ELISA: An Interpretable Hybrid Generative AI Agent for Expression-Grounded Discovery in Single-Cell Genomics

Omar Coser

q-bio.GNMar 12, 2026

ELISA: An Interpretable Hybrid Generative AI Agent for Expression-Grounded Discovery in Single-Cell Genomics

Authors: Omar Coser

Abstract

Translating single-cell RNA sequencing (scRNA-seq) data into mechanistic biological hypotheses remains a critical bottleneck, as agentic AI systems lack direct access to transcriptomic representations while expression foundation models remain opaque to natural language. Here we introduce ELISA (Embedding-Linked Interactive Single-cell Agent), an interpretable framework that unifies scGPT expression embeddings with BioBERT-based semantic retrieval and LLM-mediated interpretation for interactive single-cell discovery. An automatic query classifier routes inputs to gene marker scoring, semantic matching, or reciprocal rank fusion pipelines depending on whether the query is a gene signature, natural language concept, or mixture of both. Integrated analytical modules perform pathway activity scoringacross 60+ gene sets, ligand-receptor interaction prediction using 280+ curated pairs, condition-aware comparative analysis, and cell-type proportion estimation all operating directly on embedded data without access to the original count matrix. Benchmarked across six diverse scRNA-seq datasets spanning inflammatory lung disease, pediatric and adult cancers, organoid models, healthy tissue, and neurodevelopment, ELISA significantly outperforms CellWhisperer, a classical lexical retriever (BM25), and a random baseline in cell type retrieval (combined permutation test, $p < 2\times10^{-5}$ for each), with particularly large gains on gene-signature queries (Cohen's $d = 5.98$ for MRR). ELISA replicates published biological findings (mean composite score 0.88), and generates candidate hypotheses through grounded LLM reasoning, bridging the gap between transcriptomic data exploration and biological discovery.

Figures & tables

System	Expr.	Sem.	L–R /	Cond.	Interp.	Primary
	Emb.	Ret.	Pathway	Comp.	Report	Scope
AI Co-Scientist Gottweis et al. (2025)	–	–	–	–	✓	Hypothesis generation
Biomni Huang et al. (2025)	–	✓	–	–	–	General biomedical
GeneAgent Wang et al. (2025)	–	✓	–	–	–	Gene-set analysis
Virtual Lab Swanson et al. (2025)	–	–	–	–	✓	Multi-agent discovery
CellAgent Xiao et al. (2024)	–	–	–	–	–	scRNA-seq pipelines

Table 1: Comparison of existing AI systems for biomedical and single-cell analysis. Expr. Emb. : uses expression-derived embeddings from foundation models; Sem. Ret. : semantic retrieval over biological annotations; L–R / Pathway : ligand–receptor interaction and pathway scoring from data; Cond. Comp. : condition-aware comparative analysis; Interp. Report : automated interpretive report generation with LLM.

Figure 1: Overview of the ELISA architecture. The framework comprises three stages. In data preparation (left), a single-cell dataset undergoes standard preprocessing (normalization, log-transform, highly variable gene selection, PCA, neighbor graph construction, and Leiden clustering), after which per-cluster differential expression statistics are computed, enriched with Gene Ontology (GO) and Reactome terms, and encoded into 768-dimensional semantic embeddings via BioBERT. In parallel, cell-level expression embeddings are generated through scGPT. Both representations are fused into a single serialized embedding file (.pt). In the retrieval and analysis stage (center), a query classifier routes user input gene signatures, natural language concepts, or mixed queries to the appropriate pipeline: gene marker scoring, semantic retrieval, or hybrid retrieval via reciprocal rank fusion (RRF). Additional analytical modules perform pathway scoring, ligand–receptor interaction prediction, comparative analysis, and proportion estimation directly on the embedded data. In the interpretation stage (right), all retrieval and analysis outputs are passed to a Groq-hosted LLM (LLaMA 3.1-8B) that generates grounded biological interpretations and structured reports.

