arXiv:2607.24093v1 [q-bio.QM] 27 Jul 2026
TCellAlign: Cross-study T-cell Populations Alignment with Nomenclature-Guided Multi-Agent Workflow Pengyu Xie
Rongjia Zhou
Zhilin Ou
The Chinese University of Hong Kong, Shenzhen School of Artificial Intelligence Shenzhen, Guangdong, China [email protected]
Emory University Department of Computer Science and Biology Atlanta, GA, USA [email protected]
The Chinese University of Hong Kong, Shenzhen School of Artificial Intelligence Shenzhen, Guangdong, China [email protected]
Junyuan Zhang
Xiang Zhou
Xiaobo Sun∗
The University of Melbourne Yale University Department of Biochemistry and Department of Statistics and Data Pharmacology Science Melbourne, VIC, Australia New Haven, CT, USA [email protected] [email protected]
Emory University Department of Human Genetics Atlanta, GA, USA [email protected]
Jiaying Lu∗
Wenjing Ma∗
Emory University Center for Data Science, School of Nursing Atlanta, GA, USA [email protected]
The Chinese University of Hong Kong, Shenzhen School of Artificial Intelligence Shenzhen, Guangdong, China [email protected]
Abstract
Keywords
Cell type standardization plays a central role in integrating biological knowledge across single-cell studies. While standardized resources (e.g., Cell Ontology, Nomenclature Frameworks) provide unified vocabularies of cell populations, scientific publications and public datasets continue to use heterogeneous study-specific labels, making cross-study comparison difficult even when biologically equivalent cell populations are described. In this work, we are the first to formulate this challenge as an evidence-grounded cell population alignment problem and propose TCellAlign, a multi-agent framework that includes literature retrieval, information extraction, nomenclature-guided label alignment, and evidence-based adjudication. This modular design preserves the original terminology and supporting evidence reported by each study while producing standardized labels that can be compared across studies. We further construct a manually validated benchmark dataset linking study-specific labels, CZ CELLxGENE annotations, and standardized T-cell nomenclature across 44 manually curated, published studies (including over seven million cells) spanning four biological categories: healthy, cancer, infectious disease and inflammatory diseases. Across the evaluated tasks, TCellAlign achieves stronger semantic agreement than ontology-based baselines and maintains transcriptomic coherence with both open-source and closed-source large language models (LLM) backbones. By connecting literature, datasets, and expert’s nomenclature, TCellAlign enables consistent interpretation of T-cell subtypes and states across studies, facilitating biological knowledge integration and the development of future foundation models built upon standardized cellular representations.
T-cell nomenclature, single-cell RNA sequencing, text mining, large language models, multi-agent systems
∗ Corresponding authors: Xiaobo Sun, Jiaying Lu, and Wenjing Ma
1
INTRODUCTION
Cell types and cell states provide a fundamental basis for understanding biological development [5], tissue homeostasis [23], immune responses [43], and disease mechanisms [6]. Recent advances in single-cell transcriptomics technologies have enabled increasingly fine-grained characterization of cellular subtypes and functional states, leading to a rapidly growing body of cell atlases across diverse biological systems [17, 32, 37]. However, as the resolution of cellular characterization increases, so does the diversity of terminology used to describe biologically related cell populations. Individual studies often introduce study-specific cell labels based on distinct experimental settings and functional interpretations, making it difficult to compare, integrate, and reuse across studies. To facilitate consistent communication, the biomedical community has developed standardized ontologies [3, 8, 31], manuallycurated marker gene databases [12, 19] and nomenclature systems [28, 36, 50] for unambiguously representing cell identities. Although promising in principle, real-world practice differs substantially. In practice, cell type identification is largely performed manually and does not utilize any above resources. A standard procedure is that after clustering cells based on their transcriptomic profiles, domain experts assign biological identities by examining canonical marker genes. Depending on the biological question, annotation may stop at broad cell categories or proceed to increasingly fine-grained subtypes through iterative rounds of subclustering. For example, Saunders et al. [35] employed a two-stage clustering
KDD’27, 2027, San Jose, USA
Xie et al.
a Inputs
Information Extraction Agent
Search Agent
Mapping Agent
Semantic Scholar literature search PDF format conversion
Initial screening Candidate papers
Guidelines for T cell nomenclature
Lineage (L)
Function (F)
CD4+
Nat. Rev. Immunol.
CD8+
TH1 Treg TC2
…
…
S
D
U
B
W
R
Di erentiation (D,Q,G) N A M X G p
t
O
+
Renamed cell type
T-SNE 2
Query cells Marker gene database
2.93M
Manually annotating
T-SNE 1
Gene module inference
2.11M Query cells Reference cells
0.0
0.5
1.0
T-SNE 1
Perturbation prediction
0.95M
Inflammatory (n=6)
T-SNE 2
Reference-based annotation Infectious (n=13)
1.5
2.0
2.5
Final cell type
Drug response prediction
Marker-based annotation
Healthy (n=12)
Cross-source validation
c
1.08M
Cancer (n=13)
Renamed cell type
Judge Agent
Migration (M,R)
Dataset composition and scale
b
Cell type + contextual evidence
Evidence-supported cell types
Root
Renamed cell data
Nomenclature paper
Extract cell type names
Judge feedback
CELLxGENE data download
3.0
3.5
Our Dataset
…
Novel cell type discovery New type
Cell count (millions) ff
Figure 1: Overview of our study. (a) Overview of the proposed TCellAlign framework; (b) Statistics of our constructed benchmark dataset; (c) Potential biological applications for benchmark dataset after applying TCellAlign. strategy that identified 565 neuronal populations in the adult mouse brain, assigning study-specific labels such as “Frontal Cortex - Cluster 1 - SubCluster 5” based on biological knowledge and independent component (IC) loadings derived by independent component analysis (ICA) from single-cell transcriptomics data. Consequently, critical discrepancies exist between standardized definitions and study-specific annotations, posing a major obstacle to large-scale cross-study integration and automated knowledge discovery. Additionally, among the above resources, Cell Ontology (CL) stands out as a controlled vocabulary that describes cell types across tissues and species, promoting standardized and interoperable cell type annotations [38]. Existing computational approaches primarily formulate cell type standardization as an ontology matching problem, where lexical similarity or pretrained biomedical language models (PLMs) are used to map study-specific labels to standardized terms [10, 11, 14, 25, 29, 33, 34, 42]. However, CL primarily provides a representation of canonical cell types and may not fully capture the increasingly fine-grained cellular states revealed by single-cell transcriptomics [38, 42]. This challenge is particularly evident for immune cells, where cellular populations are often defined not only by lineage but also by dynamic functional states. Among them, T-cell subsets represent a particularly challenging example, as their populations frequently encode multiple biological dimensions, including lineage, function, migration, differentiation and antigen status. Thus, community-driven T-cell nomenclature frameworks provide a complementary reference for harmonizing fine-grained T-cell populations reported in the literature [28]. Moreover, standardized definitions available in curated single-cell data resources often provide limited context regarding the biological
rationale, marker evidence, and nomenclature decisions underlying the assigned labels [7]. These contextual details are frequently preserved in the original scientific literature, suggesting that effective T-cell nomenclature alignment requires integrating structured annotations with evidence extracted from primary studies. Motivated by recent efforts in establishing consensus T-cell nomenclature [28], which advocates a modular naming strategy for representing the complexity of T-cell biology, we formulate cross-study T-cell populations alignment as a multi-stage reasoning problem requiring literature understanding, biological evidence integration, and nomenclature-guided alignment. To address this challenge, we propose TCellAlign, a nomenclature-guided multi-agent framework for evidence-supported cross-study T-cell populations alignment (Figure 1a). Rather than treating cell type standardization as a direct matching problem, TCellAlign decomposes the task into four specialized reasoning stages: study retrieval, evidence extraction, nomenclature mapping, and mapping adjudication. Specifically, a Search Agent identifies relevant studies and associated single-cell data datasets, an Information Extraction Agent extracts paper-specific T-cell labels together with supporting evidence, a Mapping Agent aligns labels to a consensus T-cell nomenclature, and a Judge Agent verifies ambiguous mappings by integrating nomenclature standards and biological evidence. Our work makes three main contributions from the perspectives of paradigm, resource, and methodology. From a paradigm perspective, we are the first to formulate cross-study T-cell nomenclature alignment as a knowledge integration problem that unifies standardized nomenclature, study-specific annotations, and biological evidence. From a resource perspective, we construct the first
TCellAlign : Cross-study T-cell Populations Alignment with Nomenclature-Guided Multi-Agent Workflow
manually validated benchmark dataset that links paper-specific T-cell populations label, expert-curated cell ontologies [7], and standardized T-cell nomenclature across 44 studies containing over seven million cells, providing a valuable community resource for cross-study knowledge integration and downstream applications such as reference-based annotation, novel T-cell state discovery, and gene regulatory network inference (Figure 1b,c). From a methodological perspective, we develop TCellAlign, a nomenclatureguided multi-agent framework that integrates literature retrieval, information extraction, nomenclature-guided mapping, and evidencebased adjudication to achieve robust and biologically interpretable cross-study T-cell alignment. Although demonstrated on T-cell nomenclature, the framework is readily extensible to other consensus nomenclature systems and literature-grounded biological knowledge integration tasks.
2 RELATED WORK 2.1 Standards for T-Cell Population Alignment The need for consistent cell identity standardization is especially acute for T-cells, whose names often combine lineage, differentiation, activation, exhaustion, cytotoxicity, tissue residency, proliferation, clonality, and disease-context descriptors. As a result, biologically related T-cell populations are frequently described using heterogeneous terminology across studies. Several resources have been developed to facilitate standardized T-cell subtype alignment. Cell Ontology (CL) [3, 8] provides a structured and hierarchical vocabulary for representing cell types across tissues and species. However, CL is primarily designed to organize cell identities through ontology relationships and may not fully capture the context-dependent functional states and compositional descriptors increasingly used in studies involving T-cell subtypes. For example, exhausted, tissue-resident, or progenitor-like T-cell subtypes are often defined by combinations of lineage, functional state, and environmental context rather than a single fixed cell type term. Complementary resources, such as curated marker databases [12, 19], provide molecular evidence for supporting cell identity assignment but do not directly address inconsistencies in terminology across studies. To bridge these gaps, recent consensus efforts have proposed a standardized T-cell nomenclature based on compositional descriptors in a modular format, enabling a systematic representation of diverse T-cell states and functional phenotypes [28]. Compared with general-purpose ontologies, such nomenclature frameworks are better suited for harmonizing finegrained T-cell subtypes reported in the literature.
2.2
Computational Approaches to T-Cell Population Extraction and Alignment
Beyond these standardization resources, computational approaches to cell identity can be grouped into three complementary lines of work. First, biomedical text-mining models such as BioBERT [24], SciBERT [4], and PubMedBERT [14] support entity recognition, relation extraction, and other structured knowledge extraction tasks from scientific literature. However, these tasks typically treat entities or relations as their primary outputs, whereas fine-grained T-cell labels often include paper-specific compositional expressions whose interpretation depends on surrounding evidence. Second,
KDD’27, 2027, San Jose, USA
cell-typing tools and resources such as CellTypist [9], ProjecTILs [1], ImmCluster [20] and scPanKD [27] infer cell identities from expression profiles, marker genes, or reference atlases. Although effective for scalable annotation, their predictions are anchored to molecular inputs and predefined references, and do not directly reconcile heterogeneous names used for related T-cell states across papers. More recently, LLM-based approaches use biological knowledge to annotate cells or clusters from molecular evidence; scExtract also uses article-derived context, whereas CASSIA structures annotation through collaborating agents [18, 46, 47]. These advances expand both evidence sources and reasoning mechanisms for cell typing, but their principal output remains a newly assigned cell identity rather than an evidence-linked mapping from an author-reported label to an explicit T-cell nomenclature. Thus, existing work separately advances literature extraction, expression-based typing, and biological reasoning, but does not yet integrate them into a unified process for traceable cross-study T-cell label alignment.
2.3
Positioning of This Work
Our work is positioned at the intersection of single-cell annotation, biomedical information extraction, and nomenclature-guided knowledge harmonization. Unlike conventional cell annotation methods that infer cell identities directly from expression profiles, TCellAlign focuses on the complementary but unresolved challenge of harmonizing heterogeneous labels reported across scientific studies. Unlike ontology resources that provide standardized vocabularies, TCellAlign does not assume that study-specific labels conform to predefined concepts; instead, it adopts a flexible framework proposed by the T-cell nomenclature while extracting and reasoning over the biological context underlying each label. Unlike general biomedical information extraction systems that focus on retrieving isolated entities or relations, TCellAlign addresses the more challenging task of aligning context-dependent T-cell population that combine lineage and functional states. Finally, unlike recent LLM-based single-cell annotation approaches that use language models as direct predictors of cell identities from molecular features, TCellAlign leverages LLMs as components of a structured literature understanding workflow, where specialized agents perform evidence extraction, nomenclature mapping, and adjudication.
