Biomedical Knowledge Integration Enhances Language Model Cell Annotation in Single-Cell Sequencing

A computational framework pairing statistical gene expression criteria with curated tissue databases significantly improves cell type identification powered by large language models in single-cell RNA sequencing studies. Developed to resolve persistent bottlenecks in genomic analysis, the workflow, designated as CELLIA, combines cluster-level differential expression statistics with biological prior knowledge to curate informative gene markers before submitting prompts to artificial intelligence models. Across benchmarking evaluations encompassing 102 cell types—including 55 major cell identities across eight tissue settings and 25 fine-grained subtype categories—this integrative evidence-knowledge strategy demonstrated increased concordance with manual expert labels compared to differential expression ranking alone. By addressing dataset-dependent noise and statistical selection biases, the framework offers a systematic computational mechanism to streamline single-cell characterization without sacrificing biological context.

Cell type annotation serves as an indispensable step in single-cell RNA sequencing workflows, connecting cellular identity to disease pathology and targeted therapeutics. In conventional analytical pipelines, annotation occurs downstream of preprocessing and unsupervised clustering, where heterogeneous tissues are partitioned into transcriptionally distinct subsets that researchers manually classify using literature-derived canonical marker genes. However, manual curation represents an exhaustive bottleneck that is difficult to reproduce across investigators. In response, computational biologists have created diverse automated approaches, including marker-based tools such as Garnett, SCINA, and ScType, which identify cells using predefined gene profiles but struggle with incomplete tissue coverage. Complementary reference-based algorithms like scmap and SingleR match expression profiles against pre-annotated reference datasets, while supervised architectures like CellTypist and deep learning systems like scBERT deploy pre-trained representations to resolve cellular categories.

Despite the scalability of automated tools, many conventional algorithms remain sensitive to data quality and feature selection variations, often failing to capture subtle biological distinctions. The integration of large language models into single-cell analytics has introduced reasoning-based inference and context integration derived from generative pre-trained transformers, enabling models to classify identities from marker gene lists with expert-level concordance. Subsequent frameworks have expanded these capabilities: LICT applies multi-model ensembles with iterative feedback, mLLMCelltype leverages consensus strategies to quantify uncertainty, and CASSIA utilizes multi-agent reasoning modules. Nonetheless, existing language model annotation systems continue to inherit a key vulnerability from traditional bioinformatic pipelines: they rely predominantly on cluster-specific markers generated purely through statistical differential expression calculations. When clusters display high transcriptional similarity or substantial background noise, purely statistical markers can be unstable, uninformative, or devoid of critical biological context.

To overcome these limitations, the investigators structured CELLIA around an integrative evidence-knowledge paradigm that evaluates candidate genes through two complementary lenses. Within this architecture, evidence refers to quantitative differential expression parameters extracted directly from single-cell data, specifically adjusted p-values, average log2 fold-change metrics, and cluster-specific cellular detection proportions. Conversely, knowledge denotes curated tissue-specific marker resources assembled from established biological literature. Rather than relying on static or isolated databases, the researchers integrated three distinct repositories: singleCellBase, a literature-derived compendium supporting cross-species cell standardization; CellMarker 2.0, an expanded repository merging single-cell experiments with published literature; and MarkerGeneBERT, a natural language processing framework that dynamically extracts and normalizes marker associations from biomedical texts. By filtering statistically derived candidates against this multi-source biological repository, the system ensures that candidate genes delivered to the language model represent both statistically elevated signals and biologically grounded identity markers.

Operationally, CELLIA executes across four core algorithmic phases coupled with an interactive analytical environment, utilizing standard AnnData structures to maintain computational efficiency. In the primary phase, the software takes pre-clustered single-cell sequencing matrices and identifies differentially expressed genes for each cluster via Scanpy gene ranking functions driven by Wilcoxon rank-sum tests. Genes are filtered according to predefined statistical cutoffs: an adjusted p-value below 0.05, expression in more than 25 percent of cells within the target cluster, expression in fewer than 35 percent of cells across background clusters, and an average log2 fold-change exceeding 0.3. In subsequent phases, candidate genes meeting these statistical criteria are cross-referenced against the compiled tissue-specific database. From this verified intersection, the top fifteen marker genes ranked by descending log2 fold-change are selected to define the cluster, supplemented by top unmapped differentially expressed genes only if fewer than fifteen database-supported markers are available.

Once marker selection is complete, the pipeline constructs a structured prompt incorporating the top fifteen marker genes, the specific tissue background, and broader parental lineage classifications before querying the language model via an application programming interface. The system programmatically parses the resulting model output, retrieving both the predicted cellular label and an interpretable explanatory trace that highlights key diagnostic genes and their functional biological rationales. To eliminate programming barriers for researchers, the developers packaged the entire framework alongside an interactive, web-based graphical interface. The visualization dashboard displays two coordinated Uniform Manifold Approximation and Projection embeddings, dynamically rendering cell-level metadata, cluster predictions, and expression dot plots that adjust across marker list sizes ranging from 5 to 15 genes. Furthermore, every displayed marker gene features direct hyperlinks to the GeneCards database, allowing investigators to audit predicted labels and biological rationales against external genomic literature.