Figure 2: ELISA Union achieves the best retrieval performance on both query types across six datasets. Radar plots showing retrieval performance on ontology (Ont) and expression (Exp) queries for each dataset. Each plot displays six axes: Cluster Recall@ $k$ at two cutoffs and Mean Reciprocal Rank (MRR), evaluated separately on ontology and expression queries (see Supplementary Section E for metric definitions). The Recall@ $k$ cutoffs are adapted to each dataset’s number of annotated clusters and are stated in each panel title: small-cluster datasets (Breast Atlas, 8 clusters; fdAT2 Organoids, 5; High-Risk Neuroblastoma, 11) use the stringent cutoffs R@1 and R@2, whereas large-cluster datasets (CF Airways, 30; ICB Multi-Cancer, 31; First-Trimester Brain, 28) use R@5 and R@10. Because the cutoffs differ by dataset, the corresponding axes are not directly comparable across panels; each panel is internally consistent across the six retrieval modes. Higher values (further from center) indicate better performance. Six retrieval modes are compared: Random (grey dotted), BM25 (purple dash-dot), CellWhisperer (pink dashed), ELISA Semantic (blue), ELISA scGPT (orange), and ELISA Union (green). Each baseline excels on only one query type and collapses on the other: lexical (BM25) and semantic retrieval are strong on ontology but weak on expression, whereas the expression-embedding mode (scGPT) shows the reverse; Random remains near the center for large-cluster datasets and is correspondingly higher where few clusters make chance retrieval easier. ELISA Union is the only mode that consistently achieves the largest radar footprint on both ontology and expression axes, matching or exceeding every baseline on each. A combined permutation test over all 12 retrieval metrics confirms that Union significantly outperforms CellWhisperer, BM25, and a random baseline ( $p\textless 2\times 10^{-5}$ for each; see Table 2 ).

Method	vs. Baseline	$p$ (perm.)	$z$	$p$ (analytic)
Union (S+G)	CellWhisperer	$<2\times 10^{-5}$ ***	$4.95$	$4\times 10^{-7}$
Union (S+G)	BM25	$<2\times 10^{-5}$ ***	$5.11$	$2\times 10^{-7}$
Union (S+G)	Random	$<2\times 10^{-5}$ ***	$6.26$	$2\times 10^{-10}$
Semantic	CellWhisperer	$0.003$ **	$2.60$	$5\times 10^{-3}$
Semantic	BM25	$0.161$	$1.01$	$0.16$
Semantic	Random	$<2\times 10^{-5}$ ***	$5.22$	$9\times 10^{-8}$

Table 2: Combined permutation test across all 12 retrieval metrics and six datasets (50,000 sign-flip permutations; one-sided,

\text{H}_{1}

: row

>

column). ELISA Union significantly outperforms every baseline. All three Union comparisons exceed the permutation resolution and are reported as

p<2\times 10^{-5}

, the smallest value attainable with 50,000 permutations; the standardised effect size (

z

) and its analytic normal-approximation

p

-value resolve the ordering the censored permutation bound cannot, increasing from CellWhisperer (

z=4.95

) through BM25 (

z=5.11

) to Random (

z=6.26

). The analytic approximation agrees with the exact permutation test on every comparison the latter can resolve (e.g. Semantic vs. BM25,

p=0.16

for both). Among single modalities, semantic retrieval does not significantly beat BM25, confirming BM25 is a strong lexical baseline on ontology queries; only the fused Union dominates on both axes.