3 METHODS 3.1 Problem Definition Let 𝑝 ∈ P denote a qualified scientific publication, where P is the set of source papers. Within each paper, let 𝑂 = {𝑜 1, 𝑜 2, . . . , 𝑜𝑛 } denote the original label (i.e. cell type, cell state) named by its authors. For every original label 𝑜𝑖 , let E𝑖 denote the set of paperlocal evidence spans supporting the biological meaning of that label. We define each “label-evidence pair” as 𝑧𝑖 = (𝑜𝑖 , E𝑖 ). Given the Tcell nomenclature N proposed by Masopust et al. [28], our goal is to map each 𝑧𝑖 to a nomenclature-guided standardized definition: 𝑓 N (𝑧𝑖 ) = 𝑦𝑖 ,
𝑦𝑖 ∈ YN ,
(1)
where YN is the space of structured representations defined by N . Specifically, 𝑦𝑖 is represented by seven atomic nomenclature slots indexed by K = {𝐿, 𝐹, 𝑀, 𝑅, 𝐷, 𝑄, 𝐺 } corresponding to lineage,
KDD’27, 2027, San Jose, USA
Xie et al.
function, migration, migration subscript, differentiation state, differentiation state subscript, and antigen status, respectively. These slots are further organized into four biological dimensions: 𝑦𝑖 = (𝐿𝑖 , 𝐹𝑖 , (𝑀𝑖 , 𝑅𝑖 ), (𝐷𝑖 , 𝑄𝑖 , 𝐺𝑖 )) .
(2)
Although organized into four biological dimensions, the seven atomic slots can be inferred and validated independently, enabling fine-grained representation of heterogeneous T-cell identities. Each atomic slot is represented as 𝑦𝑖𝑘 = (𝑐𝑖𝑘 , 𝑣𝑖𝑘 ), 𝑘 ∈ K, where 𝑐𝑖𝑘 must be a nomenclature component defined by N and 𝑣𝑖𝑘 denotes its corresponding biological meaning. When the original label and paper-local evidence do not support the corresponding property, the corresponding slot is left empty: 𝑦𝑖𝑘 = (∅, ∅). An empty slot indicates insufficient evidence and should be distinguished from an explicit nomenclature component representing an unknown value. For example, the migration module U explicitly denotes unknown migration, whereas an empty migration slot indicates that the available evidence does not support any migration statement. Additional contextual information including tissue source, disease context, species, molecular markers, and experimental conditions may be retained as supporting evidence, but are not encoded in 𝑦𝑖 unless N defines a corresponding module. Finally, a deterministic rendering function converts the structured representations into the final aligned labels 𝑀: 𝑚𝑖 = 𝑔(𝑦𝑖 ; N ),
(3)
where 𝑔 concatenates all non-empty nomenclature components according to the ordering rules defined by N . Therefore, 𝑓 N determines the biological content of the aligned representation based on the original label and supporting evidence, whereas 𝑔 ensures a consistent nomenclature-compliant label format. The resulting alignment preserves traceability by maintaining the original label and evidence supporting each encoded biological property.
3.2
TCellAlign Model Architecture
As shown in Figure 1, TCellAlign is a four-stage workflow for transforming heterogeneous T-cell labels reported in the scientific literature into evidence-grounded structured representations under a modular nomenclature. The workflow consists of a Search Agent, an Information Extraction (IE) Agent, a Mapping Agent, and a Judge Agent. The Search Agent identifies relevant publications with available single-cell transcriptomic data. The IE Agentextracts the author-reported T-cell labels and supporting evidence while preserving the original terminology. The Mapping Agent transforms each “label-evidence pair” into a nomenclature-guided structured representation. The Judge Agent validates the mapping consistency against both source evidence and nomenclature rules.. This staged decomposes cross-study T-cell subtype alignment into four interconnected challenges: literature discovery, evidence extraction, nomenclature-guided mapping, and evidence-based judge. Literature retrieval establishes the study corpus and associated data resources; information extraction links original labels with traceable publication evidence; mapping assigns evidence-supported biological attributes to predefined nomenclature slots; and judge validates each slot-level decision against both source evidence and nomenclature constraints. By preserving the original label, supporting evidence, and structured representation throughout the process,
TCellAlign ensures that every aligned annotation remains traceable and biologically interpretable. Instead of regenerating complete labels after validation, TCellAlign performs selective slot-level refinement. Multiple judges from Judge Agent evaluate each nomenclature slot, and only inconsistent mappings are returned to the Mapping Agent with feedback, while validated slots remain unchanged. When evidence is insufficient, the corresponding information is preserved as unresolved rather than inferred. After each refinement round, the updated representation is converted into an aligned label through the deterministic nomenclature function and re-evaluated until convergence or a predefined stopping criterion is reached (Algorithm 1). Algorithm 1
TCellAlign
Input: Literature source S ; eligibility criteria Γ ; nomenclature N . Control: Maximum number of adjudication rounds 𝑇max . Output: Finalized records J containing structured mappings, aligned labels, and adjudication histories. 1 P ← SearchAgent( S, Γ) 2 for each qualified paper 𝑝 ∈ P do 3 𝑍𝑝 ← IEAgent(𝑝 ) ⊲ Extract labels and supporting evidence (0) 4 𝑌𝑝 ← MappingAgent(𝑍𝑝 , N ) ⊲ Generate seven-slot representations 5 end for (0) 6 𝑍 ← {𝑍 𝑝 }𝑝 ∈P , 𝑌 (0) ← {𝑌𝑝 }𝑝 ∈P 7 for 𝑡 ← 0, . . . ,𝑇max − 1 do 8 𝑉 (𝑡 ) ← JudgeAgent(𝑍, 𝑌 (𝑡 ) , N ) ⊲ Evidence-based slot-level adjudication 9 if AllConfirmed(𝑉 (𝑡 ) ) then break (𝑡 ) 10 𝐶 (𝑡 ) ← { (𝑖, 𝑘 ) : status(𝑉𝑖,𝑘 ) = rejected } 11 if 𝐶 (𝑡 ) = ∅ then break 12 𝑌 (𝑡 +1) ← 𝑌 (𝑡 ) ⊲ Preserve unflagged representations (𝑡 +1) (𝑡 ) (𝑡 ) 13 𝑌𝐶 ← MappingAgent(𝑍𝐶 , N, 𝑌𝐶 , 𝑉𝐶 ) ⊲ Revise rejected slots 14 if 𝑌 (𝑡 +1) = 𝑌 (𝑡 ) then break 15 end for 16 𝑀 ← 𝑔 (𝑌 (𝑡 +1) ; N ) ⊲ Generate aligned labels 17 J ← Finalize( P, 𝑍, 𝑌 (0:𝑡 ) , 𝑀, 𝑉 (0:𝑡 ) ) 18 return J
Search Agent. The Search Agent initiates the workflow from a literature source S, which can be obtained from biomedical database queries or a user-provided collection of candidate publications. It retrieves accessible candidate papers and applies the study-level eligibility criteria Γ. A publication is retained only if it reports T-cell types, subtypes, states, or populations suitable for downstream extraction and its corresponding single-cell transcriptomic data can be verified as available through CZ CELLxGENE for downstream analysis. Eligible studies are then ranked by citation count to prioritize processing. The resulting qualified paper set P retains its metadata, source PDF, screening decision, and verified data reference 𝐷 𝑝 for each study (Appendix A). Information Extraction Agent. For each publication 𝑝 ∈ P, the Information Extraction (IE) Agent first converts the source PDF into a machine-readable full-text representation and then extracts author-reported T-cell labels together with the textual evidence required to interpret their biological meanings. The extraction unit is a biologically meaningful cell population rather than individual textual mentions or cluster indentifiers. Multiple mentions referring to the same population are consolidated and paper-specific abbreviations or cluster identifiers are resolved using definitions provided within the publication, while preserving the original terminology. Each extracted label 𝑜𝑖 is paired with a set of evidence
TCellAlign : Cross-study T-cell Populations Alignment with Nomenclature-Guided Multi-Agent Workflow
spans E𝑖 , including marker gene descriptions, phenotypic descriptions, or other functional properties, to form 𝑧𝑖 = (𝑜𝑖 , E𝑖 ). For each 𝑛𝑝 publication 𝑝, these pairs form Z𝑝 = {𝑧𝑖 }𝑖=1 and are provided to the Mapping Agent for nomenclature-guided alignment. Mapping Agent. Given a “label-evidence pair” 𝑧𝑖 = (𝑜𝑖 , E𝑖 ), the Mapping Agent applies the nomenclature-guided mapping function 𝑓 N to infer an evidence-consistent structured representation: 𝑦𝑖 = 𝑓 N (𝑧𝑖 ). At refinement round 𝑡, the structured representation (𝑡 ) (𝑡 ) (𝑡 ) (𝑡 ) is denoted as 𝑦𝑖(𝑡 ) = {𝑦𝑖𝑘 }𝑘 ∈ K , where 𝑦𝑖𝑘 = (𝑐𝑖𝑘 , 𝑣𝑖𝑘 ) denotes the 𝑘-th atomic slot and K indexes the seven atomic nomenclature slots defined above. A slot is populated only when the original label 𝑜𝑖 or supporting evidence E𝑖 ) supports the corresponding biological attribute; otherwise, the slot remains empty. After inferring all supported attributes, the final aligned label is generated deterministically with 𝑚𝑖(𝑡 ) = 𝑔(𝑦𝑖(𝑡 ) ; N ). For publication 𝑝, we denote 𝑛𝑝 𝑛𝑝 𝑌𝑝(𝑡 ) = {𝑦𝑖(𝑡 ) }𝑖=1 and 𝑀𝑝(𝑡 ) = {𝑚𝑖(𝑡 ) }𝑖=1 . This separation between structured inference and label generation ensures that each mapped label is generated from an explicit and interpretable biological representation. Each record retains the original label 𝑜𝑖 , supporting evidence E𝑖 ), structured representation 𝑦𝑖 , mapped label 𝑚𝑖 , and nomenclature N . During later refinement rounds, only slots identified by the Judge Agent as inconsistent are updated, while validated slots remain unchanged. Judge Agent and Iterative Refinement. At refinement round 𝑡, the Judge Agent independently assesses whether every slot in 𝑦𝑖(𝑡 ) is consistent with the original label 𝑜𝑖 , paper evidence E𝑖 , mapped label 𝑚𝑖(𝑡 ) , and nomenclature constraints N . Three independent judges evaluate each slot, and the majority decision is recorded as either confirmed or rejected. The resulting records are de(𝑡 ) noted by 𝑉 (𝑡 ) with 𝑉𝑝,𝑖,𝑘 representing the validation outcome for slot 𝑘 of label 𝑖 in publication 𝑝. During refinement, rejected slots requiring revision are returned to the Mapping Agent with Judge feedback, while confirmed slots are preserved. The Mapping Agent then updates the flagged slots based on the original “label-evidence” pair, nomenclature constraints, and feedback, after which 𝑚𝑖(𝑡 ) is regenerated and re-evaluated. Claims that cannot be verified from the available evidence remain unresolved rather than being filled. The iterative refinement process terminates when all mappings are confirmed, no actionable feedback remains, the mapping converges, or the maximum number of refinement rounds 𝑇max is reached. All intermediate mappings and adjudication records are retained to make the refinement process traceable and reviewable.
3.3
Benchmark Dataset Construction
To evaluate cross-study T-cell populations alignment, we constructed a manually validated benchmark database linking the paper-reported T-cell subsets with both author-provided annotations and the standardized Cell Ontology (CL) annotations available in the released single-cell datasets [7]. Following the data collection procedure described in Appendix A, we screened candidate publications with Search Agent and retained 44 studies covering four biological categories: Healthy, Cancer, Infectious and Inflammatory (Appendix Table S1). Each retained study contains a T-cell-relevant H5AD dataset, an author-provided annotation list, and a manually verified
KDD’27, 2027, San Jose, USA
correspondence between paper-reported labels and data annotations (Appendix D). These 44 studies constituted the benchmark used for the subsequent alignment evaluation. For each retained study, our IE Agent first extracts the set of paper-provided T-cell labels 𝑂 = {𝑜 1, 𝑜 2, . . . , 𝑜𝑛 }. With manual curation (Appendix D), we obtained 𝑂 → 𝐴, where 𝐴 = {𝑎 1, 𝑎 2, . . . , 𝑎𝑚 } denotes the set of author-provided annotations. This manual curation ensures that the benchmark explicitly connects terminology used in the original publications with the annotations provided in the released datasets. In addition, CZ CELLxGENE provides standardized CL annotations through the cell_type field and associated ontology term identifiers cell_type_ontology_term_id. We associate each data annotation with its corresponding CL term(s), yielding 𝐴 → 𝐶𝐿 and the resulting correspondence between studyspecific terminology and standardized cell identity representations, denoted by 𝑂 → 𝐴 → 𝐶𝐿. Given extracted paper labels 𝑂, each alignment method, including TCellAlign and competing methods such as PubMedBERT [14], ZOOMA [10], BioPortal [29], etc., predicts a mapped label 𝑂 → 𝑀, where 𝑀 = {𝑚 1, 𝑚 2, . . . , 𝑚𝑛 } denotes the mapped labels generated by the method. Because the original paper labels 𝑂 serve as a shared intermediate representation, we establish an induced correspondence between predicted labels 𝑀 and reference CL annotations with 𝑀 ← 𝑂 → 𝐴 → 𝐶𝐿. This correspondence enables quantitative evaluation of alignment quality from two complementary perspectives: (i) annotation-level consistency by comparing predicted label groupings with CL-based reference groupings, and (ii) biological coherence by evaluating whether cells assigned to the same predicted label exhibit consistent transcriptomic profiles. Additional details regarding the baseline methods and evaluation metrics are provided in Appendices E and F.
