---
title: 'Cell2Text: Multimodal Expression-to-Text Model'
url: https://www.emergentmind.com/topics/cell2text
type: topic
---

# Cell2Text: Multimodal Expression-to-Text Model

Cell2Text is a multimodal generative framework that takes a single-cell RNA-seq expression profile as input and produces a structured natural-language description of the cell. Rather than emitting only a discrete label, it aims to generate a richer textual summary including cell identity, tissue of origin, disease context, developmental or donor context when available, and active biological pathways or functional state. Its central design combines a pretrained single-cell foundation model with a pretrained instruction-tuned large language model, recasting cell annotation from a closed-label prediction problem into biological description generation [2509.24840].

## 1. Conceptual framing and scope

Cell2Text is motivated by a limitation of prior single-cell foundation models such as Geneformer and scGPT: in practice they are usually used through task-specific classification heads. The paper identifies four consequences of that usage pattern: static label spaces, loss of biological nuance, limited interpretability, and information loss from pooled embeddings. In this formulation, a cell is not treated as adequately described by a single categorical output such as “T cell” or “epithelial cell,” because cells may differ simultaneously by cell state, tissue environment, disease association, developmental stage, and active pathways or processes [2509.24840].

The system therefore defines a broader target than conventional cell-type classification. Its intended output is structured biological prose synthesizing cell type or cellular identity, associated disease, tissue origin, donor development stage, and active pathways or biological processes. The paper positions this as a shift from “cell embedding → label” toward “cell expression → biological description.” This suggests a different notion of interpretability: the model is expected to expose a human-readable explanation layer rather than only a classifier score.

A related but distinct formulation appears in “CellMaster,” which treats single-cell annotation as a zero-shot “expression-to-text” problem at the cluster level rather than the single-cell level. CellMaster produces per-cluster textual labels, rationales, confidence scores, recommended next actions, dotplots, and annotated UMAP visualizations, and explicitly frames annotation as iterative reasoning over marker evidence and context rather than a one-shot classification decision [2602.13346].

## 2. Architecture

Cell2Text uses an encoder–adapter–decoder design. The cell encoder is Geneformer, specifically Geneformer-V2-316M, which is kept frozen throughout training. For each cell, Geneformer processes a tokenized gene-expression sequence and outputs a sequence of gene-level contextual embeddings,
\[
H_{g,1}, H_{g,2}, \dots, H_{g,N}.
\]
The paper emphasizes that this sequence-level representation is used instead of a single pooled cell embedding, because pooling is argued to erase subtle gene-level patterns relevant to state and function [2509.24840].

The maximum supported context length is 4096 genes. The appendix reports that Geneformer-tokenized expression sequences average 1843.2 tokens, peak between 1000–1500, and in some cases reach the 4096-token maximum. The paper argues that this upper bound is not restrictive because most cells express fewer than 4096 genes and the highest-expressed genes carry most information.

Between Geneformer and the language model sits a lightweight trainable adapter. This bridge is a two-layer feedforward network with nonlinear activation followed by L2 normalization. Its role is twofold: to resolve the dimensional mismatch between Geneformer outputs and LLM input embeddings, and to mitigate the semantic mismatch between biological latent features and linguistic token embeddings. The paper does not describe cross-attention fusion, prefix tuning, prompt tuning, or multimodal special tokens; the bridge is presented as a direct learned projection from gene-level embeddings into the LLM embedding space.

The language decoder is a pretrained instruction-tuned LLM. The reported variants use Meta-Llama-3.2-1B-Instruct and Gemma3-4B-it. Three trained variants are evaluated: Cell2Text-Llama-1B-LoRa, Cell2Text-Llama-1B, and Cell2Text-Gemma-4B. For Meta-Llama-3.2-1B-Instruct, the paper also evaluates parameter-efficient fine-tuning with LoRA inserted into self-attention modules, with rank \(r = 256\) and \(\alpha = 512\). The training prompt is explicitly instruction-following: the system message declares specialization in cell description predictions, the user message provides “Sequence embeddings: \(H_{g,1}|H_{g,2}|...|H_{g,N}\),” and the assistant output is the target cell description.