Dataset	Gene Cov.	Path. Align.	Int. Rec.	Prop.	Theme	Comp.
	(95% CI)	(95% CI)	(95% CI)	Cons.	Cov.	score
CF airway	0.80	1.00	0.20	Yes	0.85	0.71
	[0.66–0.89]	[0.70–1.00]	[0.04–0.62]
Neuroblastoma	0.84	1.00	1.00	Yes	0.88	0.93
	[0.78–0.88]	[0.81–1.00]	[0.72–1.00]
ICB Multi-Cancer	0.77	1.00	1.00	Yes	0.91	0.92

Table 3: Quantitative comparison between ELISA reports and reference single-cell studies. Scores reflect agreement between ELISA-generated biological interpretations and findings described in the main text of the corresponding publications. Gene coverage, pathway alignment, interaction recovery, and proportion consistency were computed programmatically; theme coverage was assessed independently by a domain expert as described in Section F . For the three programmatically computed recovery proportions, bracketed ranges give 95% Wilson score confidence intervals reflecting the number of reference items (

n

) evaluated per dataset. Wide intervals on the interaction metric reflect the small number of benchmark interactions available per individual study; the pooled estimate across all datasets (final row) is correspondingly tighter. See Appendix for full numbers 0 0 footnotetext: Gene Cov.: gene coverage; Path. Align.: pathway alignment; Int. Rec.: interaction recovery; Prop. Cons.: proportion consistency; Theme Cov.: theme coverage. Comp. score: unweighted mean of the four continuous metrics (Prop. Cons. coded as 1.0 when consistent). Confidence intervals are 95% Wilson score intervals on the per-dataset recovery proportions

k/n

. Theme coverage is reported as the domain-expert score and is not a

k/n

proportion, so no interval is given. Mean-row intervals for gene coverage and interaction recovery are mean

\pm\,t\,

SE across the six datasets; the pathway-alignment mean interval is the pooled estimate (83/83), as all datasets attained

1.00

. The Pooled row aggregates all gene, pathway, and interaction benchmark items across the six datasets into a single recovery proportion.

Dataset	Primary finding in reference study	ELISA candidate discovery / hypothesis
CF airway	Altered immune–structural cell crosstalk and inflammatory signaling in cystic fibrosis airway tissue	Detection of the macrophage CALR–LRP1 signaling axis, suggesting altered apoptotic cell recognition or phagocytic clearance pathways contributing to the CF lung inflammatory microenvironment
Breast Atlas	Ancestry-associated epithelial lineage variation and luminal progenitor states in healthy breast tissue	Enrichment of the Kelch-family gene KLHL29 in basal–myoepithelial cells, suggesting a potential additional marker or regulator of basal epithelial structural identity
Fetal Lung AT2	ITCH-mediated ubiquitin-dependent regulation of surfactant protein C (SFTPC) maturation in alveolar type II cells	Upregulation of TRIM21 and TRIM65 in mature AT2 cells, suggesting additional TRIM-family ubiquitin ligases may participate in surfactant protein processing and proteostasis
ICB Multi-Cancer	Tumor and immune transcriptional responses associated with immune checkpoint blockade therapy	Elevated CD163 and MRC1 expression in tumor-associated macrophages, consistent with an M2-like polarization state potentially associated with therapy-induced immune remodeling
Neuroblastoma	Therapy-induced transcriptional rewiring of tumor cell states and microenvironment interactions	Differential AP-1 transcription factor usage, with JUND enriched at diagnosis and JUNB/FOS enriched post-treatment, suggesting stress-response remodeling during therapy-induced state transitions
Brain Development Atlas	Chromatin accessibility programs defining early neuronal lineage specification	Shared transcription factor module ( TFAP2B , LHX5 , LHX1 ) across Purkinje neurons and midbrain GABAergic populations, suggesting a conserved regulatory program for inhibitory neuron specification

Table 4: Candidate regulatory signals identified by ELISA across six reference single-cell atlases. These signals were not explicitly highlighted in the original publications and represent transcriptome-derived hypotheses generated through ELISA’s discovery mode.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Metric	$\Delta$ mean	Cohen’s $d$	$p$ (paired $t$ )	W/L	$n$
Expression	MRR	+0.409	5.98	$<$ 0.001	6/0	6
Expression	Recall@5	+0.287	1.57	0.006	5/0	5
Expression	Recall@3	+0.428	5.38	0.006	3/0	3
Expression	Recall@2	+0.492	3.43	0.014	3/0	3
Expression	Recall@1	+0.442	1.84	0.043	3/0	3
Expression	Recall@10	+0.284	1.43	0.065	3/0	3