4
RESULTS
Overall, we investigate the following research questions. RQ1: Can TCellAlign generate label groupings that are consistent with the Cell Ontology annotations provided by CZ CELLxGENE? RQ2: Do the aligned labels correspond to biologically meaningful transcriptional populations compared to existing methods? RQ3: How does TCellAlign support biological interpretation? RQ4: How do individual technical components influence model performance?
4.1
Experimental Settings
Compared baselines. We adopt the following established crossstudy cell type alignment methods as baselines for comparison. (1) Lexical similarity-based methods that map original cell names to Cell Ontology (CL) terms via string matching: BM25 [34] and Levenshtein distance [33]. (2) Semantic similarity-based methods that leverage pre-trained language model embeddings: PubMedBERT [14] and SapBERT [25]. (3) Bioinformatics tools that perform multi-ontology-based annotation: BioPortal [29] and ZOOMA [10]. Further technical details on these baselines are given in Appendix E. TCellAlign implementation details. To evaluate the effect of different LLM backbones (RQ4), we first conduct component-wise ablation studies using six representative models, including the opensource models MiMo-V2.5-Pro, DeepSeek-V4-Pro, and DeepSeekV4-Flash [39, 48], and the closed-source models Gemini 3.5 Flash,
KDD’27, 2027, San Jose, USA
Xie et al.
Table 1: Performance comparison across four datasets. Higher is better for all metrics. AMI: Adjusted Mutual Information. F1: Merge F1. MC: Merge Consistency. Note: Within each IE block, bold and underlined values mark the column-best and second-best results, respectively. Underlining is omitted for tied maxima. † denotes thinking mode. All IE
Healthy
Cancer
Inflammatory
Method
AMI
F1
MC
AMI
F1
MC
AMI
F1
MC
AMI
F1
MC
AMI
F1
MC
BM25 Levenshtein PubMedBERT SapBERT BioPortal ZOOMA TCellAlign
0.254 0.368 0.518 0.524 0.532 0.197 0.647
0.347 0.350 0.479 0.439 0.519 0.381 0.586
0.264 0.352 0.407 0.345 0.361 0.252 0.378
0.427 0.311 0.424 0.447 0.775 0.285 0.671
0.511 0.463 0.525 0.590 0.734 0.422 0.845
0.313 0.381 0.446 0.399 0.537 0.295 0.460
0.206 0.381 0.441 0.417 0.535 0.280 0.474
0.184 0.211 0.116 0.022 0.291 0.381 0.304
0.226 0.215 0.194 0.060 0.146 0.161 0.320
0.201 0.447 0.628 0.767 0.403 -0.023 0.826
0.122 NA 0.118 NA 0.133 0.357 0.337
0.191 0.572 0.226 0.261 0.514 0.324 0.434
0.120 0.286 0.623 0.368 0.321 0.330 0.558
0.296 0.338 0.705 0.376 0.514 0.339 0.800
0.367 0.226 0.753 0.641 0.265 0.218 0.345
BM25 Levenshtein DeepSeek- PubMedBERT SapBERT V4-Pro† BioPortal ZOOMA TCellAlign
0.251 0.340 0.620 0.544 0.504 0.241 0.636
0.428 0.469 0.579 0.374 0.589 0.374 0.535
0.273 0.286 0.511 0.460 0.396 0.243 0.426
0.347 0.395 0.532 0.527 0.643 0.461 0.685
0.562 0.707 0.703 0.643 0.685 0.470 0.814
0.253 0.336 0.453 0.472 0.400 0.343 0.354
0.302 0.283 0.476 0.382 0.433 0.152 0.428
0.350 0.394 0.376 0.277 0.389 0.293 0.209
0.285 0.306 0.441 0.436 0.335 0.070 0.441
0.165 0.362 0.847 0.783 0.375 0.144 0.822
0.377 0.362 0.528 0.110 0.570 0.380 0.570
0.304 0.146 0.253 0.174 0.517 0.285 0.367
0.146 0.295 0.591 0.384 0.646 0.188 0.553
0.369 0.322 0.643 0.435 0.753 0.366 0.778
0.235 0.520 0.749 0.761 0.223 0.327 0.616
GPT-5.5
Claude Opus 4.8, and GPT-5.5 [2, 13, 30]. We additionally compare thinking and non-thinking variants of DeepSeek for Information Extraction, and of both DeepSeek and MiMo for Mapping. For Judge Agent ablations, the Mapping Agent is instantiated with GPT-5.5, DeepSeek-V4-Pro, or DeepSeek-V4-Flash, while the Judge ensemble is fixed to DeepSeek-V4-Pro, GLM-5.2 [51], and GPT-5.5. Detailed configurations are presented in the corresponding ablation sections. Based on the ablation results, we instantiate the complete TCellAlign workflow using GPT-5.5 and DeepSeek-V4-Pro (thinking), the best-performing closed-source and open-source backbones, respectively. TCellAlign aligns T-cell populations to a standardized T-cell nomenclature through evidence-grounded information extraction, nomenclature-guided mapping, and evidence-based adjudication. The main experiments evaluate 44 studies comprising over seven million cells and both implementations use identical paper inputs, prompt templates, nomenclature rules, and adjudication protocols, differing only in the underlying LLM backbone.
4.2
Infectious
Grouping consistency evaluation (RQ1)
As described in Section 3.3, we evaluate whether the label merging decisions produced by different methods are consistent with the Cell Ontology annotations provided by CZ CELLxGENE [7]. Given the correspondence 𝑀 ← 𝑂 → 𝐴 → 𝐶𝐿, the reference CL labels (𝐶𝐿) and aligned labels (𝑀) are merged partitions of the original data annotations (𝐴). We therefore evaluate consistency between the predicted and reference grouping structures using Adjusted Mutual Information (AMI) and Merge F1 (Appendix F). As shown in Table 1, TCellAlign with GPT-5.5 achieves the highest overall AMI and Merge F1 among all compared methods. Replacing GPT-5.5 with DeepSeek-V4-Pro (thinking) yields highly comparable results, with overall AMI and Merge F1 decreasing by only 0.011 and 0.051, respectively. The two implementations exhibit nearly identical AMI across all categories, indicating that they recover largely consistent grouping structures. The largest difference is observed in the Cancer cohort, where DS-V4-Pro yields a lower Merge F1. We attribute this reduction primarily to its tendency to extract fewer
paper-specific T-cell labels during the IE stage, producing coarser predicted partitions and consequently lower merge-level resolution. Overall, these results demonstrate that TCellAlign remains robust to the choice of LLM backbone.
4.3
Transcriptomic coherence evaluation (RQ2)
Using gene expression profiles from the CZ CELLxGENE datasets, we further evaluated whether each alignment method produces transcriptionally coherent cell populations. As shown in Table 1, TCellAlign instantiated with GPT-5.5 achieves the second-highest Merge Consistency (MC), with PubMedBERT ranking first. We hypothesize that this difference reflects the two approaches’ distinct merging strategies. Semantic similarity-based methods such as PubMedBERT and SapBERT primarily rely on local lexical or semantic similarity, and therefore tend to assign paper-specific labels to distinct ontology terms rather than merging them. Consequently, they perform fewer merges (0.727 and 0.818 per study on average, respectively), reducing the likelihood of combining transcriptionally heterogeneous populations and resulting in higher transcriptomic coherence. In contrast, TCellAlign performs evidence-guided alignment by jointly considering paper-derived biological evidence and the modular T-cell nomenclature. This produces a higher average merge rate (1.091 merges per study), yielding substantially better agreement with Cell Ontology (RQ1) while occasionally merging biologically related but transcriptionally distinguishable populations. The complementary behavior observed in RQ1 and RQ2 therefore suggests an inherent trade-off between ontology consistency and transcriptomic coherence. To better understand this trade-off, we further investigate representative alignment cases in the following section.
4.4
Structural consistency evaluation (RQ3)
Besides the quantitative evaluations, we performed a qualitative case study to examine whether the labels produced by TCellAlign preserve biologically meaningful structural relationships among T-cell populations. As discussed in Section 3.1, our nomenclature
TCellAlign : Cross-study T-cell Populations Alignment with Nomenclature-Guided Multi-Agent Workflow
a
KDD’27, 2027, San Jose, USA
b
c
d
Root C6/C11
alpha-beta T cell (107)
CD4 + CD8 + NKT C4/C7
T cell (171) C11
C10
immature alpha-beta T cell (17) mature alpha-beta T cell (88) alpha-beta intraepithelial T cell (5)
T reg T FH T H1 ⊘ ⊘ C2
C3
CD4-positive, alpha-beta T cell (36) C0/C1/C2/C3 CD8-positive, alpha-beta T cell (28) C4/C6/C7/C8
D M C1
SC0 ⊘ X C8
effector memory CD45RA-positive, alpha-beta T cell, terminally… mature NK T cell (15) C10 mucosal-associated invariant T cell (1)
Figure 2: Breast Cancer Case Study with Cluster Identifiers annotated. (a) Clustered Heatmap. (b) UMAP visualization. (c) TCellAlign nomenclature-guided standardized definitions. (d) Cell Ontology standardized definitions. represents T-cell standardized definitions with four biological dimensions covering Lineage, Function, Migration, and Differentiation and Antigen Status. An ideal alignment to nomenclature should therefore not only assign consistent names, but also group biologically related cell populations under the same dimension while preserving their functional distinctions. We selected a representative breast cancer study [44], which provides diverse T-cell populations with detailed descriptions in the original publication (Section Lymphocytes and Innate Lymphoid Cells) and corresponding validated dataset annotations. Using a GPT-5.5 backbone, the IE Agent extracted 10 T-cell populations corresponding to 12 annotated clusters (c0-c11) in the released dataset. Two clusters were excluded: cluster 5 (c5_CD8+_GZMK) was not explicitly described in the manuscript, and cluster 9 (c9_NK_cells_ AREG) represented natural killer (NK) cells rather than natural killer T (NKT) cells. Although NK and NKT cells share cytotoxic characteristics, they represent distinct immune populations, with only NKT cells included in our T-cell scope [26]. After validation, we established correspondence between paper-reported labels and data annotations using cluster identifiers. Combining this correspondence with CZ CELLxGENE CL annotations enabled direct comparison between TCellAlign-generated labels and standardized ontology terms. The resulting mappings are summarized in Table S2. Overall, the labels generated by TCellAlign exhibited structural organization consistent with both Cell Ontology annotations and transcriptomic similarity patterns measured by hierarchical clustering and UMAP visualization (Figure 2(a), (b)). For example, Clusters 0-3 were consistently assigned to the CD4+ lineage while preserving distinct functional states, including Treg and TFH , and displayed highly similar transcriptional profiles. Similarly, Clusters 4, 7, and
8 were mapped to the CD8+ lineage (Figure 2(c)) and formed a transcriptionally coherent group (Figure 2(a)). These results suggest that the proposed nomenclature alignment preserves both lineage relationships and functional distinctions among T-cell populations. An informative exception was Cluster 6 (c6_IFIT1). While CZ CELLxGENE annotated this cluster as CD8-positive, alpha-beta T cell (Figure 2(d)), the original study reported that the IFN-I signature cluster (T cells:IFIT1/c6) and the proliferative cluster (T cells:MKI67/c11) contained mixed CD4+ and CD8+ T-cell populations. Based on this evidence, the Mapping Agent assigned both clusters to the broader T-cell lineage rather than imposing a CD4+ or CD8+ designation unsupported by the source literature. These two clusters also exhibited similar transcriptional profiles, consistent with the biological interpretation from the manuscript. This example highlights that TCellAlign does not merely reproduce existing ontology annotations, but integrates publication-specific evidence to produce biologically grounded nomenclature alignment.
4.5
Ablation Study (RQ4)
To examine three component-level design choices in TCellAlign, we evaluate paper-specific label extraction, alternative reference inputs for mapping, and whether refinement reduces mapping errors. Information Extraction Agent. To evaluate paper-specific label and evidence extraction, we compare the IE Agent’s outputs with annotations from four independent domain experts on the seven-paper evaluation set listed in Appendix Table S3. Because system outputs and expert annotations are variable-length sets of free-text terms, we use the Soft Bipartite Matching F1 Score (SBMSF1) for label-only and label+evidence evaluation. We compare the LLM settings with two fixed-nomenclature baselines and report human-human agreement as a descriptive reference (Appendix F.3).
KDD’27, 2027, San Jose, USA
Xie et al.