## 3. Data construction and supervision

The training corpus is constructed from the CELLxGENE Census and contains 1,000,000 cells from 7,331 donors spanning 783 cell types, 347 tissue types, and 128 disease conditions. The split is performed at the donor level with an 80/10/10 train/validation/test partition. This donor-level design is used to reduce leakage of donor-specific signal across splits and to provide a more robust generalization test [2509.24840].

The textual supervision is not free-form human-authored per-cell annotation. Instead, descriptions are built by integrating CELLxGENE cell metadata, Cell Ontology definitions and standardization, pathway activity analysis using pySCENIC, and MSigDB Hallmark pathway definitions. The appendix further specifies that the pathway pipeline initially evaluates 50 Hallmark pathways, filters out pathways active in fewer than 0.5% of cells, retains 34 pathways, selects the top 2 enriched pathways per cell, and converts these to human-readable descriptions using pathway definitions. The resulting targets are therefore metadata-derived, ontology-enriched, and template-generated or structured, rather than unconstrained expert prose.

The paper also applies a “principled sampling strategy” to reduce biases from public datasets and improve diversity. Reported normalized Shannon diversity improves from 0.7470 to 0.8431 for cell type, from 0.6106 to 0.7035 for tissue, and from 0.3479 to 0.4957 for disease. To reduce assay confounding, several protocols are excluded, including full-length assays such as Smart-seq family, niche or proprietary protocols such as Quartz-seq and GEXSCOPE, targeted assays such as BD Rhapsody targeted mRNA, and very low-prevalence technologies including 10x Flex.

An example description given in the appendix includes a ciliated tracheobronchial epithelial cell, its lung parenchyma origin, male elderly donor context, normal disease status, and pathway/process summaries involving apoptosis and ultraviolet response. This exemplifies the intended output schema: cell identity, anatomical context, donor context, and pathway activity in a single descriptive passage.

## 4. Evaluation framework and empirical results

Cell2Text is evaluated along two axes: the quality of generated text and the accuracy of biological attributes extracted from generated text. The latter is done by parsing generated descriptions with regular expressions so that the generated outputs can be compared directly with classifier baselines. The reported text-generation metrics are Exact Match, BLEU-2, BLEU-4, ROUGE-1, ROUGE-2, ROUGE-L, BERTScore F1 with RoBERTa, and BERTScore F1 with BioBERT. For cell type, tissue, and disease, the paper reports accuracy and weighted F1-score. For pathway activity, it reports subset accuracy, Jaccard similarity, and weighted F1-score [2509.24840].

A distinctive component of the evaluation is ontology-aware similarity. Using the Cell Ontology as an undirected graph whose nodes are cell types and whose edges are `is_a` relations, the paper defines similarity between cell types \(c_i\) and \(c_j\) by
\[
S(c_i, c_j) \propto \log\left(1 + \frac{PPR(c_j \mid c_i)}{\tau}\right),
\]
where \(PPR(c_j \mid c_i)\) is a Personalized PageRank score centered on \(c_i\). This gives partial credit to biologically near-miss predictions such as subtype or parent predictions.

The reported text-generation scores indicate low exact match but high semantic fidelity. Cell2Text-Llama-1B-LoRa obtains Exact Match 5.79, BLEU-4 73.88, RoBERTa BERTScore F1 95.74, and BioBERTScore F1 93.06. Cell2Text-Llama-1B reaches Exact Match 7.02, BLEU-4 77.39, RoBERTa BERTScore F1 96.28, and BioBERTScore F1 93.9. Cell2Text-Gemma-4B reports Exact Match 6.73, BLEU-4 77.38, RoBERTa BERTScore F1 96.32, and BioBERTScore F1 93.93. The paper interprets low exact match as expected because the model paraphrases, while emphasizing the high semantic metrics.

On biological attribute recovery, Cell2Text-Llama-1B reports 76.91 cell-type accuracy, 73.35 tissue accuracy, and 77.84 disease accuracy. Cell2Text-Gemma-4B reports 77.83 cell-type accuracy, 73.04 tissue accuracy, and 77.34 disease accuracy. These results exceed the frozen-embedding baselines Geneformer+Head and Geneformer+LGBM on the main cell type, tissue, and disease tasks. The paper highlights the cell-type gain over Geneformer+Head as more than 10 percentage points: 77.83 versus 67.26.