Appendix

Table 5: Statistical comparison of ELISA Union vs. CellWhisperer retrieval performance. For each metric,

\Delta

mean reports the average improvement of Union over CellWhisperer across datasets. Cohen’s

d

is the paired effect size.

p

-values are from one-sided paired

t

-tests (

\text{H}_{1}

: Union

>

CellWhisperer). W/L counts datasets where Union won/lost. Metrics with fewer than 6 datasets reflect different Recall@

k

cutoffs used per dataset (see Supplementary Section E ). The combined permutation test aggregates all metrics simultaneously.

Category	Metric	$\Delta$ mean	Cohen’s $d$	$p$ (paired $t$ )	W/L	$n$
Expression	MRR	+0.275	1.56	0.006	6/0	6
Expression	Recall@5	+0.233	0.91	0.038	5/0	5
Expression	Recall@3	+0.298	1.26	0.081	3/0	3
Expression	Recall@2	+0.412	2.46	0.026	3/0	3
Expression	Recall@1	+0.317	1.17	0.090	3/0	3
Expression	Recall@10	+0.158	1.11	0.097	3/0	3

Appendix

Table 6: ELISA Union vs. BM25 — per-metric statistical tests. Columns as in Table 5 .

Category	Metric	$\Delta$ mean	Cohen’s $d$	$p$ (paired $t$ )	W/L	$n$
Expression	MRR	+0.460	7.82	$<$ 0.001	6/0	6
Expression	Recall@5	+0.351	1.67	0.005	5/0	5
Expression	Recall@3	+0.418	7.38	0.003	3/0	3
Expression	Recall@2	+0.490	5.58	0.005	3/0	3
Expression	Recall@1	+0.453	4.08	0.010	3/0	3
Expression	Recall@10	+0.335	2.44	0.026	3/0	3

Appendix

Table 7: ELISA Union vs. Random — per-metric statistical tests. Columns as in Table 5 .

Dataset	Gene Cov.	Path. Align.	Int. Rec.	Prop.	Theme Cov.	Comp.
	$k/n$ (95% CI)	$k/n$ (95% CI)	$k/n$ (95% CI)	Cons.	$k/n$ (95% CI)	score
CF airway	36/45	9/9	1/5	Yes	8/8	0.71
	[0.66–0.89]	[0.70–1.00]	[0.04-0.62]		[0.68-1.00]
Neuroblastoma	173/206	16/16	10/10	Yes	24/24	0.93
	[0.78-0.88]	[0.81–1.00]	[0.72-1.00]		[0.86-1.00]
ICB Multi-Cancer	133/172	16/16	10/10	Yes	20/20	0.92

Appendix

Table 8: Raw recovery counts underlying Table 3 . Each entry gives the number of recovered reference items over the total evaluated (

k/n

) together with the corresponding 95% Wilson score confidence interval. These counts are the data from which the proportions and intervals in Table 3 are derived. Proportion consistency is a binary directional criterion and is summarised as the number of datasets passing. Theme coverage is shown here as its programmatic keyword-presence count; the value reported in Table 3 is instead the independent domain-expert score (see footnote). 0 0 footnotetext:

k

: recovered reference items;

n

: total reference items evaluated. Confidence intervals are 95% Wilson score intervals. Pooled row sums

k

and

n

across all six datasets for each metric. Proportion consistency is binary per dataset (Yes/No), summarised as datasets passing (6/6); it has no

k/n

. Theme coverage counts here are the programmatic keyword-presence metric (

k/n

); the value in Table 3 is the independent domain-expert theme score, which is not a

k/n

proportion and may differ.