Table 2: Information Extraction Agent performance against expert annotations. Note: Model names are abbreviated; full configurations are described in Experimental Settings. DS denotes DeepSeek. The Human column reports mean pairwise SBMS-F1 over six expert pairs per paper, macro-averaged across seven papers. Bold numeric values denote the highest non-human score in each row. † denotes thinking mode. Open-source models
SBMS-F1 Label only Label+Evidence
Closed-source models
Baselines / reference
MiMo-Pro
DS-Pro†
DS-Flash†
DS-Pro
DS-Flash
Gemini
Claude
GPT
TCell-6
TCell-43
Human
0.8954 0.8443
0.8757 0.8373
0.8620 0.8484
0.8430 0.8315
0.8475 0.8366
0.8599 0.8449
0.8343 0.8287
0.8407 0.8209
0.5818 0.4373
0.3539 0.2721
0.9197 0.8838
All paper-contextualized LLM settings outperform the two fixednomenclature baselines on this evaluation set (Table 2). MiMo-V2.5Pro performs best for label-only extraction, whereas DeepSeek-V4Flash (thinking) performs best for label-plus-evidence extraction, closely followed by Gemini 3.5 Flash. Moreover, the thinking variants score higher on both targets for the two DeepSeek models. Table 3: Mapping Agent performance and efficiency with PDF and structured-rule references. Note: Gold-standard information is withheld from both reference inputs. Accuracy is shown as PDF→Rule; Δ is Rule minus PDF, and savings are relative to PDF. † denotes thinking mode. Model
Acc. (%)
Δ (pp)
PDF→Rule (Open-source models) MiMo-V2.5-Pro† MiMo-V2.5-Pro DeepSeek-V4-Flash† DeepSeek-V4-Flash DeepSeek-V4-Pro† DeepSeek-V4-Pro
81.2→81.2 93.8→81.2 81.2→81.2 62.5→81.2 93.8→87.5 75.0→81.2
(Closed-source models) GPT-5.5 93.8→93.8 Claude Opus 4.8 87.5→87.5 Gemini 3.5 Flash 93.8→87.5
0.0 −12.5 0.0 +18.8 −6.3 +6.3 0.0 0.0 −6.3
Saving (%) Token
Cost
66.3 76.2 51.3 75.3 54.4 75.3
54.6 69.2 34.7 67.5 37.9 67.6
56.9 72.3 61.0
30.1 45.4 28.2
Mapping Agent. We assess alternative reference inputs for the Mapping Agent through a paired comparison of the paper PDF and a structured normalization rule on the 16-case gold-standard set in Table 7 of Masopust et al. [28]. To prevent answer leakage, we remove Table 7 from the PDF and the corresponding gold-standard examples from the rule, using the withheld mappings only for evaluation. For each model setting, we evaluate case-insensitive exact-match accuracy and record token count and estimated cost using public OpenRouter list prices (Table 3). The PDF and structured rule achieve nearly identical average exact-match accuracy across model settings, while the rule substantially reduces token use and estimated cost in every setting. We therefore use the structured rule in the subsequent Judge experiment as the more efficient reference input. Judge Agent. To evaluate the corrective effect of the Judge Agent, we compare the initial mapping accuracy and final accuracy after up to three rounds of Judge-guided refinement on the same 16-case ground truth dataset. All settings use the same structured rule without examples, evaluation set, and fixed Judge ensemble with only the Mapping Agent backend varies. Judge-guided refinement corrects one additional case for each backend, corresponding
to a 6.25-percentage-point improvement: GPT-5.5 improves from 93.75% to 100.00%, while DeepSeek-V4-Pro and DeepSeek-V4-Flash in thinking mode improve from 87.50% to 93.75% and from 81.25% to 87.50%, respectively. All settings converge after the first refinement round. These results demonstrate that Judge-guided refinement consistently improves mapping accuracy across different Mapping Agent backends, while the final performance remains dependent on the quality of the initial mapping model.
5
CONCLUSION
We present TCellAlign, a nomenclature-guided multi-agent framework that integrates literature retrieval, information extraction, nomenclature-guided mapping, and evidence-based adjudication to align heterogeneous T-cell populations across studies. Experiments on 44 studies demonstrate that TCellAlign achieves superior evaluation performance, supporting transparent and biologically grounded cross-study T-cell nomenclature alignment. In the future, we plan to further improve the reasoning capabilities of the framework and expand the data resource to enable immunological discovery.
6
Limitations and Ethical Considerations
Limitations. While TCellAlign demonstrates strong performance in cross-study T-cell nomenclature alignment, several limitations remain. First, the benchmark includes only studies with accessible publications from CZ CELLXGENE, which may limit coverage and introduce selection bias. Second, CL annotations as evaluation may not capture fine-grained or context-dependent T-cell populations. Lastly, our ablation studies are limited and broader validation across additional studies and nomenclature frameworks is needed to establish generalizability. Ethical Considerations. Our study relies exclusively on publicly available, anonymized biomedical literature and single-cell transcriptomics datasets. Because no personally identifiable information is present or collected, the work raises no privacy concerns. The resulting computational resource for cell-type nomenclature mapping is intended solely for basic scientific research and should not be applied in clinical decision-making.
7
Generative AI Usage
Generative AI models were used to develop and evaluate TCellAlign, as described in the Methods and Experimental Settings, and for manuscript language editing. The authors reviewed and verified all AI-generated outputs and take full responsibility for the content of this manuscript.
TCellAlign : Cross-study T-cell Populations Alignment with Nomenclature-Guided Multi-Agent Workflow
References [1] Massimo Andreatta, Jesus Corria-Osorio, Soren Muller, Rafael Cubas, George Coukos, and Santiago J. Carmona. 2021. Interpretation of T cell states from single-cell transcriptomics data using reference atlases. Nature Communications 12 (2021), 2965. doi:10.1038/s41467-021-23324-4 [2] Anthropic. 2026. System Card: Claude Opus 4.8. https://www.anthropic.com/ claude-opus-4-8-system-card Released May 28, 2026. [3] Jonathan Bard, Seung Y Rhee, and Michael Ashburner. 2005. An ontology for cell types. Genome biology 6, 2 (2005), R21. [4] Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. SciBERT: A Pretrained Language Model for Scientific Text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing. 3613–3618. doi:10.18653/v1/D19-1371 [5] Junyue Cao, Malte Spielmann, Xiaojie Qiu, et al. 2019. The single-cell transcriptional landscape of mammalian organogenesis. Nature 566, 7745 (2019), 496–502. doi:10.1038/s41586-019-0969-x [6] Yanshuo Chu, Enyu Dai, Yating Li, et al. 2023. Pan-cancer T cell atlas links a cellular stress response state to immunotherapy resistance. Nature Medicine 29, 6 (2023), 1550–1562. doi:10.1038/s41591-023-02371-y [7] CZ CELLxGENE Discover Consortium, Siraj Abdulla, Brian Aevermann, Paulo Assis, Sarah Badajoz, et al. 2025. CZ CELLxGENE Discover: a single-cell data platform for scalable exploration, analysis and modeling of aggregated data. Nucleic Acids Research 53, D1 (2025), D886–D900. doi:10.1093/nar/gkae1142 [8] Alexander D. Diehl, Terrence F. Meehan, Yvonne M. Bradford, Matthew H. Brush, Wasila M. Dahdul, David S. Dougall, Yongqun He, David Osumi-Sutherland, Alan Ruttenberg, Sirarat Sarntivijai, Ceri E. Van Slyke, Nicole A. Vasilevsky, Melissa A. Haendel, Judith A. Blake, Christopher J. Mungall, and Suzanna E. Lewis. 2016. The Cell Ontology 2016: enhanced content, modularization, and ontology interoperability. Journal of Biomedical Semantics 7 (2016), 44. doi:10.1186/s13326016-0088-7 [9] Cecilia Dominguez Conde et al. 2022. Cross-tissue immune cell analysis reveals tissue-specific features in humans. Science 376, 6594 (2022), eabl5197. doi:10. 1126/science.abl5197 [10] EMBL-EBI. Accessed 2026. ZOOMA: Ontology annotation tool. https://www.ebi. ac.uk/spot/zooma/. [11] Can Ergen, Galen Xing, Chenling Xu, Martin Kim, Michael Jayasuriya, Erin McGeever, Angela Oliveira Pisco, Aaron Streets, et al. 2024. Consensus prediction of cell type labels in single-cell data with popV. Nature Genetics 56 (2024), 2731– 2738. doi:10.1038/s41588-024-01993-3 [12] Oscar Franzén, Li-Ming Gan, and Johan LM Björkegren. 2019. PanglaoDB: a web server for exploration of mouse and human single-cell RNA sequencing data. Database 2019 (2019), baz046. [13] Google DeepMind. 2026. Gemini 3.5 Flash Model Card. https://deepmind.google/ models/model-cards/gemini-3-5-flash/ Released May 19, 2026. [14] Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong Liu, Tristan Naumann, Jianfeng Gao, and Hoifung Poon. 2022. Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare 3, 1 (2022), 1–23. doi:10.1145/3458754 [15] Mercedes Guerrero-Murillo, Aina Rill-Hinarejos, Juan L. Trincado, et al. 2024. Integrative single-cell multi-omics of CD19-CARpos and CARneg T cells suggest drivers of immunotherapy response in B-cell neoplasias. bioRxiv (2024). doi:10. 1101/2024.01.23.576878 [16] Guangchun Han, Qing Deng, Mario L. Marques-Piubelli, et al. 2022. Follicular Lymphoma Microenvironment Characteristics Associated with Tumor Cell Mutations and MHC Class II Expression. Blood Cancer Discovery 3, 5 (2022), 428–443. doi:10.1158/2643-3230.BCD-21-0075 [17] Xiaoping Han, Renying Wang, Yincong Zhou, et al. 2018. Mapping the Mouse Cell Atlas by Microwell-Seq. Cell 172, 5 (2018), 1091–1107.e17. doi:10.1016/j.cell. 2018.02.001 [18] Wenpin Hou and Zhicheng Ji. 2024. Assessing GPT-4 for cell type annotation in single-cell RNA-seq analysis. Nature Methods 21 (2024), 1462–1465. doi:10.1038/ s41592-024-02235-4 [19] Congxue Hu, Tengyue Li, Yingqi Xu, Xinxin Zhang, Feng Li, Jing Bai, Jing Chen, Wenqi Jiang, Kaiyue Yang, Qi Ou, et al. 2023. CellMarker 2.0: an updated database of manually curated cell markers in human/mouse and web tools based on scRNAseq data. Nucleic acids research 51, D1 (2023), D870–D876. [20] Tiantongfei Jiang, Weiwei Zhou, Qi Sheng, Jiaxin Yu, Yunjin Xie, Na Ding, Yunpeng Zhang, Juan Xu, and Yongsheng Li. 2023. ImmCluster: an ensemble resource for immunology cell type clustering and annotations in normal and cancerous tissues. Nucleic Acids Research 51, D1 (2023), D1325–D1332. doi:10.1093/nar/gkac922 [21] Harold W. Kuhn. 1955. The Hungarian method for the assignment problem. Naval Research Logistics Quarterly 2, 1-2 (1955), 83–97. doi:10.1002/nav.3800020109 [22] Sillarine Kurkalang, Sumitava Roy, Arunima Acharya, et al. 2023. Single-cell transcriptomic analysis of gingivo-buccal oral cancer reveals two dominant cellular programs. Cancer Science 114, 12 (2023), 4732–4746. doi:10.1111/cas.15979 [23] Blue B. Lake, Rajasree Menon, Seth Winfree, et al. 2023. An atlas of healthy and injured cell states and niches in the human kidney. Nature 619, 7970 (2023),
KDD’27, 2027, San Jose, USA