The research team benchmarked CELLIA across nine publicly available single-cell datasets spanning eight distinct physiological and pathological tissue environments. These evaluation benchmarks included colorectal cancer, non-small cell lung cancer, breast tissue, peripheral blood mononuclear cells from both the Human Immune Health Atlas and 10X Genomics platforms, liver, kidney, skin, and pancreas, representing 55 unique major cell types. To evaluate fine-grained resolution, the researchers also examined immune cell subsets within the Human Immune Health Atlas peripheral blood mononuclear cell dataset, evaluating 19 distinct subcategories spanning CD4 positive, double-negative, and regulatory T cells, cytotoxic CD8 positive, gamma-delta, and mucosal-associated invariant T cells, dendritic cells, and B cell lineages. A non-immune stromal subset was additionally analyzed using non-small cell lung cancer fibroblasts from the perturbed-state fibroblast atlas, establishing a combined subtype evaluation suite of 25 unique cell populations.

To conduct quantitative evaluations against manual expert labels, the authors established a standardized three-level agreement scheme anchored to unambiguous Cell Ontology terminology. Under this framework, a predicted annotation received a full-match score of 1.0 if it matched the original study label, shared the exact Cell Ontology concept, or provided a more granular subtype description. Predictions sharing broad functional lineage but differing in specific terminology received a partial-match score of 0.5, while predictions assigning incorrect lineages received a mismatch score of 0.0. Averaging these scores across all evaluated clusters yielded overall dataset agreement metrics. When benchmarked against traditional computational methods, CELLIA consistently achieved superior agreement scores. Notably, reference-based and supervised tools faced severe computational hurdles on massive datasets: Seurat and SingleR failed during R object format conversion on the Human Immune Health Atlas dataset, and CellTypist crashed with process-level errors, reflecting memory exhaustion on a 1-terabyte compute node.

When benchmarked against existing language model annotation systems—including GPTCelltype, LICT, mLLMCelltype, and CASSIA—using models such as GPT-4.1 and Gemini 2.5 Flash, CELLIA maintained consistently high agreement with expert manual annotations. While overall differences among language model platforms remained modest across several standard tissues, CELLIA demonstrated clear advantages in complex anatomical settings such as the kidney and specialized immune cell cohorts. Comprehensive ablation experiments confirmed that combining statistical metrics with curated biological databases was critical to this performance. Removing either the tissue database filter or the cellular expression proportion metric caused measurable declines in annotation agreement across most benchmarked tissues. Removing both components simultaneously, reducing the input to raw p-values and log2 fold-changes, triggered the steepest drops in concordance. In kidney tissue, for example, ablated configurations frequently misidentified Loop of Henle cell populations, whereas CELLIA achieved full agreement across every evaluated cluster.

In granular subtype evaluations using adjusted p-values below 0.1, cellular detection proportions above 0.25 in target clusters and below 0.7 in non-target clusters, and log2 fold-changes above 0.3, the system proved equally robust. The authors evaluated four marker configurations: a basic setup, configurations incorporating major cell type markers, configurations incorporating cluster-specific negative markers, and full marker complements. In immune populations, performance remained stable across configurations, with CELLIA accurately identifying challenging T cell and dendritic cell subsets where competing language model frameworks frequently generated partial matches or outright errors. In contrast, non-immune fibroblast subtypes showed distinct configuration sensitivity, with the configuration omitting major markers achieving the highest agreement score. Sensitivity testing on top-ranked marker counts ranging from 10 to 100 genes demonstrated broad stability, though minor performance declines appeared in breast and kidney tissues at higher marker quantities, confirming 15 markers as an optimal operational target.

Repeated inference assessments demonstrated high reproducibility across ten independent iterations across all nine benchmark datasets, with identical predictions returned for the vast majority of evaluated clusters. Minor variations were confined almost entirely to synonymous nomenclature choices, such as alternating between natural killer cell designations, or adjacent hierarchical classifications like monocytes versus classical monocytes, directly mirroring inconsistencies documented in human expert annotations. These consistent annotation behaviors persisted when evaluating multiple model generations, including GPT-5.2, GPT-5.4, GPT-5.4 mini, and GPT-5.4 nano. Nevertheless, the authors emphasized fundamental operational boundaries: CELLIA operates on pre-clustered inputs and remains sensitive to clustering resolutions, as demonstrated across Leiden parameter variations from 0.05 to 0.3. Furthermore, because the framework relies on curated databases, it is explicitly designed for knowledge-guided annotation rather than de novo discovery of unknown cell types, requiring caution when applied to uncharacterized tumor microenvironments or non-model organisms.

Subject of Research: Automated cell type annotation in single-cell RNA sequencing using large language models and curated marker databases

Article Title: Integrative evidence-knowledge marker selection enhances LLM-based cell type annotation in single-cell RNA-seq analysis

Article References: Shin, J., Jeong, S., Kim, H. J., & Oh, M. (2026). Integrative evidence-knowledge marker selection enhances LLM-based cell type annotation in single-cell RNA-seq analysis. BioData Mining, 19(1), Article 76. https://doi.org/10.1186/s13040-026-00593-7

Image Credits: AI Generated

DOI: 10.1186/s13040-026-00593-7

Keywords: single-cell RNA sequencing, cell type annotation, large language models, bioinformatics, gene expression, computational biology, differentially expressed genes, CELLIA, Integrative, evidence-knowledge, marker, selection