ID	Tissue	Disease context	Reference	Approx.	Cell	Conditions
				cells	types
D1	Lung (bronchial)	Cystic fibrosis	Berg et al. Berg et al. [2025]	$\sim$ 96k	30	CF vs. Ctrl
D2	Adrenal / tumor	Neuroblastoma	Yu et al. Yu et al. [2025]	$\sim$ 372k	20+	Pre- vs. post-chemo
D3	Multi-cancer	ICB response	Gondal et al. Gondal et al. [2025]	$\sim$ 356k	25+	R vs. NR; 9 cancers
D4	Lung (fetal)	AT2 organoid model	Lim et al. Lim et al. [2025]	$\sim$ 9.6k	8	fdAT2 organoid lines
D5	Breast	Healthy tissue atlas	Bhat-Nakshatri et al. Bhat-Nakshatri et al. [2024]	$\sim$ 51k	13	Healthy only

Appendix

Table 9: Summary of scRNA-seq datasets used for ELISA validation. Approx. cells : approximate number of cells or nuclei profiled after quality control. Cell types : number of annotated major cell types. Conditions : experimental groups or treatment arms.

Parameter	Value	Description
Preprocessing (Scanpy)
target_sum	10,000	Library-size normalization target
n_top_genes	3,000	HVGs selected (Seurat v3)
max_value	10	Z-score clipping threshold
n_comps	50	PCA components
Leiden resolution	1.0	Used only if no annotations exist

Appendix

Table 10: Data preprocessing and embedding generation parameters.

Parameter	Value	Description
Query classification
Gene threshold	$\geq$ 60%	Token fraction to route as gene query
Mixed threshold	$\geq$ 20% each	Gene + NL tokens for mixed routing
Gene pattern	A–Z, 2–15 chars	Regex for gene symbol detection
Gene marker scoring
Score function	$(0.5+\|\log_{2}\text{FC}\|)\times(0.3+\max(\text{pct}_{\text{in}}-\text{pct}_{\text{out}},\,0))$

Appendix

Table 11: Hybrid retrieval engine parameters.

Parameter	Value	Description
Ligand–receptor interactions
Database size	280+ pairs	From CellChat, CellPhoneDB, NicheNet
Pathway categories	25	Signaling annotations
min_ligand_pct	0.10	Min ligand expr. in source
min_receptor_pct	0.05	Min receptor expr. in target
Score	$\text{pct}_{\text{in}}(L)\times\text{pct}_{\text{in}}(R)$	Expression fraction product

Appendix

Table 12: Analytical module parameters.

Parameter	Value	Description
LLM configuration
Default provider	Groq	Free tier, 500K tokens/day
Default model	LLaMA-3.1-8B	Via Groq Cloud API
Supported	4 providers	Groq, Gemini, OpenAI, Claude
Temperature	0.2	Low for reproducibility
Prompt limit	18,000 chars	$\approx$ 4,500 tokens

Appendix

Table 13: LLM interpretation parameters.

Figure 3: Cell-level UMAP of the cystic fibrosis airway dataset (D1) colored by Cell Ontology annotation. Approximately 96,000 cells are shown across 30 annotated cell types spanning immune (T cells, B cells, NK cells, macrophages, monocytes, dendritic cells, mast cells), epithelial (basal, suprabasal, multiciliated, secretory, goblet, club, ionocyte, neuroendocrine), and stromal (fibroblasts, pericytes, endocardial cells) compartments. Labels are placed at cluster centroids with iterative repulsion to minimize overlap.

Figure 4: Expression of HLA-E projected onto the cell-level UMAP of the cystic fibrosis airway dataset (D1). Color intensity (purple gradient) indicates normalized expression level, with non-expressing cells shown in grey. HLA-E is most highly expressed in immune cell clusters, particularly CD8 + T cells and NK cells, consistent with its role as a ligand for the NKG2A inhibitory receptor. Moderate expression is observed across epithelial populations including basal cells, supporting the HLA-E/NKG2A immune checkpoint axis identified by Berg et al.

AI Agents for Scientific Discovery

AI agents for scientific discovery use language models, tools, and experiments to generate, test, and refine scientific knowledge.