585–594. doi:10.1038/s41586-023-05769-3 [24] Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. 2020. BioBERT: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics 36, 4 (2020), 1234–1240. doi:10.1093/bioinformatics/btz682 [25] Fangyu Liu, Ehsan Shareghi, Zaiqiao Meng, Marco Basaldella, and Nigel Collier. 2021. Self-Alignment Pretraining for Biomedical Entity Representations. In NAACL. [26] Xia Liu, Lingyun Li, Fusheng Si, Lan Huang, Yangjing Zhao, Chenchen Zhang, Daniel F. Hoft, and Guangyong Peng. 2021. NK and NKT cells have distinct properties and functions in cancer. Oncogene 40, 27 (2021), 4521–4537. doi:10. 1038/s41388-021-01880-9 [27] Wenjing Ma, Xiaoqing Yu, Jiaying Lu, Jing Zhang, Xiang Zhou, and Xuefeng Wang. 2025. ScPanKD: Distilling Pan-Cancer Knowledge for Enhanced T Cell Subtypes Annotation in Single-Cell Transcriptomics Data. In 2025 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 532–537. [28] David Masopust, Amit Awasthi, Remy Bosselut, et al. 2026. Guidelines for T cell nomenclature. Nature Reviews Immunology 26 (2026), 298–313. doi:10.1038/ s41577-025-01238-2 [29] Natalya F. Noy, Nigam H. Shah, Patricia L. Whetzel, Benjamin Dai, Michael Dorf, Nicholas Griffith, Clement Jonquet, Daniel L. Rubin, Margaret-Anne Storey, Christopher G. Chute, and Mark A. Musen. 2009. BioPortal: ontologies and integrated data resources at the click of a mouse. Nucleic Acids Research 37, suppl_2 (07 2009), W170–W173. doi:10.1093/nar/gkp440 [30] OpenAI. 2026. GPT-5.5 System Card. Technical Report. OpenAI. https: //deploymentsafety.openai.com/gpt-5-5/gpt-5-5.pdf Released April 23, 2026. [31] David Osumi-Sutherland, Chuan Xu, Maria Keays, Adam P Levine, Peter V Kharchenko, Aviv Regev, Ed Lein, and Sarah A Teichmann. 2021. Cell type ontologies of the Human Cell Atlas. Nature cell biology 23, 11 (2021), 1129–1135. [32] Aviv Regev, Sarah A. Teichmann, Eric S. Lander, et al. 2017. The Human Cell Atlas. eLife 6 (2017), e27041. doi:10.7554/eLife.27041 [33] Eric Sven Ristad and Peter N Yianilos. 1998. Learning string-edit distance. TPAMI (1998). [34] Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends® in Information Retrieval (2009). [35] Arpiar Saunders, Evan Z Macosko, Alec Wysoker, Melissa Goldman, Fenna M Krienen, Heather de Rivera, Elizabeth Bien, Matthew Baum, Laura Bortolin, Shuyu Wang, et al. 2018. Molecular diversity and specializations among the cells of the adult mouse brain. Cell 174, 4 (2018), 1015–1030. [36] Hergen Spits, David Artis, Marco Colonna, Andreas Diefenbach, James P Di Santo, Gerard Eberl, Shigeo Koyasu, Richard M Locksley, Andrew NJ McKenzie, Reina E Mebius, et al. 2013. Innate lymphoid cells—a proposal for uniform nomenclature. Nature reviews immunology 13, 2 (2013), 145–149. [37] Tabula Muris Consortium. 2018. Single-cell transcriptomics of 20 mouse organs creates a Tabula Muris. Nature 562, 7727 (2018), 367–372. doi:10.1038/s41586018-0590-4 [38] Shawn Z. K. Tan, Alba Puig-Barbe, David Goutte-Gattat, Benjamin S. Eastwood, Brian D. Aevermann, et al. 2026. The Cell Ontology in the age of single-cell omics. Scientific Data (2026). doi:10.1038/s41597-026-07173-8 [39] Xiaomi MiMo Team, Anqi Liu, Aoxin Ma, Bo Chen, Bo Yang, Chen Wang, Chen Zhang, Chengda Tang, Chengwei Wang, Chiheng Lou, Depeng Yan, Fuli Luo, Gang Wang, Hailin Zhang, Jiale Sun, Kang Zhou, Rui Huang, Shaohui Liu, Shen Huang, Shijie Cao, Shuaishuai Fan, Tianling Zhou, Xiangwei Deng, Xueyang Xie, Xuli Wang, Yingchun Lai, Yu Yang, Yuan Zhang, Zhen Tang, Zhonghua Deng, and Zihan Jiang. 2026. Full-Pipeline Inference Optimization for MiMo-V2.5 Series: Pushing Hybrid SWA Efficiency to the Limit. arXiv:2607.13095 [cs.AR] https://arxiv.org/abs/2607.13095 [40] Ignacio Vázquez-García, Florian Uhlitz, Nicholas Ceglia, et al. 2022. Ovarian cancer mutational processes drive site-specific immune evasion. Nature 612, 7941 (2022), 778–786. doi:10.1038/s41586-022-05496-1 [41] Nguyen Xuan Vinh, Julien Epps, and James Bailey. 2010. Information Theoretic Measures for Clusterings Comparison: Variants, Properties, Normalization and Correction for Chance. Journal of Machine Learning Research 11, 95 (2010), 2837–2854. https://www.jmlr.org/papers/v11/vinh10a.html [42] Sheng Wang, Angela Oliveira Pisco, Aaron McGeever, Maria Brbic, Marinka Zitnik, Spyros Darmanis, Jure Leskovec, Jim Karkanias, and Russ B. Altman. 2021. Leveraging the Cell Ontology to classify unseen cell types. Nature Communications 12 (2021), 5556. doi:10.1038/s41467-021-25725-x [43] Aaron J. Wilk, Arjun Rustagi, Nancy Q. Zhao, et al. 2020. A single-cell atlas of the peripheral immune response in patients with severe COVID-19. Nature Medicine 26, 7 (2020), 1070–1076. doi:10.1038/s41591-020-0944-y [44] Sunny Z. Wu, Ghamdan Al-Eryani, Daniel Lee Roden, et al. 2021. A single-cell and spatially resolved atlas of human breast cancers. Nature Genetics 53, 9 (2021), 1334–1347. doi:10.1038/s41588-021-00911-1 [45] Sunny Z. Wu, Daniel L. Roden, Chenfei Wang, et al. 2020. Stromal cell diversity associated with immune evasion in human triple-negative breast cancer. The EMBO Journal 39, 19 (2020). doi:10.15252/embj.2019104063
KDD’27, 2027, San Jose, USA
[46] Yuxuan Wu and Fuchou Tang. 2025. scExtract: leveraging large language models for fully automated single-cell RNA-seq data annotation and prior-informed multi-dataset integration. Genome Biology 26, 1 (2025), 174. doi:10.1186/s13059025-03639-x [47] Elliot Xie, Lingxin Cheng, Jack Shireman, Yujia Cai, Jihua Liu, et al. 2026. CASSIA: a multi-agent large language model for automated and interpretable cell annotation. Nature Communications 17 (2026), 389. doi:10.1038/s41467-025-67084-x [48] Anyi Xu, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chenchen Ling, et al. 2026. Deepseekv4: Towards highly efficient million-token context intelligence. arXiv preprint arXiv:2606.19348 (2026). [49] Nicholas D. Youngblut, Christopher Carpenter, Arshia Nayebnazar, Abhinav Adduri, Rohan Shah, Chiara Ricci-Tam, Jaanak Prashar, Rajesh Ilango, Noam Teyssier, Silvana Konermann, Patrick D. Hsu, Alexander Dobin, Dave P. Burke, Hani Goodarzi, and Yusuf H. Roohani. 2025. scBaseCount: an AI agent-curated, uniformly processed, and autonomously updated single cell data repository. bioRxiv (2025). doi:10.1101/2025.02.27.640494 [50] Rafael Yuste, Michael Hawrylycz, Nadia Aalling, Argel Aguilar-Valles, Detlev Arendt, Ruben Armañanzas, Giorgio A Ascoli, Concha Bielza, Vahid Bokharaie, Tobias Borgtoft Bergmann, et al. 2020. A community-based transcriptomics classification and nomenclature of neocortical cell types. Nature neuroscience 23, 12 (2020), 1456–1468. [51] Aohan Zeng, Xin Lv, Zhenyu Hou, Zhengxiao Du, Qinkai Zheng, Bin Chen, Da Yin, Chendi Ge, Chenghua Huang, Chengxing Xie, et al. 2026. Glm-5: from vibe coding to agentic engineering. arXiv preprint arXiv:2602.15763 (2026).
Xie et al.
TCellAlign : Cross-study T-cell Populations Alignment with Nomenclature-Guided Multi-Agent Workflow
A
Literature Retrieval and Single-Cell Data Acquisition
To evaluate the consistency of T-cell nomenclature across physiological and disease contexts, we constructed a benchmark linking scientific publications to publicly available single-cell datasets. Inspired by the large-scale single-cell data discovery and curation workflow of scBaseCount [49], we focused our data collection on four study contexts: Healthy, Cancer, Infectious, and Inflammatory. We first used a data-first strategy to construct the Healthy reference set. Starting from 361 publications indexed in CZ CELLxGENE and linked to DOIs, we filtered studies by organism, assay modality, health status, and the presence of T-cell annotations, yielding 65 candidate papers. Of these, 21 passed full-text screening, and the 20 most highly cited studies were retained. For the three disease contexts of Cancer, Infectious, and Inflammatory, we searched the Gene Expression Omnibus (GEO) using combinations of disease-specific, T-cell, and single-cell transcriptomics terms. We further screened the retrieved studies for paper relevance and duplicated DOIs, and matched eligible studies to their corresponding H5AD datasets in CZ CELLxGENE. Together with the Healthy reference set, 66 papers proceeded to manual review. We then manually reviewed the 66 in-scope papers and their matched H5AD data. We retained a study only when its H5AD data corresponded to the T-cell analysis reported in the paper and included an appropriate author-provided annotation field. We additionally required a verifiable correspondence between the paperreported labels and the dataset annotations. In total, 44 studies met these criteria and formed the final benchmark, comprising 12 Healthy, 13 Cancer, 13 Infectious, and 6 Inflammatory studies.
B
KDD’27, 2027, San Jose, USA
A paper should return true only when ALL of the following conditions are satisfied: 1. The paper analyzes T cells / T lymphocytes as part of its own biological results. 2. The paper describes one or more specific T-cell types, subtypes, states, populations, clusters, or functionally distinct T-cell groups. 3. The T-cell categories are described with paper-specific biological information, such as marker genes, marker proteins, functional programs, phenotypic states, activation or exhaustion features, tissue or tumor context, spatial localization, abundance changes, clinical associations, or experimental findings. 4. The T-cell information is specific enough that a structured extractor could list the T-cell categories and attach supporting evidence from the paper text. Return exactly one JSON object with exactly one key: {"paper_passes_screening": true} or {"paper_passes_screening": false} Full-paper Markdown: <paper_markdown> <FULL-PAPER MARKDOWN> </paper_markdown>
C.1.2 Single-Cell Data-Clue Extraction Prompt. Single-cell data-clue extraction
Information Extraction Evaluation Set
Stage 1 was evaluated on the seven papers listed in Table S3, with four domain experts independently annotating each paper for its reported T-cell population labels and supporting evidence. This evaluation set is separate from the 44-study benchmark used for nomenclature alignment, although four papers occur in both.
C
Prompt Templates Used in TCellAlign
The following templates present the instruction text used in the workflow. Angle-bracket fields denote paper- or case-specific content substituted at runtime and were not sent literally; implementation-only file labels are omitted.
C.1
Search Agent Prompts
C.1.1 Paper-Level Screening Prompt. Paper-level screening [System] Return valid JSON only. [User] You are a paper-level screening classifier. Your task is to decide whether the input paper contains extractable, paper-specific T-cell type, subtype, state, or population content.
[System] Return valid JSON only. [User] You are a Stage 0 scRNA-seq h5ad data clue extractor. Your task is to read the full-paper Markdown and extract only the clues needed for code to verify whether this paper has usable scRNA-seq h5ad data through CELLxGENE. Do not decide whether the paper passes Stage 0. Only extract verification clues. Extract paper identity when available. These fields are verification clues: code will use them to search CELLxGENE metadata even when the paper does not explicitly mention CELLxGENE. - title - DOI - PMID - journal - year Extract general scRNA-seq evidence from the paper whenever the paper describes scRNA-seq or single-cell transcriptomic RNA expression data, even if the paper does not explicitly mention CELLxGENE: - assay hint, always "scRNA-seq" - relation to this paper - short evidence text copied or closely paraphrased from the paper Do not extract GEO, SRA, ENA, EGA, FASTQ, count matrix, or other rawdata repository clues as passing evidence. Raw data availability is not enough for this Stage 0 condition.
KDD’27, 2027, San Jose, USA
Extract CELLxGENE clues only when the CELLxGENE collection or dataset is described as containing scRNA-seq or single-cell transcriptomic RNA expression data: - CELLxGENE URLs - collection names - dataset names - collection DOI or dataset DOI - assay hint, always "scRNA-seq" - relation to this paper
Xie et al.
C.2
Information Extraction Agent Prompt
C.2.1 T-Cell Label Extraction Prompt. T-cell label extraction [System] Return valid JSON only. Do not include markdown fences.
If a paper is multi-modal, extract the clue only when it explicitly includes a scRNA-seq or single-cell RNA expression component. Do not extract clues for records that are only snRNA-seq, spatial transcriptomics, CITE-seq, bulk RNA-seq, ATAC-seq, proteomics, imaging, or other non-scRNA-seq assays.
[User] You are an immunology literature extraction assistant. Read the provided PDF and extract T cell subtypes / T cell states / T cell populations that are explicitly supported by evidence in the article.
Use this allowed value for assay_hint: - scRNA-seq
Core extraction principle: The output unit is the biological T cell subtype/state/population, not the individual cluster.
