs. GPTCelltype-Qwen	$+0.372$	$+14.21$	29	$<0.0001$ ***

Appendix

Table 19: Mouse Dataset: Paired T-test Results (DEG View)

Comparison	Human (d)	Mouse (d)	Monkey (d)
GPT-4.1	3.21 (very large)	17.01 (very large)	1.26 (large)
Gemini-2.5-Flash	5.29 (very large)	14.04 (very large)	0.53 (medium)
DeepSeek-V3	11.48 (very large)	31.18 (very large)	6.07 (very large)
GPTCelltype-Qwen-3-30B	26.01 (very large)	59.38 (very large)	8.86 (very large)

Appendix

Table 20: Effect sizes and post-hoc power analysis for MAT-Cell against representative baselines.

LLM-Assisted Annotation

Large language model-assisted annotation uses AI to label, classify, and structure data, often with human review to improve scale and consistency.

See all 202 papers

SC-Arena: A Natural Language Benchmark for Single-Cell Reasoning with Knowledge-Augmented Evaluation
Feb 26, 2026Jiahao Zhao, Feng Jiang, Shaowei Qin +5LLM Evaluation Scientific Reasoning in Language Models
Software College, Northeastern University, China · Shenzhen University of Advanced Technology, Shenzhen, China · Shenzhen Key Laboratory for High Performance Data Mining, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, China +2
GATHER: Convergence-Centric Hyper-Entity Retrieval for Zero-Shot Cell-Type Annotation
May 7, 2026Zhonghui Zhang, Feng Jiang, Shaowei Qin +2Zero-Shot Learning Graph Retrieval-Augmented Generation
Artificial Intelligence Research Institute, Shenzhen University of Advanced Technology Shenzhen, China · Software College, Northeastern University Shenyang, China
CellxPert: Inference-Time MCMC Steering of a Multi-Omics Single-Cell Foundation Model for In-Silico Perturbation
Apr 30, 2026Andac Demir, Erik W. Anderson, Jeremy L. Jenkins +1Single-Cell Perturbation Modeling Genomics
Novartis Biomedical Research

Feb 26, 2026J/K move · Enter open · S save

SC-Arena: A Natural Language Benchmark for Single-Cell Reasoning with Knowledge-Augmented Evaluation

Jiahao Zhao, Feng Jiang, Shaowei Qin, Zhonghui Zhang+4

Software College, Northeastern University, China · Shenzhen University of Advanced Technology, Shenzhen, China · Shenzhen Key Laboratory for High Performance Data Mining, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, China+2

LLM Evaluation Scientific Reasoning in Language Models Single-Cell Transcriptomics

Open paper AskPDF arXiv

Large language models (LLMs) are increasingly applied in scientific research, offering new capabilities for knowledge discovery and reasoning. In single-cell biology, however, evaluation practices for both general and specialized LLMs remain inadequate: existing benchmarks are fragmented across tasks, adopt formats such as multiple-choice classification that diverge from real-world usage, and rely on metrics lacking interpretability and biological grounding. We present SC-ARENA, a natural language evaluation framework tailored to single-cell foundation models. SC-ARENA formalizes a virtual cell abstraction that unifies evaluation targets by representing both intrinsic attributes and gene-level interactions. Within this paradigm, we define five natural language tasks (cell type annotation, captioning, generation, perturbation prediction, and scientific QA) that probe core reasoning capabilities in cellular biology. To overcome the limitations of brittle string-matching metrics, we introduce knowledge-augmented evaluation, which incorporates external ontologies, marker databases, and scientific literature to support biologically faithful and interpretable judgments. Experiments and analysis across both general-purpose and domain-specialized LLMs demonstrate that (i) under the Virtual Cell unified evaluation paradigm, current models achieve uneven performance on biologically complex tasks, particularly those demanding mechanistic or causal understanding; and (ii) our knowledge-augmented evaluation framework ensures biological correctness, provides interpretable, evidence-grounded rationales, and achieves high discriminative capacity, overcoming the brittleness and opacity of conventional metrics. SC-Arena thus provides a unified and interpretable framework for assessing LLMs in single-cell biology, pointing toward the development of biology-aligned, generalizable foundation models.

Abstract

LLM-Assisted Annotation

Explore similar work

SC-Arena: A Natural Language Benchmark for Single-Cell Reasoning with Knowledge-Augmented Evaluation

GATHER: Convergence-Centric Hyper-Entity Retrieval for Zero-Shot Cell-Type Annotation

CellxPert: Inference-Time MCMC Steering of a Multi-Omics Single-Cell Foundation Model for In-Silico Perturbation