For ontology-aware evaluation, Geneformer+Head scores 80.62, Geneformer+LGBM 63.7, Cell2Text-Llama-1B-LoRa 75.57, Cell2Text-Llama-1B 85.31, and Cell2Text-Gemma-4B 85.62. This indicates that even when Cell2Text misses the exact label, its predictions tend to be more ontologically proximate. On pathway activity identification, however, the picture is more mixed: Cell2Text beats Geneformer+Head but Geneformer+LGBM remains strongest on subset accuracy, Jaccard similarity, and weighted F1. The paper interprets this as a trade-off, since Cell2Text is not directly optimized for pathway ranking or classification.

Ablation evidence within the reported variants is centered on LLM choice and fine-tuning strategy. LoRA underperforms full fine-tuning on all major tasks. Gemma-4B tends to be slightly strongest on cell type and ontology similarity, while full-tuned Llama-1B is slightly strongest on disease.

## 5. Relation to adjacent structured-to-text systems

Within single-cell biology, the most direct adjacent system is CellMaster. CellMaster operates primarily at the cluster level and uses a four-stage iterative loop—hypothesis generation, marker selection, expression analysis, and result evaluation—implemented with a three-agent architecture and optional human-in-the-loop refinement. Its evidence representation is not raw count matrices but cluster-level summaries familiar to human annotators, including differentially expressed genes, top marker candidates, dotplot summaries of expression level and fraction-expressing cells, previous labels, tissue or developmental context, and optional rules about contamination or doublets [2602.13346].

The contrast with Cell2Text is methodologically sharp. Cell2Text starts from a single cell, uses Geneformer gene-level embeddings, projects them into an instruction-tuned LLM, and generates one integrated description. CellMaster instead begins from clustered data, uses differential expression analysis and dotplots as evidence, and employs iterative reasoning to produce labels, rationales, confidence scores, and recommendations. A plausible implication is that the two systems instantiate different granularities of “expression-to-text”: Cell2Text is cell-native and embedding-centric, whereas CellMaster is cluster-centric and marker-rationale-centric.

There is also a broader structured-to-text lineage outside single-cell biology. Table-to-text systems such as “Table-to-Text: Describing Table Region with Natural Language” [1805.11234], “Table-To-Text generation and pre-training with TabT5” [2210.09162], and “PixT3: Pixel-based Table-To-Text Generation” [2311.09808] emphasize field-aware encoding, copying of rare values, row/column embeddings, and multimodal conditioning over structured input. This suggests that Cell2Text belongs to a wider research pattern in which structured evidence is converted into generated text rather than only into labels.

## 6. Limitations, interpretability, and significance

Several limitations are explicit or strongly implied in the reported setup. First, the quality of generated text depends on the correctness of metadata, ontology coverage, and pathway annotation quality, because the training descriptions are constructed from those sources rather than written directly by experts. Second, the supervision is largely structured or template-like, which may constrain stylistic diversity and deeper reasoning. Third, the appendix shows substantial skew in tissues, diseases, assays, and pathways, so learned behavior may reflect atlas composition as well as biology [2509.24840].

The generalization claims are also bounded. The experimental split is donor-level, but the paper does not report dedicated unseen-tissue, unseen-disease, or held-out ontology-branch evaluations. Likewise, it does not provide a detailed hallucination analysis, uncertainty calibration, phrase-level faithfulness attribution linking statements back to genes, or explicit confidence estimates for generated descriptions. Normalization details for the expression input are not fully specified, and low-level architectural details such as adapter hidden sizes and exact decoding settings are also omitted.

Even with those constraints, the paper’s significance is clear within its stated scope. It shows that coupling expression data with natural language can produce outputs that are simultaneously descriptive and predictive: the generated descriptions are human-readable, recover structured biological attributes at strong accuracy, and achieve high ontology-aware similarity. The core claim is therefore not merely that a classifier can be replaced by prose, but that a multimodal encoder-to-LLM system can provide stronger predictive performance than frozen-embedding baselines while exposing a richer description of cellular identity, tissue context, disease association, and pathway activity.

Source: https://www.emergentmind.com/topics/cell2text