Use these allowed values for relation_to_paper: - generated_by_this_paper - reused_external_data - unclear Use "generated_by_this_paper" when the paper appears to have generated or directly analyzed the dataset. Use "reused_external_data" when the paper only reuses an external scRNAseq dataset. Use "unclear" only when the relation cannot be determined from the Markdown. If no scRNA-seq evidence is found anywhere in the paper, return " scrnaseq_evidence": []. If no scRNA-seq CELLxGENE clues are found, return "cellxgene_clues": []. Return "mentioned_repositories" only for CELLxGENE mentions tied to scRNA-seq h5ad data. If no scRNA-seq CELLxGENE repository is mentioned, return " mentioned_repositories": []. Return exactly one JSON object using this schema: { "paper_identity": { "title": null, "doi": null, "pmid": null, "journal": null, "year": null }, "scrnaseq_evidence": [ { "assay_hint": "scRNA-seq", "relation_to_paper": "generated_by_this_paper", "evidence_text": "The scRNA-seq data from this study were deposited in repository X." } ], "cellxgene_clues": [ { "url": null, "collection_name": null, "dataset_name": null, "doi": null, "assay_hint": "scRNA-seq", "relation_to_paper": "generated_by_this_paper" } ], "mentioned_repositories": ["CELLxGENE"] } Full-paper Markdown: <paper_markdown> <FULL-PAPER MARKDOWN> </paper_markdown>
Extraction criteria: Each final item must satisfy all of the following criteria: 1. It is a T cell-related subtype, state, or population. 2. It explicitly appears in the paper or is explicitly annotated/defined by the authors. 3. It has supporting evidence: gene expression or descriptive words. 4. If both a broad parent label and a more specific biological subtype/ state are present, keep the more specific supported subtype/state/ population as the final item. Original label wording rule: - original_label must be a complete, self-contained biological T cell subtype/state/population label. - Keep original_label as close as possible to the paper's original wording. - Prefer a complete biological label over a local shorthand, abbreviation-only label, cluster-only label, or subcluster-only label. - A local shorthand label is incomplete if it only contains an abbreviation, cluster ID, subcluster ID, or state ID, but its biological meaning is defined elsewhere in the paper. - If the paper defines a local shorthand, abbreviation, cluster ID, or subcluster ID under a broader biological label, use the broader biological label to construct a complete original_label, and preserve the shorthand/ID in parentheses. - Use only biological meanings stated or clearly defined in the paper. - Do not expand abbreviations using external immunology knowledge. - Do not output bare local shorthand labels as final original_label when the paper provides the biological meaning elsewhere. - When merging several equivalent mentions, choose the clearest representative wording from the paper. - If multiple cluster IDs belong to the same biological subtype/state/ population, include the merged cluster IDs in original_label when needed to preserve the paper's context. - Do not normalize the label into an external ontology or rewrite it into a standard name not used by the paper. Merge rule: - Merge mentions that refer to the same biological T cell subtype/state/ population. - Do not create separate final items only because the same biological subtype/state/population appears in multiple clusters, samples, figures, sections, or analysis views. - Keep cluster/subcluster/state IDs inside original_label only when they are part of the paper's wording or needed to preserve context. - If two labels have different biological meanings in the paper, keep them as separate final items. Evidence rule: - Each evidence item must be gene expression or descriptive words. - Keep every evidence item strictly in the paper's original wording. - Do not rewrite, paraphrase, translate, summarize, or normalize evidence. - Evidence should be copied as short original phrases from the article. - The same original_label may have multiple evidence items.
TCellAlign : Cross-study T-cell Populations Alignment with Nomenclature-Guided Multi-Agent Workflow
- If the same cell subtype/state/population is supported by multiple distinct evidence phrases, include all relevant evidence phrases in the same evidence array. Exclusion rules: - Do not extract non-T cell types. - Do not extract labels that appear without gene expression or descriptive words. - Do not invent subtypes from general immunology knowledge; use only information from the paper. - Do not treat marker genes themselves as cell subtypes. - Do not include T cell names that are mentioned only in background, introduction, references, or general discussion without being part of the paper's own results or annotations. Pre-output label check: Before returning JSON, review every original_label. For each item: 1. Ask whether the label is only a local shorthand, abbreviation, cluster ID, subcluster ID, or state ID. 2. If yes, look for the biological label that the paper uses to define that shorthand. 3. If a biological label is available in the paper, rewrite original_label as a complete, self-contained, paper-faithful biological label that includes the shorthand/ID in parentheses. 4. If no biological label is available in the paper, keep the paper's local shorthand label unchanged. 5. Never use external knowledge to expand or rename the label. Output requirements: Return valid JSON only. Do not output markdown or explanations. JSON schema: { "title": "", "items": [ { "original_label": "A complete, self-contained, paper-faithful biological T cell subtype/state/population label. Include paper abbreviation, cluster ID, subcluster ID, or state ID in parentheses when needed.", "evidence": [ "Exact original gene expression or descriptive words from the paper", "Another exact original evidence phrase for the same biological cell subtype/state/population, if available" ] } ] } Stage 0 paper markdown input: Use only the supplied full-paper markdown below. SUPPLIED TEXT START <FULL-PAPER MARKDOWN> SUPPLIED TEXT END
C.3
KDD’27, 2027, San Jose, USA
Do not return mapped_label. The final mapped_label will be derived by code from mapped_slots. For each supported slot, return: - module: the final naming module inferred from the reference. - semantic_value: a concise snake_case semantic key describing the biological meaning of that module, inferred from the reference. If the evidence does not support a slot, return: {"module": null, "semantic_value": null} Do not invent unsupported biological properties. Do not fill a slot only because it is familiar from the label name. Prefer less specific correct slots over more specific unsupported slots. Use the same semantic_value string consistently for the same biological meaning across cases. REFERENCE: <<< <T-CELL NOMENCLATURE REFERENCE> >>> CASES: <MAPPING CASES JSON> Output schema: [ { "case_id": 0, "mapped_slots": { "lineage": { "module": "CD4+", "semantic_value": "cd4_conventional" }, "function": { "module": null, "semantic_value": null }, "migration": { "module": null, "semantic_value": null }, "migration_subscript": { "module": null, "semantic_value": null }, "differentiation": { "module": null, "semantic_value": null }, "differentiation_subscript": { "module": null, "semantic_value": null }, "antigen": { "module": null, "semantic_value": null } } } ]
Mapping Agent Prompt
C.3.1 Nomenclature-Guided Mapping Prompt. Nomenclature-guided mapping [User] Use the reference to complete structured T cell nomenclature slot filling for the following T cell cases.
C.3.2 T-Cell Nomenclature Reference. The following reference text was supplied to the Mapping Agent through the reference field shown above. # T Cell Type Nomenclature Rule ## 1. Core Principle
Return only valid JSON. Do not include explanations. Return one output object per input case. Do not copy original_label or evidence into the output.
Use a **modular T cell name** when a paper describes a T cell type, state, subset, or population.
KDD’27, 2027, San Jose, USA
The name should encode only the biological properties that are actually stated or reasonably supported by the paper: `[lineage] T[function][migration][migration-subscript][differentiationstate][state-subscript][antigen-status]` If a property is not stated or not measured, do **not** infer it from a familiar subset label. ## 2. Core Naming Syntax Use this order: 1. Lineage 2. `T` root 3. Functional module 4. Migration module 5. Migration subscript 6. Differentiation-state module 7. Differentiation-state subscript 8. Antigen-status suffix 9. No terminal standalone word `cell` after the modular code
Xie et al.
| `TCTL` | Cytotoxic T cell function | | `TC2`, `TC17`, `TC22` | CD8 T cells with helper-like cytokine/function programs | Rules: - Use functional labels only when the paper states functional identity, transcription-factor evidence, cytokine evidence, or an explicit author claim. - Keep entrenched `TH` and `Treg` labels; do not erase them merely because lineage is already `CD4+`. - Use `TCTL` for cytotoxic function when cytotoxicity is the intended property. Do not use `C` alone to mean cytotoxic, because `C` is historically ambiguous. - Avoid introducing `C` or `E` as new modular shorthand for central/ effector migration categories, because they are easily confused with `TCM` and `TEM`. ## 5. Migration Module Migration is encoded separately from memory, activation, exhaustion, or function.
## 3. Lineage Module
| Module | Meaning | Typical evidence/proxy | |---|---|---| | `S` | Can migrate from blood into uninflamed secondary lymphoid organs | CD62L+ and/or CCR7+ | | `D` | Disseminated; does not tend to enter uninflamed secondary lymphoid organs from blood | CD62L- and/or CCR7- | | `U` | Migration properties unknown | No migration assay or no homingreceptor evidence |
Lineage is the broad, relatively stable cell identity.
Migration subscripts:
Allowed lineage labels include:
| Subscript | Meaning | |---|---| | `B` | Isolated from blood; no further migration claim | | `R` | Resident; parked within an organ or restricted vascular compartment | | `W` | Widespread; recirculates through non-lymphoid tissues |
For plain-text LLM input, serialize subscripts inline in the canonical order. Do not reorder modules for readability. Keep function before migration, migration before differentiation state, and antigen status last.
| Module | Meaning | |---|---| | `CD4+` | Conventional CD4-positive T cell | | `CD8+` | Conventional CD8-positive T cell | | `gamma-delta+` or `gamma-delta T` | gamma-delta T cell lineage | | `NKT` / `iNKT` | Natural killer T cell / invariant NKT cell | | `MAIT` | Mucosal-associated invariant T cell | | `CD8aa+ TCRab` | CD8 alpha-alpha-positive TCR alpha-beta T cell | | `DN T` | Double-negative T cell | | `DP T` | Extrathymic double-positive T cell | Rule: - Treat stable lineage identities such as `CD4+`, `CD8+`, `gamma-delta T `, `MAIT`, and `iNKT` as cell-type-like labels. - Do not convert transient activation, memory, exhaustion, or tissuelocation descriptions into lineage labels.
Rules: - Use `S`, `D`, or `U` only for migration behavior or its accepted proxy markers. - Use `D` for CD62L- and/or CCR7- cells when the paper only supports failure to enter uninflamed SLOs. - Do not call a cell `TRM`, resident, or `R` merely because it was isolated from a tissue. Tissue source alone is not residence. - If a cell is from blood and no further migration is known, use `B`. - If migration is unknown but blood origin is known, use `UB`, not `DB`. ## 6. Differentiation-State Module
## 4. Functional Module
Differentiation state follows migration.
Functional modules follow the `T` root and precede migration/state modules.
| Module | Meaning | |---|---| | `N` | Naive | | `A` | Activated | | `M` | Memory | | `X` | Exhausted | | `G` | Anergic |
Common functional modules: | Module | Meaning | |---|---| | `TH1` | CD4 T cell biased toward IFN-gamma-associated type 1 helper function | | `TH2` | CD4 T cell biased toward IL-4/IL-5/IL-13-associated type 2 helper function | | `TH9` | CD4 T cell biased toward IL-9-associated helper function | | `TH17` | CD4 T cell biased toward IL-17-associated helper function | | `TFH` | Follicular helper T cell function | | `Treg` | Regulatory T cell function | | `tTreg` | Thymus-derived regulatory T cell | | `pTreg` | Peripherally derived regulatory T cell | | `eTreg` | Effector regulatory T cell | | `iTreg` | In vitro-induced regulatory T cell | | `TFR` | Follicular regulatory T cell |
State subscripts: | Subscript | Meaning | |---|---| | `p` | Progenitor / precursor | | `t` | Terminal | Rules: - Use `N` for naive T cells. - Use `A` for recently activated T cells. - Use `M` for memory T cells.
TCellAlign : Cross-study T-cell Populations Alignment with Nomenclature-Guided Multi-Agent Workflow
- Use `X` for exhausted T cells. - Use `G` for anergic T cells. - Use `p` or `t` only when progenitor/precursor or terminal status is stated or supported. - Map current labels as follows when evidence supports the mapping: - `TPEX` -> `TXp` - terminal exhausted T cell / `TEX-term` -> `TXt` - stem cell memory T cell / `TSCM` -> `TMp` - memory precursor effector cell / `MPEC` -> `TAp` - short-lived effector cell / `SLEC` -> `TAt` - Do not use `effector` as a precise state if the paper only says effector-like molecules are expressed. Prefer `A`, `M`, or leave the state unspecified depending on evidence. ## 7. Antigen-Status Suffix Antigen status is optional and attaches after the differentiation-state module. | Suffix | Meaning | |---|---| | `+` | Cognate antigen is claimed to persist in the organism | | `0` | Cognate antigen is claimed cleared or irrelevant to the described cell | | no suffix | Antigen status not stated or not needed | Rules: - Use `+` only for explicit or strongly supported persistent-antigen contexts, such as chronic infection or tumor antigen persistence. - Use `0` only when antigen clearance or irrelevance is stated or reasonably claimed. - Do not infer antigen status from `memory`, `exhausted`, `TRM`, or tissue location alone. - Plain `M` does not imply antigen cleared. ## 8. Recommended Combination Order
KDD’27, 2027, San Jose, USA
- `Effector T cell` is semantically broad. Map to `A`, `At`, cytotoxic function, or leave the state unspecified if the paper does not give enough evidence. - `Memory T cell` does not itself imply antigen clearance. - `Exhausted T cell` usually implies chronic antigen context, but still use `+` only when persistence is claimed or contextually supported . - Marker-only descriptions should be converted only as far as the markers support. If no claim is made, keep markers in evidence/ context. - If migration markers are absent, use `U` only when the naming task requires migration to be represented; otherwise omit migration. - If activation recency is unknown, do not add `A`. - If residence, recirculation, antigen persistence, or terminal/ progenitor status is not stated, omit those modules. ## 11. examples Use these source-paper examples as examples of how method-defined criteria support modular names. ### Example 1 Results text: ```text We compared populations of naive CD4+ T cells (CD4+ TN) with activated CD4+ T cells (CD4+ TA) and memory CD4+ T cells (CD4+ TM). ``` Methods criteria: ```text CD4+ TN were CD44low, CD62L+. CD4+ TA were CD44hi, CD62L-, and known to have been stimulated by antigen within the last week. CD4+ TM were CD44hi and the specific pathogen is thought to have been cleared at least 1 month before analysis. ```
Canonical order: ### Example 2 `[lineage] T[function][S/D/U][B/R/W][N/A/M/X/G][p/t][+/0]` Results text: ## 9. Content to Preserve as Evidence/Context, Not Force Into the Name Keep the following outside the mapped label unless a defined module exists and the paper makes the relevant claim: - Tissue or anatomical source: `lung`, `liver`, `blood`, `tumor`, `lymph node` - Species: human, mouse, non-human primate - Disease or perturbation context - Antigen specificity, clone, tetramer identity, or TCR specificity - Marker evidence used to justify a module - Cytokines, transcription factors, and effector molecules when they are evidence rather than the intended functional label - Proliferation potential - Differentiation potential beyond `p` or `t` - Longevity - Developmental plasticity - Epigenetic profile - Assay type or gating strategy - Uncertainty notes such as `claimed`, `not measured`, `proxy only` ## 10. Boundaries: Do Not Over-Standardize Do not force a familiar subset label into a stronger modular claim than the paper supports. Rules: - `TEM` does not automatically mean blood-derived, tissue-recirculating, resident, terminal, cytotoxic, or antigen-cleared. - `TRM` requires a residence claim; tissue isolation alone is insufficient. - `TEMRA` is human-specific and should not be projected onto mouse data.