See all 287 papers

Jul 26, 2026

SCTA: An Agentic Framework for Stable and Interpretable Target Gene Discovery from Single-Cell RNA Sequencing

Identifying therapeutic target genes from single-cell RNA sequencing (scRNA-seq) data remains a fundamental challenge in translational biology. Unlike bulk assays, scRNA-seq captures heterogeneous cellular states and rare subpopulations, but this same heterogeneity makes target discovery highly sensitive to analytical choices throughout the pipeline, including preprocessing, cell population selection, differential expression analysis, and downstream biological interpretation. As a result, existing workflows and general-purpose analysis agents often produce unstable or difficult-to-interpret target hypotheses, limiting their reliability for disease-focused discovery. We present SCTA (Single-Cell Target Agent), a decision-centric agentic framework for stable and interpretable target gene discovery from scRNA-seq data. Rather than treating analysis as a single general-purpose reasoning task, SCTA decomposes target discovery into specialized agents aligned with key decision points in the single-cell pipeline and constrains downstream reasoning with structured biological evidence. In a representative ablation study on hereditary chronic pancreatitis, we demonstrate that SCTA's full evidence integration yields the most stable target selection across independent runs among the tested configurations, while recovering biologically coherent, disease-relevant mechanisms validated in prior studies. These results suggest that decision-aware agent orchestration tailored to the structure of single-cell analysis can improve the robustness, interpretability, and practical utility of target discovery in precision medicine.

Shuyu Chen, Chen Zhu, Ye Zhang +3

Siebel School of Computing and Data Science University of Illinois Urbana-Champaign Champaign, IL, USA · School of Information Sciences University of Illinois Urbana-Champaign Champaign, IL, USA · Novartis Biomedical Research Biomedical Research Cambridge, MA, USA +1

AI Agents for Scientific Discovery Scientific Workflow Automation Single-Cell RNA Sequencing Drug Discovery

Ask PDF

Jun 25, 2026

scBench-Long: Verifiable Benchmarking of Long-Horizon Single-Cell Biology

Single-cell studies require analysts to convert raw measurements into specific biological claims through multi-step workflows and integration of metadata, assay context, and auxiliary evidence. Existing AI-biology benchmarks largely measure broad knowledge, executable workflows, or local analysis steps. We introduce scBench-Long, a benchmark for long-horizon single-cell biology in which agents must recover scientific conclusions from raw or near-raw data without prescribed methods. The benchmark contains 21 evaluations spanning melanoma CD8 T-cell reactivity, CD8 RNA+ATAC regulatory inference, human--monkey chimera development, KRAS-driven lung tumor aging, and lethal COVID-19 lung pathology. Tasks cover paired scRNA/TCR sequencing, RNA and chromatin profiling, cross-species transcriptomics, combinatorial scRNA-seq, single-nucleus RNA-seq, immune repertoires, ortholog maps, ligand--receptor resources, and validation evidence. Candidate claims are reproduced, reviewed, and converted into controlled answer vocabularies with deterministic grading and trajectory rubrics. Across 1,068 completed trajectories, the strongest model--harness pair passes 16/63 runs (25.4%). scBench-Long evaluates whether agents can move beyond local analysis steps and make complex scientific claims that are supported by single-cell data.

Ian Diks, Zhen Yang, Arjun Banerjee +2

LatchBio, San Francisco, CA, USA

AI Agents for Scientific Discovery Long-Horizon Agent Evaluation Single-Cell Transcriptomics Agent Benchmarks

Ask PDF

Feb 26, 2026

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

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.

Jiahao Zhao, Feng Jiang, Shaowei Qin +5

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

Ask PDF

Abstract

AI Agents for Scientific Discovery

Explore similar work

SCTA: An Agentic Framework for Stable and Interpretable Target Gene Discovery from Single-Cell RNA Sequencing

scBench-Long: Verifiable Benchmarking of Long-Horizon Single-Cell Biology

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