```text We studied influenza virus-specific disseminated resident memory CD8+ T cells (CD8+ TDRM) in the mediastinal lymph node. ``` Methods criteria: ```text We used a CD69+, CD62L- flow cytometry phenotype to define CD8+ TDRM. Influenza-specific CD69+, CD62L- CD8+ T cells were previously shown to be resident based on minimal equilibration in parabionts. ``` ### Example 3 Results text: ```text We profiled exhausted progenitor CD8+ T cells (TXp) by single-cell RNA sequencing. ``` Methods criteria: ```text All CD8+ T cells within a cluster that was distinguished by elevated expression of TCF7, PDCD1 and CXCR5 were defined as TXp. ``` ## 12. Decision Rule When normalizing T cell population names:
KDD’27, 2027, San Jose, USA
Xie et al.
"lineage": {"verdict": "yes", "error_type": null, "reason": null}, "function": {"verdict": "yes", "error_type": null, "reason": null}, "migration": {"verdict": "yes", "error_type": null, "reason": null}, "migration_subscript": {"verdict": "yes", "error_type": null, " reason": null}, "differentiation": {"verdict": "yes", "error_type": null, "reason": null}, "differentiation_subscript": {"verdict": "yes", "error_type": null, "reason": null}, "antigen": {"verdict": "yes", "error_type": null, "reason": null}
1. Preserve the original mention. 2. Extract only stated or supported biological properties. 3. Build the modular name in canonical order. 4. Store unsupported implications as uncertainty, not as modules. 5. Omit a final generic `cell` suffix from modular mapped labels. 6. Prefer a less specific correct name over a more specific overinferred name.
C.4
}
Judge Agent Prompt
}
C.4.1 Slot-Level Adjudication Prompt. Slot-level adjudication
NOMENCLATURE RULE: <<< <T-CELL NOMENCLATURE REFERENCE> >>>
[System] Return only valid JSON matching the requested schema. Do not include markdown fences or any text outside the JSON object. [User] Judge whether every structured slot in the candidate T cell mapping is valid under the provided nomenclature rule and paper-local evidence. Evaluate all seven slots, including slots whose module and semantic_value are null: - lineage - function - migration - migration_subscript - differentiation - differentiation_subscript - antigen A null slot is correct when neither the original label nor the supplied evidence supports a value for that slot. Do not use insufficient_evidence merely because a slot is null; use it only when the supplied evidence is genuinely ambiguous or inadequate to verify a biologically relevant claim. For each slot, return exactly these fields: - verdict: "yes" or "no" - error_type: null when verdict is "yes"; otherwise exactly one allowed error type - reason: null when verdict is "yes"; otherwise one concise explanation grounded in the rule and evidence Allowed error types: - unsupported_value: the mapped slot adds a biological property not supported by the original label or evidence - missing_value: the evidence supports a property that the mapped slot omits - incorrect_normalization: the biological meaning is supported, but module or semantic_value does not follow the rule - wrong_slot: supported information is assigned to the wrong slot - insufficient_evidence: the available paper-local evidence is insufficient to verify the slot
MAPPING RECORD: <MAPPING RECORD JSON>
D
Paper Extracted Label to Data Annotation Prompt Annotation Matching
You are matching PDF-extracted cell type names to data annotation labels from the same paper. MATCHING RULES, in priority order: 1. CLUSTER ID (authoritative). If the extracted label and a candidate both contain a cluster identifier (e.g. "c0", "c11", "cluster 3"), match on that ID alone. It overrides all semantic similarity. Never match differing IDs. 2. LINEAGE CONSTRAINT (hard block). If the extracted label specifies a lineage(CD4, CD8, Treg, TFH, NKT, MAIT, gamma-delta) and no candidate shares it, return NO_MATCH. Never map across a specified lineage. 3. MARKER EVIDENCE. Candidate labels often embed markers (CCR7, FOXP3, GZMK, MKI67). Match these against the extracted label's text and evidence. 4. SEMANTIC CORRESPONDENCE. Prefer specific biological correspondence over loose family-level similarity. Choose EXACTLY ONE candidate, or NO_MATCH. Copy the chosen candidate label EXACTLY as written. NO_MATCH is a valid, expected answer. Return NO_MATCH when: - no candidate plausibly refers to the same population, or - the only candidates conflict on lineage (rule 2), or - the population was likely filtered out of this data object. Do NOT force a match. A wrong match is worse than NO_MATCH.
When information is assigned to the wrong slot, mark that populated slot as wrong_slot and mark the slot that should contain the information as missing_value. This ensures both affected slots are identifiable.
WHEN THE EXTRACTED LABEL SPANS SEVERAL CANDIDATES (e.g. "proliferating T cells" vs four cycling subclusters), still choose ONE: the closest biological representative, preferring (a) shared cluster ID, then (b) the candidate whose markers appear in the evidence, then (c) the largest/most canonical member of that program. Set " spans_multiple": true and name the others in "reason".
Do not generate revised modules, revised semantic values, a corrected mapped label, or any correction instructions. Judge only the supplied mapping.
Ground every decision in the extracted label text and its evidence.
Return only one valid JSON object with the exact top-level field " slot_verdicts". Return every slot exactly once and do not add fields. Output schema: { "slot_verdicts": {
Return JSON only: {"items":[{"original_label": "...", "matched_data_label": "..." | null, "match_status": "cluster_id" | "marker" | "semantic" | " no_match", "spans_multiple": true | false, "confidence": "high" | "medium" | "low", "reason": "short, concrete"}]}
TCellAlign : Cross-study T-cell Populations Alignment with Nomenclature-Guided Multi-Agent Workflow
E
Technical Details of Compared Baselines
We compare TCellAlign with existing cross-study cell population alignment approaches. BM25. BM25 (Best Matching 25) [34] is a ranking function used in information retrieval to determine how relevant a document is to a given query. We cast cell type normalization as retrieval over the Cell Ontology. We treat each paper-local label as a query and each CL surface form as a document. Every label is scored against all CL surface forms and the top-ranked term is returned as its normalized label. Levenshtein. Levenshtein distance [33] measures the difference between two strings as the minimum number of character edits required to transform one string into the other. We compute the distance between each extracted paper-local labels and every CL surface form, and convert it into a length-normalized similarity to ensure candidate names of different lengths remain comparable. The CL term with highest similarity is returned as its normalized label. PubMedBERT. PubMedBERT [14] is a domain-specific language model developed by Microsoft Research pretrained on biomedical text from PubMed abstracts and full-text articles. We use it as a frozen encoder without task-specific fine-tuning. Each extracted paper-local label and each CL label is encoded into a fixed-length embedding, and every label is assigned to the CL term with the highest cosine similarity in the shared embedding space. SapBERT. SapBERT [25] is a specialized language model that links different names of the same medical concept by aligning their vector representations in the embedding space. It is initialized from PubMedBERT and further pretrained on synonym pairs drawn from the UMLS. Similar to PubMedBERT, we use SapBERT to encode paper-local labels and CL surface forms and assign each label to the CL term with the highest cosine similarity. The two baselines therefore differ only in this additional self-alignment pretraining, and their comparison isolates its effect on cell type normalization. ZOOMA. ZOOMA [10] is an ontology annotation service developed and maintained by the European Bioinformatics Institute (EMBL-EBI). It integrates dictionary-based lookup, curated annotation resources, ontology synonyms, and semantic matching to retrieve candidate ontology terms for a given biological entity. We apply ZOOMA to the extracted paper-local labels and configure the target ontologies to include Cell Ontology, Cell Line Ontology, Human Cell Atlas Ontology, and Provisional Cell Ontology. BioPortal. BioPortal [29] provides ontology-aware search across a large collection of biomedical ontologies. Given an extracted paperlocal label, BioPortal retrieves candidate ontology concepts based on lexical similarity and ontology metadata. We query BioPortal against multiple cell-related ontologies, including Cell Ontology, Cell Line Ontology, Placental Cell Type Ontology, Brain Region & Cell Type Terminology, and Breast Tissue Cell Lines Ontology.
F Technical Details of Evaluation Metrics F.1 Annotation-based Evaluation Given the predicted normalization mapping 𝑓pred : 𝑂 → 𝑀, where 𝑀 = {𝑚 1, 𝑚 2, . . . , 𝑚𝑛 } denotes the normalized labels generated by a mapping method, and the validated reference mapping 𝑓ref : 𝑂 → 𝐶𝐿 obtained from the standardized Cell Ontology annotations in
KDD’27, 2027, San Jose, USA
CZ CELLxGENE, we propagate both mappings through the shared paper-local labels. Consequently, each paper-local label is associated with both a reference Cell Ontology label 𝐶𝐿 and a predicted normalized label 𝑀. Rather than evaluating exact string matching, we assess whether the predicted normalization induces a partition that is consistent with the reference Cell Ontology annotations. Adjusted Mutual Information (AMI). We first compute the Adjusted Mutual Information (AMI) [41] between the predicted labels 𝑀 and the reference Cell Ontology labels 𝐶𝐿. Let 𝐼 (𝐶𝐿, 𝑀) denote the mutual information between the two partitions. AMI is defined as 𝐼 (𝐶𝐿, 𝑀) − E[𝐼 (𝐶𝐿, 𝑀)] AMI(𝐶𝐿, 𝑀) = , (4) max{𝐻 (𝐶𝐿), 𝐻 (𝑀)} − E[𝐼 (𝐶𝐿, 𝑀)] where 𝐻 (·) denotes entropy and E[𝐼 (𝐶𝐿, 𝑀)] is the expected mutual information under random assignments. AMI can take negative values when agreement is below chance, with larger values indicating stronger agreement after correcting for chance. Merge F1. While AMI evaluates the overall agreement between two partitions, it does not directly assess whether a normalization method makes biologically correct merge decisions. We therefore propose to evaluate the correctness of merge operations by considering every pair of paper-local labels. For every pair (𝑜𝑖 , 𝑜 𝑗 ), we define ( 1, 𝑓pred (𝑜𝑖 ) = 𝑓pred (𝑜 𝑗 ), 𝛿 𝑀 (𝑜𝑖 , 𝑜 𝑗 ) = (5) 0, otherwise, , and similarly ( 𝛿𝐶𝐿 (𝑜𝑖 , 𝑜 𝑗 ) =
1, 0,
𝑓ref (𝑜𝑖 ) = 𝑓ref (𝑜 𝑗 ), otherwise.
(6)
Based on these pairwise relationships, we define • True Positive (TP): 𝛿 𝑀 = 1 and 𝛿𝐶𝐿 = 1; • False Positive (FP): 𝛿 𝑀 = 1 but 𝛿𝐶𝐿 = 0; • False Negative (FN): 𝛿 𝑀 = 0 but 𝛿𝐶𝐿 = 1. Precision, Recall, and Merge F1 are computed as 𝑇𝑃 , 𝑇𝑃 + 𝐹𝑃 𝑇𝑃 Recall = , 𝑇𝑃 + 𝐹𝑁 2 · Precision · Recall Merge F1 = . Precision + Recall Precision =
(7a) (7b) (7c)
Unlike AMI, Merge F1 directly evaluates whether a normalization method correctly merges biologically equivalent paper-local labels while avoiding incorrect merges across distinct Cell Ontology classes.
F.2
Expression-based Evaluation.
Although the annotation-based metrics quantify semantic agreement with Cell Ontology, they do not evaluate whether the merged labels correspond to transcriptionally coherent cellular populations. We therefore perform a secondary evaluation using gene expression embeddings. For each predicted normalized label 𝑚, suppose it contains 𝑘 author-provided annotations, 𝐴(𝑚) = {𝑎 1, 𝑎 2, . . . , 𝑎𝑘 }.
(8)
KDD’27, 2027, San Jose, USA
Xie et al.
For each annotation 𝑎𝑖 , we compute its centroid in the embedding space, 1 ∑︁ 𝑥. (9) 𝑐𝑖 = |𝑎𝑖 | 𝑥 ∈𝑎 𝑖
The biological consistency of the merged label is quantified by the average pairwise cosine similarity among these annotation centroids, ∑︁ 2 MC(𝑚) = cos(𝑐𝑖 , 𝑐 𝑗 ), (10) 𝑘 (𝑘 − 1) 𝑖< 𝑗 where cos(·, ·) denotes cosine similarity. Finally, the overall Merge Consistency score is computed as 1 ∑︁ MergeConsistency = MC(𝑚), (11) |M| 𝑚∈ M
where M denotes the set of normalized labels that merge two or more author-provided annotations. Higher Merge Consistency indicates that the author-provided annotations merged by a normalization method exhibit high transcriptomic similarity, providing biological evidence that the merge is plausible.
F.3
Soft Bipartite Matching F1 Score
SBMS-F1 proceeds as follows. Let the gold-standard cell type set be 𝑂 = 𝑜 1, 𝑜 2, . . . , 𝑜𝑛 , and the human or model extracted cell type set be 𝑃 = 𝑝 1, 𝑝 2, . . . , 𝑝𝑚 . Each cell type surface text is first embedded using Qwen3-Embedding-8B. A pairwise cosine similarity matrix S ∈ [0, 1] 𝑛×𝑚 is constructed with entries 𝑃 𝑠𝑖 𝑗 = cos(e𝑂 𝑖 , e 𝑗 ),
(12)
𝑃 where e𝑂 𝑖 and e 𝑗 are the embedding vectors of the i-th gold term and the j-th predicted term, respectively. Next, the Hungarian algorithm [21] is applied to find the optimal one-to-one assignment that maximizes the total similarity, i.e., 𝑛 ∑︁ 𝜋 ∗ = arg max 𝑠𝑖,𝜋 (𝑖 ) , (13) 𝜋 𝑖
where 𝜋 is a matching that pairs each gold term with at most one predicted term. This transforms the evaluation into a bipartite graph matching problem. Finally, the summed similarity of the matched pairs is normalized to obtain precision, recall, and the harmonic mean F1 score: 𝑛 1 ∑︁ SBMS-P = 𝑠𝑖,𝜋 ∗ (𝑖 ) , (14a) 𝑚 𝑖 𝑛
SBMS-R =
1 ∑︁ 𝑠𝑖,𝜋 ∗ (𝑖 ) , 𝑛 𝑖
(14b)
2 · SBMS-P · SBMS-R , (14c) SBMS-P + SBMS-R with the convention SBMS-P = 0 if 𝑚 = 0, and SBMS-R = 0 if 𝑛 = 0. This design allows SBMS-F1 to reward semantically correct but lexically different outputs (e.g., “natural killer T cells” vs. “NKT-like cells”) while penalizing missing or spurious extractions. The metric is inherently soft, bounded between 0 and 1, and reduces to the conventional exact-match 𝐹 1 when all matched pairs are identical (i.e., when every 𝑠𝑖 𝑗 = 1 for the assigned pairs). SBMS-F1 =
TCellAlign : Cross-study T-cell Populations Alignment with Nomenclature-Guided Multi-Agent Workflow
KDD’27, 2027, San Jose, USA
Table S1: Per-study data scale and T-cell label counts across the final 44-paper analysis set. TCellAlign-GPT No. Paper title 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44
A single-cell and spatially resolved atlas of human breast cancers. Spatially organized multicellular immune hubs in human colorectal cancer. Single-cell transcriptomics of human T cells reveals tissue and activation signatures in health and disease. Characteristics of anti-CD19 CAR T cell infusion products associated with efficacy and toxicity in patients with large B cell lymphomas. High-resolution single-cell atlas reveals diversity and plasticity of tissue-resident neutrophils in non-small cell lung cancer. Signatures of plasticity, metastasis, and immunosuppression in an atlas of human small cell lung cancer. Stromal cell diversity associated with immune evasion in human triple-negative breast cancer. Single-cell sequencing links multiregional immune landscapes and tissue-resident T cells in ccRCC to tumor topology and therapy efficacy. Ovarian cancer mutational processes drive site-specific immune evasion. Mapping single-cell transcriptomes in the intra-tumoral and associated territories of kidney cancer. Single-cell proteo-genomic reference maps of the hematopoietic system enable the purification and massive profiling of precisely defined cell states. A single-cell atlas enables mapping of homeostatic cellular shifts in the adult human breast. Follicular Lymphoma Microenvironment Characteristics Associated with Tumor Cell Mutations and MHC Class II Expression. Single cell RNA sequencing of human liver reveals distinct intrahepatic macrophage populations Cross-tissue immune cell analysis reveals tissue-specific features in humans Distinct microbial and immune niches of the human colon A spatially resolved single-cell genomic atlas of the adult human breast Human skeletal muscle aging atlas Blood and immune development in human fetal bone marrow and Down syndrome Asian diversity in human immune cells. A spatial human thymus cell atlas mapped to a continuous tissue axis Multi-omic profiling reveals age-related immune dynamics in healthy adults Single-cell and spatial mapping identify cell types and signaling networks in the human ureter Multimodal profiling reveals tissue-directed signatures of human immune cells altered with age Cellular heterogeneity and dynamics of the human uterus in healthy premenopausal women A single-cell atlas of the peripheral immune response in patients with severe COVID-19. Immunophenotyping of COVID-19 and influenza highlights the role of type I interferons in development of severe COVID-19. COVID-19 immune features revealed by a large-scale single-cell transcriptome atlas. Single-cell multi-omics analysis of the immune response in COVID-19. Local and systemic responses to SARS-CoV-2 infection in children and adults. A blood atlas of COVID-19 defines hallmarks of disease severity and specificity. Time-resolved systems immunology reveals a late juncture linked to fatal COVID-19. Type I interferon autoantibodies are associated with systemic immune alterations in patients with COVID-19. Impaired local intrinsic immunity to SARS-CoV-2 infection in severe COVID-19. Early human lung immune cell development and its role in epithelial cell fate. Interstitial macrophages are a focus of viral takeover and inflammation in COVID-19 initiation in human lung. An interactive single cell web portal identifies gene and cell networks in COVID-19 host responses. mRNA COVID-19 vaccine elicits potent adaptive immune response without the acute inflammation of SARS-CoV-2 infection. Intra- and Inter-cellular Rewiring of the Human Colon during Ulcerative Colitis. Cells of the human intestinal tract mapped across space and time. Single-cell RNA-seq reveals cell type-specific molecular and genetic associations to lupus. The landscape of immune dysregulation in Crohn’s disease revealed through single-cell transcriptomic profiling in the ileum and colon. Single-cell integration reveals metaplasia in inflammatory gut diseases. Single-cell and spatially resolved interactomics of tooth-associated keratinocytes in periodontitis.
Category
TCellAlign-DS
Cells Annotated cell (𝑛 ) types (𝑛 ) Extracted Mapped Extracted Mapped
Cancer Cancer Cancer Cancer
35,214 76,965 51,876 133,405
Cancer
230,697
Cancer Cancer Cancer
46,140 7,990 73,806
Cancer Cancer Cancer
221,315 141,106 4,048
Cancer Cancer
12 26 6 15
10 9 18 6
8 6 14 5
10 7 10 3
8 3 10 3
9
8
7
8
7
6 7 10
6 8 10
6 7 8
6 6 10
6 6 10
41 26 9
18 16 11
13 12 10
18 9 10
12 7 5
17,483 35,483
6 11
9 11
7 11
6 7
5 7
Healthy Healthy Healthy Healthy Healthy Healthy Healthy Healthy Healthy Healthy Healthy Healthy Infectious Infectious
1,786 216,611 18,574 76,567 6,875 1,349 630,055 357,049 997,874 6,045 610,429 3,503 15,607 15,790
3 18 9 14 6 8 25 26 35 8 48 3 6 4
3 16 14 12 4 4 11 16 48 8 15 4 4 7
2 15 11 12 3 4 10 8 27 6 11 4 4 5
3 12 13 10 4 4 8 14 43 7 13 4 3 6
3 11 10 10 3 4 8 5 22 6 11 4 3 6
Infectious Infectious Infectious Infectious Infectious Infectious Infectious Infectious Infectious Infectious Infectious
634,595 286,085 10,385 402,103 204,133 275,099 1,475 11,123 18,272 109,995 127,655
28 17 16 21 11 14 3 6 10 15 13
15 17 21 28 14 6 3 6 9 11 5
8 17 17 17 12 6 3 6 9 11 5
13 17 23 20 13 3 3 6 8 11 5
8 14 15 13 9 3 3 6 8 11 5
Inflammatory 1,293 Inflammatory 34,764 Inflammatory 551,388 Inflammatory 78,818
9 18 7 8
16 18 7 6
9 13 7 5
11 17 7 5
7 12 7 5
Inflammatory 262,642 Inflammatory 19,840
17 7
11 10
10 10
10 10
9 10
Notes. Cells are observations in the selected T-cell-focused H5AD dataset for each study. “Annotated cell types” denotes the number of distinct non-missing values in the author-provided annotation field used for analysis. Extracted counts are distinct Stage 1 paper-local labels after exact duplicate removal. Mapped counts are distinct non-empty Stage 2 normalized labels; several extracted labels may map to one normalized label. GPT and DS denote the final GPT-5.5 and DeepSeek-V4-Pro (thinking) configurations, respectively.
KDD’27, 2027, San Jose, USA
Xie et al.
Table S2: Example mappings between paper-reported labels, dataset annotations, normalized labels, and Cell Ontology terms. Paper Label
Dataset Annotation
Mapped Label
naïve/central memory CD4+ T cells (CCR7/c0) Th1 CD4 effector memory T cells (IL7R/c1) FOXP3+ regulatory T cells (FOXP3/c2) T follicular helper cells (CXCL13/c3) Chemokine-expressing CD8+ T cells (ZFP36/c4) Type I interferon signature T cells (IFIT1/c6) PDCD1low CD8+ T cells expressing IFNG and TNF (IFNG/c7) Exhausted CD8 T cells (LAG3/c8) NKT-like cells (FCGR3A/c10) Proliferating T cells (MKI67/c11)
T_cells_c0_CD4+_CCR7
CD4+ TS
CD4-positive, alpha-beta T cell
CL:0000624
T_cells_c1_CD4+_IL7R
CD4+ TH1DM
CD4-positive, alpha-beta T cell
CL:0000624
CD4 Treg CD4 TFH CD8+ T
CD4-positive, alpha-beta T cell CD4-positive, alpha-beta T cell CD8-positive, alpha-beta T cell
CL:0000624 CL:0000624 CL:0000625
T
CD8-positive, alpha-beta T cell
CL:0000625
T_cells_c7_CD8+_IFNG
CD8+ T
CD8-positive, alpha-beta T cell
CL:0000625
T_cells_c8_CD8+_LAG3 T_cells_c10_NKT_cells_FCGR3A T_cells_c11_MKI67
CD8+ TX NKT T
CD8-positive, alpha-beta T cell Mature NK T cell T cell
CL:0000625 CL:0000814 CL:0000084
T_cells_c2_CD4+_T-regs_FOXP3 T_cells_c3_CD4+_Tfh_CXCL13 T_cells_c4_CD8+_ZFP36 T_cells_c6_IFIT1
Cell Ontology Label
CL ID
Table S3: Seven-paper evaluation set for the Stage 1 Information Extraction Agent.
No. 1 2 3 4 5 6 7
In 44-study benchmark
Paper title
Venue
Year
DOI
A single-cell and spatially resolved atlas of human breast cancers [44] Integrative single-cell multi-omics of CD19-CARpos and CARneg T cells suggest drivers of immunotherapy response in B-cell neoplasias [15] Pan-cancer T cell atlas links a cellular stress response state to immunotherapy resistance [6] Follicular Lymphoma Microenvironment Characteristics Associated with Tumor Cell Mutations and MHC Class II Expression [16] Single-cell transcriptomic analysis of gingivo-buccal oral cancer reveals two dominant cellular programs [22] Ovarian cancer mutational processes drive site-specific immune evasion [40] Stromal cell diversity associated with immune evasion in human triple-negative breast cancer [45]
Nature Genetics bioRxiv
2021 2024
https://doi.org/10.1038/s41588-021-00911-1 https://doi.org/10.1101/2024.01.23.576878
Yes No
Nature Medicine Blood Cancer Discovery
2023 2022
https://doi.org/10.1038/s41591-023-02371-y https://doi.org/10.1158/2643-3230.BCD-21-0075
No Yes
Cancer Science
2023
https://doi.org/10.1111/cas.15979
No
Nature The EMBO Journal
2022 2020
https://doi.org/10.1038/s41586-022-05496-1 https://doi.org/10.15252/embj.2019104063
Yes Yes